Tool life prediction method and system for CNC machine tools based on deep learning

Through deep learning technology, dynamic mutant autoencoder and ADNN model are used, combined with predictive feedback loops, the problems of adaptability and real-time updates in tool life prediction of CNC machine tools are solved, and efficient and accurate life prediction is achieved.

CN118378373BActive Publication Date: 2024-11-29SHENZHEN XIBOER TECH CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410409354.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-07
Publication Date
2024-11-29
Estimated Expiration
2044-04-07

AI Technical Summary

Technical Problem

The tool life prediction technology of existing CNC machine tools is difficult to adapt to different processing conditions and tool types, and has insufficient generalization capabilities and fails to make full use of the complex information in the monitoring data, resulting in low prediction accuracy and inability to update in real time, which cannot accurately reflect the tool wear process. Dynamic.

Method used

Using a deep learning-based method, using dynamic variant autoencoder and adaptive dynamic neural network (ADNN) model, through a predictive feedback loop mechanism, the key features in the monitoring data are automatically extracted, the network structure and depth are dynamically adjusted, and real-time updates are achieved. and optimize to improve the adaptability and accuracy of the model.

Benefits of technology

It significantly improves the accuracy and practicality of tool life prediction, can timely reflect the actual wear status of the tool, enhances the generalization ability and adaptability of the model, and provides an efficient and accurate life prediction path.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118378373B_ABST
    Figure CN118378373B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of intelligent application technology of mechanical engineering, and in particular to a method and system for predicting the life of a CNC machine tool tool based on deep learning. The present invention can automatically extract key features from complex monitoring data and capture the inherent abnormal patterns and nonlinear characteristics of the data by using a dynamic variation autoencoder; an adaptive dynamic neural network (ADNN) is used to dynamically adjust the model structure according to data characteristics to improve the model's adaptability to different processing conditions and tool types; a predictive feedback loop is introduced to achieve self-optimization of the model during the prediction process, ensuring that the prediction results accurately reflect the wear state of the tool in real time. The present invention significantly improves the accuracy and practicality of life prediction, and provides effective technical support for the efficient management and use of CNC machine tool tools.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent application technology of mechanical engineering, and in particular to a method and system for predicting tool life of CNC machine tools based on deep learning. Background Art

[0002] In modern manufacturing, the life prediction technology of CNC machine tool tools plays a vital role. The life of tools not only directly affects the processing quality and production efficiency, but also is related to cost control and production safety. Accurate prediction of tool life can reasonably arrange the replacement time of tools, avoid the scrapping of workpieces due to excessive tool wear, and improve production efficiency and reduce costs. However, tool life is affected by many factors, including tool material, processing material, cutting parameters (such as speed, feed rate and cutting depth), and machine tool working conditions. The complex interaction of these factors makes life prediction a challenging problem.

[0003] There are several major problems in predicting tool life for CNC machine tools: tool life is affected by many factors, and each factor has its variability. It is difficult to extract useful information for life prediction from complex data; existing prediction models are often difficult to adapt to different processing conditions and tool types, resulting in insufficient generalization ability of the model and inability to accurately predict tool life under new processing conditions; many existing prediction methods fail to fully consider the dynamics and real-time changes of the tool wear process, resulting in deviations between the prediction results and the actual situation.

[0004] The existing CNC machine tool tool life prediction technology (Chinese invention patent, publication number: CN116533063A), although it can achieve life prediction to a certain extent, still has the following shortcomings or defects:

[0005] Existing technologies often rely on empirically selected features or simple statistical data, failing to fully utilize the complex information contained in tool monitoring data, resulting in low prediction accuracy;

[0006] Many prediction models are difficult to automatically adapt to different machining conditions or tool types and require manual adjustment of model parameters, which limits the scope of application and efficiency of the models.

[0007] The lack of effective methods to update the prediction model in real time so that it reflects the latest situation in the tool wear process causes the prediction results to lag behind the actual wear status. Summary of the invention

[0008] In view of the many problems existing in the above-mentioned prior art, the present invention provides a method and system for predicting the tool life of CNC machine tools based on deep learning. The present invention adopts a dynamic variation autoencoder, which can automatically extract key features that are helpful for life prediction from complex monitoring data, thereby improving the prediction accuracy; through an adaptive dynamic neural network (ADNN) model, the network structure and depth can be dynamically adjusted according to the characteristics of the input data, thereby improving the generalization ability and adaptability of the model; a predictive feedback loop mechanism is introduced to realize real-time updating and dynamic optimization of the model, so that the prediction results can timely reflect the actual wear status of the tool. The present invention can effectively solve the problems and challenges encountered by the prior art in predicting the tool life of CNC machine tools, and provides a new technical path for efficient and accurate tool life prediction.

[0009] A method for predicting tool life of CNC machine tools based on deep learning, comprising the following steps:

[0010] Collecting the monitoring data of the tool and performing standardization and denoising preprocessing, wherein the monitoring data includes: vibration data, sound data and temperature data;

[0011] The dynamic variation autoencoder is used to extract and encode the monitoring data, capture abnormal patterns and nonlinear characteristics, and generate an encoded feature data set; the feature data set is input into the ADNN model, and the network structure and depth are dynamically adjusted according to the characteristics of the input data to generate the initial life prediction results;

[0012] Through a predictive feedback loop, the initial wear prediction results are compared with the actual observed tool life, and the ADNN model is self-adjusted based on the comparison results to produce an optimized ADNN model and update the initial life prediction results;

[0013] Before the predictive feedback loop is started, the hybrid data source ensemble learning method is applied to fuse and weight the feature data set, obtain the fused feature data and input it into the optimized ADNN model for final life prediction, and output the final life prediction result of the tool;

[0014] The actual tool usage data and the final life prediction results are comprehensively considered for performance evaluation. The feature extraction, ADNN model adjustment and feature fusion weights are adjusted based on the evaluation results to comprehensively optimize the life prediction performance.

[0015] Preferably, the dynamic variation autoencoder is used to identify and encode abnormal patterns and nonlinear characteristics in the monitoring data, and by inputting the monitoring data into the dynamic variation autoencoder, part of the high-dimensional monitoring data is converted into a low-dimensional potential representation, and an encoded feature data set is generated to complete feature extraction;

[0016] The feature data set includes vibration coding features, sound coding features and temperature coding features.

[0017] Preferably, the dynamically adjusting the network structure and depth according to the characteristics of the input data includes:

[0018] The ADNN model evaluates the key characteristics of the data in the feature data set and outputs the evaluation results, wherein the key characteristics include: the dimension, complexity, variability, abnormal patterns and nonlinear characteristics of the data;

[0019] Based on the evaluation results and the initial network structure, a dynamic adjustment strategy is formulated. The dynamic adjustment strategy is used to define when and how to adjust the structure and depth of the network.

[0020] Preferably, in the ADNN model that completes dynamic structural adjustment, the feature data set is received through the input layer; in each layer of the ADNN model, the input data is first matrix multiplied with the weight of the layer, and then a bias term is added, and finally a nonlinear transformation is performed through an activation function, and this process is repeated in all layers until the output layer; in the output layer, after the last round of weighted and activation operations, an initial life prediction result is generated.

[0021] Preferably, the predictive feedback loop comprises:

[0022] Generate initial lifespan prediction results by training the ADNN model using monitoring data;

[0023] Compare the initial life prediction results with the actual observed tool life to identify deviations in life prediction;

[0024] Based on the deviation in life span prediction, the ADNN model is self-adjusted, and the adjusted ADNN model is trained and predicted again;

[0025] Repeat the above process, and each time the optimized ADNN model is used to update the initial life prediction results. The updated initial life prediction results are compared with the actual observed tool life to continue to guide the further adjustment of the ADNN model;

[0026] The feedback loop continues until a preset number of iterations is completed or the accuracy of the prediction results meets the requirements.

[0027] Preferably, the self-adjusting of the ADNN model by comparing the results includes:

[0028] By comparing the results, the prediction deviation of the ADNN model is analyzed, and the prediction deviation includes: the size of the deviation, overestimation or underestimation and its cause;

[0029] According to the results of the deviation analysis, a model adjustment strategy is formulated, wherein the model adjustment strategy includes: adjusting the network structure, optimizing the learning rate, modifying the loss function, or adjusting the processing method of the input features;

[0030] Based on the model adjustment strategy, the ADNN model is adjusted;

[0031] During the iterative training process of the ADNN model, the adjusted ADNN model refines its parameters through repeated forward propagation, loss calculation, back propagation, and parameter updating.

[0032] Preferably, the application of the hybrid data source integrated learning method to fuse and weight the feature data set includes:

[0033] Perform importance analysis on the features in the feature data set to determine the contribution of each feature to the prediction task; based on the results of the importance analysis, assign corresponding weights to each feature;

[0034] According to the determined weights, all features are weighted fused to obtain fused feature data.

[0035] Preferably, the comprehensive consideration of actual tool usage data and final tool life prediction results to perform performance evaluation includes:

[0036] Collect tool usage records and actual life data, wherein the tool usage records include: usage conditions, working environment and tool replacement time points;

[0037] The final life prediction result is compared with the actual life data, and the accuracy and error of the final life prediction result are calculated;

[0038] Analyze tool usage records when the prediction effect is good or insufficient, and determine the corresponding influencing factors.

[0039] A system for implementing the deep learning-based CNC machine tool tool life prediction method, comprising:

[0040] Monitoring hardware, the monitoring hardware includes: a vibration sensor for collecting vibration data of the tool in real time; a sound sensor for capturing sound data when the machine tool is running; a temperature sensor for monitoring the temperature of the tool or the processing area;

[0041] A data preprocessing unit, the data preprocessing unit is used to receive raw data from the sensor and perform standardization and denoising;

[0042] A computing unit, wherein the high-performance computing unit is used to run a dynamic variation autoencoder and an ADNN model;

[0043] A database server, which is used to store preprocessed data, feature data sets, and model prediction results, and supports data reading, writing, and retrieval;

[0044] A central processing unit, which is responsible for executing the predictive feedback loop, including: comparative analysis, ADNN model self-adjustment and weight update.

[0045] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0046] Through the dynamic variation autoencoder technology, the present invention realizes efficient feature extraction and encoding, automatically captures abnormal patterns and nonlinear characteristics in monitoring data, and effectively improves the accuracy of life prediction;

[0047] Adopting adaptive dynamic neural network (ADNN), the network structure and depth are dynamically adjusted according to the characteristics of input data. The present invention enhances the generalization ability and adaptability of the model.

[0048] The introduction of a predictive feedback loop mechanism enables the model to be updated and optimized in real time to reflect the latest situation in the tool wear process. The present invention further improves the accuracy and practicality of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the process of the present invention;

[0050] Figure 2 It is a schematic diagram of the data processing flow in the present invention;

[0051] Figure 3 It is a structural schematic diagram of the ADNN model in the present invention;

[0052] Figure 4 Schematic diagram of the predictive feedback loop in the present invention;

[0053] Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0054] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0055] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0056] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0057] In the case of using expressions such as "at least one of A, B, and C, etc.", it should generally be interpreted in accordance with the meaning of the expression generally understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.). In the case of using expressions such as "at least one of A, B, or C, etc.", it should generally be interpreted in accordance with the meaning of the expression generally understood by those skilled in the art (for example, "a system having at least one of A, B, or C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0058] Some block diagrams and / or flow charts are shown in the accompanying drawings. It should be understood that some boxes or combinations thereof in the block diagrams and / or flow charts can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that these instructions can create a device for implementing the functions / operations described in these block diagrams and / or flow charts when executed by the processor. The technology of the present disclosure can be implemented in the form of hardware and / or software (including firmware, microcode, etc.). In addition, the technology of the present disclosure can take the form of a computer program product on a computer-readable storage medium storing instructions, which can be used by an instruction execution system or used in conjunction with an instruction execution system.

[0059] like Figure 1-Figure 2 As shown, a CNC machine tool tool life prediction method based on deep learning includes the following steps:

[0060] Collecting the monitoring data of the tool and performing standardization and denoising preprocessing, wherein the monitoring data includes: vibration data, sound data and temperature data;

[0061] The vibration data in the present invention reflects the dynamic performance of the tool during the machining process, and is collected in real time from the CNC machine tool through a vibration sensor. The frequency and amplitude changes of the vibration signal can reveal the degree of tool wear and possible faults.

[0062] The sound data is captured by microphones during machine tool operation. Different tool wear levels and cutting conditions will produce different sound characteristics, and the analysis of sound signals helps monitor tool status.

[0063] Temperature sensors are used to monitor the temperature of the tool or machining area. Changes in temperature reflect the thermal load on the tool and are critical for predicting tool life and avoiding damage caused by overheating.

[0064] Convert all monitoring data to a uniform magnitude to ensure that the model does not favor certain features due to differences in data scale. For example, standardize all data to between 0 and 1, or make the data conform to a standard normal distribution.

[0065] Monitoring data will be affected by noise during the acquisition process, and denoising removes random noise in the data through filters or digital signal processing technology. For example, a low-pass filter is used to remove high-frequency noise in vibration data and retain the low-frequency part that is sensitive to tool wear.

[0066] The present invention makes the pre-processed data cleaner and more standardized, and can be more effectively used by the deep learning model. Standardization ensures the stability of model training, and denoising improves the signal-to-noise ratio of the data, so that the key information of the tool state can be captured more accurately in the feature extraction stage.

[0067] In one embodiment, the frequency of the vibration signal generated by a tool during machining increases as the degree of wear increases. This frequency change trend can be observed by real-time monitoring and analysis of vibration data. In the preprocessing stage, the vibration signal is first standardized so that the magnitude of all vibration data is consistent, and then the background noise is removed through denoising, retaining the key frequency components reflecting tool wear. These preprocessed data are then used to train a deep learning model, which learns the relationship between increased vibration frequency and tool wear, so that tool life can be predicted based on vibration data.

[0068] The collection and preprocessing of monitoring data is a crucial first step in the deep learning-based CNC machine tool tool life prediction method. By standardizing and denoising the vibration, sound, and temperature data, the accuracy and efficiency of the subsequent deep learning model in predicting tool life can be significantly improved. The effective execution of this stage provides high-quality and reliable input data for the deep learning model, which is the basis for achieving accurate life prediction.

[0069] The dynamic variation autoencoder is used to extract and encode the monitoring data, capture abnormal patterns and nonlinear characteristics, and generate an encoded feature data set; the feature data set is input into the ADNN model, and the network structure and depth are dynamically adjusted according to the characteristics of the input data to generate the initial life prediction results;

[0070] like Figure 3 As shown in the figure, the structure diagram of the adaptive dynamic neural network (ADNN) model shows the data input layer, preprocessing layer, feature extraction layer (feature extraction through dynamic variation autoencoder DVAE), and multiple fully connected layers that are adaptively adjusted (using ReLU activation function) until the final output layer (using Sigmoid activation function for life prediction). In addition, the figure also includes a dynamic adjustment process, showing how the network dynamically adjusts the structure and depth according to the characteristics of the input data.

[0071] In this invention, the use of dynamic variation autoencoder (DVAE) to extract and encode features, and then input these feature data sets into the adaptive dynamic neural network (ADNN) model for life prediction is an efficient and advanced method. This process involves complex data processing and model training steps, each of which is based on in-depth principles and aims to improve the accuracy and efficiency of prediction.

[0072] Dynamic variational autoencoder is a deep learning model based on autoencoders that can learn compressed representations of input data. Unlike traditional autoencoders, DVAE introduces dynamic elements and variational inference to more flexibly capture the nonlinear characteristics and abnormal patterns of input data. DVAE learns the deep characteristics of the data by minimizing the difference between the input data and its reconstruction while introducing regularization of the latent representation. Feature extraction using DVAE can obtain highly compressed and information-rich feature representations that can accurately reflect key information in the monitoring data, such as specific patterns of tool wear. By identifying these patterns, the model can more accurately predict tool life.

[0073] In one example, there is a certain vibration pattern in the tool monitoring data that only occurs when the tool is near the end of its life. By analyzing the vibration data, DVAE is able to learn a compressed representation of this pattern and encode it as a feature in the feature data set. This enables subsequent life prediction models to use this key information to make more accurate predictions.

[0074] The core feature of the ADNN model is its ability to dynamically adjust its network structure and depth according to the characteristics of the input feature data set. This means that the model is not static, but can adjust itself according to the complexity, dimension, variability, and abnormal patterns and nonlinear characteristics of the data. ADNN adapts to the characteristics of the data in the best way through internal mechanisms such as adding or reducing network layers, adjusting the number of neurons, or modifying the connection method.

[0075] This dynamic adjustment enables ADNN to configure the most appropriate network structure for different data characteristics, thereby improving the learning efficiency and prediction accuracy of the model. For example, for high-dimensional feature data sets containing complex patterns, ADNN will increase the depth of the network to capture deeper data features; for simple pattern data, it will reduce the complexity of the network to avoid overfitting.

[0076] Through the adjusted network structure, the ADNN model uses the encoded feature data set for forward propagation and finally generates the initial life prediction result. This result is the best estimate based on the data characteristics and patterns currently learned by the model. The initial life prediction result provides the basis for the subsequent predictive feedback loop. By comparing with the actual observed tool life, the model can be further adjusted and optimized to gradually reduce the prediction error and improve the accuracy of the prediction results.

[0077] In summary, the use of DVAE for feature extraction and encoding, combined with the dynamic adjustment capability of the ADNN model, provides an efficient and accurate method for predicting tool life of CNC machine tools. The present invention can generate high-quality feature data sets by deeply analyzing the complex patterns and characteristics in the monitoring data, and configure the most appropriate network structure for each data characteristic through the dynamic adjustment mechanism of the ADNN model, thereby achieving accurate life prediction.

[0078] Preferably, the dynamic variation autoencoder is used to identify and encode abnormal patterns and nonlinear characteristics in the monitoring data, and by inputting the monitoring data into the dynamic variation autoencoder, part of the high-dimensional monitoring data is converted into a low-dimensional potential representation, and an encoded feature data set is generated to complete feature extraction;

[0079] The feature data set includes vibration coding features, sound coding features and temperature coding features.

[0080] Dynamic variational autoencoder (DVAE) is an innovation based on the traditional variational autoencoder (VAE), introducing the ability to dynamically adapt to data characteristics. VAE learns the high-dimensional distribution of data by building a probabilistic model of the input data, mapping the data to a latent space through the encoder, and then reconstructing the data through the decoder. On this basis, DVAE adds the ability to dynamically adjust the network structure or parameters to more finely capture and represent the complexity and variability in the data.

[0081] Adaptive network structure: DVAE dynamically adjusts its network structure according to the complexity and pattern of the input data. For example, when complex nonlinear characteristics are detected in the data, DVAE increases the depth of the network or adjusts the width of the hidden layer to provide sufficient model capacity to capture these characteristics.

[0082] Parameter adjustment: During the training process, DVAE not only optimizes the weights and bias parameters, but may also adjust the learning rate or other hyperparameters of the network to improve the learning process and enhance the generalization ability of the model.

[0083] DVAE uses a combination of reconstruction loss (making the reconstructed data as close as possible to the original data) and regularization loss (ensuring that the distribution of the latent representation has good properties, such as normal distribution) to train the model. The dynamic adjustment strategy also considers the changes in these losses to guide the adjustment of the structure and parameters.

[0084] DVAE optimizes the weights and structure of the network through back-propagation algorithm and gradient descent (or its variants). The dynamic adjustment strategy makes adjustment decisions based on gradient information and model performance indicators.

[0085] In this paper, the dynamic variation autoencoder combines the data reconstruction ability of the autoencoder and the probabilistic generation model characteristics of the variational autoencoder (VAE), providing a method to dynamically adjust its structure to adapt to the characteristics of the data. By introducing latent variables and inferring the encoding distribution, DVAE can capture the complex distribution and structure inherent in the data.

[0086] The abnormal modes in the present invention are for example:

[0087] Instantaneous vibration increase: In CNC machining, the tool may experience a sudden increase in vibration during the machining process due to wear or damage. This instantaneous vibration change is an abnormal pattern and may indicate that the tool is about to fail.

[0088] Abnormal changes in sound frequency: Normally, the sound produced by a tool when machining a specific material should be relatively stable. If there is a sudden change in sound frequency, it may be due to tool wear or changes in the workpiece material. This abnormal change in sound data needs to be captured.

[0089] Nonlinear characteristics in the invention include:

[0090] Nonlinear relationship between temperature and wear: There may be a nonlinear relationship between the temperature change of the tool during use and its wear state. As the wear deepens, the temperature rise may accelerate. This nonlinear change is very important for predicting the remaining life of the tool.

[0091] Vibration characteristics of composite material machining: When machining composite materials, the vibration patterns experienced by the tool can be very complex and very different from those when machining traditional materials, exhibiting significant nonlinear characteristics that require the model to accurately capture and understand.

[0092] By identifying and encoding these abnormal patterns and nonlinear characteristics, DVAE provides a rich and detailed feature representation for the subsequent ADNN model, greatly improving the accuracy and efficiency of tool life prediction for CNC machine tools. This approach enables the model to predict possible tool failures at an early stage, thereby avoiding potential downtime in the production process and ensuring production efficiency and product quality.

[0093] DVAE is particularly suitable for identifying abnormal patterns and nonlinear characteristics in monitoring data. Abnormal patterns may indicate abnormal wear of the tool or signals before failure, while nonlinear characteristics reflect the complex dynamic changes during tool use. By learning the high-dimensional structure of the data, DVAE is able to convert these complex and subtle patterns into low-dimensional representations that are easier to analyze and understand.

[0094] DVAE uses an encoder-decoder architecture to implement the conversion process from high-dimensional data to low-dimensional potential representation. This conversion not only reduces the dimensionality of the data, but also captures the most important features in the data that are most useful for subsequent tasks (such as prediction). The following is a detailed explanation of this conversion process:

[0095] The encoder of DVAE receives high-dimensional monitoring data (such as vibration, sound, and temperature data) as input.

[0096] The encoder processes the data through a series of neural network layers (including fully connected layers, convolutional layers, etc.), gradually reducing the dimension of the data. Each layer of the network can be regarded as a functional block that converts the input data into a higher-level feature representation.

[0097] At the final stage of the encoder, the data is converted into one or more low-dimensional latent vectors that capture the core information and structure of the input data.

[0098] The training process of DVAE involves minimizing the reconstruction error (i.e., the difference between the input data and the data reconstructed by the decoder) and regularizing the latent representation (usually achieved through KL divergence). This ensures that the latent representation is not only able to effectively reconstruct the original data, but also follows a certain predefined distribution (such as Gaussian distribution), which contributes to the continuity and smoothness of the latent space. DVAE is able to dynamically adjust the structure of the encoder (such as network depth and layer width) according to the characteristics of the data. This adaptive ability allows the model to respond to different data characteristics and better capture and retain key information.

[0099] In one example, a CNC machine tool generates a vibration data set during machining. This data is high-dimensional and complex, and contains important information about the wear state of the tool. Through the DVAE encoder, this vibration data is compressed into a low-dimensional latent representation. For example, if the original vibration data dimension is 1000 dimensions, the encoder may compress it into a 10-dimensional latent vector. Despite the significant reduction in dimensionality, this 10-dimensional vector can still reflect the key characteristics of tool wear - such as the enhancement of specific frequencies or changes in patterns - because the DVAE learns during training which features are necessary to reconstruct the original data and which can be ignored.

[0100] Through the encoder part, DVAE achieves an effective conversion from high-dimensional monitoring data to low-dimensional latent representations while retaining information that is critical to the prediction task. The core of this process lies in the design of the loss function, which not only forces the model to learn compressed representations that can reconstruct the original data, but also ensures that these representations have good properties, allowing the model to effectively generalize on unseen data. The dynamic adjustment mechanism further enhances the adaptability of the model, enabling it to adjust its structure for data sets with different characteristics, thereby more accurately capturing and encoding key information in the data.

[0101] After DVAE processing, the monitoring data is converted into a feature data set containing vibration coding features, sound coding features and temperature coding features. These encoded features are more compact and information-rich, reducing the redundancy and noise of the original data while retaining information that is critical to the prediction task.

[0102] In one embodiment, during the cutting process of a CNC machine tool, the vibration signal may become more intense due to increased tool wear; at the same time, the worn tool may cause changes in the frequency and intensity of the sound signal generated during the processing; the temperature data may also show different change patterns due to tool wear. The dynamic variation autoencoder can extract low-dimensional features reflecting the tool wear status from these high-dimensional data respectively, and these features are encoded into encoding features of vibration, sound and temperature to form a feature data set.

[0103] By inputting this feature data set into an adaptive dynamic neural network (ADNN), the system is able to make more accurate tool life predictions based on these compressed and information-rich features. The dynamic adjustment capability of the ADNN model ensures that it can adjust its network structure and depth according to the specific characteristics of the input feature data set (such as the dimension, complexity and inherent variability of the data), optimizing the model for the best prediction performance.

[0104] The application of dynamic variation autoencoders in the tool life prediction scheme for CNC machine tools provides strong data support for subsequent life prediction by accurately capturing and encoding key information in monitoring data. This method not only improves the accuracy of prediction, but also provides an effective strategy for deep learning models to process complex monitoring data.

[0105] Preferably, the dynamically adjusting the network structure and depth according to the characteristics of the input data includes:

[0106] The ADNN model evaluates the key characteristics of the data in the feature data set and outputs the evaluation results, wherein the key characteristics include: the dimension, complexity, variability, abnormal patterns and nonlinear characteristics of the data;

[0107] Based on the evaluation results and the initial network structure, a dynamic adjustment strategy is formulated. The dynamic adjustment strategy is used to define when and how to adjust the structure and depth of the network.

[0108] ADNN first analyzes the input feature data set and evaluates its key characteristics, including the dimension, complexity, variability, and presence of abnormal patterns and nonlinear features. These characteristics directly affect the model's ability and method to process data:

[0109] Data dimensionality: refers to the number of features in a dataset. High-dimensional data may require deeper or more complex network structures to capture the relationships between features.

[0110] Complexity: reflects the structural complexity of the data. Highly complex data may contain more nonlinear relationships and patterns, requiring the model to have better feature extraction and learning capabilities.

[0111] Variability: The range of variation in the data. High variability may mean that there is more noise or a wider distribution in the data, and the model needs to be able to effectively extract information from it.

[0112] Unusual patterns and nonlinear features: Specifically refer to unusual or difficult to directly observe patterns that may exist in the data, which are often key factors in prediction.

[0113] Based on the evaluation results of these key characteristics, the ADNN model formulates a dynamic adjustment strategy. This strategy defines how the model should adjust its network structure (e.g., number of layers, number of neurons in each layer) and depth according to the data characteristics to most effectively process the input data.

[0114] Through this dynamic adjustment mechanism, the present invention enables ADNN to adaptively optimize its structure and provide more accurate predictions for specific data sets. This not only improves the generalization ability of the model, but also enables the model to maintain high performance when facing variable data characteristics.

[0115] In one embodiment, there are two different sets of monitoring data sets used for tool life prediction:

[0116] The first set of data mainly contains vibration and temperature data. The relationship between the features is relatively simple, and the data dimension and complexity are low.

[0117] The second set of data includes vibration, sound and temperature data. The spectral analysis of the sound data introduces high-dimensional features, which significantly increases the complexity and variability of the entire data set.

[0118] For the first set of data, ADNN uses a relatively shallow network structure because the data has low dimensionality and complexity and does not require an overly complex model to capture the relationship between features. For the second set of data, ADNN increases the depth of the network or adjusts the connections between layers to adapt to the high dimensionality and complexity of the data, especially to handle the nonlinear features and high variability in sound data.

[0119] This dynamic adjustment based on data characteristics not only improves the model's ability to handle different data sets, but also provides strong support for more accurate tool life prediction.

[0120] The steps to monitor and evaluate key characteristics of a feature dataset in real time include:

[0121] At the beginning of each epoch of model training, the feature data set is monitored in real time by specific statistical methods or algorithms. For example, principal component analysis (PCA) is used to evaluate the dimensionality and complexity of the data, the variability is quantified by calculating the standard deviation and coefficient of variation of the data set, and anomaly detection algorithms are used to identify abnormal patterns. Based on the monitoring results, pre-defined rules or machine learning models are used to evaluate the key characteristics of the data. This includes classification of data complexity (low, medium, high), quantitative evaluation of variability, and identification of nonlinear characteristics and abnormal patterns.

[0122] Based on the evaluation results of the complexity of the data, if the data complexity is high, it may be necessary to increase the number of network layers to improve the expressiveness of the model; conversely, if the data complexity is low, reduce the number of network layers to avoid overfitting. Adjust according to the dimension and variability of the data. For high-dimensional data with large variability, increase the number of neurons in each layer to better capture the data characteristics; for low-dimensional data with small variability, reduce the number of neurons to improve model efficiency. Based on the evaluation results of abnormal patterns and nonlinear characteristics, optimize the connection pattern within the network. For example, for data with nonlinear characteristics, introduce or enhance nonlinear activation functions in the network; for data with obvious abnormal patterns, a specific network architecture such as an attention mechanism may be adopted so that the model can pay more attention to these patterns.

[0123] In one embodiment, in a tool life prediction task, the initial ADNN model is designed to have 3 hidden layers, with 100 neurons in each layer. During the training process, the feature data set monitored in real time shows that the data has a high degree of complexity and significant nonlinear characteristics:

[0124] The model decides to add 2 hidden layers to increase the depth of the model to better capture the complex relationships in the data.

[0125] The ReLU activation function is used instead of the linear activation function in the newly added hidden layer to enhance the model's ability to capture nonlinear relationships.

[0126] Observing the large variability in the data, the model increased the number of neurons in each hidden layer by 20% to improve the learning ability and adaptability of the model.

[0127] Through such dynamic adjustment, the ADNN model can more accurately reflect and adapt to the real characteristics of the feature data set, thereby improving the accuracy of tool life prediction and the generalization ability of the model.

[0128] Preferably, in the ADNN model that completes dynamic structural adjustment, the feature data set is received through the input layer; in each layer of the ADNN model, the input data is first matrix multiplied with the weight of the layer, and then a bias term is added, and finally a nonlinear transformation is performed through an activation function, and this process is repeated in all layers until the output layer; in the output layer, after the last round of weighted and activation operations, an initial life prediction result is generated.

[0129] The first step of the ADNN model is to receive a feature dataset through its input layer. These datasets are usually generated by a preprocessing step (such as processing by a dynamic variation autoencoder) and contain encoded features of monitoring data, such as vibration, sound, and temperature encoded features, which have been converted into a form that the model can directly process.

[0130] As the input data passes through each layer of the model, it is first matrix multiplied with the weights of that layer, and then the bias term is added. The purpose of this step is to linearly transform the input data into a new space. Subsequently, it is nonlinearly transformed by an activation function, allowing the model to capture nonlinear relationships and patterns in the data. This combination of linear-nonlinear transformations is repeated in all layers of the model.

[0131] The role of the activation function in a neural network is to introduce nonlinearity, allowing the network to learn and simulate complex nonlinear relationships. Without an activation function, no matter how many layers the network has, it can eventually be simplified into a linear model, which greatly limits the model's expressive power. In tasks such as tool life prediction for CNC machine tools, the relationship between data is often nonlinear, so the choice of activation function has a significant impact on the performance of the model.

[0132] ReLU (Rectified Linear Unit): It is one of the most commonly used activation functions because of its simple calculation and non-saturated gradient (in the positive range). For the CNC machine tool tool life prediction task, ReLU helps to accelerate the training of the model, and it has a certain resistance to the gradient vanishing problem, which makes it suitable for deep networks.

[0133] Sigmoid: It can compress the output to between 0 and 1 and is often used in the output layer for binary classification. However, when used in the hidden layer, Sigmoid may cause the gradient disappearance problem, which limits its application in deep networks.

[0134] Tanh: The output range is -1 to 1, which is wider than the output range of the Sigmoid function, so it performs better in some cases. But it still has the problem of gradient disappearance, similar to Sigmoid, which limits its use in deep networks.

[0135] For the CNC machine tool tool life prediction in the present invention, ReLU and its variants (such as Leaky ReLU, PReLU, etc.) are usually better choices because they help alleviate the gradient vanishing problem and are suitable for learning complex data patterns.

[0136] Weights and biases are parameters in a neural network that are continuously updated during training by an optimization algorithm to minimize the loss function, i.e., reduce the difference between the model's predicted values ​​and the true values. This process is mainly achieved through backpropagation and gradient descent (or other optimization algorithms).

[0137] Backpropagation: Calculate the gradient of the loss function with respect to each parameter (weight and bias). This process starts at the output layer and propagates backwards through the network, using the chain rule to calculate the gradient of each layer parameter.

[0138] Gradient Descent: Once the gradient is calculated, use gradient descent (or other optimization algorithms such as Adam, RMSprop, etc.) to update each parameter. The update formula for the parameters is:

[0139] parameter 丶 = parameter - learning rate * gradient 丶

[0140] Among them, the learning rate defines the step size of the update.

[0141] In the task of predicting tool life for CNC machine tools, this optimization process helps the model gradually learn how to accurately predict tool life based on input features (such as vibration, sound, and temperature encoding features). Through iterative updates, the model is able to find the optimal parameter settings to minimize prediction errors and improve prediction accuracy.

[0142] In one embodiment, assume that in the initial training stage, the tool life predicted by the model is significantly different from the actual life, and the loss function value is high. Through back propagation, the model calculates the gradient of the loss function for each weight and bias term, and then updates these parameters through the gradient descent method. If the relationship between a certain feature and tool life is underestimated by the model, the corresponding weight will increase, giving the feature a greater influence in subsequent predictions. As training progresses, the model gradually adjusts the weights and biases, the loss function value decreases, the difference between the tool life predicted by the model and the actual life gradually decreases, and the prediction performance is improved.

[0143] After multiple layers of processing, the data reaches the output layer. In this layer, the data undergoes the last linear transformation and nonlinear activation, and the initial life prediction result is output. This result represents the model's prediction of tool life based on the current learned features and weight configuration.

[0144] Through this complex data processing process, the ADNN model of the present invention can comprehensively consider the various information in the input feature data set and capture the complex relationships and patterns between these features through the learning mechanism within the model. This enables the model to not only predict the life of the tool, but also to understand the factors that affect the life to a certain extent.

[0145] In one embodiment, if there is a certain frequency pattern in the monitored vibration data, this pattern is closely related to rapid tool wear. Through the multiple layers of the ADNN model, the model is able to identify this pattern and learn its relationship with tool wear. At the output layer, this relationship is converted into a specific prediction of tool life, such as predicting that the tool may fail within a certain period of time in the future.

[0146] The ADNN model can generate accurate tool life prediction results by deeply processing and analyzing the input feature data set. This process relies on the complex structure and nonlinear processing capabilities within the model, which enables the model to not only predict tool life, but also provide a deep understanding of the tool wear process. This deep learning-based prediction method provides a powerful tool for the maintenance and management of CNC machine tools, helping to achieve a more efficient and reliable production process.

[0147] Through a predictive feedback loop, the initial wear prediction results are compared with the actual observed tool life, and the ADNN model is self-adjusted based on the comparison results to produce an optimized ADNN model and update the initial life prediction results;

[0148] like Figure 4 As shown in Figure 1, the predictive feedback loop (PFL) is a key mechanism that allows the ADNN model to self-adjust and optimize the prediction performance by comparing the initial wear prediction results with the actual observed tool life. The core principle of this process is to learn based on the difference between the model output and the actual results, continuously iteratively improve the model, and thus improve the accuracy and reliability of the prediction.

[0149] After the model completes a forward pass, it generates initial tool life predictions. These predictions are then compared with actual observed tool life data to assess the model’s prediction accuracy.

[0150] Based on the comparison results, the model adjusts its internal parameters (such as weights and biases) to reduce the difference between the predicted results and the actual life span. This adjustment process relies on back propagation and gradient descent (or other optimization algorithms) to minimize the prediction error.

[0151] The predictive feedback loop in the present invention enables the ADNN model to adaptively optimize its structure and parameters, thereby improving the accuracy of tool life prediction. The effects of this self-adjusting mechanism include:

[0152] By continuously learning from actual observed data, the model can better understand the patterns and relationships in the data, thereby improving the accuracy of predictions; through continuous comparison and adjustment with actual data, the model can resist overfitting and enhance its performance on unseen data; the model can adapt to changes in the data (such as new tool types or new processing materials) because it continuously uses the latest observed data to update itself.

[0153] In one embodiment, assume that in the initial stage, the ADNN model predicts that the life of a tool is 300 hours based on vibration data and temperature data. However, actual observations show that the tool failed at 250 hours. This significant difference will be fed back to the model, triggering the self-adjustment process of the model parameters. The model will find that a certain characteristic frequency segment of the vibration data is significantly enhanced near the moment of failure, and this pattern has a low weight in the initial model. Therefore, the model will adjust the weight so that in future predictions, it will give more weight to the characteristics of this frequency segment. With multiple such feedback cycles, the model gradually learns to correctly interpret the key signals in the data, so that it can more accurately estimate the tool life in subsequent predictions.

[0154] The predictive feedback loop provides a dynamic and adaptive optimization mechanism for tool life prediction of CNC machine tools. By continuously comparing the predicted results with the actual life data and adjusting itself accordingly, the ADNN model can continuously improve the accuracy and reliability of its predictions, adapt to the changing machining conditions and tool characteristics, and thus play a key role in practical applications.

[0155] Preferably, the predictive feedback loop comprises:

[0156] Generate initial lifespan prediction results by training the ADNN model using monitoring data;

[0157] Compare the initial life prediction results with the actual observed tool life to identify deviations in life prediction;

[0158] Based on the deviation in life span prediction, the ADNN model is self-adjusted, and the adjusted ADNN model is trained and predicted again;

[0159] Repeat the above process, and each time the optimized ADNN model is used to update the initial life prediction results. The updated initial life prediction results are compared with the actual observed tool life to continue to guide the further adjustment of the ADNN model;

[0160] The feedback loop continues until a preset number of iterations is completed or the accuracy of the prediction results meets the requirements.

[0161] The ADNN model is trained using monitoring data (such as vibration, sound, and temperature data) to generate initial life prediction results. This step is based on the model's current network structure and parameter settings. The initial prediction is based on the patterns and relationships that the model has learned from the data. These relationships may be nonlinear and may be affected by a variety of factors, such as the condition of the tool, material type, etc.

[0162] Deviation identification is accomplished by comparing the model’s initial life predictions with the actual observed tool life. This process can be accomplished using a variety of methods, including but not limited to:

[0163] Difference calculation: The most direct method is to calculate the difference between the predicted lifespan and the actual lifespan. This method is simple and intuitive, but it may be affected by data noise, resulting in inaccurate error assessment.

[0164] Statistical analysis methods: In order to improve the accuracy of deviation assessment, more complex statistical analysis methods can be used, such as calculating the standard deviation of the prediction error, mean square error (MSE) or mean absolute error (MAE). These methods can more comprehensively reflect the difference between the prediction results and the actual values.

[0165] Dealing with data noise and outliers: During the comparison process, the impact of noise can be reduced through data preprocessing steps (such as filtering, denoising, etc.). For outliers, anomaly detection algorithms can be used to identify and exclude them, or robust statistical methods (such as median absolute deviation) can be used to reduce the impact of outliers on deviation assessment.

[0166] The initial life predictions are compared with the actual observed tool life to identify deviations in the predictions. The ADNN model is then self-adjusted based on these deviations. The identification of deviations allows the model to understand where its predictions do not match reality. Self-adjustment involves adjusting the weights of the network, modifying the network structure (such as adding or reducing the number of layers), or adjusting hyperparameters such as the learning rate. These adjustments are designed to make the model better fit the actual distribution and inherent relationships of the data.

[0167] The self-tuning mechanism allows the ADNN model to adjust its structure and parameters based on the identified biases to optimize the prediction performance. This process is achieved by:

[0168] The model can decide the direction and magnitude of the adjustment based on the size and direction of the deviation. For example, if predicted lifespans are generally higher than actual lifespans, the model may need to adjust weights to reduce the response to certain features. The magnitude of the adjustment is usually proportional to the size of the deviation, with larger deviations likely resulting in more significant adjustments.

[0169] There are many techniques or algorithms that can be used in the self-tuning process, such as gradient descent, adaptive learning rate adjustment (such as Adam, RMSprop, etc.), and weight regularization techniques. These methods help to fine-tune model parameters, avoid overfitting, and ensure the effectiveness and efficiency of model optimization.

[0170] For the adjustment of network structure, the model can dynamically add or delete layers or adjust the number of neurons in the layer according to the feedback of deviation. This adjustment is usually based on preset rules or achieved through network pruning, network growth and other techniques.

[0171] In one embodiment, if the ADNN model finds that its predicted tool life is generally lower than the actual life during the initial training phase, this may indicate that the model fails to adequately capture the factors that lead to longer life. In this case, the model can be adjusted by increasing its responsiveness to relevant features, such as the vibration intensity at a specific frequency. Specific operations include increasing the number of neurons in a specific layer, increasing the weight of relevant features, or introducing new network layers to capture more complex data patterns. Through such self-adjustment, the model gradually improves the accuracy of its prediction of tool life.

[0172] The predictive feedback loop provides a dynamic optimization mechanism for the ADNN model by comparing, identifying deviations, and self-adjusting the model based on the deviations, thereby achieving higher accuracy and robustness in the tool life prediction task of CNC machine tools.

[0173] The adjusted ADNN model is trained and predicted again. This process is repeated, each time using the optimized model to update the initial life prediction results, which are then compared again with the actual observed values. Iterative training allows the model to have a chance to "learn" its errors after each feedback loop and improve accordingly. As the number of iterations increases, the model's accuracy in predicting tool life gradually increases and the deviation gradually decreases.

[0174] As the feedback loop of the present invention proceeds, the model is able to more accurately predict tool life because it continuously learns and adapts to the data distribution and complexity of the real world; through repeated self-adjustment and optimization, the model exhibits better generalization ability when processing new, unseen data.

[0175] In one example, an ADNN model initially predicted an average tool life of 200 hours for a batch of specific materials and specific conditions at a manufacturing plant. However, the actual observed average life was 180 hours. Through a predictive feedback loop, the model recognizes this 20-hour deviation and adjusts its internal parameters to reduce this difference. After a few iterations, the model's prediction may be adjusted to closer to 180 hours, significantly improving the accuracy of the prediction.

[0176] The predictive feedback loop provides a dynamic and effective mechanism that enables the deep learning-based CNC machine tool tool life prediction model to self-optimize and improve prediction accuracy. Through continuous iterative training and self-adjustment, the model can more accurately capture the complex relationships in the data, thereby achieving more accurate tool life prediction in practical applications.

[0177] Preferably, the self-adjusting of the ADNN model by comparing the results includes:

[0178] By comparing the results, the prediction deviation of the ADNN model is analyzed, and the prediction deviation includes: the size of the deviation, overestimation or underestimation and its cause;

[0179] According to the results of the deviation analysis, a model adjustment strategy is formulated, wherein the model adjustment strategy includes: adjusting the network structure, optimizing the learning rate, modifying the loss function, or adjusting the processing method of the input features;

[0180] Based on the model adjustment strategy, the ADNN model is adjusted;

[0181] During the iterative training process of the ADNN model, the adjusted ADNN model refines its parameters through repeated forward propagation, loss calculation, back propagation, and parameter updating.

[0182] In the task of tool life prediction for CNC machine tools, it is crucial to accurately identify and analyze the nature and causes of prediction deviations. The following is a detailed description of how to implement this process:

[0183] The magnitude of the deviation is quantitatively assessed by calculating the difference between the predicted results and the actual values ​​(e.g., the sum of squared errors, mean squared error MSE, mean absolute error MAE). Using visualization techniques such as error distribution graphs and scatter plots of predicted values ​​versus actual values, patterns of deviations, such as systematic overestimation or underestimation, can be qualitatively identified.

[0184] Use model interpretation tools (such as SHAP values, LIME) to identify and explain which input features contribute most to the bias. These tools can reveal the decision logic of the model and the importance of features, helping to understand why a specific bias occurs.

[0185] By comparing the model's predictions with actual observations, the deviations in the model's predictions are analyzed, including the magnitude of the deviation and whether the predictions overestimate or underestimate the actual life. Deviation analysis helps identify deficiencies in the model's predictive capabilities. For example, if the prediction system consistently underestimates the life of a tool, this may indicate that the model fails to adequately capture the positive factors that affect tool life.

[0186] Based on the results of the deviation analysis, formulate a model adjustment strategy, including adjusting the network structure, optimizing the learning rate, modifying the loss function, or adjusting the input feature processing method. Each adjustment strategy is designed to solve a specific problem identified in the deviation analysis. For example, adjusting the network structure may involve increasing the number of layers or adjusting the width of the layer to increase the complexity of the model and increase the model's fitting ability for the problem of underestimation.

[0187] Specific example 1 of the model adjustment strategy:

[0188] Problem: The tool life predicted by the model is generally higher than the actual observed value.

[0189] Analysis: Using the model interpretation tool, we found that the model over-relies on a feature that is strongly associated with a longer lifespan, while ignoring other factors that affect a shorter lifespan.

[0190] Adjustment strategy: Increase the number of network layers and neurons to improve the model's ability to capture complex relationships; introduce feature engineering to more clearly encode factors that affect life shortening, such as the hardness of the processed material, cutting speed, etc.; introduce regularization terms (such as L1 regularization) to avoid the model's over-reliance on a single or a few features.

[0191] Specific embodiment 2 of the model adjustment strategy:

[0192] Problem: The model consistently underestimates tool life predictions for certain materials.

[0193] Analysis: The model failed to adequately learn the unique durability characteristics of this material.

[0194] Adjustment strategy: Adopt adaptive learning rate adjustment method (such as Adam algorithm) to improve the sensitivity of the model when learning the data of this type of tool; increase data samples for tools of this material and improve the training effect of the model through data enhancement technology; develop new feature extraction technology or deep feature learning method based on the characteristics of this material to ensure that the model can capture the key factors affecting life.

[0195] Based on the above specific strategy implementation examples, it can be seen that the identification and analysis of prediction deviations, as well as the self-adjustment of the model based on the analysis results, are the key to improving the accuracy of tool life prediction for CNC machine tools. This requires a deep understanding of the data and the flexible use of data analysis, model evaluation, and machine learning techniques to achieve continuous optimization and improvement of model performance.

[0196] Here, we provide successful and failed model adjustment cases in practical applications and analyze their differences to help understand the effectiveness and applicability of model adjustment strategies in different situations:

[0197] Case 1: Based on the technology of the present invention, when processing the tool life prediction of a set of specific materials, by increasing the network layer depth and adjusting the loss function, the model successfully captured the nonlinear relationship between material hardness and life, significantly improving the prediction accuracy.

[0198] Accurate problem diagnosis, targeted model adjustment strategies, appropriate learning rate selection, and effective training monitoring.

[0199] Case 2: Based on conventional wisdom, an attempt was made to improve prediction performance by significantly increasing the complexity of the neural network. As a result, the model overfitted on the training set, and its performance in the validation set and actual applications decreased.

[0200] Failure to correctly assess the nature of the problem, lack of effective monitoring of the training process, and over-adjustment of model complexity lead to decreased generalization ability.

[0201] The main differences between successful and failed cases lie in the choice of model adjustment strategy, the monitoring method of the training process, and the accurate understanding of the nature of the problem. The successful cases using the technical solution of the present invention involve precise adjustments after in-depth analysis of the problem, and avoid overfitting through strategies such as continuous monitoring and early stopping to ensure the generalization ability of the model. Failure cases based on conventional cognition often lack a correct understanding of the problem, blindly increase the complexity of the model, or ignore training monitoring, resulting in poor performance of the model in actual applications.

[0202] These comparisons emphasize that when adjusting the model, it is necessary to comprehensively consider the specific circumstances of the problem, the structure of the model and the parameter adjustment strategy, as well as the monitoring and evaluation methods of the training process to achieve effective model optimization and improve prediction performance.

[0203] The ADNN model is adjusted according to the formulated model adjustment strategy, and then the model refines its parameters during the iterative training process, including through forward propagation, loss calculation, back propagation, and parameter update. The model adjusts and refines parameters to better adapt to the real distribution of the data, thereby reducing prediction bias and improving prediction accuracy.

[0204] Once the ADNN model is tuned using the tuning strategy, its iterative training process involves the following key steps to refine the parameters and gradually reduce the prediction error:

[0205] (1) The adjusted model processes the input data using the updated network structure and parameters. The data is passed through each layer of the model, and each layer applies the corresponding weights and activation functions to calculate the output. This process continues until the output layer is reached and a prediction result is generated.

[0206] (2) Calculate the difference between the predicted result and the actual value, usually quantified using a loss function (such as mean square error MSE). The smaller the loss function value, the closer the model's prediction is to the actual value.

[0207] (3) Use the gradient of the loss function to adjust the model parameters (weights and biases). This step involves calculating the derivative of the loss function with respect to each parameter and using these derivatives to guide the update direction of the parameters to reduce the prediction error.

[0208] (4) Use optimization algorithms (such as gradient descent, Adam, etc.) to update the weights and biases of the model according to the gradient calculated by back propagation. The learning rate determines the size of the update step, and an appropriate learning rate can help the model converge effectively.

[0209] (5) To avoid overfitting and ensure convergence, monitor key indicators during training, such as training and validation loss, accuracy, etc. Use strategies such as early stopping to stop training when the performance on the validation set no longer improves over multiple consecutive iterations.

[0210] Through targeted model adjustment, the present invention enables ADNN to more accurately predict tool life and improve production efficiency and cost-effectiveness; the self-adjustment mechanism enables the ADNN model to adapt to changes in data distribution and enhance its generalization ability under different working conditions.

[0211] In one example, in a production environment, the ADNN model initially predicted that the average life of a group of tools was 100 hours, while the actual observed value was 120 hours. Analysis found that the model systematically underestimated the tool life, probably because it did not fully consider the impact of certain key features (such as the hardness of the tool material) on the life. Based on this analysis, model adjustment strategies include increasing the weights of these features or introducing new network layers to capture more complex data relationships. After adjustment and retraining, the model's predictions are closer to the actual observed values, reducing the deviation.

[0212] Through the self-adjustment mechanism, the ADNN model can optimize itself based on the deviation between the predicted results and the actual observed values, and continuously improve its accuracy in predicting tool life. This mechanism provides strong adaptability and flexibility for tool life prediction of CNC machine tools, which helps to improve the reliability and efficiency of the manufacturing process.

[0213] Before the predictive feedback loop is started, the hybrid data source ensemble learning method is applied to fuse and weight the feature data set, obtain the fused feature data and input it into the optimized ADNN model for final life prediction, and output the final life prediction result of the tool;

[0214] In the hybrid data source ensemble learning (HDIL) method, feature data sets from different data sources are first integrated together. These data sources include vibration data, sound data, and temperature data, each of which can reflect the usage and wear of the tool from different perspectives.

[0215] The features of different data sources contribute differently to life expectancy prediction. Therefore, the HDIL method analyzes the influence of each feature on the prediction results and assigns different weights to them. This weighting process ensures that the model can pay more attention to those features that have a greater impact on the prediction results.

[0216] After weighting, the features of different data sources are fused into a comprehensive feature dataset. This fusion process involves feature selection, dimensionality reduction techniques, or complex feature engineering techniques to capture complex relationships and patterns across data sources.

[0217] By integrating multiple data sources and weighting the features, the HDIL method can provide more comprehensive and accurate tool wear and life information, thereby significantly improving the accuracy of life prediction; the fusion of feature data sets makes the model no longer dependent on a single data source or a specific type of feature, thereby increasing the model's adaptability to different working conditions and abnormal situations.

[0218] In one embodiment, vibration data can provide information about the tool load, sound data can reflect the stability of the tool-material contact, and temperature data can indicate the thermal state of the cutting area. Independently, these data sources can provide useful life prediction information. However, when these data sources are integrated and weighted through the HDIL method, the fused feature data set not only contains the information of all single data sources, but also emphasizes the features with the most predictive value through a weighting mechanism. For example, if the analysis finds that the predictive value of vibration data is higher than that of sound and temperature data under specific cutting conditions, the HDIL method can assign higher weights to the vibration features accordingly. In this way, when these fused feature data are input into the optimized ADNN model, the model can more accurately predict tool life.

[0219] By applying a hybrid data source ensemble learning method to fuse and weight feature datasets before the predictive feedback loop is initiated, the deep learning-based CNC machine tool tool life prediction method can significantly improve the prediction accuracy and model robustness. This method makes full use of the complementary information of multiple data sources and emphasizes the most important features through intelligent weighting, providing a powerful and flexible solution for tool life prediction.

[0220] Preferably, the application of the hybrid data source integrated learning method to fuse and weight the feature data set includes:

[0221] Perform importance analysis on the features in the feature data set to determine the contribution of each feature to the prediction task; based on the results of the importance analysis, assign corresponding weights to each feature;

[0222] According to the determined weights, all features are weighted fused to obtain fused feature data.

[0223] In a multi-source data environment, different data sources (such as vibration data, sound data, and temperature data) contribute differently to the prediction task. By applying various data analysis techniques, such as model-based feature importance evaluation (such as feature importance indicators provided by machine learning models such as random forests and gradient boosting trees), the impact of each feature on the prediction result can be quantified. Feature importance analysis is not only based on statistical measures, but also considers the actual role of features in model prediction, identifying which features are more directly related to the target prediction value (tool life).

[0224] In the hybrid data source ensemble learning method, feature importance analysis is a key step to identify which features contribute most to the prediction task. The following are specific methods and tools for feature importance analysis:

[0225] Random Forest:

[0226] Random forest is an ensemble learning method that performs classification or regression by building multiple decision trees and aggregating their predictions. During training, it automatically evaluates the importance of each feature. In feature importance analysis, random forest evaluates the importance of a feature by calculating the degree of reduction in model performance when a feature is randomly permuted. This method is intuitive and easy to implement, and is suitable for high-dimensional data and data sets with complex nonlinear relationships.

[0227] Gradient Boosting Decision Tree (GBDT):

[0228] GBDT is a method to improve model performance by gradually adding decision trees, with each new tree reducing the residual error of the previous tree. GBDT can also evaluate the importance of features. Using the GBDT model for feature importance analysis, the importance of each feature can be evaluated by observing its contribution in reducing the model error. GBDT is particularly suitable for handling prediction tasks with strong feature interactions.

[0229] Based on the analysis results of feature importance, each feature will be assigned a weight, which reflects the relative importance of the feature in the prediction task. Features with high weights will have a greater influence in the final feature data set. The basis for assigning weights can be direct feature importance scores, or they can be dynamically adjusted through optimization algorithms (such as genetic algorithms, particle swarm optimization, etc.) to achieve the best prediction performance.

[0230] The process of determining feature weights is based on the results of the above importance analysis. The following is the algorithm and optimization process for weight assignment:

[0231] Dynamic weight adjustment:

[0232] Based on the evaluation results of feature importance, the weights of features are adjusted dynamically. This can be done by converting feature importance scores into weights through a simple linear proportional distribution method, or more complex methods can be used, such as automatically adjusting weights based on optimization algorithms to maximize the performance of the prediction model. There are automated optimization methods such as genetic algorithms and particle swarm optimization (PSO), which can find the optimal weight configuration in a wider solution space, thereby improving the accuracy and robustness of the prediction model.

[0233] Evaluate the performance of the model under different weight configurations through methods such as cross-validation to verify the effectiveness of weight allocation. Ensure that the weight allocation does not overly favor certain features and ignore other equally important features. By setting upper and lower limits for weight allocation, avoid excessive influence of a single feature on the model prediction results and ensure that the model comprehensively considers all relevant features.

[0234] This method of dynamic weight adjustment based on feature importance allows the model to adaptively adjust feature weights in different prediction tasks, thereby effectively capturing complex data structures and optimizing prediction performance. By comprehensively applying these advanced analysis tools and algorithms, the accuracy and reliability of tool life prediction for CNC machine tools can be significantly improved.

[0235] Through weighted fusion, all features are integrated according to their weights to generate a comprehensive, weighted feature data set. This step usually involves mathematical weighted summation or more complex data fusion techniques, such as feature cascading, feature embedding, etc. The fused feature data set can more comprehensively reflect the use and wear of the tool, providing richer and more detailed input information for the prediction model.

[0236] Weighted fusion technology is an important part of ensemble learning. It can effectively integrate information from different data sources and improve the performance of prediction models. The following are several commonly used weighted fusion technologies:

[0237] Mathematically weighted average:

[0238] The most direct fusion method is to perform weighted averaging on the features, where the weight of each feature reflects the importance of the feature. Weighted averaging can reduce the impact of noise features while enhancing the role of features that contribute more to the prediction results. Through weighted averaging, the information of all features is retained, while the contribution of important features is strengthened through weight adjustment, reducing information loss.

[0239] Feature Cascade:

[0240] Feature concatenation is to directly concatenate different feature data sets together to form a larger feature vector. The cascaded feature set can provide more comprehensive information, but it may increase the complexity of the model. Feature concatenation maintains the integrity of all original features, but requires the model to be able to handle high-dimensional data and requires a more complex model structure to ensure effective learning.

[0241] Advanced Fusion Methods:

[0242] These include, but are not limited to, feature selection and dimensionality reduction techniques (such as principal component analysis (PCA)), and automatic feature extraction in deep learning (such as through autoencoders). These methods aim to reduce redundant information while retaining key features. Advanced fusion methods extract and enhance key information through technical means, reduce the interference of non-critical features, and help maintain the representativeness of the data while reducing information loss.

[0243] The weighted fusion feature data has a significant impact on the performance of the ADNN model. The following is an analysis based on actual cases:

[0244] By weighting and fusing feature data, the model can comprehensively consider information from different data sources, making the prediction results more accurate and reliable. For example, in a specific tool life prediction case, the accuracy of the ADNN model is significantly improved by weighting and fusing vibration, sound and temperature data, because the model can use richer and more comprehensive information for learning and prediction.

[0245] The impact of different fusion strategies:

[0246] The weighted average strategy emphasizes the role of some key features and is suitable for scenarios where the correlation between features is not high. It can effectively improve the accuracy of prediction.

[0247] The feature cascading strategy provides more information to the model by expanding the feature space, but it also requires the model to have good feature extraction capabilities to avoid the dimensionality disaster.

[0248] Advanced fusion strategies, such as using PCA or autoencoders for feature dimensionality reduction and extraction, can reduce feature dimensions while retaining the most critical information, which helps improve the generalization ability of the model.

[0249] In summary, weighted fusion of feature data is a key technology to improve the accuracy of tool life prediction for CNC machine tools. By rationally selecting fusion strategies and techniques, the prediction ability of the model can be greatly enhanced, and efficient and accurate life prediction can be achieved.

[0250] The present invention allows the model to pay more attention to those features that have a greater impact on the prediction results when performing life prediction through weighted fusion, thereby improving the accuracy of the prediction; fusing features from multiple data sources increases the information richness of the model, allowing the model to better adapt to different working conditions and environmental changes.

[0251] In one embodiment, in a specific CNC machine tool tool life prediction scenario, vibration data, sound data, and temperature data are analyzed and their respective importance is determined. For example, if the features of vibration data are found to be most important in tool wear prediction, these features will be given higher weights. Subsequently, these weighted features are fused into a feature data set and input into the optimized ADNN model for life prediction. In this way, even subtle vibration changes can be fully considered in life prediction, thereby significantly improving the accuracy of the prediction.

[0252] By deeply understanding and implementing these key steps of the hybrid data source integrated learning method, the accuracy and efficiency of CNC machine tool tool life prediction will be significantly improved, helping companies to manage and maintain their production equipment more effectively.

[0253] The actual tool usage data and the final life prediction results are comprehensively considered for performance evaluation. The feature extraction, ADNN model adjustment and feature fusion weights are adjusted based on the evaluation results to comprehensively optimize the life prediction performance.

[0254] Collect actual tool usage data, including the tool's working environment, working hours, load conditions, and the actual life of the tool. Then, compare these actual data with the model's prediction results to evaluate the accuracy and reliability of the prediction results. Performance evaluation usually uses multiple indicators, such as prediction accuracy, mean square error (MSE), R square value, etc., to quantify the effect of model prediction.

[0255] If it is found that the impact of certain features on the prediction results is not as expected, it is necessary to return to the feature extraction stage and adjust the feature extraction method or parameters, such as changing the feature selection strategy or optimizing the structure of the dynamic variation autoencoder.

[0256] According to the results of performance evaluation, the ADNN model is adjusted, including adjusting the network structure (such as adding or removing layers, changing activation functions), adjusting the learning rate, introducing regularization, etc., to improve the accuracy and generalization ability of the model.

[0257] If it is found that the current weight configuration of the weighted fusion feature data is not optimal, the weight distribution can be adjusted according to the performance evaluation results to ensure that the key features can be fully utilized in the model.

[0258] The present invention can enable the model to more accurately capture the key factors affecting tool life through optimization of feature extraction and model structure, thereby improving the accuracy of prediction; adjusting model parameters and structure, especially introducing appropriate regularization means, can reduce overfitting and enhance the model's generalization ability for unseen data; and through careful adjustment of feature fusion weights, it is ensured that the model can comprehensively utilize all relevant information to further improve the prediction performance.

[0259] In one embodiment, in a tool life prediction project, the initial model prediction results deviated from the actual life. Performance evaluation showed that the feature importance of the sound data was underestimated, resulting in inaccurate predictions. Based on this finding, the research team adjusted the dynamic variation autoencoder, increased the weight of the sound data features, and adjusted several key layers of the ADNN model, such as increasing the number of layers and adjusting the activation function. After these adjustments, the model's prediction accuracy was significantly improved and the prediction deviation was significantly reduced, demonstrating the effectiveness of the performance optimization process.

[0260] Preferably, the comprehensive consideration of actual tool usage data and final tool life prediction results to perform performance evaluation includes:

[0261] Collect tool usage records and actual life data, wherein the tool usage records include: usage conditions, working environment and tool replacement time points;

[0262] The final life prediction result is compared with the actual life data, and the accuracy and error of the final life prediction result are calculated;

[0263] Analyze tool usage records when the prediction effect is good or insufficient, and determine the corresponding influencing factors.

[0264] Actual tool usage data, including usage conditions (such as cutting speed, feed rate, cutting depth, etc.), working environment (such as temperature, humidity, etc.), and the time of tool replacement, are the basis for evaluating model performance. These data directly reflect the wear rate and life of the tool during actual use and are the key basis for verifying the model prediction results.

[0265] Comparative analysis is a direct way to evaluate the accuracy of the model. By calculating the accuracy and error between the final life prediction results of the model and the actual life data, you can intuitively understand the prediction ability of the model. The smaller the error and the higher the accuracy, the better the prediction performance of the model.

[0266] The identification of influencing factors is a key step in the evaluation and optimization of deep learning models. In the context of tool life prediction for CNC machine tools, determining which factors affect the prediction results is crucial for model improvement. The following is a detailed description of how to use advanced statistical analysis methods, machine learning algorithms or causal inference techniques to identify the key factors that affect the accuracy of tool life prediction:

[0267] (1) Advanced statistical analysis methods

[0268] Multivariate regression analysis: Multivariate regression analysis is used to explore how different tool usage conditions (such as cutting speed, feed rate) and working environment (such as temperature, humidity) affect the actual life of the tool. The factors that have a significant impact on tool life can be identified through the size and statistical significance of the regression coefficient.

[0269] Analysis of variance (ANOVA): When the data are divided into several fixed groups (for example, different tool materials or different tool types), ANOVA can be used to determine whether there are significant life differences between these groups, thereby identifying key classification factors that may affect life.

[0270] (2) Machine Learning Algorithms

[0271] Feature importance assessment: Use the feature importance assessment function of ensemble learning algorithms such as random forests and gradient boosting machines (GBM) to determine which features contribute most to the predictive ability of the tool life prediction model. This method not only identifies important features, but also reveals the relationship pattern between features and prediction targets.

[0272] Feature extraction and selection in deep learning: Deep neural networks (such as convolutional neural networks (CNN) and recurrent neural networks (RNN)) can automatically learn and extract complex features from large amounts of raw data. Using these features, we can gain a deeper understanding of the patterns contained in the data and their impact on the prediction results.

[0273] (3) Causal Inference Technology

[0274] Matching and propensity score analysis: Identify similar pairs of instances in the observational data and infer the causal effect of a specific operation by comparing the effects of different operations on tool life under similar conditions (such as the timing of tool replacement).

[0275] Causal diagrams and structural equation models: Use causal diagrams to represent the hypothesized causal relationships between variables, combined with structural equation models to estimate the strength of these relationships. This approach can help researchers better understand how different factors affect tool life through complex causal paths.

[0276] Through the above methods, researchers can comprehensively consider the actual tool usage data and the final life prediction results, and accurately identify the key factors that affect the prediction accuracy. This is of great significance for adjusting and optimizing the tool life prediction model and improving its prediction performance and practicality.

[0277] By analyzing the cases where the predictions were accurate and inaccurate in detail, it is possible to identify specific situations where the model performance was good or inadequate. This analysis helps to identify the correlation between the model's predicted performance and the actual conditions in which the tool is used, and thus determine which factors have a significant impact on the accuracy of the prediction results.

[0278] Through performance evaluation and analysis, the present invention can accurately understand in which aspects the model performs well and in which aspects it needs to be improved. This provides a direction for further optimization of the model and helps to improve the accuracy of the prediction; by analyzing the comparison between the actual tool usage records and the prediction results, the performance of the model in the actual application scenario can be better understood, so as to adjust the model in a targeted manner to adapt to the actual production environment and usage conditions, and enhance the practicality of the model.

[0279] In one embodiment, the usage conditions of CNC machine tools in a manufacturing plant vary greatly. The predicted life of tool A under high-speed cutting is highly consistent with the actual life, but the predicted results under low-speed and heavy-load cutting have a large deviation. Through performance evaluation, the comparison between the predicted results and the actual life of tool A under different usage conditions was analyzed, and it was found that the model under-predicted the wear rate under certain usage conditions. In response to this finding, the research team adjusted the model, especially in terms of feature extraction and feature fusion weights, to better capture the wear characteristics of the tool under heavy load conditions. The optimized model showed higher accuracy and better adaptability in subsequent evaluations, effectively improving the overall performance of tool life prediction.

[0280] Through this systematic performance evaluation and optimization process, the accuracy and practicality of the CNC machine tool tool life prediction model based on deep learning can be significantly improved.

[0281] like Figure 5 As shown, a system for implementing the method for predicting tool life of CNC machine tools based on deep learning comprises:

[0282] Monitoring hardware, the monitoring hardware includes: a vibration sensor for collecting vibration data of the tool in real time; a sound sensor for capturing sound data when the machine tool is running; a temperature sensor for monitoring the temperature of the tool or the processing area;

[0283] A data preprocessing unit, the data preprocessing unit is used to receive raw data from the sensor and perform standardization and denoising;

[0284] A computing unit, wherein the high-performance computing unit is used to run a dynamic variation autoencoder and an ADNN model;

[0285] A database server, which is used to store preprocessed data, feature data sets, and model prediction results, and supports data reading, writing, and retrieval;

[0286] A central processing unit, which is responsible for executing the predictive feedback loop, including: comparative analysis, ADNN model self-adjustment and weight update.

[0287] Vibration sensors: such as MEMS accelerometers, are used to monitor and record the vibration frequency and amplitude generated by the tool during the machining process in real time. These data can reveal the wear status and cutting performance of the tool.

[0288] Acoustic sensors: such as condenser microphones, are used to capture the sound waves generated during the machining process. Changes in the sound often indicate changes in the condition of the tool.

[0289] Temperature sensors: such as thermocouples or infrared temperature sensors, monitor temperature changes in the tool or machining area. High temperatures may indicate tool wear or excessive load.

[0290] The data preprocessing unit uses an embedded system or a dedicated microprocessor, such as the ARM Cortex series, to receive the raw data from the sensor and perform data cleaning, standardization, and denoising operations. Preprocessing is to ensure data quality and provide accurate and clean data input for subsequent feature extraction and analysis.

[0291] Data preprocessing is a key step to ensure model accuracy. Standardization usually uses Z-score standardization to convert data into a distribution with a mean of 0 and a variance of 1, which helps speed up model training and improve accuracy. The denoising process uses a low-pass filter (such as a Butterworth filter) or a wavelet transform to remove high-frequency noise in the data, thereby retaining useful signals and reducing errors.

[0292] Compute units refer to servers equipped with GPU or TPU accelerators, which are responsible for executing computationally intensive Dynamic Variation Autoencoder (DVAE) and Adaptive Dynamic Neural Network (ADNN) models that can extract meaningful features from processed monitoring data and make lifespan predictions based on these features.

[0293] DVAE maps high-dimensional monitoring data to a low-dimensional latent space through a self-encoding mechanism, and then reconstructs the data from this space. Its dynamic nature is reflected in its ability to automatically adjust the network structure of the encoder and decoder according to the characteristics of the input data, such as changing the number of layers or adjusting the number of neurons. This process involves importance weighting or gradient-based methods to identify and encode abnormal patterns and nonlinear characteristics in the data.

[0294] Database servers, such as servers with high-speed SSD storage and high-performance processors, are used to securely store pre-processed data, feature data sets, and prediction results. They support efficient data read, write, and retrieval operations, providing support for model training and query.

[0295] The database server plays the role of a data warehouse in this system, storing all data and results from preprocessing to feature extraction and model prediction. It supports efficient data reading, writing and retrieval operations, ensuring smooth model training and prediction processes. The database design uses high concurrency and index optimization technology to cope with the processing needs of large-scale data sets.

[0296] The central processing unit (CPU) is responsible for coordinating the operation of the entire system in the present invention, including executing predictive feedback loops. This usually involves comparing the model prediction results with the actual observed data, and adjusting the model according to the deviation, such as adjusting the network weights or structure. The CPU can be provided by a high-performance computing platform, such as Intel Xeon or AMD EPYC series, which have the ability to handle complex algorithms and large-scale data sets.

[0297] The adjustment strategy of the ADNN model relies on a comprehensive evaluation of the input data, including data dimension, complexity, and variability. Based on these evaluation results, the ADNN model can dynamically increase or decrease network layers, adjust the number of neurons, or modify the connection pattern. This adaptability ensures that the model can effectively process data of various complexities while maintaining computational efficiency.

[0298] The predictive feedback loop first compares the initial life prediction results with the actual observed data. By analyzing the deviation of the prediction results, the ADNN model can adjust itself. This includes adjusting the network structure, optimizing parameters (such as learning rate), and even changing the loss function. By iterating this process, the model gradually reduces the prediction deviation and improves the prediction accuracy.

[0299] The system of the present invention uses deep learning technology to deeply analyze subtle changes in monitoring data, thereby accurately predicting the life of the tool. The dynamic variation autoencoder (DVAE) improves the quality of feature representation by learning nonlinear features and abnormal patterns in the data; the adaptive dynamic neural network (ADNN) can dynamically adjust the network structure according to the characteristics of the input data to optimize the prediction performance. The predictive feedback loop further improves the accuracy and robustness of the model by monitoring the deviation between the predicted results and the actual life in real time, and optimizing the model accordingly to enable it to continuously adapt to new data and changes in conditions. The entire system is designed to achieve efficient and accurate tool life prediction to reduce production interruptions and improve manufacturing efficiency.

[0300] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0301] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0302] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0303] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0304] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0305] Memory includes non-permanent memory in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0306] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0307] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0308] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for predicting tool life of CNC machine tools based on deep learning, It is characterized in that The following steps are involved: Collecting the monitoring data of the tool and performing standardization and denoising preprocessing, wherein the monitoring data includes: vibration data, sound data and temperature data; The dynamic variation autoencoder is used to extract and encode the monitoring data, capture abnormal patterns and nonlinear characteristics, and generate an encoded feature data set; the feature data set is input into the adaptive dynamic neural network model, and the network structure and depth are dynamically adjusted according to the characteristics of the input data to generate the initial life prediction results; By means of a predictive feedback loop, the initial wear prediction result is compared with the actually observed tool life, and the adaptive dynamic neural network model is self-adjusted by the comparison result to generate an optimized adaptive dynamic neural network model and update the initial life prediction result; the self-adjustment of the adaptive dynamic neural network model by the comparison result includes: Analyze the prediction deviation of the adaptive dynamic neural network model by comparing the results, wherein the prediction deviation includes: the size of the deviation, overestimation or underestimation and its cause; According to the results of the deviation analysis, a model adjustment strategy is formulated, wherein the model adjustment strategy includes: adjusting the network structure, optimizing the learning rate, modifying the loss function, or adjusting the processing method of the input features; Based on the model adjustment strategy, the adaptive dynamic neural network model is adjusted; During the iterative training process of the adaptive dynamic neural network model, the adjusted adaptive dynamic neural network model refines its parameters through repeated forward propagation, loss calculation, back propagation and parameter update; Before the predictive feedback loop is started, a hybrid data source integrated learning method is used to fuse and weight the feature data set, obtain the fused feature data and input it into the optimized adaptive dynamic neural network model for final life prediction, and output the final life prediction result of the tool; The actual tool usage data and the final life prediction results are comprehensively considered for performance evaluation. The feature extraction, adaptive dynamic neural network model adjustment and feature fusion weights are adjusted based on the evaluation results to comprehensively optimize the life prediction performance.

2. According to the method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that The dynamic variation autoencoder is used to identify and encode abnormal patterns and nonlinear characteristics in monitoring data. By inputting the monitoring data into the dynamic variation autoencoder, part of the high-dimensional monitoring data is converted into a low-dimensional potential representation, and an encoded feature data set is generated to complete feature extraction; The feature data set includes vibration coding features, sound coding features and temperature coding features.

3. The method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that The dynamically adjusting the network structure and depth according to the characteristics of the input data includes: The adaptive dynamic neural network model evaluates key characteristics of the data in the feature data set and outputs evaluation results, wherein the key characteristics include: dimension, complexity, variability, abnormal patterns and nonlinear characteristics of the data; Based on the evaluation results and the initial network structure, a dynamic adjustment strategy is formulated. The dynamic adjustment strategy is used to define when and how to adjust the structure and depth of the network.

4. The method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that In the adaptive dynamic neural network model that completes the dynamic structural adjustment, the feature data set is received through the input layer; in each layer of the adaptive dynamic neural network model, the input data is first matrix multiplied with the weight of the layer, and then the bias term is added, and finally a nonlinear transformation is performed through the activation function. This process is repeated in all layers until the output layer; in the output layer, after the last round of weighted and activation operations, the initial life prediction result is generated.

5. The method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that The predictive feedback loop includes: Generate initial life prediction results by training an adaptive dynamic neural network model using monitoring data; Compare the initial life prediction results with the actual observed tool life to identify deviations in life prediction; Based on the deviation in the life prediction, the adaptive dynamic neural network model is self-adjusted, and the adjusted adaptive dynamic neural network model is trained and predicted again; Repeat the above process, each time the optimized adaptive dynamic neural network model is used to update the initial life prediction result, and the updated initial life prediction result is compared with the actually observed tool life to continue to guide the further adjustment of the adaptive dynamic neural network model; The feedback loop continues until a preset number of iterations is completed or the accuracy of the prediction results meets the requirements.

6. The method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that The method of applying a hybrid data source integrated learning method to fuse and weight feature data sets includes: Perform importance analysis on the features in the feature data set to determine the contribution of each feature to the prediction task; based on the results of the importance analysis, assign corresponding weights to each feature; According to the determined weights, all features are weighted fused to obtain fused feature data.

7. The method for predicting tool life of CNC machine tools based on deep learning according to claim 1, It is characterized in that The performance evaluation by comprehensively considering the actual tool usage data and the final life prediction results includes: Collect tool usage records and actual life data, wherein the tool usage records include: usage conditions, working environment and tool replacement time points; The final life prediction result is compared with the actual life data, and the accuracy and error of the final life prediction result are calculated; Analyze tool usage records when the prediction effect is good or insufficient, and determine the corresponding influencing factors.

8. A system for implementing the method for predicting tool life of CNC machine tools based on deep learning according to any one of claims 1 to 7, It is characterized in that include: Monitoring hardware, monitoring hardware includes: vibration sensor, used to collect vibration data of the tool in real time; sound sensor, used to capture sound data when the machine tool is running; temperature sensor, used to monitor the temperature of the tool or processing area; A data preprocessing unit, the data preprocessing unit is used to receive raw data from the sensor and perform standardization and denoising; A computing unit, which is used to run a dynamic variation autoencoder and an adaptive dynamic neural network model; A database server, which is used to store preprocessed data, feature data sets, and model prediction results, and supports data reading, writing, and retrieval; A central processing unit is responsible for executing a predictive feedback loop including comparative analysis, adaptive dynamic neural network model self-adjustment and weight updating.

Citation Information

Patent Citations

  • Online prediction method for residual life of numerical control lathe tool

    CN116533063A

  • Tool wear monitoring system capable of realizing self-learning in numerical control machining state

    CN102073300A

  • Prediction system and method for life of machine tool cutter

    CN108536938A

  • Tool wear state detection method and apparatus based on convolutional neural network and device

    CN109048492A

  • Residual life prediction method of cutting tools for key equipment based on depth neural network

    CN109145319A