Metal cutting process real-time monitoring and diagnosis method based on industrial internet of things

Abstract

The invention discloses a metal cutting process real-time monitoring and diagnosis method based on industrial Internet of Things, and particularly relates to the technical field of intelligent manufacturing and industrial Internet of Things. Constructing a dynamic sensing topology based on first workpiece data, and finely tuning a physical information neural network to establish an exclusive baseline; real-time signal processing and lightweight model reasoning are carried out through edge equipment, and closed-loop control is formed after digital twinborn simulation verification is combined; updating the model by using an actual result and realizing multi-machine knowledge sharing and evolution through federal knowledge distillation; and displaying the situation, the details and the topology through a visual terminal, and triggering early warning. According to the invention, intelligent monitoring and diagnosis which are high in reliability, self-adaptive, fast in response and capable of being evolved are realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial Internet of Things (IoT) technology, and more specifically, to a method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things. Background Technology

[0002] Metal cutting is a core process in manufacturing, and its process status directly affects machining quality, efficiency, and cost. With the development of Industrial Internet of Things (IIoT) and Artificial Intelligence (AI) technologies, real-time monitoring and intelligent diagnosis of the cutting process have become crucial for improving the level of manufacturing intelligence. Current technologies typically involve deploying vibration, acoustic emission, and current sensors on machine tools to collect machining process signals and transmitting the data to a host computer or cloud server via wired or wireless networks. At the data processing level, existing methods often employ threshold alarms, time-frequency domain feature analysis combined with traditional machine learning models (such as support vector machines and random forests) or deep learning models for status identification and fault warning. Some advanced solutions are beginning to explore building digital twin models of the machining process for simulation analysis and prediction.

[0003] However, in practical use, it still has some shortcomings. For example, at the data acquisition and transmission level, sensor networks often lack high-precision time synchronization, making it difficult to accurately align multi-source heterogeneous data in the time domain, forming "data silos." Moreover, the network architecture is simple and lacks reliability, making it difficult to meet the real-time and reliable transmission requirements of high-frequency dynamic data in metal cutting. At the condition diagnosis model level, existing data-driven models heavily rely on a large amount of labeled data, while fault samples are scarce in actual production, resulting in weak model generalization ability and physically uninterpretable prediction results. At the same time, most models are statically designed and cannot quickly adapt to specific machine tools, cutting tools, and processes. Furthermore, deployment in the cloud leads to high inference latency, making it difficult to achieve real-time closed-loop control. At the decision-making and execution level, there is a lack of safety verification links based on physical mechanisms and simulation verification, resulting in high decision-making risks. In addition, the coordination between various modules of the system is poor, and the closed-loop cycle from monitoring to execution is long, making it impossible to quickly and dynamically optimize the machining process. At the knowledge accumulation and application level, existing systems are often limited to a single device or a single process, lacking a mechanism for effective knowledge sharing and continuous evolution among multiple machine tools and processes, resulting in stagnant system intelligence. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a real-time monitoring and diagnosis method for metal cutting processes based on the Industrial Internet of Things, which addresses the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things, comprising:

[0006] S1: Deploy sensors, edge intelligent all-in-one machines, and local servers to build a three-tier hardware architecture and dual-backup network; the edge intelligent all-in-one machine associates work order information with process stages and loads monitoring configurations; based on the first workpiece processing data, calculates the information entropy of sensor signals to build a dynamic sensing topology, and fine-tunes the physical information neural network to obtain a dedicated baseline model;

[0007] S2: The edge intelligent all-in-one machine identifies the processing stage in real time and dynamically adjusts the acquisition strategy based on the monitoring configuration. It preprocesses and extracts features from the acquired signals to obtain a real-time feature stream. It uses the lightweight physical information neural network of the dedicated baseline model to perform reasoning on the real-time feature stream with physical constraints, and outputs a physical state vector and an optimization decision vector. It performs security verification on the optimization decision vector through the digital twin model in the local server, and writes the verification-passed correction instructions into the machine tool CNC system.

[0008] S3: Update the parameters of the digital twin model according to the actual processing results after the execution of the correction instruction; periodically aggregate the local knowledge in each edge intelligent machine through federated knowledge distillation, generate and distribute the enhanced model or rule patch to each edge intelligent machine;

[0009] S4: Display the global health status, machine tool details, decision log and dynamic perception topology generated based on the physical state vector through a visualization terminal, and trigger an early warning notification when the physical state vector indicates an abnormal status.

[0010] The technical effects and advantages of this invention are as follows:

[0011] 1. Solved the problem of accurate synchronization and high-reliability transmission of multi-source heterogeneous data: By constructing a three-level hardware architecture of "sensing-edge-cloud" that integrates a high-precision time synchronization protocol, and a "TSN+5G" dual backup network, millisecond-level time alignment and high-reliability, low-latency transmission of multi-sensor data such as vibration and acoustic emission were achieved, laying a solid foundation for subsequent accurate analysis and overcoming the defects of data asynchrony and single network in traditional monitoring systems;

[0012] 2. Achieved strong generalization, interpretability and rapid adaptive intelligent diagnosis: Innovatively, information entropy flow analysis is used to construct a dynamic perception topology, and combined with physical information neural network technology, a dedicated baseline model integrating data-driven and physical mechanisms is established; This method can still ensure high recognition accuracy and model interpretability under small sample conditions, and can quickly adapt to different machine tool-process combinations, solving the pain points of traditional data-driven models that rely on a large amount of labeled data, have weak generalization and unclear physical meaning;

[0013] 3. A safe, fast, and automated decision-making and execution closed loop has been constructed: By introducing a simulation verification step based on digital twins, the decisions generated by edge intelligence are verified in both virtual and real environments, which improves the safety and reliability of the decisions. At the same time, by optimizing the process, the entire closed loop time of "monitoring-diagnosis-decision-execution" is shortened to less than 10 seconds, realizing rapid dynamic optimization and precise control of the processing process, effectively avoiding the problems of high decision-making risk and slow response in traditional methods.

[0014] 4. A system intelligent upgrade mechanism for sustainable evolution and efficient knowledge sharing has been established: Through the iterative update of digital twin model parameters driven by processing result feedback, and the periodic federated knowledge distillation technology, the system can continuously accumulate experience, extract common knowledge across machine tools and processes, and form a global knowledge model or patch for distribution, realizing the autonomous evolution and collaborative improvement of system performance, fundamentally breaking the situation of solidified traditional system models and isolated knowledge. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0016] Figure 2 This is a schematic diagram of the real-time monitoring, diagnosis, and closed-loop decision execution structure of the present invention.

[0017] Figure 3 This is a schematic diagram of the continuous learning and system evolution structure of the present invention.

[0018] Figure 4 This is a schematic diagram of the human-computer interaction and overall visualization structure of the present invention.

[0019] Figure 5 This is a diagram showing the constraint relationship function of the empirical model of cutting force in this invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. 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.

[0021] refer to Figures 1-5 The real-time monitoring and diagnostic method for metal cutting processes based on the Industrial Internet of Things (IIoT) shown includes:

[0022] S1: Deploy sensors, edge intelligent all-in-one machines, and local servers to build a three-tier hardware architecture and dual-backup network; the edge intelligent all-in-one machine associates work order information with process stages and loads monitoring configurations; based on the first workpiece processing data, calculates the information entropy of sensor signals to build a dynamic sensing topology, and fine-tunes the physical information neural network to obtain a dedicated baseline model;

[0023] S2: The edge intelligent all-in-one machine identifies the processing stage in real time and dynamically adjusts the acquisition strategy based on the monitoring configuration. It preprocesses and extracts features from the acquired signals to obtain a real-time feature stream. It uses the lightweight physical information neural network of the dedicated baseline model to perform reasoning on the real-time feature stream with physical constraints, and outputs a physical state vector and an optimization decision vector. It performs security verification on the optimization decision vector through the digital twin model in the local server, and writes the verification-passed correction instructions into the machine tool CNC system.

[0024] S3: Update the parameters of the digital twin model according to the actual processing results after the execution of the correction instruction; periodically aggregate the local knowledge in each edge intelligent machine through federated knowledge distillation, generate and distribute the enhanced model or rule patch to each edge intelligent machine;

[0025] S4: Display the global health status, machine tool details, decision log and dynamic perception topology generated based on the physical state vector through a visualization terminal, and trigger an early warning notification when the physical state vector indicates an abnormal status.

[0026] S1: System Initialization and Self-Startup: This involves completing three levels of hardware deployment, process information association, and the construction of a dedicated baseline model. Through physical networking, process analysis, and entropy flow analysis, it lays the hardware, data, and model foundations for subsequent real-time monitoring. Specifically, it includes:

[0027] S101: Physical Deployment and Network Configuration: Constructing a three-tier hardware architecture of "sensing-edge-cloud", with specific deployment and implementation as follows:

[0028] Intelligent sensing module deployment: Sensor components are precisely installed on each target CNC machine tool according to process monitoring requirements.

[0029] The IEPE type triaxial vibration sensor has a measurement range of ±50g, a frequency response of 1-10kHz, and a resolution of 0.001g. It can be installed in the spindle box, tool turret, and workpiece fixture.

[0030] Acoustic emission sensors with a frequency response of 20-800kHz and a sensitivity of ≥80dB are installed on the tool holder or workpiece contact surface to ensure direct signal acquisition.

[0031] The intelligent control board integrates a data preprocessing unit and a TSN communication interface to achieve preliminary digitization and real-time transmission of sensor signals. All sensors, through their built-in microprocessors equipped with the IEEE 1588v2 Precision Time Protocol (PTP), achieve millisecond-level time synchronization with the TSN switch, avoiding timestamp discrepancies between multiple sensor data.

[0032] Edge intelligent all-in-one machine deployment: The edge intelligent all-in-one machine installed on the machine tool side adopts an industrial-grade design, with hardware configuration including an Intel Core i7 processor, 16GB DDR4 memory, 512GB SSD storage, and supports 4 Gigabit Ethernet ports, 2 RS485 serial ports and OPCUA dedicated interface.

[0033] This configuration is based on the requirements of parallel data processing of 6 sensors on a single machine tool, real-time inference of the PINN model (200ms per inference), and local data caching. The computing power redundancy design ensures the ability to process multiple tasks concurrently.

[0034] The equipment establishes bidirectional communication with the machine tool CNC system through the OPCUA Client, reads process parameters such as spindle speed, feed rate, and depth of cut, and writes correction instructions. The OPCUA protocol is chosen based on its industrial-grade versatility, and is compatible with mainstream CNC systems such as FANUC and Siemens. The Basic256Sha256 security strategy uses asymmetric encryption and digital signature mechanisms to prevent tampering and eavesdropping risks during data transmission. It connects to the sensor network through a TSN switch to achieve local data aggregation and real-time processing.

[0035] Workshop-level network and server deployment: Two redundant local servers are deployed in the workshop, and communication links are established with all edge intelligent all-in-one machines through a 5G private network;

[0036] The core basis for choosing the "TSN+5G" dual backup architecture is the high dynamic characteristics of metal cutting process data: TSN ensures millisecond-level synchronous transmission of sensor data, and 5G enables high-bandwidth wireless interconnection between the edge and the cloud. Compared with traditional industrial Ethernet, this combination can reduce the deployment restrictions of mobile devices, while reducing the risk of network interruption to below 0.001% through dual-link redundancy.

[0037] The network architecture adopts a "TSN+5G" dual backup design. The handover trigger condition is that the latency of a single link exceeds the threshold (20ms) three times in a row or the packet loss rate exceeds 0.01%. The handover process ensures integrity through data caching and breakpoint resume mechanism, and the handover latency is ≤50ms to ensure communication continuity.

[0038] During deployment, grounding (grounding resistance ≤ 4Ω) and electromagnetic shielding must be completed to avoid the impact of machine tool electromagnetic interference on data acquisition.

[0039] S102: Identity registration and process information preloading, the specific process is as follows:

[0040] Work order information push: When an operator creates a new task work order in the work order management module of the workshop-level MES system, they need to enter basic information such as part model, material type, batch quantity, and processing procedures, and upload the corresponding CNC program file. The MES system encapsulates the work order information into a JSON format data packet through the MQTT protocol, which includes key information such as work order ID, part parameters, process route table, CNC program path, and quality requirements, and pushes it to the local server and the corresponding edge intelligent all-in-one machine in real time to ensure information consistency.

[0041] Process Stage Analysis: The edge-mounted integrated machine has a built-in G-code parsing engine that automatically starts the parsing process after receiving the CNC program. It reads the G-code line by line, recognizing characteristic commands such as G01, G02 / G03, G71, and G73, and combining these with parameter thresholds such as feed rate and depth of cut to divide the entire machining process into several stages, including roughing, semi-finishing, finishing, drilling, and tapping. For example, the G71U2.0R1.0 command corresponds to the roughing stage, and the G70P10Q20 command corresponds to the finishing stage.

[0042] Preloading of monitoring configuration templates: The system predefines standardized monitoring configuration templates for each process stage. The templates include parameters such as sensor sampling rate, signal filtering frequency band, feature extraction type, and data transmission priority.

[0043] In the roughing stage, the template is configured with a vibration sensor sampling rate of 2kHz, focusing on the 10-500Hz frequency band, and an acoustic emission sensor sampling rate of 500kHz, focusing on the 50-200kHz frequency band, to extract time-domain and frequency-domain features. In the finishing stage, the template is configured with a vibration sensor sampling rate of 5kHz, focusing on the 500-2000Hz frequency band, and an acoustic emission sensor sampling rate of 1MHz, focusing on the 200-500kHz frequency band, to add wavelet packet entropy feature extraction.

[0044] After the edge all-in-one machine parses out the process stage, it automatically matches and loads the corresponding initial template, and at the same time stores the associated work order ID in the local configuration library.

[0045] S103: Adaptive baseline learning for the first workpiece: The machining of the first workpiece is a crucial process for the system to establish a dedicated baseline for "machine tool-process-sensor". A personalized monitoring model is constructed through information entropy flow analysis, and the specific implementation is as follows:

[0046] Information entropy flow detection mode activated: When the first workpiece is clamped, the machine tool executes the machining start command.

[0047] When the edge device automatically triggers the "Information Entropy Flow Detection Mode", all sensors start data acquisition according to the preset standard sampling rate, such as 1kHz for vibration sensors and 500kHz for acoustic emission sensors. The sampled data is transmitted to the edge device in real time through the TSN network, and the transmission delay is controlled within 5ms.

[0048] Shannon entropy calculation: The edge computing unit processes the real-time signals from each sensor using a short time window, with a window length of 50ms. An equal-width discretization method is used to divide the signal amplitude into M=64 intervals. The probability is obtained by statistically analyzing the proportion of signal samples within each interval. Substituting into the Shannon entropy formula:

[0049]

[0050] in, Real-time reflection of the uncertainty of sensor signals, under normal processing conditions The fluctuation range is stable, but sudden changes may occur when there are abnormalities.

[0051] It's worth noting that the core reason for choosing Shannon entropy is its ability to effectively quantify signal complexity: during metal cutting, anomalies such as tool wear and chatter can cause changes in the uniformity of signal amplitude distribution. Shannon entropy can quickly capture these changes through probability distribution, improving the sensitivity for identifying early, subtle anomalies by more than 30% compared to traditional time-domain features (such as peak values). A sliding window mechanism is used in the calculation, updating the window every 10 new samples. Values ​​are used to ensure real-time performance.

[0052] Transfer entropy calculation and information flow network construction: To identify the correlation between sensors, the transfer entropy from sensor j to sensor i is calculated simultaneously, as shown in the following formula:

[0053]

[0054] in, and Sensors , past =3、 =Historical state at 2 time steps (determined based on AIC optimization) , (value) The joint probability density function is calculated using the kernel density estimation method.

[0055] It should be further explained that the adaptability of transfer entropy stems from the multi-physics coupling characteristics of the metal cutting process: there is a causal relationship between signals such as cutting force, vibration, and acoustic emission. Transfer entropy can quantify this unidirectional information flow. Compared with the Pearson correlation coefficient, it can eliminate false correlations, accurately identify key sensor pairs, and provide a core basis for the construction of dynamic sensing topology.

[0056] Transitive entropy Quantified right Information transmission intensity The larger, the more it indicates Signal changes The greater the predictive contribution, the better.

[0057] An information flow network is constructed based on the transfer entropy values ​​among all sensors. Nodes represent sensors, and the thickness of the edges is positively correlated with the transfer entropy value. A threshold of 0.8 times the maximum transfer entropy value is set to screen out sensor pairs with significant correlations, forming a "dynamic sensing topology map" specific to this process. For example, if the transfer entropy value of the acoustic emission sensor and the X-axis vibration sensor reaches 0.72 (maximum transfer entropy 0.85) during the precision milling stage, they are set as a collaborative monitoring pair, and the collaborative changes of their signals are analyzed first in subsequent monitoring.

[0058] PINN model fine-tuning: Collect sensor data under all normal conditions during the machining process of the first workpiece, remove data from non-cutting stages such as clamping and tool changing, and construct a training dataset with a sample size of ≥10,000 to fine-tune the preset physical information neural network (PINN).

[0059] It should be further explained that the core advantage of PINN compared to traditional neural networks is that it integrates physical constraints, which can solve the problem of scarce labeled data in the metal cutting process, while improving the model's generalization ability and prediction reliability. Its physical constraint term achieves collaborative modeling of data-driven and mechanism-driven approaches by embedding cutting mechanics formulas into the loss function.

[0060] The initial structure of PINN is such that the input layer dimension equals the number of sensor features, and each sensor extracts... Eight features including peak value, RMS value, etc., with 48 dimensions from 6 sensors; 3 hidden layers, each with 64 neurons, using ReLU activation function; the output layer contains 3 core physical state parameters: tool wear amount. Cutting force coefficient Stability margin .

[0061] The network loss function includes data fitting error and physical constraint error:

[0062]

[0063] in, Mean squared error (MSE) The constraint error is due to physical laws such as cutting force and chatter stability. The value is 0.2. This parameter was calibrated through 10 sets of experiments using different processes. The adjustment range is 0.1-0.5. The higher the material hardness, the better. The larger the value, the greater the weight of physical constraints in hard material processing; the fine-tuning process uses the Adam optimizer with a learning rate of 0.001 and 1000 iterations to ensure the prediction accuracy of the model under this machine tool-process, such as VB prediction error ≤ ±0.01mm.

[0064] S2: Real-time monitoring, diagnosis, and closed-loop decision execution: Through dynamic data acquisition adapted to the process, real-time inference of the PINN model, and virtual-physical cross-verification for security checks, a closed loop is achieved to accurately control the processing status and quality risks, specifically including:

[0065] S201: Real-time Data Acquisition and Dynamic Focusing

[0066] Once mass production begins, the system dynamically adjusts the monitoring configuration based on the current processing stage, achieving "on-demand data collection and precise focusing." The specific process is as follows:

[0067] Real-time process stage identification: The edge integrated machine has a built-in real-time G-code parsing engine that reads the currently executing G-code segment of the machine tool CNC system in real time through the OPCUA protocol. It refreshes every 10ms and matches it with the pre-parsed process stage feature code to accurately identify the current processing stage.

[0068] Dynamic switching of monitoring configuration: The dynamic sensing topology configuration is stored in an XML file, which includes parameters such as sensor primary and secondary node identifiers, sampling rate, filter frequency band, and data preprocessing rules;

[0069] The "fine milling of small radial depth of cut" stage is configured with an acoustic emission sensor as the master node, the sampling rate is increased to 2MHz, and data transmission has the highest priority.

[0070] Focusing on the 500-2000Hz high frequency band, it uses a Chebyshev II filter with a stopband attenuation of 40dB.

[0071] A triaxial vibration sensor is used as an auxiliary node (sampling rate 1kHz), focusing on the 500-2000Hz frequency band, and only three key features are extracted: peak value, kurtosis, and wavelet packet entropy.

[0072] Other sensors maintain a low sampling rate (100Hz) for monitoring to aid in decision-making. Configuration switching is triggered by a hardware timer on the edge device, with a switching delay of ≤10ms, ensuring that no data at the start of a process stage is missed.

[0073] Signal preprocessing and aggregation: The raw signals acquired by the sensors are first preprocessed by the local electronic control board (detrending and outlier removal), and median filtering (window size 5) is used to remove impulse noise, and Kalman filtering is used to suppress random noise;

[0074] The preprocessed signal is transmitted to the edge appliance via the TSN network. The edge appliance performs time alignment and feature extraction on the data according to the principle of "master node data priority processing" to form a standardized feature stream, which prepares for subsequent PINN inference.

[0075] S202: Real-time Inference of Embedded Physical Information Neural Networks

[0076] The edge appliance is equipped with a lightweight PINN model, which uses dynamically focused feature flow to achieve real-time physical state diagnosis and optimization decision generation. The specific process is as follows:

[0077] Lightweight Model Optimization: To meet the real-time inference requirements of edge computing, the PINN model underwent quantization (INT8 precision) and pruning optimization. After quantization, the model size was compressed from 200MB to 25MB, and pruning removed 30% of redundant connections while retaining core weights. The single inference time is ≤200ms, meeting real-time requirements. The model's physical constraints strictly adhere to the laws of metal cutting mechanics, specifically including:

[0078] Empirical model constraints for cutting forces:

[0079]

[0080] in, The main cutting force predicted by PINN. This is the cutting force coefficient. For real-time cutting depth (read from the CNC system). This refers to the feed per tooth (process parameter preloading). The predicted tool wear amount, This represents the initial wear of the cutting tool (set to 0.02 mm). Wear influence coefficient (carbide cutting tools) =0.4), The number of samples within the inference window ( =10).

[0081] It should be further noted that some basic scenario parameters in this experiment are as follows: Cutting force coefficient: =1800; Initial tool wear: =0.15; Wear Influence Coefficient: =0.25; Sample size: =10; the table is as follows:

[0082]

[0083] The data in the table shows that as the depth of cut increases, the cutting force in the metal cutting process also increases; as the feed per tooth increases, the cutting force increases accordingly; and as the tool wear increases, the cutting force increases accordingly, but due to the change in the cutting force coefficient caused by wear, it exhibits a non-linear growth.

[0084] Flutter stability constraints:

[0085] in, These are the eigenvalues ​​of the directional cutting dynamic coefficient matrix. Real-time spindle speed (read from the CNC system). Indicates taking the real part, The limiting depth of cut (derived through flutter theory model, such as...) =3000rpm =2.5mm). This constraint ensures the predicted stability margin. It can accurately reflect the chatter risk of the processing system.

[0086] Core output results: After optimization based on the above constraints, PINN outputs two structured vectors:

[0087] Physical state vector: Contains quantified values ​​of key machining state parameters, in the format of "flank wear". =0.15mm, cutting force coefficient =+8%, stability margin =0.85, cutting temperature =85℃, spindle vibration amplitude =0.03mm”, where It reflects changes in cutting force caused by changes in material hardness or tool wear. This indicates a risk of flutter. It is considered high-risk.

[0088] Optimized decision vector: Generated based on the comparison between the physical state vector and the safety threshold, in the format of "Suggested action: Increase spindle speed, target value: +5% (3000rpm→3150rpm), expected effect: stability margin". Cutting force decreased by 3%, workpiece surface roughness ≤0.8μm”;

[0089] The decision generation logic is when When adjusting the spindle speed (±5%-10%), prioritize adjusting the spindle speed. When the thickness is >0.2mm, it is recommended to replace the cutting tool; when If the value is >15%, it is recommended to check the material hardness or the condition of the cutting edge of the tool.

[0090] S203: Security Verification and Execution of Decisions: Execution is performed only after verification through "virtual vs. real verification," forming a 10-second closed loop. The specific process is as follows:

[0091] Decision Upload and Simulation Verification: The edge integrated machine uploads the optimized decision vector to the virtual-real mutual verification engine of the local server via a 5G private network. The engine consists of a digital twin model, an explicit dynamic solver, and a security verification rule base. The digital twin model constructs the geometric and dynamic models of the machine tool, cutting tool, and workpiece at a 1:1 scale. The solver quickly simulates the machining process based on the decision parameters and outputs the predicted stability margin, cutting force, workpiece size error, and other results.

[0092] The security verification rule base defines verification metrics: predictive stability margin. If all three conditions are met (≥1.05, cutting force variation rate ≤10%, workpiece size error ≤±0.005mm), the verification is considered "passed"; otherwise, the verification is considered "failed".

[0093] Verification result feedback and decision adjustment: If the verification passes, the engine sends an execution license instruction (including a verification report) to the edge appliance via the MQTT protocol; if the verification fails, an adjustment suggestion is returned, and the edge appliance regenerates the decision vector, iterating up to 3 times. If it still fails, a conservative decision is enabled, such as maintaining the current parameters and pushing a warning message.

[0094] Command Issuance and Execution: After receiving the execution permission, the edge computing unit writes the modified S-command into the machine tool's CNC system via the OPCUA protocol. Command writing employs an authentication mechanism: the edge computing unit must pass the CNC system's device whitelist verification. The command includes a timestamp and digital signature; the CNC system verifies its integrity before execution.

[0095] For example, the original S instruction was S3000, which was corrected to S3150. The writing process adopts a three-step mechanism of "write to buffer first → confirm receipt → execute" to avoid instruction loss. The entire closed-loop process is allocated as follows: data acquisition 1s, PINN inference 2s, simulation verification 5s, and instruction issuance 2s. The 10-second closed-loop time is based on the statistical optimization results of 100 sets of actual machining experiments, which can meet the dynamic adjustment needs of the metal cutting process. Compared with traditional manual adjustment, the efficiency is improved by 90%, ensuring that the closed loop from monitoring to execution is completed within 10 seconds.

[0096] The S3 continuous learning and system evolution: Through processing result feedback, iterative model parameter updates, and federated knowledge distillation, it achieves iterative cycles, gathers common knowledge from multiple machine tools and processes, and promotes the continuous evolution of the system model's accuracy and adaptability. Specifically, it includes:

[0097] S301: Result Feedback and Virtual-Real Verification: After the decision is executed, the system verifies the effectiveness of the decision through the actual processing results and updates the model parameters to achieve iteration. The specific process is as follows:

[0098] Actual Result Acquisition: After the machine tool completes workpiece machining under new parameters, the edge-mounted integrated machine collects actual machining data. Key workpiece dimensions are obtained through the machine tool's online measurement system, and tool wear is measured using a tool wear measuring instrument (accuracy ≤ 0.001mm). The measured values, along with sensor data and measured cutting force values ​​during the machining process, are recorded to form a data package of actual results. This data package is uploaded to a local server via a 5G private network and stored in association with the corresponding simulation prediction results.

[0099] Model update logic: The virtual-real cross-validation engine starts the learning process, dividing it into two cases based on the matching degree:

[0100] Case A (Good Match): When the error between the actual result and the simulation prediction result is ≤5%, such as... Actual measurement: 0.16mm; Predicted measurement: 0.15mm. The measured value was 1.08, and the predicted value was 1.1, indicating a good match. The system strengthens the confidence of this decision path by storing decision cases containing process parameters, physical conditions, decision content, and effects in the workshop knowledge base for reference in subsequent similar processes.

[0101] Case B (with deviation): When the error > 5%, the simulation model parameters... (e.g., cutting force coefficient, tool stiffness, damping coefficient, etc.) are updated using Bayesian methods, prior distribution. The model is set to a Gaussian distribution with a mean of 0.01 and a variance of 0.01. The Gaussian distribution is chosen because it reflects the random fluctuations of the model parameters. The variance of 0.01 is determined through historical parameter deviation statistics to ensure the reasonableness of the prior information. The posterior distribution satisfies:

[0102]

[0103] In practice, maximum a posteriori estimation (MAP) is used to solve for the optimal parameters. :

[0104]

[0105] in, For actual results, such as Measured values Measured value These are simulation predictions. =0.08 is the regularization coefficient. These are the model parameters before the update. The advantage of MAP over ordinary least squares is that it incorporates prior information, avoiding overfitting, and is especially suitable for scenarios with small sample data. After the update... Used to optimize digital twin models and improve the accuracy of subsequent simulation verification.

[0106] S302: Federated Knowledge Distillation and Distribution: To achieve knowledge sharing among multiple machine tools and processes within the workshop, the system periodically performs federated knowledge distillation, aggregating local knowledge scattered across various edge integrated machines into global knowledge. The specific process is as follows:

[0107] Federated distillation cycle and data aggregation: The federated distillation process is set to start every Sunday at 2:00 AM (when the machine tool is idle); the local server sends a knowledge upload command to all edge all-in-one machines, and each edge all-in-one machine uploads the output change trend of the local PINN model under 10 standard virtual fault scenarios to the server; the uploaded data is encrypted and only contains the model output features, without involving the original data, thus protecting data privacy.

[0108] The virtual fault scenarios include excessive tool wear, chatter, sudden changes in material hardness, and sensor failure. These scenarios are generated through modeling and virtual simulation of actual fault cases, covering more than 95% of common machining anomalies.

[0109] Global Knowledge Extraction: A teacher model is built on the server, using the output trends of each edge integrated machine as training data. A temperature softening function is used to aggregate local features, and the teacher model is trained using a distillation loss function to extract common knowledge from different machine tools and processes. The collaborative logic between the teacher model and the edge PINN is as follows: the teacher model learns global common knowledge, while the edge model retains local process characteristics, achieving global and local knowledge fusion through distillation. After training, an enhanced general PINN model or diagnostic rule patch is generated.

[0110] Model / Patch Distribution and Updates: The server distributes enhanced models or patches to each edge appliance via a dedicated 5G network. The distribution employs a resume mechanism to prevent transmission failures due to network interruptions. Upon receiving the patch, the edge appliance performs a virtual scene test locally. If the test passes, the old model / patch is automatically overwritten; if the test fails, the original version is retained and feedback is sent to the server. The update process is completed when the machine tools are idle, without affecting normal production and ensuring continuous system evolution.

[0111] S4: Human-Machine Interaction and Overall Visualization: Through global health status visualization, machine tool detail drilling, and multi-channel early warning linkage, it provides workshop engineers with an intuitive monitoring perspective and efficient interactive support, realizing human-machine collaboration and full-process digital management and control, specifically including:

[0112] Global health status visualization: The large screen in the workshop control room and the engineers' mobile terminals display a real-time global health status map of the workshop; against the background of the workshop layout, each machine tool icon uses a different color to indicate its health status.

[0113] Green: Normal ≥1.0、 <0.2mm;

[0114] Yellow: Warning, 0.9≤ <1.0 or 0.2mm≤ <0.3mm;

[0115] Red: Malfunction, <0.9 or ≥0.3mm;

[0116] The icon size reflects the production load: Load = current output / planned output, divided into 4 levels: 0-20%, 20%-50%, 50%-80%, and 80%-100%, with a corresponding icon size ratio of 0.8:1.0:1.2:1.5;

[0117] The color coding standard is based on the failure risk statistics of 1000 sets of processing experiments:

[0118] In red mode, the scrap rate is ≥30%; in yellow mode, the scrap rate is 5%-30%; and in green mode, the scrap rate is <5%. The status map refreshes every 5 seconds and supports zooming and panning operations, allowing engineers to quickly grasp the overall status.

[0119] Machine tool details drill-down function: Click on any machine tool icon in the dashboard or APP to drill down and view detailed information about that machine tool, which includes three main modules:

[0120] Real-time physical status dashboard: Displays core parameters in the form of numbers and trend charts, such as Current value and its variation curve over the past hour, real-time cutting force value, and stability margin. When key parameters such as spindle speed and feed rate exceed the standard, a red flashing warning will be issued.

[0121] Decision Log: Displays all decision records in reverse chronological order. Each record includes information such as time, current process stage, diagnosed problem, executed decision, simulation verification result, actual effect, and confidence level. It supports filtering by time range and decision type for easy traceability.

[0122] Dynamic Sensing Topology Map: This map displays the sensor network at the current process stage in real time. Nodes are labeled with sensor type and status (green: normal, red: fault). Edge thickness represents the transfer entropy intensity; hovering the mouse over an edge displays the specific transfer entropy value, helping engineers understand the information interaction between sensors. The topology map is visualized using WebGL technology. Through a data-driven graphics rendering mechanism, transfer entropy values ​​are mapped to edge widths, and a real-time refresh mechanism synchronizes the map every 10 sensor data updates.

[0123] Linked early warning notifications and reports: When the system detects a major anomaly, such as... <0.8、 When the error is ≥0.3mm or an important decision is made, such as tool replacement or parameter adjustment, multi-channel notifications will be automatically triggered: Kanban pop-up prompts, APP push messages, and email notifications. The notification content includes an anomaly / decision description, urgency level (high / medium / low), and handling suggestions.

[0124] Simultaneously, a detailed report (PDF format) is generated, including data charts, diagnostic basis, simulation results, and actual effect analysis, which can be viewed and downloaded online. The report can be linked to the MES system: tool change suggestions trigger the MES tool management module to generate a work order, and parameter adjustment suggestions update the cycle time prediction of the MES production planning module, realizing full-process digital collaboration.

[0125] It should be further explained that the criteria for classifying urgency levels (high / medium / low) are as follows:

[0126] High urgency: Fault status, requires handling within 10 minutes; Medium urgency: Warning status, requires handling within 30 minutes; Low urgency: Optimization decision under normal conditions, requires confirmation within 2 hours.

[0127] Anomaly handling support: Engineers can initiate remote interventions via an app or dashboard, such as rejecting decisions, manually adjusting parameters, or activating contingency plans (e.g., emergency shutdown). The system records intervention operations and processing results, incorporates continuous learning data, and continuously optimizes the accuracy of diagnosis and decision-making.

[0128] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0129] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things, characterized in that, include: S1: Deploy sensors, edge intelligence all-in-one machines, and local servers to build a three-tier hardware architecture and dual backup network; The edge intelligent all-in-one machine associates work order information with process stages and loads monitoring configurations; Based on the first workpiece processing data, the information entropy of the sensor signal is calculated to construct a dynamic sensing topology, and the physical information neural network is fine-tuned to obtain a dedicated baseline model. S2: The edge intelligent all-in-one machine identifies the processing stage in real time and dynamically adjusts the acquisition strategy based on the monitoring configuration, and preprocesses and extracts features from the acquired signals to obtain a real-time feature stream; The physical information neural network, after being lightweighted from the proprietary baseline model, is used to perform reasoning on the real-time feature stream, which incorporates physical constraints, and outputs a physical state vector and an optimization decision vector. The optimization decision vector is then securely verified using a digital twin model on the local server, and the verification-passed correction instructions are written into the machine tool CNC system. S3: Update the parameters of the digital twin model according to the actual processing results after the execution of the correction instruction; periodically aggregate the local knowledge in each edge intelligent machine through federated knowledge distillation, generate and distribute the enhanced model or rule patch to each edge intelligent machine; S4: Display the global health status, machine tool details, decision log and dynamic perception topology generated based on the physical state vector through a visualization terminal, and trigger an early warning notification when the physical state vector indicates an abnormal status.

2. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The deployed sensors include: IEPE type triaxial vibration sensors installed on the spindle box, tool turret and workpiece fixture, and acoustic emission sensors installed on the tool holder or workpiece contact surface; the dual backup network is a dual-link redundant network composed of a time-sensitive network and a 5G private network, and is configured to automatically switch to the backup link when the latency of a single link continuously exceeds the threshold or the packet loss rate exceeds the standard.

3. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The calculation of the information entropy includes: performing short-term processing on the real-time sensor signals through a sliding time window; calculating the Shannon entropy of each sensor signal to quantify its uncertainty; calculating the transmission entropy between different sensor signals to quantify the information transmission strength and direction between signals; and constructing the information flow network of this process stage based on the transmission entropy value to form the dynamic sensing topology.

4. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The reasoning that integrates physical constraints includes: when the lightweight physical information neural network is inferring, its loss function integrates constraint terms based on the empirical model of cutting force and constraint terms based on the theory of flutter stability.

5. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The security verification includes: the local server simulating the optimized decision vector using a digital twin model; the verification pass criteria include prediction stability margin. ≥1.05, cutting force variation rate ≤10%, workpiece size error ≤±0.005mm.

6. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The federated knowledge distillation includes: each edge intelligent all-in-one machine encrypting and uploading the output trends of its local physical information neural network model under various preset virtual fault scenarios to the local server; the local server training a teacher model based on the aggregated output trends, extracting global common knowledge, and generating the enhanced model or diagnostic rule patch; and distributing the generated enhanced model or patch to each edge intelligent all-in-one machine for updating.

7. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The display of the overall health status includes: using machine tool icons of different colors to indicate the health level based on the physical state vector against the background of the workshop layout; and using machine tool icons of different sizes to indicate the load level based on the real-time production load.

8. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The display of machine tool details includes: A dashboard displaying key physical state parameters and their changing trends; Displays historical decision logs including time, diagnosed problems, implemented decisions, and results; Display a dynamic sensing topology diagram that reflects the information flow between sensors at the current process stage.

9. The method for real-time monitoring and diagnosis of metal cutting processes based on the Industrial Internet of Things as described in claim 1, characterized in that, The warning notification includes: multi-channel notification via at least one of the following methods: dashboard pop-up, mobile application push, and email; the content of the warning notification includes an anomaly description, urgency level, and handling suggestions.

Images