Numerical control machine tool machining environment monitoring system

Abstract

The invention discloses a numerical control machine tool machining environment monitoring system, and belongs to the technical field of intelligent monitoring. Comprising the following modules: a multi-dimensional sensing monitoring module for realizing omnibearing data acquisition of an interaction state of a cutter and a workpiece in a machining process; the feature extraction module is used for converting the time sequence data into a key feature set for representing the state of the cutter; the wear type identification module is used for accurately identifying and classifying the wear type of the cutter based on a deep learning classifier and judging the specific wear type; the wear progress prediction module calls a corresponding special prediction model according to the identified specific wear type, quantifies the wear progress rate and estimates the residual life of the cutter; the decision support module is used for integrating the tool wear state and the prediction result, balancing the production efficiency, the machining quality and the tool cost, and providing optimization suggestions of parameter adjustment and tool changing opportunities; and the self-learning optimization module is used for continuously collecting actual production data to carry out model evaluation and incremental learning so as to realize multi-machine knowledge sharing.

Description

Technical Field

[0001] The present application relates to the field of intelligent monitoring technology, and more specifically, to a CNC machine tool processing environment monitoring system. Background Art

[0002] With the continuous development of modern manufacturing, CNC machine tools (CNC) have become essential equipment in industrial production, playing an irreplaceable role in improving production efficiency, machining accuracy, and automation levels. However, due to the complexity and variability of the machining environment, CNC machine tools often face problems such as tool wear, machining errors, and equipment failures during long-term use. These issues not only affect product quality but also lead to reduced production efficiency and premature equipment wear, thereby increasing maintenance costs.

[0003] Traditional CNC machine tool monitoring systems rely heavily on manual inspections and scheduled maintenance. This approach is ineffective in identifying potential machine tool issues in a timely and effective manner and lacks real-time monitoring and early warning mechanisms for the machining process. Furthermore, existing monitoring technologies primarily focus on collecting and analyzing data from a single dimension, such as temperature and vibration, and lack multi-dimensional comprehensive perception and intelligent analysis.

[0004] In recent years, with the advancement of sensor technology, the Internet of Things, artificial intelligence (AI), and big data analytics, intelligent monitoring of industrial automation equipment has become a key means of improving production efficiency and reducing failure rates. By real-time monitoring of various environmental parameters of machine tools, combined with big data analysis and AI technologies, it is possible to effectively identify equipment operating status, predict equipment failures, and achieve intelligent maintenance and precise management. However, achieving comprehensive, accurate, and real-time monitoring of the CNC machine tool machining environment, and based on this, effective tool life prediction and machining quality control, remains a challenge in current technological development.

[0005] In summary, how to achieve multi-dimensional, high-precision, real-time data acquisition and analysis in the processing environment of CNC machine tools, and combine intelligent algorithms for tool wear prediction, fault diagnosis and process optimization, has become a technical problem that needs to be solved urgently. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, the present application aims to provide a CNC machine tool processing environment monitoring system, which includes the following modules: The multi-dimensional sensing monitoring module collects all-round data on the interaction between the tool and the workpiece during machining through an optimized layout of cutting force, vibration, acoustic emission, and temperature sensors. The feature extraction module processes the raw sensor signals and converts the time series data into a set of key features that characterize the tool status; Wear type recognition module, which accurately identifies and classifies tool wear types based on a deep learning classifier to determine the specific wear type; The wear progression prediction module uses a dedicated prediction model for the specific wear type identified to quantify the wear progression rate and estimate the remaining life of the tool. The decision support module integrates tool wear status and prediction results, balances production efficiency, processing quality and tool cost, and provides optimization suggestions for parameter adjustment and tool change timing; The self-learning optimization module continuously collects actual production data for model evaluation and incremental learning, enables multi-machine knowledge sharing, and continuously improves prediction accuracy and adaptability.

[0007] Furthermore, the optimized sensor layout in the multi-dimensional sensing monitoring module includes: The cutting force sensor is set between the workpiece fixture base and the machine tool table. It adopts a piezoelectric dynamic force sensor array, which consists of at least four high-sensitivity force sensing units arranged along the three orthogonal directions of X, Y, and Z to simultaneously measure the main cutting force, feed force and back force. The vibration sensor uses a three-axis acceleration sensor, which is installed on the tool holder close to the cutting area to capture the tiny vibration changes in the cutting process in real time; The acoustic emission sensor uses a broadband piezoelectric sensor, which is installed no more than 50 mm away from the cutting area and is equipped with a preamplifier with adjustable gain to capture microscopic damage signals; Temperature sensing monitoring uses a combination of infrared thermal imager and thermocouple to achieve non-contact temperature field distribution monitoring of the cutting area; A K-type thermocouple is embedded near the contact area between the tool and the workpiece to form a point-surface temperature monitoring network.

[0008] Furthermore, the multi-dimensional sensing monitoring module also includes an environmental factor monitoring unit, including a temperature and humidity sensor, a particulate matter sensor, a noise sensor, and an air quality sensor, wherein: The temperature and humidity sensors use a digital high-precision sensor array with no less than 6 measurement points set around the processing area. The measurement range is -40°C to 125°C and 0-100%RH, with measurement accuracy of ±0.1°C and ±1.5%RH respectively. The sampling frequency is 1Hz, and the temperature and humidity distribution and changes of the processing environment are monitored in real time. The particle sensor uses a laser scattering sensor to simultaneously monitor the concentration of particulate matter in three particle size ranges: PM1.0, PM2.5, and PM10. The measurement range is 0-1000μg / m 3 , with a resolution of 1 μg / m 3, with an accuracy of ±10% of reading and a sampling frequency of 0.5Hz, it is used to monitor metal dust and suspended particles generated during the processing; The noise sensor uses a high-precision digital sound level meter with a measurement range of 30-130dB(A), an accuracy of ±1.5dB, a frequency response range of 20Hz-20kHz, and a sampling frequency of 10Hz. It is used to analyze the noise distribution in different frequency bands and evaluate the noise level in the working environment. The air quality sensor integrates TVOC, CO, CO2 and O3 concentration detection functions to monitor harmful gases generated during the processing process. The TVOC measurement range is 0-5ppm with an accuracy of ±5% of reading; the CO measurement range is 0-1000ppm with an accuracy of ±3ppm; the CO2 measurement range is 0-5000ppm with an accuracy of ±30ppm; the O3 measurement range is 0-5ppm with an accuracy of ±0.02ppm, and the sampling frequency is 0.1Hz. Environmental monitoring data is transmitted to the edge computing gateway in real time via a low-power wireless network, and uploaded to the cloud platform after data fusion and pre-processing; An environmental impact assessment model is constructed based on environmental monitoring data to quantitatively analyze the impact of environmental factors on tool life and processing quality. The parameters of the wear prediction model are automatically adjusted according to the assessment results to improve prediction accuracy. At the same time, by establishing an environmental risk early warning mechanism, when environmental parameters exceed the safety threshold, an alarm is automatically issued and improvement suggestions are provided to ensure the health and safety of operators and the stable operation of equipment.

[0009] Furthermore, the signal processing flow of the feature extraction module includes: Multi-stage adaptive filtering is performed on the raw sensor signals, using a method that combines a Butterworth high-pass filter with wavelet decomposition and reconstruction to remove low-frequency noise and machine tool vibration interference. A moving average filter is used to smooth sudden changes in the cutting force signal. A Kalman filter algorithm is used to improve the signal-to-noise ratio of the vibration signal. The Hilbert-Huang transform is used to perform time-frequency analysis on the acoustic emission signal and extract instantaneous frequency features. The preprocessed time domain signal is subjected to fast Fourier transform to obtain spectral features, wavelet packet transform to obtain time-frequency features, and empirical mode decomposition to obtain modal features; The mean, standard deviation, peak-to-peak value, skewness, kurtosis, waveform factor, and crest factor features were extracted from the time domain. The energy distribution, main frequency, spectrum centroid, and spectrum entropy features of each frequency band were extracted from the frequency domain. The wavelet energy entropy, subband energy ratio, and modal energy ratio features were extracted from the time-frequency domain. These features were combined to form an initial feature vector space. Subsequently, a combination of principal component analysis and independent component analysis was applied to reduce feature dimensionality and reduce redundancy between features. Finally, an optimized feature set with a dimension of 30-50 was formed and normalized to the interval [-1,1]. A hybrid feature selection strategy based on the combination of packaging and filtering methods is adopted to adaptively select the most discriminative feature subsets for different wear types, forming a dedicated feature set for each type of wear mode.

[0010] Furthermore, the feature extraction module also includes a cross-domain feature fusion unit, which adopts a multi-level feature fusion strategy: The first level is sensor-level fusion, which aligns data from different sensors through a time synchronization mechanism and uses a Kalman filter to perform state estimation and data fusion on multi-source signals to improve measurement accuracy and anti-interference capabilities. The second level is feature-level fusion, which uses a combination of principal component analysis and independent component analysis to extract shared representations of multimodal data. A simplified matrix decomposition is used to discover the underlying feature structure. The contribution of different feature sources is dynamically adjusted through an empirical weight-based fusion method, and the fusion strategy is automatically optimized based on current processing conditions and historical performance. The third level is decision-level fusion, which designs dedicated post-fusion networks for different decision-making objectives and uses a stacking ensemble method combined with a weighted voting mechanism to generate the final decision result; Uncertainty assessment based on statistical methods is introduced into the fusion process, and the credibility of the prediction results is evaluated through variance analysis. For high-uncertainty areas, a deep analysis process is triggered, and a more sophisticated analysis model is called or manual confirmation is requested.

[0011] Furthermore, the deep learning classifier adopts a multi-stream hybrid neural network structure, including: The backbone network adopts a simplified architecture based on the combination of LSTM and CNN. The LSTM part consists of 2-3 layers of bidirectional LSTM units, each layer contains 128-256 hidden nodes, and the dropout rate is 0.3, which is used to capture long-term dependencies in temporal features. The convolutional neural network part adopts a lightweight CNN structure, including 3-4 convolutional blocks. Each convolutional block contains two convolutional layers and a residual connection. The convolution kernel sizes are 1×3, 1×5, and 1×7, and the number of convolution kernels is 64, 128, 256, and 384, respectively. The activation function uses LeakyReLU, combined with batch normalization layers and maximum pooling layers to extract multi-scale local features. Temporal features and multi-scale local features are fused through an attention mechanism, and the weights are dynamically adjusted. The model is trained in a two-stage strategy: first, basic feature extraction capabilities are trained on an existing dataset, and then fine-tuned for data specific to machining processes, materials, and tool types. Both stages use the Adam optimizer and a learning rate decay strategy. The loss function uses a weighted combination of cross-entropy loss and focal loss to address the imbalanced distribution of wear types. The wear types identified by the deep learning classifier include: rake face wear, flank face wear, tool tip wear, edge chipping, plastic deformation, adhesive wear, diffusion wear, oxidation wear and fatigue cracks; in addition, an anomaly detection mechanism is introduced to mark abnormal patterns that do not belong to known wear types.

[0012] Furthermore, the wear type identification module further includes an abnormal wear detection unit, which is implemented by the following steps: First, the latent space representation of normal wear patterns is learned based on the autoencoder, and the reconstruction error is calculated as the anomaly score; A simplified contrastive learning framework is introduced to generate positive and negative sample pairs through basic data augmentation techniques, and discriminative feature representations are learned, so that similar samples are clustered in the feature space and different samples are kept away from each other. Establish a density-based local anomaly factor model and combine it with traditional anomaly detection algorithms to evaluate the degree of sample anomaly from multiple perspectives; Build a change point detector based on statistical control charts to monitor sudden changes in wear patterns in real time. When an abnormality is detected that lasts longer than a preset threshold and the abnormality score is higher than the threshold, an abnormality alarm is triggered. Abnormal samples will be stored in a dedicated database and manually reviewed and labeled regularly to form new wear categories or subcategories.

[0013] Furthermore, the wear progression prediction module includes a dedicated prediction model library, which configures corresponding models for different wear types, including: a simplified LSTM model for uniform wear on the rake face; a prediction model based on time series analysis for crater wear on the rake face; a model based on a recurrent neural network for flank wear; a two-stage predictor for tool tip wear; and a simplified network with an attention mechanism for edge chipping. The prediction model selection mechanism is based on the wear type identification results and process parameters, and uses a rule-based decision framework to automatically call the most suitable prediction model combination; For composite wear types, a decomposition-combination strategy is adopted to first predict the progress of each single wear type separately, and then comprehensively evaluate the overall wear progress through the weighted average method; The remaining life assessment uses a statistical model based on historical data to quantify the prediction uncertainty as confidence intervals and output a remaining life estimate with a 90% confidence interval.

[0014] Furthermore, the decision support module adopts a multi-objective optimization framework, and its decision process includes: Construct a tool-process-quality mapping model based on experimental data, combining empirical rules and data statistics to quantify the impact of wear status on machining quality; Establish a production cost model that comprehensively considers tool cost, processing time cost, quality loss cost, and downtime loss cost; Based on the weighted summation method, a multi-objective optimization algorithm is designed to find the optimal balance between production efficiency and tool utilization while ensuring processing quality requirements; Three types of optimization suggestions are generated: the first type is real-time machining parameter adjustment suggestions. When the wear reaches the warning threshold but does not meet the replacement standard, the optimized cutting speed, feed rate and cutting depth are automatically calculated to slow down the wear process by reducing the cutting load; the second type is tool replacement timing suggestions. Based on the remaining life prediction and production plan, the economically optimal tool change point is calculated and an early warning is issued so that the workshop can reasonably arrange tool change operations and minimize unplanned downtime; the third type is process optimization suggestions. For specific wear types that occur frequently, the causes are analyzed and improvement suggestions are provided for tool material selection, geometric parameter optimization, cutting fluid application solutions and machining path planning.

[0015] Furthermore, the self-learning optimization module adopts a knowledge update architecture implemented in stages, including: The data collection and preprocessing layer continuously collects processing data locally, uses data desensitization to process sensitive information, combines basic data augmentation techniques to expand training samples, and corrects class imbalance problems using a combination of upsampling and downsampling methods; The local model training layer deploys an optimized training framework on the edge server and fine-tunes the model based on the characteristics of local data. The training process uses an early stopping strategy and adaptive learning rate adjustment, focusing on fine-tuning parameters in the latter layers of the model to reduce computing resource consumption. The model update and parameter sharing layer adopts a regular update strategy, weights contributions based on model performance increments, and adaptively adjusts the aggregation period. It ensures the security of the model parameter transmission process through a secure communication protocol, and uses data desensitization technology to prevent the model from inferring the original data. The performance evaluation and degradation detection layer establishes a comprehensive evaluation system including accuracy, precision, recall, F1 score, AUC value, and confusion matrix. It sets performance monitoring thresholds to trigger the model checking mechanism and locates model weaknesses through feature importance analysis and sensitivity testing. The knowledge base construction and application layer establishes a process knowledge base to store model parameters and performance characteristics under different processing conditions, realizes knowledge reuse under similar process conditions, and adapts to new scenarios using case-based reasoning methods combined with expert experience rules to quickly adapt to new processing conditions.

[0016] Compared with the prior art, this application has the following beneficial effects: This application uses a comprehensive application of multi-dimensional sensor monitoring, deep learning classification, wear progression prediction, and self-learning optimization modules to monitor and accurately predict CNC machine tool wear in real time, thereby optimizing the balance between tool usage, processing quality, and production efficiency, and achieving intelligent knowledge sharing and adaptive optimization in a multi-machine environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic structural diagram of a CNC machine tool processing environment monitoring system disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the present invention more apparent, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Throughout the drawings, identical or similar reference numerals represent identical or similar elements or elements having identical or similar functions. The described embodiments are only some, not all, of the embodiments of the present invention.

[0019] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0020] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0021] like Figure 1 As shown, a CNC machine tool processing environment monitoring system includes the following modules: The multi-dimensional sensing monitoring module collects all-round data on the interaction between the tool and the workpiece during machining through an optimized layout of cutting force, vibration, acoustic emission, and temperature sensors. The feature extraction module processes the raw sensor signals and converts the time series data into a set of key features that characterize the tool status; Wear type recognition module, which accurately identifies and classifies tool wear types based on a deep learning classifier to determine the specific wear type; The wear progression prediction module uses a dedicated prediction model for the specific wear type identified to quantify the wear progression rate and estimate the remaining life of the tool. The decision support module integrates tool wear status and prediction results, balances production efficiency, processing quality and tool cost, and provides optimization suggestions for parameter adjustment and tool change timing; The self-learning optimization module continuously collects actual production data for model evaluation and incremental learning, enables multi-machine knowledge sharing, and continuously improves prediction accuracy and adaptability.

[0022] In summary, this CNC machine tool machining environment monitoring system, through the collaborative operation of multiple modules, aims to monitor the interaction between the tool and the workpiece in real time and accurately diagnose and predict wear, thereby optimizing the machining process and improving production efficiency and quality. First, the multi-dimensional sensing monitoring module comprehensively collects data during the machining process using an optimized layout of cutting force, vibration, acoustic emission, and temperature sensors, ensuring comprehensive monitoring of tool conditions. The feature extraction module further processes the raw sensor signals, converting the time-series data into a set of key features reflecting the tool's condition, providing a reliable basis for subsequent analysis. The wear type identification module utilizes a deep learning classifier to accurately identify and classify tool wear types, determining the specific wear state and providing basic data for wear progression prediction. The wear progression prediction module uses a dedicated prediction model to quantify the wear progression rate for different wear types and estimate the remaining tool life, helping users prepare for tool replacement in advance. The decision support module integrates the tool wear state with the prediction results to provide a balanced solution between production efficiency, machining quality, and tool cost, optimizing parameter adjustments and tool change timing. Finally, the self-learning optimization module continuously collects actual production data, performs model evaluation and incremental learning, thereby enabling continuous optimization, improving prediction accuracy and adaptability, and supporting knowledge sharing among multiple machines.

[0023] Furthermore, the optimized sensor layout in the multi-dimensional sensing monitoring module includes: The cutting force sensor is set between the workpiece fixture base and the machine tool table. It adopts a piezoelectric dynamic force sensor array, which consists of at least four high-sensitivity force sensing units arranged along the three orthogonal directions of X, Y, and Z to simultaneously measure the main cutting force, feed force and back force. The vibration sensor uses a three-axis acceleration sensor, which is installed on the tool holder close to the cutting area to capture the tiny vibration changes in the cutting process in real time; The acoustic emission sensor uses a broadband piezoelectric sensor, which is installed no more than 50 mm away from the cutting area and is equipped with a preamplifier with adjustable gain to capture microscopic damage signals; Temperature sensing monitoring uses a combination of infrared thermal imager and thermocouple to achieve non-contact temperature field distribution monitoring of the cutting area; A K-type thermocouple is embedded near the contact area between the tool and the workpiece to form a point-surface temperature monitoring network.

[0024] The optimized layout of sensors precisely integrates different types of sensors to achieve comprehensive monitoring of the interaction between the tool and the workpiece during machining. First, the cutting force sensor is placed between the workpiece clamp base and the machine tool table, using a piezoelectric dynamic force sensor array, and at least four high-sensitivity force sensing units are arranged along the X, Y, and Z orthogonal directions. This layout can simultaneously measure the main cutting force, feed force, and back force, providing accurate real-time data for the tool stress state. Second, the vibration sensor is installed on the tool holder near the cutting area, using a three-axis acceleration sensor, which can capture real-time micro-vibration changes during cutting, reflecting the micro-dynamic changes when the tool contacts the workpiece, thereby effectively monitoring abnormal conditions during machining.

[0025] In terms of acoustic emission monitoring, a wideband piezoelectric sensor is used and installed within 50mm of the cutting area, equipped with a preamplifier with adjustable gain. This sensor can capture micro-damage signals during cutting, helping to monitor tool wear or damage in real time, thereby providing data support for wear type identification. Temperature monitoring uses a combination of infrared thermography and thermocouples. Infrared thermography can achieve non-contact temperature field distribution monitoring of the cutting area, while K-type thermocouples embedded in the tool and workpiece contact area form a point-surface combined temperature monitoring network, effectively capturing local temperature changes during cutting, thereby providing reliable basis for tool wear prediction and optimization decision-making.

[0026] Through the above optimized layout, comprehensive monitoring of cutting force, vibration, acoustic emission, and temperature during machining can be achieved, providing multi-dimensional and high-precision data support for tool wear diagnosis, progress prediction, and decision support modules, thereby improving overall machining efficiency and quality.5556 Further, the cutting force sensor further includes a distributed pressure mapping layer composed of at least 16x16 pressure sensitive units, each unit area not exceeding 0.5mm², and working pressure range of 0-2000N, to draw a pressure distribution heat map of the cutting area in real time, thereby accurately locating the position and pressure pattern of the wear point.

[0027] Further, the multi-dimensional sensing monitoring module further includes an environmental factor monitoring unit, including temperature and humidity sensors, particulate matter sensors, noise sensors, and air quality sensors, wherein: The temperature and humidity sensor uses a digital high-precision sensor array, with at least 6 measurement points set around the machining area, measuring range of -40℃ to 125℃ and 0-100%RH, measurement accuracy of ±0.1℃ and ±1.5%RH respectively, sampling frequency of 1Hz, real-time monitoring of temperature and humidity distribution and changes in the machining environment; The particulate matter sensor adopts a laser scattering sensor, which simultaneously monitors the concentrations of particulate matter in three particle size ranges of PM1.0, PM2.5 and PM10, with a measurement range of 0-1000 μg / m 3 , a resolution of 1 μg / m 3 , an accuracy of ±10% reading, and a sampling frequency of 0.5 Hz, which is used to monitor the metal dust and suspended particulate matter generated in the machining process. The noise sensor adopts a high-precision digital sound level meter, with a measurement range of 30-130 dB(A), an accuracy of ±1.5 dB, a frequency response range of 20 Hz-20 kHz, and a sampling frequency of 10 Hz, which is used to analyze the noise distribution in different frequency bands and evaluate the noise level in the working environment. The air quality sensor integrates TVOC, CO, CO2 and O3 concentration detection functions, which are used to monitor harmful gases generated in the machining process. The TVOC measurement range is 0-5 ppm, the accuracy is ±5% reading; the CO measurement range is 0-1000 ppm, the accuracy is ±3 ppm; the CO2 measurement range is 0-5000 ppm, the accuracy is ±30 ppm; the O3 measurement range is 0-5 ppm, the accuracy is ±0.02 ppm, and the sampling frequency is 0.1 Hz. The environmental monitoring data is transmitted in real time to the edge computing gateway through a low-power wireless network, and after data fusion and preprocessing, it is uploaded to the cloud platform. Based on the environmental monitoring data, an environmental impact evaluation model is constructed to quantitatively analyze the influence of environmental factors on tool life and machining quality. According to the evaluation results, the parameters of the wear prediction model are automatically adjusted to improve the prediction accuracy. At the same time, through the establishment of an environmental risk early warning mechanism, when the environmental parameters exceed the safety threshold, an alarm is automatically sent and improvement suggestions are provided to protect the health and safety of operators and the stable operation of equipment.

[0028] Further, the environmental factor monitoring unit also includes an electrostatic field intensity monitoring sensor with a measurement range of ±100 kV / m, an accuracy of ±0.5 kV / m, and a sampling frequency of 2 Hz. Combined with the relative humidity data, an electrostatic safety evaluation model is constructed. When the detected electrostatic field intensity exceeds the set threshold and the environmental relative humidity is lower than 30%, the ion fan is automatically started to neutralize static electricity, preventing the accumulation of metal dust and electrostatic discharge from damaging electronic components.

[0029] Further, the signal processing procedure of the feature extraction module includes: Multi-stage adaptive filtering is performed on the raw sensor signals, using a method that combines a Butterworth high-pass filter with wavelet decomposition and reconstruction to remove low-frequency noise and machine tool vibration interference. A moving average filter is used to smooth sudden changes in the cutting force signal. A Kalman filter algorithm is used to improve the signal-to-noise ratio of the vibration signal. The Hilbert-Huang transform is used to perform time-frequency analysis on the acoustic emission signal and extract instantaneous frequency features. The preprocessed time domain signal is subjected to fast Fourier transform to obtain spectral features, wavelet packet transform to obtain time-frequency features, and empirical mode decomposition to obtain modal features; The mean, standard deviation, peak-to-peak value, skewness, kurtosis, waveform factor, and crest factor features were extracted from the time domain. The energy distribution, main frequency, spectrum centroid, and spectrum entropy features of each frequency band were extracted from the frequency domain. The wavelet energy entropy, subband energy ratio, and modal energy ratio features were extracted from the time-frequency domain. These features were combined to form an initial feature vector space. Subsequently, a combination of principal component analysis and independent component analysis was applied to reduce feature dimensionality and reduce redundancy between features. Finally, an optimized feature set with a dimension of 30-50 was formed and normalized to the interval [-1,1]. A hybrid feature selection strategy based on the combination of packaging and filtering methods is adopted to adaptively select the most discriminative feature subsets for different wear types, forming a dedicated feature set for each type of wear mode.

[0030] Furthermore, the feature extraction module also includes a decay response feature extraction unit, which applies standardized micro-impacts to the tool during the cutting gap through a triggered micro-pulse excitation method, records the attenuation curve of the response signal, extracts the damping ratio, natural frequency offset and energy attenuation rate parameters, and forms a tool structure health index for evaluating the internal microcracks and structural fatigue state of the tool.

[0031] Furthermore, the feature extraction module also includes a cross-domain feature fusion unit, which adopts a multi-level feature fusion strategy: The first level is sensor-level fusion, which aligns data from different sensors through a time synchronization mechanism and uses a Kalman filter to perform state estimation and data fusion on multi-source signals to improve measurement accuracy and anti-interference capabilities. The second level is feature-level fusion, which uses a combination of principal component analysis and independent component analysis to extract shared representations of multimodal data. A simplified matrix decomposition is used to discover the underlying feature structure. The contribution of different feature sources is dynamically adjusted through an empirical weight-based fusion method, and the fusion strategy is automatically optimized based on current processing conditions and historical performance. The third level is decision-level fusion, which designs dedicated post-fusion networks for different decision-making objectives and uses a stacking integration method combined with a weighted voting mechanism to generate the final decision result; Uncertainty assessment based on statistical methods is introduced into the fusion process, and the credibility of the prediction results is evaluated through variance analysis. For high-uncertainty areas, a deep analysis process is triggered, and a more sophisticated analysis model is called or manual confirmation is requested.

[0032] Furthermore, the cross-domain feature fusion unit adopts tensor decomposition technology to realize multimodal data fusion, specifically including: organizing multi-sensor data into a three-order tensor structure, including three dimensions of time, frequency and sensor type; using Tucker decomposition method to extract core tensors and factor matrices, and the core tensor size is set to 30%-50% of the original dimension; determining key modal combinations based on the energy distribution of core tensor elements, and retaining 85%-95% of the information through adaptive threshold screening; and finally obtaining a noise-reduced and information-concentrated fusion feature representation through tensor reconstruction.

[0033] Furthermore, the deep learning classifier adopts a multi-stream hybrid neural network structure, including: The backbone network adopts a simplified architecture based on the combination of LSTM and CNN. The LSTM part consists of 2-3 layers of bidirectional LSTM units, each layer contains 128-256 hidden nodes, and the dropout rate is 0.3, which is used to capture long-term dependencies in temporal features. The convolutional neural network part adopts a lightweight CNN structure, including 3-4 convolutional blocks. Each convolutional block contains two convolutional layers and a residual connection. The convolution kernel sizes are 1×3, 1×5, and 1×7, and the number of convolution kernels is 64, 128, 256, and 384, respectively. The activation function uses LeakyReLU, combined with batch normalization layers and maximum pooling layers to extract multi-scale local features. Temporal features and multi-scale local features are fused through an attention mechanism, and the weights are dynamically adjusted. The model is trained in a two-stage strategy: first, basic feature extraction capabilities are trained on an existing dataset, and then fine-tuned for data specific to machining processes, materials, and tool types. Both stages use the Adam optimizer and a learning rate decay strategy. The loss function uses a weighted combination of cross-entropy loss and focal loss to address the imbalanced distribution of wear types. The wear types identified by the deep learning classifier include: rake face wear, flank face wear, tool tip wear, edge chipping, plastic deformation, adhesive wear, diffusion wear, oxidation wear and fatigue cracks; in addition, an anomaly detection mechanism is introduced to mark abnormal patterns that do not belong to known wear types.

[0034] Furthermore, the deep learning classifier also includes a gradient wear fine classification layer, which subdivides each main wear type into three stages: early, middle and late. It adopts a hierarchical classification structure based on Focal Loss to first identify the main wear type and then determine the stage it is in, forming a fine classification system of 9 main wear types × 3 stages; and introduces a data enhancement strategy based on Mixup to generate transition samples between different stages, thereby improving the model's ability to recognize boundary states.

[0035] Furthermore, the wear type identification module further includes an abnormal wear detection unit, which is implemented by the following steps: First, the latent space representation of normal wear patterns is learned based on the autoencoder, and the reconstruction error is calculated as the anomaly score; A simplified contrastive learning framework is introduced to generate positive and negative sample pairs through basic data augmentation techniques, and discriminative feature representations are learned, so that similar samples are clustered in the feature space and different samples are kept away from each other. Establish a density-based local anomaly factor model and combine it with traditional anomaly detection algorithms to evaluate the degree of sample anomaly from multiple perspectives; Build a change point detector based on statistical control charts to monitor sudden changes in wear patterns in real time. When an abnormality is detected that lasts longer than a preset threshold and the abnormality score is higher than the threshold, an abnormality alarm is triggered. Abnormal samples will be stored in a dedicated database and manually reviewed and labeled regularly to form new wear categories or subcategories.

[0036] Furthermore, the abnormal wear detection unit also includes a spatiotemporal pattern deviation detector, which constructs a prediction band of the normal wear progression trajectory based on the time series prediction model, calculates the deviation between the observed value and the prediction band in real time, and designs an adaptive threshold mechanism. When the deviation of more than three consecutive data points exceeds 3σ or the deviation of any data point exceeds 5σ, an abnormal pattern warning is triggered; at the same time, a normal area envelope surface is constructed in the feature vector space, and the degree of deviation of the sample point from the normal area is measured by the Mahalanobis distance to achieve early detection of high-dimensional abnormal patterns.

[0037] Furthermore, the wear progression prediction module includes a dedicated prediction model library, which configures corresponding models for different wear types, including: a simplified LSTM model for uniform wear on the rake face; a prediction model based on time series analysis for crater wear on the rake face; a model based on a recurrent neural network for flank wear; a two-stage predictor for tool tip wear; and a simplified network with an attention mechanism for edge chipping. The prediction model selection mechanism is based on the wear type identification results and process parameters, and uses a rule-based decision framework to automatically call the most suitable prediction model combination; For composite wear types, a decomposition-combination strategy is adopted to first predict the progress of each single wear type separately, and then comprehensively evaluate the overall wear progress through the weighted average method; The remaining life assessment uses a statistical model based on historical data to quantify the prediction uncertainty as confidence intervals and output a remaining life estimate with a 90% confidence interval.

[0038] Furthermore, the wear progression prediction module also includes an uncertainty quantification unit, which adopts an integrated learning framework and specifically includes: The infrastructure uses a multi-model ensemble approach to combine the outputs of multiple basic prediction models. Each basic model uses a different subset of training data and features. The network structure uses a 3-4 layer feedforward neural network, with 128, 64, 32, and 16 nodes per layer, respectively. The activation function is soft tangent, and the parameters are initialized using the standard normal distribution. The training process uses the bootstrap method to generate multiple training sets, and optimization is performed by minimizing the mean square error. The model outputs are mean and prediction intervals. The uncertainty of the prediction is estimated based on the discrete output of the ensemble model. The prediction intervals are obtained using the quantile regression method. Uncertainty is simplified to overall uncertainty, and reliability is assessed through prediction deviation analysis. Decision strategies are adaptively adjusted based on the uncertainty level, triggering additional data collection mechanisms for high-uncertainty areas, prioritizing the collection of labeled data in these areas to improve the model. A multi-sensor validation process is triggered for areas of high data uncertainty, cross-checking different sensor signals to improve reliability.

[0039] Furthermore, the uncertainty quantification unit adopts a strategy to separate epistemic uncertainty from stochastic uncertainty, obtains epistemic uncertainty estimates through the Dropout network, and obtains stochastic uncertainty estimates by modeling observation noise with a Gaussian mixture model; automatically adjusts the data acquisition strategy according to the ratio of the two types of uncertainty: for areas with high epistemic uncertainty, active learning sampling is prioritized, and the most informative samples are selected for labeling through an uncertainty-guided query strategy; for areas with high stochastic uncertainty, sensor redundancy and sampling frequency are increased; and the confidence interval of the prediction results is dynamically adjusted based on the level of epistemic uncertainty.

[0040] Furthermore, the decision support module adopts a multi-objective optimization framework, and its decision process includes: Construct a tool-process-quality mapping model based on experimental data, combining empirical rules and data statistics to quantify the impact of wear status on machining quality; Establish a production cost model that comprehensively considers tool cost, processing time cost, quality loss cost, and downtime loss cost; Based on the weighted summation method, a multi-objective optimization algorithm is designed to find the optimal balance between production efficiency and tool utilization while ensuring processing quality requirements; Three types of optimization suggestions are generated: the first type is real-time machining parameter adjustment suggestions. When the wear reaches the warning threshold but does not meet the replacement standard, the optimized cutting speed, feed rate and cutting depth are automatically calculated to slow down the wear process by reducing the cutting load; the second type is tool replacement timing suggestions. Based on the remaining life prediction and production plan, the economically optimal tool change point is calculated and an early warning is issued so that the workshop can reasonably arrange tool change operations and minimize unplanned downtime; the third type is process optimization suggestions. For specific wear types that occur frequently, the causes are analyzed and improvement suggestions are provided for tool material selection, geometric parameter optimization, cutting fluid application solutions and machining path planning.

[0041] Furthermore, the multi-objective optimization framework also includes a processing surface quality prediction unit, which predicts the surface quality indicators under the current wear state in real time by establishing a mapping relationship between the wear state and the surface quality parameters; at the same time, a parameter optimization engine based on the particle swarm algorithm is introduced to search for the optimal process parameter combination in the parameter space of four dimensions of cutting speed, feed speed, cutting depth and coolant flow rate for specific quality requirements. The search process is limited to the process safety constraints and balances the processing efficiency and surface quality requirements.

[0042] Furthermore, the decision support module also includes a multi-machine collaborative optimization unit, which analyzes the wear data of machine tools of the same model to establish a correlation model between tool wear patterns and machine tool vibration characteristics, structural stiffness and spindle dynamic balance status; based on the wear differences of the same batch of tools on different machine tools, the tool adaptability index is calculated to provide tool-machine matching recommendations, and customized cutting parameters are adjusted according to the unique characteristics of each machine tool to achieve overall optimization of tool utilization efficiency.

[0043] Furthermore, the self-learning optimization module adopts a knowledge update architecture implemented in stages, including: The data collection and preprocessing layer continuously collects processing data locally, uses data desensitization to process sensitive information, combines basic data augmentation techniques to expand training samples, and corrects class imbalance problems using a combination of upsampling and downsampling methods; The local model training layer deploys an optimized training framework on the edge server and fine-tunes the model based on the characteristics of local data. The training process uses an early stopping strategy and adaptive learning rate adjustment, focusing on fine-tuning parameters in the latter layers of the model to reduce computing resource consumption. The model update and parameter sharing layer adopts a regular update strategy, weights contributions based on model performance increments, and adaptively adjusts the aggregation period. It ensures the security of the model parameter transmission process through a secure communication protocol, and uses data desensitization technology to prevent the model from inferring the original data. The performance evaluation and degradation detection layer establishes a comprehensive evaluation system including accuracy, precision, recall, F1 score, AUC value, and confusion matrix. It sets performance monitoring thresholds to trigger the model checking mechanism and locates model weaknesses through feature importance analysis and sensitivity testing. The knowledge base construction and application layer establishes a process knowledge base to store model parameters and performance characteristics under different processing conditions, realizes knowledge reuse under similar process conditions, and adapts to new scenarios using case-based reasoning methods combined with expert experience rules to quickly adapt to new processing conditions.

[0044] Furthermore, the model update of the self-learning optimization module adopts a transfer learning framework, and solves the distribution offset problem between different machine tools, different materials and different tool types through domain adaptation technology; the specific implementation includes: designing a domain-invariant feature extractor, and making the feature representation insensitive to the machine tool type through adversarial training methods; adopting a progressive parameter fine-tuning strategy, first freezing the feature extraction layer, only fine-tuning the decision layer, and then determining whether to fine-tune deeper layers based on the degree of performance improvement; reducing the feature distribution difference between the source domain and the target domain through the maximum mean difference minimization method; thereby achieving a performance level equivalent to complete retraining under new working conditions with only 10%-20% labeled data.

[0045] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art will appreciate that modifications may be made to the technical solutions described in the above embodiments, or that some of the technical features may be replaced with equivalents; such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A CNC machine tool processing environment monitoring system, characterized in that: Includes the following modules: The multi-dimensional sensing monitoring module collects all-round data on the interaction between the tool and the workpiece during machining through an optimized layout of cutting force, vibration, acoustic emission, and temperature sensors. The feature extraction module processes the raw sensor signals and converts the time series data into a set of key features that characterize the tool status; Wear type recognition module, which accurately identifies and classifies tool wear types based on a deep learning classifier to determine the specific wear type; The wear progression prediction module uses a dedicated prediction model for the specific wear type identified to quantify the wear progression rate and estimate the remaining life of the tool. The decision support module integrates tool wear status and prediction results, balances production efficiency, processing quality and tool cost, and provides optimization suggestions for parameter adjustment and tool change timing; The self-learning optimization module continuously collects actual production data for model evaluation and incremental learning, enables multi-machine knowledge sharing, and continuously improves prediction accuracy and adaptability.

2. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The optimized sensor layout in the multi-dimensional sensing monitoring module includes: The cutting force sensor is set between the workpiece fixture base and the machine tool table. It adopts a piezoelectric dynamic force sensor array, which consists of at least four high-sensitivity force sensing units arranged along the three orthogonal directions of X, Y, and Z to simultaneously measure the main cutting force, feed force and back force. The vibration sensor uses a three-axis acceleration sensor, which is installed on the tool holder close to the cutting area to capture the tiny vibration changes in the cutting process in real time; The acoustic emission sensor uses a broadband piezoelectric sensor, which is installed no more than 50 mm away from the cutting area and is equipped with a preamplifier with adjustable gain to capture microscopic damage signals; Temperature sensing monitoring uses a combination of infrared thermal imager and thermocouple to achieve non-contact temperature field distribution monitoring of the cutting area; A K-type thermocouple is embedded near the contact area between the tool and the workpiece to form a point-surface temperature monitoring network.

3. A CNC machine tool processing environment monitoring system according to claim 2, characterized in that: The multi-dimensional sensing monitoring module also includes an environmental factor monitoring unit, including a temperature and humidity sensor, a particulate matter sensor, a noise sensor, and an air quality sensor, wherein: The temperature and humidity sensors use a digital high-precision sensor array with no less than 6 measurement points set around the processing area. The measurement range is -40°C to 125°C and 0-100%RH, with measurement accuracy of ±0.1°C and ±1.5%RH respectively. The sampling frequency is 1Hz, and the temperature and humidity distribution and changes of the processing environment are monitored in real time. The particle sensor uses a laser scattering sensor to simultaneously monitor the concentration of particulate matter in three particle size ranges: PM1.0, PM2.5, and PM10. The measurement range is 0-1000μg / m 3 , with a resolution of 1 μg / m 3 , with an accuracy of ±10% of reading and a sampling frequency of 0.5Hz, it is used to monitor metal dust and suspended particles generated during the processing; The noise sensor uses a high-precision digital sound level meter with a measurement range of 30-130dB(A), an accuracy of ±1.5dB, a frequency response range of 20Hz-20kHz, and a sampling frequency of 10Hz. It is used to analyze the noise distribution in different frequency bands and evaluate the noise level in the working environment. The air quality sensor integrates TVOC, CO, CO2 and O3 concentration detection functions to monitor harmful gases generated during the processing process. The TVOC measurement range is 0-5ppm with an accuracy of ±5% of reading; the CO measurement range is 0-1000ppm with an accuracy of ±3ppm; the CO2 measurement range is 0-5000ppm with an accuracy of ±30ppm; the O3 measurement range is 0-5ppm with an accuracy of ±0.02ppm, and the sampling frequency is 0.1Hz. Environmental monitoring data is transmitted to the edge computing gateway in real time via a low-power wireless network, and uploaded to the cloud platform after data fusion and pre-processing; An environmental impact assessment model is constructed based on environmental monitoring data to quantitatively analyze the impact of environmental factors on tool life and processing quality. The parameters of the wear prediction model are automatically adjusted according to the assessment results to improve prediction accuracy. At the same time, by establishing an environmental risk early warning mechanism, when environmental parameters exceed the safety threshold, an alarm is automatically issued and improvement suggestions are provided to ensure the health and safety of operators and the stable operation of equipment.

4. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The signal processing flow of the feature extraction module includes: Multi-stage adaptive filtering is performed on the raw sensor signals, using a method that combines a Butterworth high-pass filter with wavelet decomposition and reconstruction to remove low-frequency noise and machine tool vibration interference. A moving average filter is used to smooth sudden changes in the cutting force signal. A Kalman filter algorithm is used to improve the signal-to-noise ratio of the vibration signal. The Hilbert-Huang transform is used to perform time-frequency analysis on the acoustic emission signal and extract instantaneous frequency features. The preprocessed time domain signal is subjected to fast Fourier transform to obtain spectral features, wavelet packet transform to obtain time-frequency features, and empirical mode decomposition to obtain modal features; The mean, standard deviation, peak-to-peak value, skewness, kurtosis, waveform factor, and crest factor features were extracted from the time domain. The energy distribution, main frequency, spectrum centroid, and spectrum entropy features of each frequency band were extracted from the frequency domain. The wavelet energy entropy, subband energy ratio, and modal energy ratio features were extracted from the time-frequency domain. These features were combined to form an initial feature vector space. Subsequently, a combination of principal component analysis and independent component analysis was applied to reduce feature dimensionality and reduce redundancy between features. Finally, an optimized feature set with a dimension of 30-50 was formed and normalized to the interval [-1,1]. A hybrid feature selection strategy based on the combination of packaging and filtering methods is adopted to adaptively select the most discriminative feature subsets for different wear types, forming a dedicated feature set for each type of wear mode.

5. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The feature extraction module also includes a cross-domain feature fusion unit, which adopts a multi-level feature fusion strategy: The first level is sensor-level fusion, which aligns data from different sensors through a time synchronization mechanism and uses a Kalman filter to perform state estimation and data fusion on multi-source signals to improve measurement accuracy and anti-interference capabilities. The second level is feature-level fusion, which uses a combination of principal component analysis and independent component analysis to extract shared representations of multimodal data. A simplified matrix decomposition is used to discover the underlying feature structure. The contribution of different feature sources is dynamically adjusted through an empirical weight-based fusion method, and the fusion strategy is automatically optimized based on current processing conditions and historical performance. The third level is decision-level fusion, which designs dedicated post-fusion networks for different decision-making objectives and uses a stacking integration method combined with a weighted voting mechanism to generate the final decision result; Uncertainty assessment based on statistical methods is introduced into the fusion process, and the credibility of the prediction results is evaluated through variance analysis. For high-uncertainty areas, a deep analysis process is triggered, and a more sophisticated analysis model is called or manual confirmation is requested.

6. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The deep learning classifier adopts a multi-stream hybrid neural network structure, including: The backbone network adopts a simplified architecture based on the combination of LSTM and CNN. The LSTM part consists of 2-3 layers of bidirectional LSTM units, each layer contains 128-256 hidden nodes, and the dropout rate is 0.3, which is used to capture long-term dependencies in temporal features. The convolutional neural network part adopts a lightweight CNN structure, including 3-4 convolutional blocks. Each convolutional block contains two convolutional layers and a residual connection. The convolution kernel sizes are 1×3, 1×5, and 1×7, and the number of convolution kernels is 64, 128, 256, and 384, respectively. The activation function uses LeakyReLU, combined with batch normalization layers and maximum pooling layers to extract multi-scale local features. Temporal features and multi-scale local features are fused through an attention mechanism, and the weights are dynamically adjusted. The model is trained in a two-stage strategy: first, basic feature extraction capabilities are trained on an existing dataset, and then fine-tuned for data specific to machining processes, materials, and tool types. Both stages use the Adam optimizer and a learning rate decay strategy. The loss function uses a weighted combination of cross-entropy loss and focal loss to address the imbalanced distribution of wear types. The wear types identified by the deep learning classifier include: rake face wear, flank face wear, tool tip wear, edge chipping, plastic deformation, adhesive wear, diffusion wear, oxidation wear and fatigue cracks; in addition, an anomaly detection mechanism is introduced to mark abnormal patterns that do not belong to known wear types.

7. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The wear type identification module also includes an abnormal wear detection unit, which is implemented by the following steps: First, the latent space representation of normal wear patterns is learned based on the autoencoder, and the reconstruction error is calculated as the anomaly score; A simplified contrastive learning framework is introduced to generate positive and negative sample pairs through basic data augmentation techniques, and discriminative feature representations are learned, so that similar samples are clustered in the feature space and different samples are kept away from each other. Establish a density-based local anomaly factor model and combine it with traditional anomaly detection algorithms to evaluate the degree of sample anomaly from multiple perspectives; Build a change point detector based on statistical control charts to monitor sudden changes in wear patterns in real time. When an abnormality is detected that lasts longer than a preset threshold and the abnormality score is higher than the threshold, an abnormality alarm is triggered. Abnormal samples will be stored in a dedicated database and manually reviewed and labeled regularly to form new wear categories or subcategories.

8. The CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The wear progression prediction module includes a dedicated prediction model library with corresponding models configured for different wear types, including: a simplified LSTM model for uniform wear on the rake face; a prediction model based on time series analysis for crater wear on the rake face; a model based on a recurrent neural network for flank wear; a two-stage predictor for tool tip wear; and a simplified network with an attention mechanism for edge chipping. The prediction model selection mechanism is based on the wear type identification results and process parameters, and uses a rule-based decision framework to automatically call the most suitable prediction model combination; For composite wear types, a decomposition-combination strategy is adopted to first predict the progress of each single wear type separately, and then comprehensively evaluate the overall wear progress through the weighted average method; The remaining life assessment uses a statistical model based on historical data to quantify the prediction uncertainty as confidence intervals and output a remaining life estimate with a 90% confidence interval.

9. A CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The decision support module adopts a multi-objective optimization framework, and its decision process includes: Construct a tool-process-quality mapping model based on experimental data, combining empirical rules and data statistics to quantify the impact of wear status on machining quality; Establish a production cost model that comprehensively considers tool cost, processing time cost, quality loss cost, and downtime loss cost; Based on the weighted summation method, a multi-objective optimization algorithm is designed to find the optimal balance between production efficiency and tool utilization while ensuring processing quality requirements; Three types of optimization suggestions are generated: the first type is real-time machining parameter adjustment suggestions. When the wear reaches the warning threshold but does not meet the replacement standard, the optimized cutting speed, feed rate and cutting depth are automatically calculated to delay the wear process by reducing the cutting load; the second type is tool replacement timing suggestions. Based on the remaining life prediction and production plan, the economically optimal tool change point is calculated and an early warning is issued so that the workshop can reasonably arrange tool change operations and minimize unplanned downtime; the third type is process optimization suggestions. For specific wear types that occur frequently, the causes are analyzed and improvement suggestions are provided for tool material selection, geometric parameter optimization, cutting fluid application solutions and machining path planning.

10. The CNC machine tool processing environment monitoring system according to claim 1, characterized in that: The self-learning optimization module adopts a knowledge update architecture implemented in stages, including: The data collection and preprocessing layer continuously collects processing data locally, uses data desensitization to process sensitive information, combines basic data augmentation techniques to expand training samples, and corrects class imbalance problems using a combination of upsampling and downsampling methods; The local model training layer deploys an optimized training framework on the edge server and fine-tunes the model based on the characteristics of local data. The training process uses an early stopping strategy and adaptive learning rate adjustment, focusing on fine-tuning parameters in the latter layers of the model to reduce computing resource consumption. The model update and parameter sharing layer adopts a regular update strategy, weights contributions based on model performance increments, and adaptively adjusts the aggregation period. It ensures the security of the model parameter transmission process through a secure communication protocol, and uses data desensitization technology to prevent the model from inferring the original data. The performance evaluation and degradation detection layer establishes a comprehensive evaluation system including accuracy, precision, recall, F1 score, AUC value, and confusion matrix. It sets performance monitoring thresholds to trigger the model checking mechanism and locates model weaknesses through feature importance analysis and sensitivity testing. The knowledge base construction and application layer establishes a process knowledge base to store model parameters and performance characteristics under different processing conditions, realizes knowledge reuse under similar process conditions, and adapts to new scenarios using case-based reasoning methods combined with expert experience rules to quickly adapt to new processing conditions.

Images