A method and system for predicting the life of CNC tools considering real-world scenarios.

Abstract

This invention relates to a method and system for predicting the lifespan of CNC tools considering real-world scenarios. The prediction method runs on an edge device that is communicatively connected to a cloud platform. The prediction method includes the following steps: acquiring basic data of the real-world scenario, extracting features of the real-world scenario, and matching the features of the real-world scenario with preset scenario features; obtaining the adaptation evaluation results of the lifespan prediction models in the model library of the cloud platform with the current real-world scenario based on the matching results; generating a corresponding lifespan prediction model based on the adaptation evaluation results, and predicting the lifespan of CNC tools under the current real-world scenario. Compared with existing technologies, this invention combines model recommendation and optimization algorithms to achieve scenario-consistent prediction of remaining tool lifespan, and has advantages such as scenario adaptability and high prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent tool life prediction technology, and in particular to a method and system for predicting the life of CNC tools that takes into account real-world scenarios. Background Technology

[0002] With the rapid development of modern equipment manufacturing, various industries are placing higher demands on the quality and production costs of key components. CNC machine tools, due to their advantages of high machining accuracy and high degree of automation, have become crucial equipment in the increasingly digital and intelligent development of industry. As the main component of the CNC machine tool machining process, the quality of CNC cutting tools directly affects the machining quality of parts and the production efficiency of enterprises. Therefore, predicting their remaining service life has become an urgent problem to be solved.

[0003] Currently, deep learning-based methods for predicting remaining tool life have been widely applied. For example, patent application CN113762182A discloses a tool wear state prediction method based on deep network adaptation. However, the source domain dataset and target domain dataset of the deep network adaptation method should be collected in the same environment, and the target domain dataset must be labeled data, making the application conditions of this method quite demanding. Another patent application CN113867263A discloses a tool life prediction method based on cloud-edge collaboration and machine learning. This patent is applicable to a relatively limited scenario, only effective for current datasets. Furthermore, the optimization method in the cloud training module uses a quadratic convolutional neural network for optimization, which is prone to overfitting and reduces the accuracy of the model.

[0004] However, existing methods suffer from the problem of the tool usage scenario and the corresponding prediction model. That is, a prediction model is typically only effective for the specific scenario from which the training data is sourced. If the scenario changes, the model becomes unsuitable and requires re-evaluation and retraining. Due to the complexity of CNC machining scenarios, changes in part materials, tool types, and process parameters can alter the scenario definition, changing the model's training dataset and consequently the original training model. Furthermore, there is a lack of effective management methods for many existing lifespan prediction models, and effective model recommendation mechanisms are urgently needed to address real-world machining scenarios. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a CNC tool life prediction method and system that can achieve scene adaptation and improve prediction accuracy by taking into account real-world scenarios.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for predicting the lifespan of CNC tools considering real-world scenarios, characterized in that the method operates on an edge device, which is communicatively connected to a cloud platform, and the prediction method includes the following steps:

[0008] Acquire basic data of real-world scenarios, extract real-world scenario features from the basic data of real-world scenarios, and match the real-world scenario features with preset scenario features. The preset scenario features are stored in the scenario library of the cloud platform. Each preset scenario feature corresponds to multiple category tags, and each real-world scenario feature corresponds to one real-world tag. The matching process is performed based on the relationship between the real-world tag of the real-world scenario feature and each category tag of the corresponding preset scenario feature.

[0009] Based on the matching results, obtain the adaptation evaluation results between the lifetime prediction model in the model library of the cloud platform and the current real-world scenario;

[0010] Based on the adaptation evaluation results, a corresponding life prediction model is generated to predict the life of CNC tools in the current real-world scenario.

[0011] Furthermore, the matching process specifically includes the following steps:

[0012] 101) Compare the reality label of the reality scene feature with the corresponding classification label of different existing scenes in the scene library to determine whether the current reality scene is the same as a certain existing scene in the scene library. If yes, output the matching result of the current reality scene and the existing scene. If no, proceed to step 102).

[0013] 102) Compare the real-world labels of the real-world scene features with each category label of the corresponding preset scene features in the scene library, and determine whether the real-world labels of all real-world scene features are the same as a certain category label of the corresponding preset scene features. If yes, output the first part of the matching result; if no, output the second part of the matching result.

[0014] Furthermore, when the matching result is the result of matching the current real scene with the existing scene, the adaptation evaluation result is obtained as a complete adaptation. The life prediction model corresponding to the existing scene is used as the final life prediction model to predict the life of CNC tools in the current real scene.

[0015] Furthermore, when the matching result is the first part of the matching result, the adaptation evaluation result is obtained through the following steps:

[0016] Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features;

[0017] Encoding is performed based on the comparison between the real-world label of each real-world scene feature and the category label of each existing scene. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and the category label corresponding to each existing scene in the scene library is calculated, and the maximum value is assigned at the position corresponding to the category label of the existing scene to generate a real-world scene encoding matrix. The corresponding adaptation evaluation result is obtained based on the real-world scene encoding matrix.

[0018] Furthermore, each existing scene in the scene library has a corresponding scene encoding matrix pre-constructed through one-hot encoding to facilitate comparison with real-world scenes.

[0019] Furthermore, when the matching result is the second part of the matching result, the adaptation evaluation result is obtained through the following steps:

[0020] Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features;

[0021] Encoding is performed based on the comparison between the real-world label of each real-world scene feature and each category label of the corresponding preset scene feature. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and each category label of the corresponding preset scene feature is calculated, and the maximum value is assigned at the position corresponding to the category label to generate a real-world scene encoding matrix. The corresponding adaptation evaluation result is obtained based on the real-world scene encoding matrix.

[0022] Furthermore, obtaining the corresponding adaptation evaluation result based on the real-world scene encoding matrix specifically involves:

[0023] The formula for converting the real-world scene encoding matrix into a percentage system is as follows:

[0024]

[0025] In the formula, This represents the sum of scores for real-world scene features that share the same category label. This represents the sum of scores for real-world scene features for which no matching category label was found. , These are the weights of the features in each scene. , This represents the maximum similarity.

[0026] The score range is For scores below a certain threshold, the fit assessment result is defined as partial fit, and scores below a certain threshold are considered part of the fit. If the score is not met, the fit assessment result will be deemed unsuitable.

[0027] Furthermore, if the adaptation evaluation result is partial adaptation, the K-nearest neighbor algorithm is used to obtain the existing scenario that best matches the current real-world scenario. The life prediction model corresponding to the existing scenario is then optimized by the cloud platform and used as the final life prediction model to predict the life of CNC tools in the current real-world scenario.

[0028] Furthermore, the formula used in the K-nearest neighbor algorithm is:

[0029]

[0030] in, and These represent the expanded vectors of the scene encoding matrices of scenes already stored in the scene library and the expanded vectors of the actual scene encoding matrices, respectively. Indicates the index of the expanded vector. Indicates the dimension of the expanded vector. Represents a distance metric.

[0031] Furthermore, if the adaptation assessment result is that it is not compatible, the life prediction model is reconstructed by the cloud platform based on the historical data of the current real-world scenario, and used as the final life prediction model to predict the life of CNC tools in the current real-world scenario.

[0032] The present invention also provides a CNC tool life prediction system that takes into account real-world scenarios, characterized in that it includes a cloud platform and an edge device connected by communication, wherein the edge device executes the CNC tool life prediction method that takes into account real-world scenarios as described above.

[0033] This invention addresses the characteristics of common tool processing scenarios by performing feature modeling of these scenarios and establishing a model library by matching corresponding tool remaining service life prediction models within each scenario. Combining model recommendation and optimization algorithms, it achieves self-consistent tool remaining service life prediction for each scenario. This solves the problems of lacking effective management methods for a large number of tool life prediction models and the difficulty in selecting suitable life prediction models for real-world scenarios. Compared with existing technologies, this invention has the following advantages:

[0034] 1. This invention evaluates the actual machining scenario and the scenario corresponding to the model in the model library, confirms the fit based on the evaluation results, and recommends or reconstructs the tool remaining service life prediction model based on the fit. It can quickly identify and recommend a suitable tool remaining service life prediction model when encountering complex scenarios.

[0035] 2. When the actual machining scenario is fully adapted, the tool remaining service life prediction model is directly applied. When it is partially adapted, the optimized model is applied through model optimization. When it is not adapted, model reconstruction is considered, which can make the actual machining scenario more likely to match the corresponding tool remaining service life prediction model.

[0036] 3. This invention adopts an edge-cloud collaborative architecture, which is implemented through edge devices and a cloud platform via communication connections. The scene library, model library, and model optimization are deployed on the cloud platform, while model matching, evaluation, and recommendation are deployed on the edge devices. Because the cloud platform has advantages such as large storage capacity and high online computing efficiency, it can improve the coverage of the scene and model libraries. Furthermore, the edge devices can share the computing pressure of the cloud platform, reduce latency between local and cloud environments, and improve the operating efficiency of the edge-side model recommendation system.

[0037] 4. The scenario library of this invention includes scenario data based on multiple real-world scenarios. Based on scenario attributes such as the properties of the processed parts, tool properties, process properties, equipment properties, and other attributes, the scenario is feature-modeled, which can better express the characteristics of the scenario. At the same time, a "scenario-model" relationship is constructed, and the corresponding tool remaining service life prediction models in each scenario are organized to build a tool remaining service life prediction model library. This allows the relevant scenarios to be quickly matched with models, improving the efficiency of obtaining the appropriate model. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the structure of a CNC tool life prediction system according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of a model recommendation and optimization application process in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of a scene feature evaluation process in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0042] This invention summarizes and models the features of common tool processing scenarios, matches the corresponding tool remaining service life prediction model in the scenario, establishes a "scenario-model" relationship, and combines the model recommendation and optimization process to achieve scenario-consistent prediction of tool remaining service life.

[0043] One specific embodiment of the present invention provides a method for predicting the lifespan of CNC tools considering real-world scenarios. This method operates on an edge device that is communicatively connected to a cloud platform. The prediction method includes the following steps:

[0044] S1. Obtain basic data of the real scene, extract real scene features from the basic data of the real scene, and match the real scene features with preset scene features. The preset scene features are stored in the scene library of the cloud platform. Each preset scene feature corresponds to multiple category tags, and each real scene feature corresponds to one real tag. The matching process is performed based on the relationship between the real tag of the real scene feature and each category tag of the corresponding preset scene feature.

[0045] S2. Based on the matching results, obtain the adaptation evaluation results between the lifetime prediction model in the model library of the cloud platform and the current real-world scenario;

[0046] S3. Generate a corresponding life prediction model based on the adaptation evaluation results, and perform CNC tool life prediction in the current real-world scenario.

[0047] In the above steps, the scene features involve attributes including multiple attributes such as the processed part, tool, process, equipment, and others. Each attribute has one or more category features, meaning several preset scene features belong to the same scene attribute, and each preset scene feature has several classification labels. It's important to note that the classification labels are represented only in text form. The subsequent encoding process compares the text versions of the classification labels in the two scenes to be compared. If they match, they are assigned a value of 1; otherwise, a similarity score is used, and the rest are filled with 0s. The significance of encoding is to facilitate the model's recommendation module in recognizing and quantifying the similarity between different scenes.

[0048] The scene library contains several existing scenes. Each category feature of each existing scene corresponds to a specific classification label, which is the specific feature value of that category feature. Similarly, the reality label of a reality scene feature is also the specific feature value of the corresponding reality scene feature, and they all follow the classification rules of the same classification label. For each existing scene in the scene library, a corresponding scene encoding matrix is ​​pre-constructed through one-hot encoding based on the comparison of the above classification labels.

[0049] In step S1, the matching process specifically includes the following steps:

[0050] 101) Compare the reality label of the reality scene feature with the corresponding classification label of different existing scenes in the scene library to determine whether the current reality scene is the same as a certain existing scene in the scene library. If yes, output the matching result of the current reality scene and the existing scene. If no, proceed to step 102).

[0051] 102) Compare the real-world labels of the real-world scene features with each category label of the corresponding preset scene features in the scene library, and determine whether the real-world labels of all real-world scene features are the same as a certain category label of the corresponding preset scene features. If yes, output the first part of the matching result; if no, output the second part of the matching result.

[0052] Based on the matching results obtained in step S1, the adaptation evaluation results determined in step S2 are categorized as fully adapted, partially adapted, and misfit. When the actual machining scenario is fully adapted, the tool remaining life prediction model corresponding to the fully adapted scenario is directly applied. When it is partially adapted, the optimized model is applied using model optimization methods. When it is misfit, model reconstruction is considered. This approach increases the probability that the actual machining scenario will match the corresponding tool remaining life prediction model.

[0053] In steps S2 and S3, when the matching result is the result of matching the current real scene with the existing scene, the adaptation evaluation result is obtained as a complete fit, and the life prediction model corresponding to the existing scene is used as the final life prediction model to predict the life of CNC tools in the current real scene.

[0054] When the matching result is the first part of the matching result, that is, the real-world scene feature is in the specific category label contained in the scene library, but does not belong to any existing scene, the adaptation evaluation result is obtained based on the similarity between the real-world label of the real-world scene feature and other category labels other than the current real-world label. Specifically:

[0055] Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features;

[0056] Feature encoding is performed based on the comparison between the real-world label of each real-world scene feature and the category label of each existing scene. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and the category label corresponding to each existing scene in the scene library is calculated, and the maximum value is assigned at the position corresponding to the category label of the existing scene to generate a real-world scene encoding matrix.

[0057] The real-world scene encoding matrix is ​​converted into a percentage system, with the score range as follows: For scores below a certain threshold, the fit assessment result is defined as partial fit, and scores below a certain threshold are considered part of the fit. If the score is not met, the fit assessment result will be deemed unsuitable.

[0058] When the matching result is the second part of the matching result, that is, the real-world scene feature is not in any of the classification labels included in the scene library, the adaptation evaluation result is obtained based on the similarity between the real-world scene feature's real-world label and all specific classification labels in the scene library, specifically:

[0059] Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features;

[0060] Feature encoding is performed based on the comparison between the real-world label of each real-world scene feature and each category label of the corresponding preset scene feature. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and each category label of the corresponding preset scene feature is calculated, and the maximum value is assigned at the position corresponding to the category label to generate a real-world scene encoding matrix.

[0061] After model training, the classification labels of the preset scenes can be expanded so that when encountering new real-world scenes, the similarity calculation process can be skipped and the value 1 can be directly assigned to the corresponding position of the classification label.

[0062] The real-world scene encoding matrix is ​​converted into a percentage system, with the score range as follows: For scores below a certain threshold, the fit assessment result is defined as partial fit, and scores below a certain threshold are considered part of the fit. If the score is not met, the fit assessment result will be deemed unsuitable.

[0063] The above similarity is calculated using cosine similarity.

[0064] Based on the above results, if the adaptation evaluation result is partial adaptation, the K-nearest neighbor algorithm is used to obtain the existing scenario that best fits the current real-world scenario. The cloud platform then optimizes the life prediction model corresponding to this existing scenario, and this optimized model becomes the final life prediction model for predicting the life of CNC tools in the current real-world scenario. If the adaptation evaluation result is no adaptation, the cloud platform reconstructs the life prediction model based on historical data from the current real-world scenario, and this reconstructed model becomes the final life prediction model for predicting the life of CNC tools in the current real-world scenario.

[0065] Through the above steps, the edge device realizes the recommendation of the optimal life prediction model considering the real-world scenario and the prediction of CNC tool life. At the same time, the cloud platform realizes the construction and storage of the scene library and model library, as well as model optimization and reconstruction. It takes advantage of the cloud platform's advantages of large storage capacity and high online computing efficiency, which can improve the coverage of the scene library and model library. Meanwhile, the edge device can share the computing pressure of the cloud platform, reduce the latency between local and cloud, and improve the operating efficiency of the edge model recommendation system.

[0066] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] Another specific embodiment of the present invention provides a CNC tool life prediction system based on an edge-cloud collaborative architecture that considers real-world scenarios, including a cloud platform and edge devices with communication connections, such as... Figure 1 As shown, the edge device executes the CNC tool life prediction method that considers real-world scenarios, as described above. The system deploys the scenario library, model library, and model optimization on the cloud platform, and deploys model recommendations on the edge device; this edge-cloud collaboration improves the overall system's operating efficiency.

[0068] (1) Cloud platform

[0069] In this embodiment, as Figure 1As shown, the cloud platform includes a scene library B, a model library A, a cloud-based model training module C, and a model optimization module D. The cloud platform mainly serves to store model library and scene library data, train tool remaining service life prediction models, and optimize service life prediction models.

[0070] 1) Model Library

[0071] Model library A stores the structure of tool remaining service life prediction models, along with the results and weight vectors after model execution. Model attribute definitions in the library may include:

[0072] Model attributes = {model structure attributes, model training result attributes, weight attributes}

[0073] Model structure attributes: By searching relevant literature on tool remaining service life prediction models and combining them with the algorithm models studied in actual experiments, the structure of tool remaining service life prediction models is collected.

[0074] Model training result attributes: The remaining tool life prediction model is put into the model training module for training, and the training results are stored in the model library.

[0075] Weight attributes: After the tool remaining service life prediction model finishes running, the parameters connecting the nodes of each layer in the neural network are saved, which is to obtain a weight vector, and the weight is stored in the model library.

[0076] 2) Scene Library

[0077] The purpose of scenario library B is to store the definitions of various processing scenarios, the scenario matching conditions of each category of features in the scenario attributes, and the addresses of the remaining useful life prediction models corresponding to each scenario in the model library. This makes it easy to retrieve the corresponding scenario based on the actual processing information and further find the corresponding model in the model library.

[0078] The attribute definitions of scenes in the scene library can include:

[0079] Scene attributes = {Workpiece attributes, tool attributes, process attributes, equipment attributes, other attributes}

[0080] The tool processing scenario is defined according to the above attributes, and this information is classified and encoded using one-hot encoding. The specific definitions of the processed part attributes, tool attributes, process attributes, equipment attributes, and other attributes can be extended as necessary based on the characteristics of the enterprise and model on the cloud platform. The above description only provides a typical attribute composition.

[0081] It should be noted that the attribute values ​​described below are just examples; companies can select specific values ​​based on their own data characteristics.

[0082] ① Attributes of the part being processed = {part material, part size, blank supplier, etc.}

[0083] Parts materials: Commonly used parts materials in CNC machining processes are organized, such as high-temperature alloys, titanium alloys, aluminum, composite materials, plastics and pure metals.

[0084] Part Dimensions: Based on the collected part materials, the most common part dimensions for each material are listed, such as 100×100×10, 80×80×5, 60×80×10, 60×60×8, etc., in mm. The first number in the part dimension information represents the length of the part, the second number represents the width, and the third number represents the depth.

[0085] Raw material suppliers: By summarizing information, we have compiled a list of the most common raw material suppliers for various parts of the company, such as Baosteel, Ansteel, Wuhan Iron and Steel, Xiangtan Iron and Steel, Jiangxi Iron and Steel, Hebei Iron and Steel, Changgang Steel, Shanghai Iron and Steel, etc.

[0086] ② Tool attributes = { Tool category, tool model, tool material, tool supplier, tool batch, etc.}

[0087] Cutting tools categories: Common cutting tool types are organized, such as: milling cutters, boring tools, turning tools, drills, etc.

[0088] Tool Models: Common tool models are listed, such as: CNMG120408PS, CNMG432PS, TNMG160404L-C TN60, R840-0960-50-A1A 1220, etc.

[0089] Tooling materials: This section summarizes information on common tooling materials used in existing tooling scenarios, such as high-speed steel, cemented carbide, diamond, cubic boron nitride, and ceramics.

[0090] Tooling suppliers: We have compiled information on the most common CNC tooling suppliers, such as Coromant, Kennametal, Iscar, Seco, Walter, Mitsubishi Materials, Kyocera, etc.

[0091] Tool batches: The tool information compiled above is divided into different batches, such as the first batch, the second batch, and the third batch, and can be numbered, such as 1, 2, 3, 4, 5.

[0092] ③ Process attributes = {CNC program number, spindle speed, feed rate, cutting method, etc.}

[0093] CNC program numbering: Common CNC machining programs are organized and numbered, such as 1, 2, 3, 4, 5, etc.

[0094] Spindle speed: Common spindle speeds are listed, such as 500, 1000, 2000, 3000, 5000, etc., in r / min.

[0095] Feed rate: Common feed rates are listed, such as 10, 20, 50, 100, etc., in mm / min.

[0096] Cutting methods: Common cutting methods for tool processing are organized, such as milling, drilling, sawing, grooving, tapping, broaching, polishing, etc.

[0097] ④ Equipment attributes = {Equipment model, equipment supplier, equipment age, equipment owner, etc.}

[0098] Equipment Models: A list of common CNC equipment models, such as: CKD6140, CKD6150, CK5112, CK5116, CK6140, CJK6130, VMC855, α-D-21SiA, etc.

[0099] Equipment Suppliers: This section compiles information on the most common CNC equipment suppliers, such as FANUC, SIEMENS, YASKAWA, ABB, KUKA, Yangtie, DELTA, etc.

[0100] Equipment age: Common equipment age information is compiled, such as 7, 8, 9, 10, etc., in units of years.

[0101] Equipment Ownership: Organize the companies that own the CNC equipment and label them, such as: 1, 2, 3, 4, 5, etc.

[0102] ⑤ Other attributes = {Workplace temperature, workplace humidity, etc.}

[0103] Workplace temperature: The workplace temperature is divided into certain intervals, such as 5℃, 10, 15, 20, 25, 30, etc., with the unit being ℃.

[0104] Workplace humidity: The humidity of the workplace is divided into certain intervals, such as 30, 40, 50, 60, 70, 80, etc., with the unit being %rh.

[0105] ⑥ Scene matching conditions for each category feature in scene attributes: The classification labels under each category feature in the real scene may not be completely consistent with the classification labels in the scene library. Therefore, corresponding scene matching conditions are needed to evaluate the real scene and the scene in the scene library.

[0106] ⑦ Scene Features - Model Corresponding Attributes: The tool processing scene and the corresponding model are bound like key-value pairs in a dictionary. The specific tool processing scene is the key in the key-value pair, and the remaining service life prediction model is the value in the key-value pair, with a one-to-one correspondence.

[0107] 3) Cloud-based model training module

[0108] The main function of the cloud-based model training module C is to train the lifespan prediction model. Specifically, it trains the collected remaining lifespan prediction models in this module and obtains the training results and weight vectors.

[0109] 4) Model Optimization Module

[0110] The model optimization module D is used to: when the real scene and the scene in the scene library are partially adapted, obtain the structure and corresponding weights of the tool remaining service life prediction model from the model library recommended by the edge device, and train it on the dataset obtained from the real scene through fine-tuning to achieve the optimization effect. The optimized model is used as the final model for edge device prediction.

[0111] (2) Edge devices

[0112] Edge devices primarily serve as the model recommendation tool at the edge, and use the recommended models for lifetime prediction. The edge device includes an edge model recommendation module E, which comprises a scene matching module E1, a scene evaluation standard module E2, and a recommendation module E3.

[0113] Scene matching module E1: Matches scene features in the real scene with scene features in the scene library to obtain the matching results between the real scene and the scene library.

[0114] Scene evaluation standard module E2: Based on the matching situation reflected in the evaluation module, the real scene is scored and the scoring result of the real scene is obtained. The scoring result has three situations: fully adapted, partially adapted, and unsuitable.

[0115] Recommendation Module E3: Based on the real-world scenario scoring results obtained from the evaluation criteria module, the recommendation algorithm recommends the most suitable tool remaining service life prediction model from the model library for the real-world scenario.

[0116] Based on the above structure, the CNC tool life prediction system of this embodiment mainly realizes two major functions: model management and model recommendation.

[0117] Model Management: In the cloud platform, a library of common CNC tool processing scenarios is established based on the characteristics of different CNC machining scenarios. Training models are built in the cloud platform, and the remaining tool life prediction model is trained on the datasets collected in different scenarios. The trained models are then bound to the scenarios.

[0118] Model Recommendation: A remaining tool life prediction model recommendation module is deployed on edge devices to recommend suitable tool life prediction models based on the matching between real-world scene features and scene features in the scene library. If the real-world scene and the scene in the scene library are only partially compatible, the model is optimized using the model optimization module in the cloud platform before being applied to the real-world scene.

[0119] (2) Model Management Process

[0120] 1) Establish a model library

[0121] Establishing a model library refers to collecting remaining useful life prediction models and storing them in the library, while assigning an address to the location where each model is stored, ensuring a one-to-one correspondence between the model and the address. Specifically, this includes:

[0122] S101: By searching the literature on tool remaining service prediction models and combining them with the algorithm models studied in actual experiments, tool remaining service prediction models are collected.

[0123] S102: Store the collected models into a model library and assign an address to the location where each model is stored, so that there is a one-to-one correspondence between the model and the address.

[0124] 2) Establish a scene library

[0125] Establishing a scene library refers to taking the dataset trained by the collected tool remaining service life prediction model, describing the scene features based on the scene characteristics of the collected dataset, referring to the aforementioned scene feature description methods, and storing the description results along with the model's address in the model library into the scene library. Specifically, this includes:

[0126] S201: Describe the scene features corresponding to the training dataset of the lifespan prediction model, referring to the aforementioned scene feature description method. The description result will be represented in the form of one-hot encoding. The scene encoding is a two-dimensional matrix, where rows represent category features and columns represent category labels. 1 indicates the status of that attribute in the scene, and 0 represents the rest.

[0127] For example, the scene features in the scene library include five types: workpiece attributes, tool attributes, process attributes, equipment attributes, and other attributes. The specific classification is shown in Table 1.

[0128] Table 1 Examples of scene features in the scene library

[0129]

[0130] Therefore, for the scenario {Part material: high-temperature alloy, Tool category: boring tool, CNC program code: 3, Equipment model: 1, Workplace temperature: 15}, the unique thermal code is: .

[0131] S202: The scenario description results and the model's address in the model library are stored together in the scenario library. This allows users to input real-world scenario features, index the corresponding scenario in the scenario library, and obtain the corresponding remaining useful life prediction model from the model library using the model address provided by that scenario. It is not excluded that a model may be applicable to multiple scenarios; therefore, the model addresses provided by different scenarios in the scenario library may be the same.

[0132] (3) Model recommendation process

[0133] The model recommendation process includes acquiring real-world scene feature information and the model recommendation and optimization application process, specifically including:

[0134] S30: Acquire feature information of the real scene, including feature information of tool material, tool size, data acquisition type and material being processed in the real tool processing scene, and classify and encode the feature information.

[0135] S40: Recommend suitable remaining useful life prediction models for real-world processing scenarios through a model recommendation and optimization system.

[0136] 1) Process for obtaining feature information of real-world scenes

[0137] Obtaining real-world scene feature information refers to the need to obtain information on the parts being processed, tools, processes, equipment, etc., that are matched in the scene library before applying the data collected in the real-world scene to the model recommendation and optimization system.

[0138] The detailed steps are as follows:

[0139] S301: Obtain specific scene feature information in the real-world processing scenario.

[0140] S302: Classify and encode real-world scenes using one-hot encoding.

[0141] Scene attributes consist of the following properties:

[0142] {Workpiece attributes, tool attributes, process attributes, equipment attributes, other attributes}

[0143] It should be noted that the specific scenario attributes can be supplemented and defined according to the respective processing scenarios of each enterprise.

[0144] STEP 1: Count the number of category features and classification labels in the statistical scene attributes.

[0145] STEP 2: Encode the features of the real-world scene, using 1 to represent the case where the real-world scene matches the scene in the scene library; if they do not match, use cosine similarity. Calculate the maximum result obtained and fill in the corresponding category label position; use 0 to represent the rest. (Cosine similarity formula) This represents the evaluation vector of this category of features in the real-world scene. Evaluation vectors representing features of the same category in the scene library.

[0146] The meaning of the parameters in the evaluation vector can be defined according to different category features. The scenario matching conditions for the classification labels in the category features are explained in Table 2. The scenario matching conditions described in Table 2 are just examples. Enterprises can select specific scenario matching conditions based on the category features of their own scenarios.

[0147] Table 2 Examples of Scene Matching Conditions in the Scene Library

[0148]

[0149] If a scene category feature in a real-world scenario does not exist in the scene library, an evaluation vector is constructed by combining the real-world scenario's category feature with each category label in the scene library. If the scene matches the scene matching criteria given in the scene library, the corresponding value in the vector is assigned 1; otherwise, it is assigned -1. The correlation between the evaluation vectors built from the real-world scenario and those built from the scene library is calculated using cosine similarity to quantify the degree of correlation between the real-world scenario category features and the scene category features in the scene library.

[0150] Because the combination of classification labels among the numerous category features in the scene library does not necessarily have a complete corresponding model in the model library. For example, a real-world scene might be {Part Material: High-Temperature Alloy, Tool Category: Boring Tool, CNC Program Code: 3, Equipment Model: 1, Workplace Temperature: 15}, where "High-Temperature Alloy" and "Boiling Tool" are real-world labels for various features in this scene, and corresponding models exist in the model library. However, a scene like {Part Material: Titanium Alloy, Tool Category: Boring Tool, CNC Program Code: 3, Equipment Model: 1, Workplace Temperature: 15} might not have a corresponding model in the model library. However, during one-hot encoding, a value of 1 would be assigned to {Part Material: Titanium Alloy}. In this case, the method of establishing evaluation vectors and calculating cosine similarity can be used to calculate the category features. Evaluation vectors are established for all classification labels in this category feature in the scene library, and the evaluation vectors established for the real-world scene are compared with those established in the scene library using cosine similarity. Here, the evaluation vectors established in the scene library should exclude evaluation vectors that are identical to the classification labels of the real-world scene.

[0151] For example, the scene attributes in the scene library include five types: workpiece attributes, tool attributes, process attributes, equipment attributes, and other attributes. The specific classification is consistent with that described in S201.

[0152] Therefore, for the scenario {Part Material: Pure Metal, Tool Category: Boring Tool, CNC Program Code: 3, Equipment Model: 1, Workplace Temperature: 15}, the corresponding feature information for tool category, CNC program code, equipment model, and workplace temperature can be found in the scenario library. Only the evaluation score obtained for the part material in the real-world scenario needs to be calculated. An evaluation vector is constructed based on whether the part is metal, whether it is a synthetic material, and whether it can be used in a high-temperature environment. If the above conditions are met, the corresponding value in the vector is assigned 1; otherwise, it is assigned -1. For the evaluation vector of the part material in the real-world scenario, for pure metal: The evaluation vector for the part material in the scene library is: high-temperature alloy: ,plastic: Therefore, the maximum value of the cosine similarity calculation result is... The final real-world scene features are encoded as follows: .

[0153] 2) Model Recommendation and Optimization Application Process

[0154] Based on the evaluation results of real-world scene characteristics and scene characteristics in the scene database, a tool remaining service life prediction model suitable for the real-world scene is recommended, and then optimized by an optimization algorithm, such as... Figure 2 As shown, it specifically includes:

[0155] S401: Evaluate the characteristics of the real-world scene. Based on the evaluation results, the scene's adaptability will be divided into three categories: fully adapted, partially adapted, and unadapted.

[0156] Scene feature evaluation process as follows Figure 3 As shown, it includes:

[0157] STEP 1. Evaluate the matching between real-world scene features and scenes in the scene library, that is, collect the category feature information of each category of features in the real-world scene and match the features with scenes in the scene library.

[0158] STEP 2: The above evaluation results will be summarized in percentage form. The initial score for the scene is 0, and the score will be calculated according to formula (1). If the category features in the real scene are consistent with the category features in the scene library, a score will be obtained. If there is an inconsistency, the matching degree between the real scene category features and the scene category features in the scene library will be calculated based on the cosine similarity, and then a score will be given; if the cosine similarity calculation result is 0, then each category label of that category feature will be assigned a very low value q (for example, q can be 0.1).

[0159] (1)

[0160] in and Representing the The first category feature and the first There are 100 category labels, and X points are because the scene library has a total of Y category attributes. The 100 points are divided equally into Y parts, i.e. Then, weights are assigned based on the different importance of scene characteristics. . The maximum probability value obtained from the cosine similarity calculation among the category features is, i.e. . This represents the summation of scores for cases where category features in the real-world scenario match category features in the scenario library. This means that when there are inconsistencies in the classification labels of the category features, scores are calculated and summed for the inconsistent category features.

[0161] STEP 3. Based on the final evaluation results, those with a score of 100 are classified as perfectly matched; those with a score range of... The cases that are divided into partial adaptations are classified as such. Generally in In between, without losing generality, You can take 75; the score range is lower than Those who are mismatched are categorized as such.

[0162] For example, the same as in S302 STEP 2. Consistent. Based on the cosine similarity calculation results, we can obtain... For ease of calculation, the weights here are... All are set to 1. Therefore, the final evaluation score can be calculated as follows: .

[0163] S402: If the scenario assessment results are not suitable, then consider model reconstruction.

[0164] S403: If the real-world scenario is a perfect fit, the recommended model will be applied directly.

[0165] S404: If the real-world scenario is a partially adapted case, the classification algorithm should be applied to S302. STEP The matrix calculated in step 2 is classified into scene matrices in the scene library to achieve the matching between real-world scenes and scenes in the scene library. The tool remaining service life prediction model corresponding to the matched scene is then optimized through transfer learning and applied to real-world scenes.

[0166] In this embodiment, the classification algorithm flow is as follows:

[0167] STEP 1: The real-world scene encoding matrix obtained through step S302 is expanded into row vectors and compressed into a one-dimensional space.

[0168] STEP 2: Obtain the corresponding tool remaining service life prediction model in the model library through the K-nearest neighbor algorithm, optimize the model using transfer learning, and then apply it to real-world scenarios.

[0169] (2)

[0170] in, and These represent the expanded vectors of the scene encoding matrix in the scene library and the expanded vectors of the real-world scene encoding matrix, respectively. The index represents the expanded vector. As a distance metric, since the input is primarily in discrete form, therefore The usual value is 2.

[0171] This embodiment uses a real-world tool processing scenario to illustrate a typical implementation process:

[0172] The CNC tool life prediction method in this embodiment, which considers real-world scenarios, can match real-world scenario features with scenario features in a scenario library. Relying on model recommendation and optimization algorithms, it can recommend tool life prediction models for real-world tool processing scenarios on edge devices. This solves the problem of the difficulty in selecting a suitable tool life prediction model from a large number of tool life prediction models for different processing scenarios, and realizes the management and recommendation of CNC tool life prediction models based on an edge-cloud collaborative architecture.

[0173] This embodiment uses a simplified real-world tool machining scenario as an example to demonstrate the specific process for predicting the remaining tool life. The process is as follows:

[0174] (1) Collect the remaining useful life prediction models by literature search and combined with the algorithm models studied in actual experiments, store them in the model library, and assign an address to the location where each model is stored.

[0175] (2) Describe the scene features corresponding to the training dataset of the lifespan prediction model, and encode each scene using one-hot encoding. Store the encoding along with the corresponding model address in the scene library. The specific classification of scene attribute features is shown in Table 3.

[0176] Table 3 Examples of Scene Features in the Scene Library

[0177]

[0178] Assume the scene library already contains 3 scenes, meaning the model library already contains models corresponding to these 3 scenes. The specific scene attribute definitions are as follows:

[0179] Scene 1:

[0180] Table 4. Feature Examples of Scenario 1

[0181]

[0182] Scene 2:

[0183] Table 5. Feature Examples for Scenario 2

[0184]

[0185] Scene 3:

[0186] Table 6 Examples of Features in Scenario 3

[0187]

[0188] Specific scenario attributes, category characteristics, and classification labels should be supplemented and defined according to the actual processing scenarios of each enterprise.

[0189] (3) Obtain category features in real-world scenarios. There are two situations for category features in real-world scenarios: one is that the real-world scenario features are not in the specific category labels listed in Table 3, and the other is that the real-world scenario features are in the specific category labels listed in Table 3, but are not in the existing scenarios 1 to 3 in the scenario library.

[0190] 1) Real-world scene characteristics are not included in the specific category labels listed in Table 3.

[0191] ① For example, the characteristics of a real-world processing scenario are:

[0192] Part material: high-temperature alloy; part size: 60×60×5 mm; blank supplier: Wuhan Iron and Steel Group.

[0193] Tool Category: Lathe Tools; Tool Model: TNMG160404L-C TN60; Tool Material: Ceramic; Tool Supplier: Kyocera; Tool Batch: 5

[0194] CNC program number: 1, spindle speed: 1000, feed rate: 50, cutting mode: turning.

[0195] Equipment Model: CJK6130, Equipment Supplier: KAWASAKI, Equipment Age: 10 years, Equipment Owner: 2

[0196] Workplace temperature: 15°C, workplace humidity: 50%. The obtained real-world scene features are then classified and coded. Since the scene feature information for part material, blank supplier, tool category, tool model, tool material, tool supplier, tool batch, CNC program number, spindle speed, feed rate, cutting method, equipment model, equipment supplier, equipment age, equipment manufacturer, workplace temperature, and workplace humidity are all in the scene library, only the part size information does not perfectly match the feature information in the scene library. Therefore, it is necessary to calculate the matching between the real-world scene feature information and the scene feature information in the scene library using the cosine similarity method. Specifically, an evaluation vector for the part size is constructed based on whether the length exceeds 70mm, the width exceeds 70mm, and the height exceeds 7mm. This yields... , , , , The calculated maximum similarity to the feature information of the fourth scenario is: .

[0197] Therefore, the final feature encoding matrix of the real-world scene can be obtained as follows: The row vectors are the classification labels, the column vectors are the category features, and the features represented by each row and column are consistent with the information shown in the process of obtaining real-world scene feature information described above.

[0198] ② The real-world scene is evaluated using a scene evaluation algorithm. According to formula (1), for ease of calculation, all weights are set to 1, and the real-world scene evaluation score can be calculated using this formula. .

[0199] ③ The evaluation results are divided into three categories: fully adapted, partially adapted, and unsuitable. Based on the evaluation scores in the above scenarios, the real-world scenario is classified as partially adapted, and the model recommended by the recommendation algorithm needs to be optimized before application.

[0200] ④ Using the K-nearest neighbor algorithm, the scene that best matches the real-world scenario and its corresponding lifetime prediction model can be obtained from the scene library. This scene is encoded as follows: The scene corresponding to this scene code in the scene library is scene 1 in the scene library, that is:

[0201] Part material: high-temperature alloy; part dimensions: 60×60×8 mm; blank supplier: Wuhan Iron and Steel Group.

[0202] Tool Category: Lathe Tools; Tool Model: TNMG160404L-C TN60; Tool Material: Ceramic; Tool Supplier: Kyocera; Tool Batch: 5

[0203] CNC program number: 1, spindle speed: 1000, feed rate: 50, cutting mode: turning.

[0204] Equipment Model: CJK6130, Equipment Supplier: Yang Tie, Equipment Age: 10 years, Equipment Owner: 2 companies

[0205] Workplace temperature: 15°C, Workplace humidity: 50%.

[0206] ⑤ Optimize the lifetime prediction model recommended by the K-nearest neighbor algorithm and recommend the optimized model to users.

[0207] 2) The real-world scenario is not in scenarios 1 through 3.

[0208] ① The real-world scene features are listed in the specific classification labels in Table 3, but not in Scenes 1 to 3. For example, the features of a real-world processing scene are:

[0209] Part material: high-temperature alloy; part dimensions: 80×80×5 mm; blank supplier: Wuhan Iron and Steel Group.

[0210] Tool Category: Lathe Tools; Tool Model: TNMG160404L-C TN60; Tool Material: Ceramic; Tool Supplier: Kyocera; Tool Batch: 5

[0211] CNC program number: 1, spindle speed: 1000, feed rate: 50, cutting mode: turning.

[0212] Equipment Model: CJK6130, Equipment Supplier: KAWASAKI, Equipment Age: 10 years, Equipment Owner: 2

[0213] Workplace temperature: 15°C, workplace humidity: 50%. All category features are listed in the specific feature examples in Table 3, but this real-world scenario is not in scenarios 1-3. Therefore, the previously obtained real-world scenario features will also be classified and encoded. Since only the part size information does not perfectly match the feature information of scenarios 1-3 in the scenario library, it is necessary to calculate the matching between the real-world scenario feature information and the scene feature information in the scenario library using the cosine similarity method. Specifically, an evaluation vector for the part size is constructed based on whether the length exceeds 70mm, the width exceeds 70mm, and the height exceeds 7mm. This vector... Only the category labels that match the category labels of the scene examples in Table 3 need to be calculated. This yields... , , , Calculated Vectors are calculated here. The remaining category labels that match the category labels of the scene examples in Table 3 were skipped, therefore The subscripts will differ from the indexes of the category labels for the scene examples in Table 3.

[0214] Therefore, the final feature encoding matrix of the real-world scene can be obtained as follows: .

[0215] ② The real-world scene is evaluated using a scene evaluation algorithm. According to formula (1), for ease of calculation, all weights are set to 1, and the real-world scene evaluation score can be calculated using this formula. .

[0216] ③ Based on the evaluation scores in the above scenarios, the real-world scenarios are classified as partially adapted. The model recommended by the recommendation algorithm needs to be optimized before it can be applied.

[0217] ④ Using the K-nearest neighbor algorithm, the scene that best matches the real-world scenario and its corresponding lifetime prediction model can be obtained from the scene library. This scene is encoded as follows: The scene corresponding to this scene code in the scene library is scene 1 in the scene library, that is:

[0218] Part material: high-temperature alloy; part dimensions: 60×60×8 mm; blank supplier: Wuhan Iron and Steel Group.

[0219] Tool Category: Lathe Tools; Tool Model: TNMG160404L-C TN60; Tool Material: Ceramic; Tool Supplier: Kyocera; Tool Batch: 5

[0220] CNC program number: 1, spindle speed: 1000, feed rate: 50, cutting mode: turning.

[0221] Equipment Model: CJK6130, Equipment Supplier: Yang Tie, Equipment Age: 10 years, Equipment Owner: 2 companies

[0222] Workplace temperature: 15°C, Workplace humidity: 50%.

[0223] ⑤ Optimize the lifetime prediction model recommended by the K-nearest neighbor algorithm and recommend the optimized model to users.

[0224] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting the life of CNC tools considering real-world scenarios, characterized in that, This method operates on an edge device that is communicatively connected to a cloud platform. The prediction method includes the following steps: Acquire basic data of real-world scenarios, extract real-world scenario features from the basic data of real-world scenarios, and match the real-world scenario features with preset scenario features. The preset scenario features are stored in the scenario library of the cloud platform. Each preset scenario feature corresponds to multiple category tags, and each real-world scenario feature corresponds to one real-world tag. The matching process is performed based on the relationship between the real-world tag of the real-world scenario feature and each category tag of the corresponding preset scenario feature. Based on the matching results, obtain the adaptation evaluation results between the lifetime prediction model in the model library of the cloud platform and the current real-world scenario; Based on the adaptation evaluation results, a corresponding life prediction model is generated to predict the life of CNC tools in the current real-world scenario. The matching process specifically includes the following steps: 101) Compare the reality label of the reality scene feature with the corresponding classification label of different existing scenes in the scene library to determine whether the current reality scene is the same as a certain existing scene in the scene library. If yes, output the matching result of the current reality scene and the existing scene. If no, proceed to step 102). 102) Compare the real-world labels of the real-world scene features with each category label of the corresponding preset scene features in the scene library, and determine whether the real-world labels of all real-world scene features are the same as a certain category label of the corresponding preset scene features. If yes, output the first part of the matching result; if no, output the second part of the matching result. When the matching result is the first part of the matching result, the adaptation evaluation result is obtained through the following steps: Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features; Encoding is performed based on the comparison between the real-world label of each real-world scene feature and the category label of each existing scene. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and the category label corresponding to each existing scene in the scene library is calculated, and the maximum value is assigned at the position corresponding to the category label of the existing scene to generate a real-world scene encoding matrix. The corresponding adaptation evaluation result is obtained based on the real-world scene encoding matrix.

2. The CNC tool life prediction method considering real-world scenarios according to claim 1, characterized in that, When the matching result is the result of matching the current real scene with the existing scene, the adaptation evaluation result is obtained as a complete adaptation. The life prediction model corresponding to the existing scene is used as the final life prediction model to predict the life of CNC tools in the current real scene.

3. The CNC tool life prediction method considering real-world scenarios according to claim 1, characterized in that, When the matching result is the second part of the matching result, the adaptation evaluation result is obtained through the following steps: Construct one Y × N The zero matrix, Y The number of scene features. N The maximum value of the classification label among all scene features; Encoding is performed based on the comparison between the real-world label of each real-world scene feature and each category label of the corresponding preset scene feature. For real-world scene features with the same category label, a value of 1 is assigned at the position corresponding to the same category label. For real-world scene features without the same category label, the maximum similarity between the real-world label of the real-world scene feature and each category label of the corresponding preset scene feature is calculated, and the maximum value is assigned at the position corresponding to the category label to generate a real-world scene encoding matrix. The corresponding adaptation evaluation result is obtained based on the real-world scene encoding matrix.

4. The CNC tool life prediction method considering real-world scenarios according to claim 1 or 3, characterized in that, The specific steps for obtaining the corresponding adaptation evaluation result based on the aforementioned real-world scene encoding matrix are as follows: The formula for converting the real-world scene encoding matrix into a percentage system is as follows: In the formula, This represents the sum of scores for real-world scene features that share the same category label. This represents the sum of scores for real-world scene features for which no matching category label was found. , These are the weights of the features in each scene. , This represents the maximum similarity. The score range is For scores below a certain threshold, the fit assessment result is defined as partial fit, and scores below a certain threshold are considered part of the fit. Those who score below a certain threshold will be deemed unsuitable based on the fit assessment results.

5. The CNC tool life prediction method considering real-world scenarios according to claim 4, characterized in that, If the adaptation evaluation result is partial adaptation, the K-nearest neighbor algorithm is used to obtain the existing scenario that best matches the current real-world scenario. The life prediction model corresponding to the existing scenario is then optimized by the cloud platform and used as the final life prediction model to predict the life of CNC tools in the current real-world scenario.

6. The CNC tool life prediction method considering real-world scenarios according to claim 5, characterized in that, The formula used in the K-nearest neighbor algorithm is: in, and These represent the expanded vectors of the scene encoding matrices of scenes already stored in the scene library and the expanded vectors of the actual scene encoding matrices, respectively. Indicates the index of the expanded vector. Indicates the dimension of the expanded vector. Represents a distance metric.

7. The CNC tool life prediction method considering real-world scenarios according to claim 1 or 3, characterized in that, If the adaptation assessment result is not suitable, the life prediction model will be reconstructed by the cloud platform based on the historical data of the current real-world scenario, and used as the final life prediction model to predict the life of CNC tools in the current real-world scenario.

8. A CNC tool life prediction system considering real-world scenarios, characterized in that, The method includes a cloud platform and an edge device with communication connectivity, wherein the edge device performs the CNC tool life prediction method considering real-world scenarios as described in any one of claims 1-7.

Images