Tool wear state online monitoring method based on physical guidance deep learning network

Abstract

The application discloses a tool wear state online monitoring method based on a physically guided deep learning network, which is combined with physical laws in a tool wear degradation process and deep learning technology, and comprises the following steps: 1, collecting machining experimental data and constructing an actual training data set; 2, expanding training data samples by using a physical model and constructing a simulation training data set; 3, adding a periodic feature extraction unit, a monotone increasing unit and a physically guided module in a deep learning network to construct a physically guided deep learning tool wear state online monitoring network; 4, constructing a physically guided deep learning loss function; 5, pre-training the model by using simulation experimental data, then fine-tuning and optimizing the model by using a small amount of actual data, and obtaining an optimal tool wear state monitoring network, so that the current tool wear can be predicted. The method has strong adaptability and good robustness, and can accurately monitor the tool state in a complex and changeable machining environment.

Description

Technical Field

[0001] This invention belongs to the field of tool monitoring technology, specifically relating to an online tool wear state monitoring method based on a physical guided deep learning network, used to monitor and evaluate the wear state of tools in real time during machining. Background Technology

[0002] With the rapid development of intelligent manufacturing and Industry 4.0, tool wear monitoring technology has gradually become one of the key technologies for improving machining quality and production efficiency. Tool wear monitoring can not only optimize the machining process and extend tool life, but also effectively prevent workpiece scrap and equipment damage caused by excessive tool wear, thereby achieving the goals of high-efficiency production and reduced production costs. Especially in fields where high-speed cutting technology is widely used, such as aerospace, precision electronics, and medical device manufacturing, the requirements for machining accuracy and surface quality are extremely high, making tool wear monitoring particularly important.

[0003] However, traditional methods for monitoring tool wear typically rely on periodic offline inspections or online monitoring based on physical sensors, such as vibration sensors, acceleration sensors, or acoustic emission sensors. These methods have certain limitations in practical applications. First, periodic offline inspections require pausing the machining process, leading to decreased production efficiency. While online monitoring based on physical sensors can achieve a certain degree of real-time performance, it is highly dependent on the monitoring environment and easily affected by machining noise and environmental interference, resulting in insufficient monitoring accuracy. Furthermore, these methods require significant manual intervention for data analysis and decision-making, increasing labor costs in the production process.

[0004] In recent years, deep learning technology has demonstrated powerful nonlinear fitting capabilities in processing time-series data. By inputting data from the machining process, such as tool vibration signals and cutting force signals, into a deep learning network, accurate prediction of tool wear conditions can be achieved. However, deep learning-based tool monitoring methods also face some challenges. First, training a high-precision, highly generalizable deep learning model typically requires a large number of high-quality training samples, including complete tool wear life data under different working conditions. However, in actual production environments, due to the variability of machining parameters and the complexity of the data acquisition environment, it is difficult to obtain a sufficient number of high-quality samples. Especially for diverse machining scenarios, the generalization ability of deep learning models may be insufficient, making it difficult to adapt to the prediction of tool wear conditions under new working conditions. Furthermore, deep learning models are often "black box" models, making it difficult to interpret and understand the model's prediction results, which may limit their further promotion and application in certain scenarios. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies by proposing an online monitoring method for tool wear state based on a physics-guided deep learning network. The goal is to improve the generalization of the model by incorporating the guidance of physical laws, thereby achieving high-precision prediction and real-time monitoring of tool wear state and overcoming the deficiencies of traditional methods.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] The present invention provides an online monitoring method for tool wear status based on a physical-guided deep learning network, characterized by the following steps:

[0008] Step 1: Collect the actual cutting force signal of the tool during the machining process online using a force sensor installed on the worktable. ,in, This represents the cutting force signal of the tool in stage i. This represents the total number of stages in the actual cutting force signal;

[0009] Obtain the cutting force signal of the tool in stage i. Corresponding tool wear information And serve as the actual wear label; thus constructing the actual training dataset. ;

[0010] Collect machining information from the CNC machine tool and extract the cutting parameters for the corresponding i-th stage. ;

[0011] Step 2: Apply the actual training set The samples were augmented to obtain an expanded simulation training dataset. ,in, Indicates the cutting tool in the first... Simulated cutting force signals at each stage; express Corresponding simulated wear labels; This indicates the total number of stages in the simulated cutting force signal;

[0012] Step 3: Construct a physical-guided deep learning-based online monitoring network for tool wear status, including: a deep learning module and a physical guidance module;

[0013] The deep learning module described in step 3.1 includes: a feature extraction block and a head decision block;

[0014] Step 3.1.1 Feature extraction block pair The process is performed to obtain the cutting tool at the [number]th [year]. Global characteristics of the stage and local features ;

[0015] Step 3.1.2 describes the head decision block pair. , and Processing yields the first... Predicted wear value for the stage ;

[0016] Step 3.2 The physical guiding block obtains the first... Physically predicted wear rate of the stage ;

[0017] (1)

[0018] In equation (1), and The first Early and late wear rates of the phase. and There are two amplitude coefficients. For deviation parameters, and There are two constant terms. For the first At a certain moment in a phase;

[0019] Step 4: Construct the physical-guided deep learning loss function using equation (2). ;

[0020] (2)

[0021] In equation (2), This represents the predicted wear value output by the deep learning module. Indicates whether the wear label is real or simulated. Represents the loss function for data items. The physical term loss function representing the relationship between tool wear rate and physical term. The coefficient representing the difference scale adjustment;

[0022] Step 5: Set the simulation training set Input a physical-guided deep learning-based online tool wear monitoring network and calculate the loss function. With the initial weights of the network Perform pre-training until The network converges until the pre-trained network weights are obtained. and its corresponding pre-trained online monitoring model for tool wear status;

[0023] Step 6: Use the actual training dataset The pre-trained online tool status monitoring model is trained, and the loss function is calculated. To Make fine adjustments until... The network converges until the network weights are obtained, thus yielding the fine-tuned network weights. and its corresponding optimal tool wear condition monitoring model;

[0024] Step 7: Collect the cutting force signal generated during the machining process in real time, input the collected real-time cutting force signal into the optimal tool wear condition monitoring model for processing, and output the current tool wear prediction value;

[0025] Step 8: If the current tool wear prediction value reaches the predetermined tool wear critical threshold, stop the machine and replace the tool; otherwise, return to step 7 to continue monitoring.

[0026] The online monitoring method for tool wear state based on a physical guided deep learning network described in this invention is characterized in that the feature extraction block in step 3.1.1 sequentially includes: a periodic attention mechanism unit, a segmented attention mechanism unit, and a two-layer Bi-LSTM network;

[0027] Step a1, The input is fed into the periodic attention mechanism unit for feature extraction, obtaining the tool's position in the [number]th [phase]. Periodic characteristics of the stage ;

[0028] Step a2, and The feature extraction is performed on the segmented attention mechanism unit to obtain the tool's position in the first step. Segmentation characteristics of the stage ;

[0029] Step a3, The input is processed in a two-layer Bi-LSTM network to generate the first... Global characteristics of the stage and local features .

[0030] Furthermore, the head decision block in step 3.1.2 consists of an LSTM network and a monotonicity constraint unit;

[0031] Step b1, , and After concatenation, the data is input into an LSTM network for processing to obtain the... Hidden predicted wear value of the stage ;

[0032] Step b2, will Input the monotonicity constraint unit, and then use equation (3) to obtain the first... Predicted wear value for the stage ;

[0033] (3)

[0034] In equation (3), Indicates the first Hidden predicted wear value for the stage, when When =1, let =0, Choose a function for larger values. It is an absolute value function.

[0035] Furthermore, step 4 includes:

[0036] Step 4.1: Construct the data item loss function using equation (4). :

[0037] (4)

[0038] In equation (4), It is the first The actual wear label or simulated wear label for each stage, where n represents the total number of stages;

[0039] Step 4.2: Construct the physical term loss function using equation (5). :

[0040] (5)

[0041] In equation (5), yes The derivative;

[0042] Step 4.3: Calculate the difference scale adjustment coefficient using equation (6). ;

[0043] (6)

[0044] In equation (6), epoch represents the number of training iterations. This represents the activation function.

[0045] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the online tool wear condition monitoring method, and the processor is configured to execute the program stored in the memory.

[0046] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the online tool wear state monitoring method.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1) This invention achieves high-precision prediction of tool wear status by combining the physical laws of tool wear degradation with a deep learning model. The physics guidance module constrains the deep learning network, enabling the model to not only learn complex nonlinear features but also follow the physical characteristics of the machining process during prediction, thereby significantly improving the model's adaptability and robustness in variable machining environments.

[0049] 2) This invention uses simulation experimental data for pre-training to initially learn the changing patterns of tool wear. Then, a small amount of actual experimental data is used to fine-tune and optimize the model. This effectively solves the problem of deep learning models' dependence on large-scale, high-quality training data, significantly reduces the cost of experimental data acquisition, shortens the model's training cycle, and enhances the ability of the online tool condition monitoring model to be promoted and applied in actual production environments.

[0050] 3) In this invention, a two-layer Bi-LSTM network is used as the feature extraction backbone in the deep learning model structure, combined with periodic constraint blocks and monotonic constraint blocks, to perform in-depth analysis and processing of time series data. The two-layer Bi-LSTM network can extract complex time series features from consecutive time steps, while the introduction of periodic constraint blocks enables the model to better capture the periodic changes in tool wear, thereby improving the prediction accuracy of tool condition and ensuring the real-time performance and accuracy of the monitoring system. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall implementation framework of the present invention;

[0052] Figure 2 This is a flowchart of an online monitoring method for tool wear status based on a physical guided deep learning network according to the present invention;

[0053] Figure 3 This is a schematic diagram of the experimental platform architecture of the present invention;

[0054] Figure 4 This is a structural diagram of the online monitoring network for tool wear status of the present invention;

[0055] Figure 5 This is a comparison chart of the wear prediction value and the actual value of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings of the embodiments thereof. These embodiments are merely examples and are not intended to limit the present invention.

[0057] In this embodiment, a method for online monitoring of tool wear status based on a physical guided deep learning network achieves high-precision and high-generalization online monitoring of tool wear status by combining a physical model and a deep learning model. Figure 1 This is a schematic diagram of the overall implementation framework of the present invention. Figure 2 This is a flowchart of an embodiment of the present invention, which is specifically carried out according to the following steps:

[0058] Step 1: The experimental platform in this embodiment is as follows Figure 3 As shown, the actual cutting force signal of the tool during the machining process is collected online by a force sensor mounted on the worktable. ,in, This represents the cutting force signal of the tool in stage i. This represents the total number of stages of the actual cutting force signal. In this embodiment, the cutting experiment was conducted on a JDGR50 milling machining center, with a total of 12 cutting tests (S=12). The cutting parameters (such as feed rate, spindle speed, and depth of cut) varied between different groups. Data acquisition was performed using a Kistler 9265B with a sampling frequency of 12kHz. Each test group contained 300 samples, and each sample was set to include the cutting force in the X, Y, and Z directions. The cutting tool used was a two-tooth end mill with a diameter of 0.6 mm, a cutting edge length of 2.0 mm, and a helix angle of 35°.

[0059] The cutting force signal of the tool in the i-th stage was obtained by magnifying the cutting edge and working surface of the tool using a high-powered microscope. Corresponding tool wear information And serve as the actual wear label; thus constructing the actual training dataset. ;

[0060] Collect machining information from the CNC machine tool and extract the cutting parameters for the corresponding i-th stage. .

[0061] Step 2: Apply the actual training set The samples were augmented to obtain an expanded simulation training dataset. ,in, Indicates the cutting tool in the first... Simulated cutting force signals at each stage; express Corresponding simulated wear labels; This indicates the total number of stages in the simulated cutting force signal.

[0062] Sample augmentation is achieved by using equation (1) to obtain the first... Segment simulation experiment milling force Using formula (2) for its tool wear label Thus, the simulation training dataset is obtained. ;

[0063] (1)

[0064] (2)

[0065] In equation (1), T is the direction transformation matrix of the coordinate system, and K is the specific force coefficient matrix. For the first The uncut thickness of the j-th cutting edge at height z in stage j is Let z be the cutting depth of the tool. Number of blades This represents the upper limit of the cutting height of the j-th cutting edge on the tool during machining. This is the lower limit of the height of the cutting portion of the j-th cutting edge on the tool during machining;

[0066] In equation (2), A, B, C, and H are tool wear coefficients; It is a logarithmic function.

[0067] Step 3: Construct a physical-guided deep learning-based online monitoring network for tool wear status, including: a deep learning module and a physical guidance module;

[0068] Step 3.1 Deep learning module as follows Figure 4 Includes: feature extraction block and head decision block;

[0069] Step 3.1.1 Feature Extraction Block Pairs The process is performed to obtain the cutting tool at the [number]th ... Global characteristics of the stage and local features ;

[0070] The feature extraction block includes, in sequence: a periodic attention mechanism unit, a segmented attention mechanism unit, and a two-layer Bi-LSTM network, and feature extraction is completed using equations (3)-(5).

[0071] Step a1, Input to the periodic attention mechanism unit Feature extraction is performed to obtain the tool in the [missing information]. Periodic characteristics of the stage ;

[0072] Step a2, and Input segmented attention mechanism unit Feature extraction is performed to obtain the tool in the [missing information]. Segmentation characteristics of the stage .

[0073] Step a3, Input two-layer Bi-LSTM network Processing is performed to generate the first... Global characteristics of the stage and local features .

[0074] (3)

[0075] (4)

[0076] (5)

[0077] Step 3.1.2 Head Decision Block Pair , and Processing yields the first... Predicted wear value for the stage .

[0078] The head decision block consists of an improved Bi-LSTM network and a monotonicity constraint unit, and uses equations (6) and (7) to output the wear value.

[0079] Step b1, , and spliced ​​as The input is processed in an LSTM network to obtain the first... Hidden predicted wear value of the stage ;

[0080] Step b2, will Input the monotonicity constraint unit, and then use equation (3) to obtain the first... Predicted wear value for the stage ;

[0081] (6)

[0082] (7)

[0083] In formula (6) For the first Forget gate of staged LSTM For the first Phase input, For the first Hidden state of the stage, It is the Sigmoid activation function. For the first The input gate output of a staged LSTM It is the hyperbolic tangent function. For the first Candidate memory units for a stage For the first The state of memory units at each stage For the first The output gate of the staged LSTM It is a weight matrix. For bias terms, For the first Stage-based hidden predicted wear values.

[0084] In equation (7), Indicates the first Hidden predicted wear value for the stage, when When =1, let =0, Choose a function for larger values. It is an absolute value function.

[0085] Step 3.2 The physical guide block uses equation (8) to obtain the first... Physically predicted wear rate of the stage ;

[0086] (8)

[0087] In equation (8), and The first Early and late wear rates of the phase. and There are two amplitude coefficients. For deviation parameters, and There are two constant terms. For the first At a certain moment in a phase.

[0088] Step 4: Construct the physical-guided deep learning loss function using equation (9). ;

[0089] (9)

[0090] In equation (9), This represents the predicted wear value output by the deep learning module. Indicates whether the wear label is real or simulated. Represents the loss function for data items. The physical term loss function representing the relationship between tool wear rate and physical term. This represents the differential scale adjustment coefficient.

[0091] Step 4.1: Construct the data item loss function using equation (10). :

[0092] (10)

[0093] In equation (10), It is the first The actual wear label or simulated wear label for each stage, where n represents the total number of stages;

[0094] Step 4.2: Construct the physical term loss function using equation (11). :

[0095] (11)

[0096] In equation (11), yes The derivative of .

[0097] Step 4.3: Calculate the difference scale adjustment coefficient using equation (6). ;

[0098] (12)

[0099] In equation (12), epoch represents the number of training iterations. This represents the activation function.

[0100] Step 5: Set the simulation training set Input a physical-guided deep learning-based online tool wear monitoring network and calculate the loss function. With the initial weights of the network Perform pre-training until The network converges until the pre-trained network weights are obtained. The corresponding pre-trained online monitoring model for tool wear status; the optimization algorithm in this embodiment can be Adam, SGD, MSprop or Adadelta.

[0101] Construct the loss function in pre-training using equation (13) Thus, by using equation (14), we can obtain ;

[0102] (13)

[0103] (14)

[0104] In equations (13) and (14), The initial weights for the tool wear condition monitoring network are: The weights for changes in the tool wear condition monitoring network, To find the weights that minimize the loss function , The weights of the pre-trained network for the online monitoring network of tool wear status.

[0105] Step 6: Use the actual training dataset The pre-trained online tool status monitoring model is trained, and the loss function is calculated. To Make fine adjustments until... The network converges until the network weights are obtained, thus yielding the fine-tuned network weights. and its corresponding optimal tool wear condition monitoring model;

[0106] The loss function in the fine-tuning process is constructed using equation (15). Thus, by using equation (16), we can obtain :

[0107] (15)

[0108] (16)

[0109] In equation (16), The network weights are the fine-tuned weights for the online monitoring network of tool wear status.

[0110] Step 7: Real-time acquisition of cutting force signals generated during machining, inputting the acquired real-time cutting force signals into the optimal tool wear state monitoring model for processing, outputting the current tool wear prediction value and visualizing it; in this embodiment, the visualization result comparing the wear prediction value and the actual value is as follows: Figure 5 As shown;

[0111] Step 8: If the current tool wear prediction value reaches the predetermined tool wear critical threshold, stop the machine and replace the tool; otherwise, return to step 7 to continue monitoring.

[0112] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor in executing the above-described online tool wear condition monitoring method. The processor is configured to execute the program stored in the memory.

[0113] In this embodiment, a computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above-described online tool wear condition monitoring method.

[0114] In summary, this invention, by combining physical models and data-driven deep learning technology, achieves high-precision prediction and real-time monitoring of tool wear conditions, solving the problem of insufficient accuracy in traditional wear monitoring methods due to over-reliance on empirical formulas or single data-driven models. This invention introduces a physical guidance module into the deep learning network, enabling the model to learn not only complex nonlinear features but also to follow the physical characteristics of tool wear degradation. This invention pre-trains the model using simulation experimental data to learn the changing patterns of tool wear, and then fine-tunes and optimizes the model using a small amount of actual experimental data. This strategy effectively reduces the dependence on large-scale, high-quality training data and reduces the cost of experimental data acquisition. This invention uses a two-layer Bi-LSTM network as the backbone model for feature extraction, combined with periodic constraint blocks and monotonic constraint blocks, to deeply analyze and process time-series data. This invention achieves significant improvements in high-precision prediction, reduced data requirements, and innovative deep learning network structure, contributing to increased production efficiency, reduced tool replacement costs, and improved tool management in intelligent manufacturing processes.

Claims

1. A method for online monitoring of tool wear state based on a physical-guided deep learning network, characterized in that, It is done according to the following steps: Step 1: Collect the actual cutting force signal of the tool during the machining process online using a force sensor installed on the worktable. ,in, This represents the cutting force signal of the tool in stage i. This represents the total number of stages in the actual cutting force signal; Obtain the cutting force signal of the tool in stage i. Corresponding tool wear information And serve as the actual wear label; thus constructing the actual training dataset. ; Collect machining information from the CNC machine tool and extract the cutting parameters for the corresponding i-th stage. ; Step 2: Apply the actual training set The samples were augmented to obtain an expanded simulation training dataset. ,in, Indicates the cutting tool in the first... Simulated cutting force signals at each stage; express Corresponding simulated wear labels; This indicates the total number of stages in the simulated cutting force signal; Step 3: Construct a physical-guided deep learning-based online monitoring network for tool wear status, including: a deep learning module and a physical guidance module; The deep learning module described in step 3.1 includes: a feature extraction block and a head decision block; Step 3.1.1 Feature extraction block pair The process is performed to obtain the cutting tool at the [number]th [year]. Global characteristics of the stage and local features ; Step 3.1.2 describes the head decision block pair. , and Processing yields the first... Predicted wear value for the stage ; Step 3.2 The physical guiding block is obtained using equation (1) Physically predicted wear rate of the stage ; (1) In equation (1), and The first Early and late wear rates of the phase. and There are two amplitude coefficients. For deviation parameters, and There are two constant terms. For the first At a certain moment in a phase; Step 4: Construct the physical-guided deep learning loss function using equation (2). ; (2) In equation (2), This represents the predicted wear value output by the deep learning module. Indicates whether the wear label is real or simulated. Represents the loss function for data items. The physical term loss function representing the relationship between tool wear rate and physical term. The coefficient representing the difference scale adjustment; Step 5: Set the simulation training set Input a physical-guided deep learning-based online tool wear monitoring network and calculate the loss function. With the initial weights of the network Perform pre-training until The network converges until the pre-trained network weights are obtained. and its corresponding pre-trained online monitoring model for tool wear status; Step 6: Use the actual training dataset The pre-trained online tool status monitoring model is trained, and the loss function is calculated. To Make fine adjustments until... The network converges until the network weights are obtained, thus yielding the fine-tuned network weights. and its corresponding optimal tool wear condition monitoring model; Step 7: Collect the cutting force signal generated during the machining process in real time, input the collected real-time cutting force signal into the optimal tool wear condition monitoring model for processing, and output the current tool wear prediction value; Step 8: If the current tool wear prediction value reaches the predetermined tool wear critical threshold, stop the machine and replace the tool; otherwise, return to step 7 to continue monitoring.

2. The method for online monitoring of tool wear state based on a physical guided deep learning network according to claim 1, characterized in that, The feature extraction block in step 3.1.1 includes, in sequence: a periodic attention mechanism unit, a segmented attention mechanism unit, and a two-layer Bi-LSTM network; Step a1, The input is fed into the periodic attention mechanism unit for feature extraction, obtaining the tool's position in the [number]th [phase]. Periodic characteristics of the stage ; Step a2, and The feature extraction is performed on the segmented attention mechanism unit to obtain the tool's position in the first step. Segmentation characteristics of the stage ; Step a3, The input is processed in a two-layer Bi-LSTM network to generate the first... Global characteristics of the stage and local features .

3. The method for online monitoring of tool wear state based on a physical guided deep learning network according to claim 2, characterized in that, The head decision block in step 3.1.2 consists of an LSTM network and a monotonicity constraint unit; Step b1, , and After concatenation, the data is input into an LSTM network for processing to obtain the... Hidden predicted wear value of the stage ; Step b2, will Input the monotonicity constraint unit, and then use equation (3) to obtain the first... Predicted wear value for the stage ; (3) In equation (3), Indicates the first Hidden predicted wear value for the stage, when When =1, let =0, Choose a function for larger values. It is an absolute value function.

4. The method for online monitoring of tool wear state based on a physical guided deep learning network according to claim 1, characterized in that, Step 4 includes: Step 4.1: Construct the data item loss function using equation (4). : (4) In equation (4), It is the first The actual wear label or simulated wear label for each stage, where n represents the total number of stages; Step 4.2: Construct the physical term loss function using equation (5). : (5) In equation (5), yes The derivative; Step 4.3: Calculate the difference scale adjustment coefficient using equation (6). ; (6) In equation (6), epoch represents the number of training iterations. This represents the activation function.

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the online tool wear condition monitoring method according to any one of claims 1-4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the online tool wear condition monitoring method according to any one of claims 1-4.

Images