Multi-loss joint optimization tool wear monitoring method and system and storage medium

Abstract

The invention discloses a multi-loss joint optimization tool wear monitoring method and system and a storage medium, and belongs to the technical field of machining state monitoring, tool vibration signals are collected and preprocessed to obtain a training set and a test set, and the tool wear level of each sample is marked as a label; constructing a tetrad training set based on the training set, and training a CNN network by using the tetrad training set to obtain a feature extraction model; constructing a tool wear monitoring model based on the feature extraction model and the softmax classification layer, and training the tool wear monitoring model by adopting the training set; and predicting the wear state category of the tool based on the trained tool wear monitoring model. According to the method, a joint optimization mechanism of a tetrad loss function and a classification loss function is introduced, the minimization and maximization targets of the inter-class distance in a local embedding space are considered, the supervision constraint of global class discrimination information is reserved, the learning ability and the monitoring accuracy of the model are remarkably improved, and the method is suitable for large-scale popularization and application. The intra-class consistency and the inter-class difference of feature expression are effectively enhanced, the robustness is high, and the discrimination capability is strong.

Description

Technical Field

[0001] This invention belongs to the technical field of machining condition monitoring, specifically relating to a multi-loss joint optimization method, system, and storage medium for tool wear monitoring. Background Technology

[0002] In the field of machining, tool wear directly affects machining accuracy, surface quality, and production efficiency. In aircraft assembly, drilling not only relates to the reliability of structural connections but also directly impacts manufacturing efficiency and product quality. Due to the diversity of materials, high requirements for hole precision, and stringent demands for machining stability, real-time monitoring of drilling tool wear has become a key technological means to ensure high-quality drilling and a crucial support for promoting the intelligent and reliable development of aerospace manufacturing. Traditional monitoring methods typically require a large amount of wear data and struggle to simultaneously achieve compactness within categories and differentiation between categories, resulting in low monitoring accuracy and poor generalization ability.

[0003] For example, Chinese patent CN201810516646 discloses a signal segmentation method for real-time monitoring of drilling processing status. First, wavelet packet decomposition is applied to the drilling processing monitoring signal to calculate the normalized Shannon energy of each layer of signal. Then, the Shannon envelope of the layer with the maximum Shannon energy is reconstructed as the drilling monitoring signal envelope. Next, the signal envelope is smoothed using a moving average algorithm, and dual-threshold adaptive signal detection and segmentation is achieved by continuously updating the noise signal estimate in real time. Finally, the theoretical length of the cutting signal segment is calculated based on the drilling processing parameters, and this theoretical length is introduced into the detection as an additional constraint condition for decision-making to correct the results.

[0004] Chinese patent CN202210807766 discloses a method for predicting tool wear during drilling based on power signals. This method can effectively predict drill wear during the drilling process and includes the following steps: Step 1: Select a brand-new drill bit for continuous drilling, measure and record the power signal of each hole and the drill wear amount VB(1, 2, …, n) after drilling each hole; Step 2: Calculate the change in total power consumption between any two holes using the data obtained in Step 1; Step 3: Use the theoretical model from Step 2; Step 4: Collect the power signal during the drilling process and predict the drill wear during the drilling process using the drill wear model from Step 3. This method predicts the drill wear amount based on the power signal during the drilling process.

[0005] However, the aforementioned monitoring methods typically require a large amount of wear data and struggle to simultaneously achieve both intra-category compactness and inter-category differentiation, resulting in low monitoring accuracy and poor generalization ability. Therefore, a robust and highly discriminative drilling tool wear monitoring scheme is urgently needed. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-loss joint optimization method, system, and storage medium for tool wear monitoring, aiming to solve the above-mentioned problems.

[0007] This invention is mainly achieved through the following technical solutions: A multi-loss joint optimization method for tool wear monitoring includes the following steps: Step S1: Data acquisition and preprocessing; acquire tool vibration signals and obtain vibration datasets after preprocessing. Divide the vibration datasets into training and testing sets, and label the tool wear level of each sample as a label. Step S2: Construct quadruplet sample pairs based on the training set to form a quadruplet training set; Step S3: Train the CNN network using the quadruples training set to obtain the feature extraction model; Step S4: Construct a tool wear monitoring model based on the feature extraction model and the softmax classification layer, and train the tool wear monitoring model using the training set; Step S5: Input the collected tool vibration signal into the trained tool wear monitoring model and output the tool wear state category.

[0008] To better implement the present invention, step S1 further includes the following steps: Step S11: Collect the raw vibration signal of the tool and perform signal slicing to obtain the vibration data of the tool; Step S12: Perform bandpass filter processing on the vibration data to filter out low-frequency mechanical interference and high-frequency noise in the processing environment; Step S13: Use the Min-Max normalization method to normalize the filtered vibration data, and map the data uniformly to the [0,1] interval to obtain the vibration dataset; Step S14: Divide the vibration dataset into a training set and a test set, and label each sample with the tool wear level as a label.

[0009] To better realize the present invention, step S2 further includes the following steps: S21: Randomly select a sample from the training set as the anchor sample. ; S22: Randomly select one sample from other samples of the same class as the anchor sample as the positive sample. ; S23: Randomly select two samples from samples of a different class than the anchor sample as negative samples. and negative samples To form a quadruple sample ( , , , ); S24: Repeat steps S21 to S23 until the required number of quadruplet samples are generated to obtain the quadruplet training set.

[0010] To better realize the present invention, step S3 further includes the following steps: Step S31: Input the quadruple samples into the CNN network for forward propagation and extract anchor samples respectively. Positive samples negative samples and negative samples Embedding features , , and ; Step S32: Embedded features , , and Input to the quadruple loss function and classification loss function ; Step S33: Based on the quadruplet loss function (i.e., the loss function of the feature extraction model), reduce the distance between anchor samples and positive samples, and increase the distance between anchor samples and negative samples; Step S34: Classification Loss Function Cross-entropy loss is used to provide a supervisory signal for the overall classification of samples; Step S35: Using the joint loss function (i.e., the loss function of the tool wear monitoring model) performs backpropagation of network parameters to optimize the model until convergence, thereby achieving the compression of intra-class features and the differentiation of inter-class features; ; in: This is the weighting factor.

[0011] To better implement the present invention, further, in step S33, the quadruplet loss function for: ; in: and This is the interval parameter.

[0012] To better implement the present invention, further, in step S34, the classification loss function... for: ; in, This is a real label; The probability predicted by the model output; C represents the total number of samples.

[0013] This invention is mainly achieved through the following technical solutions: A multi-loss joint optimization tool wear monitoring system, based on the above-mentioned multi-loss joint optimization tool wear monitoring method, includes a data acquisition and preprocessing module, a quadruplet construction module, a feature extraction module, a monitoring training module, and a tool wear monitoring module; The data acquisition and preprocessing module is used to acquire data and form a vibration dataset; The quadruplet construction module is used to form a quadruplet training set based on the vibration dataset; The feature extraction module is used to train a CNN network using a quadruplet training set to obtain a feature extraction model. The monitoring and training module is used to build a tool wear monitoring model based on a feature extraction model and a softmax classification layer, and to train the tool wear monitoring model using a vibration dataset. The tool wear monitoring module is used to predict the wear state category of the output tool based on the trained tool wear monitoring model.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned multi-loss joint optimization method for tool wear monitoring.

[0015] The beneficial effects of this invention are as follows: (1) This invention introduces a joint optimization mechanism of quadruplet loss function and classification loss function to address key problems in drilling tool wear monitoring such as "intra-class non-compactness, indistinct inter-class distinction, strong sample dependence, and poor recognition robustness," and improves the sample construction method and training constraint structure of traditional monitoring models. Based on existing deep metric learning models, a dual supervision framework that takes into account both local embedding relationships and global classification objectives is constructed, thereby realizing a tool wear monitoring method that still has high discriminative power and strong generalization ability under small sample conditions.

[0016] (2) This invention optimizes the quadruplet loss and classification loss together, using the quadruplet loss and cross-entropy classification loss as the total loss function. This balances the minimization and maximization of inter-class distance in the local embedding space while retaining the supervision constraint of global class discrimination information, significantly improving the model's learning ability and monitoring accuracy. It effectively enhances the intra-class consistency and inter-class differences of feature expression, improves the discrimination ability of tool wear monitoring, and exhibits high robustness and strong discrimination ability. Secondly, this invention expands the combination space of the training set by combining and recombining labeled vibration samples to form quadruplet units, improving the training effect while keeping the label information unchanged, and significantly reducing the dependence on large-scale labeled data. Compared with traditional methods, this invention has significant advantages in small sample learning—even with a small total number of samples, a large number of quadruplet training units can still be constructed between samples inside and outside the class through different combination methods, realizing the "recombination amplification" of data.

[0017] (3) Based on existing quadruplet loss functions and classification loss functions, this invention applies them for the first time to the scenario of drilling tool wear monitoring, and makes targeted optimizations to address key issues such as small sample size, dispersed intra-class features, low inter-class discrimination, and strong working condition disturbances in this scenario. By constructing a quadruplet training unit containing anchor samples, positive samples, and two negative samples, and combining it with the feature extraction capabilities of CNN networks, this invention proposes a training mechanism suitable for unstable vibration signals, effectively enhancing intra-class feature consistency and inter-class boundary separation capabilities. Furthermore, this invention introduces a joint optimization strategy of quadruplet loss and cross-entropy classification loss, which improves local feature discrimination capabilities while maintaining global classification accuracy. This significantly improves the monitoring robustness and model generalization ability of traditional methods under conditions of limited data, strong interference, and blurred boundaries, and has good engineering adaptability and promotion value.

[0018] (4) This invention introduces a quadruplet loss function for the first time in the drilling vibration signal analysis scenario, aiming to simultaneously enhance the intra-class compactness and inter-class separability of feature representation, and overcome the problems of weak inter-class discrimination ability, scattered intra-class features, and unclear boundary sample identification in traditional tool wear monitoring methods. By constructing a quadruplet structure containing anchor samples, positive samples, and two negative samples, the training constraint range is significantly expanded, the coverage of training supervision constraints is further expanded, the discrimination ability of sample distribution boundaries is improved, and the feature robustness and generalization ability of the model are effectively enhanced. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multi-loss joint optimization tool wear monitoring method of the present invention; Figure 2 A block diagram illustrating the principle of calculating the quadruple loss; Figure 3The curves showing the changes in loss and accuracy during model training; Figure 4 This is a schematic diagram of the confusion matrix of the monitoring results. Detailed Implementation

[0020] Example 1: A multi-loss joint optimization method for tool wear monitoring, such as Figure 1 As shown, it includes the following steps: 1: Data acquisition and preprocessing; During the machining process, an accelerometer is used to collect tool vibration signals. Filtering and normalization methods are employed to reduce signal interference, effectively addressing changes in working conditions and signal disturbances in the actual drilling environment. Training and test sets are constructed.

[0021] 2: Construct training samples for quadruples; From the training set with class labels, for each anchor sample x a Randomly select a sample of the same type but different individuals as the positive sample x. P Then, select two samples that are different from the anchor point and the positive sample category respectively as negative samples x. n1 and x n2 This forms a quadruple ( , , , ).

[0022] This construction method has significant advantages in few-shot learning: even with a small total number of samples, a large number of four-tuple training units can still be constructed between in-class and out-of-class samples through different combinations, achieving "recombination and amplification" of data and greatly expanding the training sample space. This significantly expands the training sample space, reduces dependence on large-scale labeled data, and specifically reflects an innovative data construction and training mechanism, thereby improving the model's feature learning ability and discriminative performance.

[0023] 3: Construction of a feature extraction model based on CNN; A feature extraction model based on a convolutional neural network (CNN) is constructed. The training samples, consisting of four tuples constructed in step 2, are input, and feature representations are extracted through the CNN module to obtain the embedding vectors of the samples. This improves the model's feature robustness and generalization ability, demonstrating an innovative algorithmic structure design.

[0024] 4: Introduce quadruples loss and classification loss for model training; like Figure 2As shown, the embedded features extracted by the feature extraction model are input into the quadruplet loss function and the classification loss function. By simultaneously optimizing the embedding distance between anchor samples and positive and negative samples, and combining the global supervision signal provided by the classification loss, the inter-class discriminativeness is improved while the intra-class feature consistency is enhanced, thus training the network until convergence. This approach balances the goals of compressing intra-class distances and expanding inter-class distances in the local embedding space, while retaining the supervision constraint of global class discrimination information. It significantly improves the model's learning ability and monitoring accuracy under small sample conditions, specifically demonstrating the innovation of the joint loss optimization strategy.

[0025] The definition of quadruple loss is as follows: (1) in, , , and These are the embedding features of the anchor sample, the positive sample, and two negative samples, respectively. and This is the interval parameter.

[0026] The classification loss uses cross-entropy loss, defined as follows: (2) The class probabilities output by the model are ,in ,and ; The one-hot encoding of the real label is ,in ,and ; in, For real labels, The model outputs the predicted probability, where C is the total number of samples.

[0027] The total loss function is: (3) in, This is a weighting factor used to balance the impact of the quadruple loss and the classification loss.

[0028] 5: Construction of a tool wear monitoring model; A softmax classification layer is added at the end of the feature extraction model. The training set is input, the network parameters are fine-tuned, and the optimal weights of the classification layer are determined to build a tool wear monitoring model.

[0029] 6: Tool status assessment; The test set data is input into the tool wear monitoring model, and the tool wear status results are output in real time.

[0030] Example 2: A multi-loss joint optimization method for tool wear monitoring, such as Figure 1 As shown, it includes the following steps: 1: Data acquisition and preprocessing; 11: A three-axis accelerometer is used to collect vibration data in the X, Y, and Z directions in real time during the drilling process. The sensor is fixedly installed on the top of the spindle of the machining center, and the data sampling frequency is set to 1000Hz. 12: Vibration data is processed by a bandpass filter with passband frequency ranging from 20Hz to 500Hz to filter out low-frequency mechanical interference and high-frequency noise in the processing environment, effectively reducing environmental noise and signal interference, and ensuring data reliability and robustness. 13: Normalize the filtered data using the Min-Max normalization method to map the data uniformly to the [0,1] interval; 14: The processed data is divided into training set and test set, with the training set accounting for 80% of the total data and the test set accounting for 20%. At the same time, the tool wear level of each sample is labeled as a tag.

[0031] 2: Construct a training set of quadruplets; 21: Randomly select a sample from the training set as the anchor sample. ; 22: Randomly select one sample from other samples of the same category as the anchor sample as the positive sample. ; 23: Randomly select two samples from the class different from the anchor sample as negative samples. and To form a quadruple sample ( , , , ); 24: Repeat the above steps until the required number of quadruplet samples are generated. Through repeated iterations and recombination, the training sample space is significantly expanded, reducing the dependence on large-scale labeled data, demonstrating the innovation of this invention in data construction and training mechanisms.

[0032] 3: For example Figure 2 As shown, a feature extraction model is built based on CNN; 31: Construct a CNN network, the network structure of which includes 3 convolutional layers (Conv), 3 pooling layers (Pooling) and 1 fully connected layer; 32: Input the quadruple samples into the CNN network for forward propagation and extract the embedding features of each sample. , , and ; 33: The dimension of the embedded features is uniformly fixed at 128 for subsequent loss function calculation.

[0033] 4: Model training and optimization; 41: The embedded features output by the feature extraction model are input into the quadruple loss function. and classification loss function ; 42: Quadruple Loss Function The aim is to reduce the distance between anchor samples and positive samples, and increase the distance between anchor samples and negative samples; 43: Classification Loss Function The cross-entropy loss function is used to provide a supervisory signal for the overall classification of the samples; 44: Through the joint loss function Backpropagation of network parameters is performed to optimize the model until convergence. This achieves feature compression within classes and feature differentiation between classes, improving monitoring accuracy and generalization ability, demonstrating the innovation of the joint loss optimization strategy in this invention.

[0034] The definition of quadruple loss is as follows: (1) in, , , and These are the embedding features of the anchor sample, the positive sample, and two negative samples, respectively. and This is the interval parameter.

[0035] The classification loss uses cross-entropy loss, defined as follows: (2) The class probabilities output by the model are ,in ,and ; The one-hot encoding of the real label is ,in ,and ; in, For real labels, The model outputs the predicted probability, where C is the total number of samples.

[0036] The total loss function is: (3) in, This is a weighting factor used to balance the impact of the quadruple loss and the classification loss.

[0037] 5: Construction of a tool wear monitoring model; 51: Add a softmax classification layer at the end of the feature extraction model as the classification decision layer for the tool wear monitoring model; 52: Train the network using the training set, fine-tune the network parameters, and see the loss and accuracy change curves during the model training process. Figure 3 As shown; 53: Optimize the weight W and bias parameter b of the classification layer to construct the final tool wear monitoring model.

[0038] 6: Tool status assessment; 61: Input the test set into the constructed tool wear monitoring model for real-time analysis; 62: The model outputs the current tool wear status category. The confusion matrix diagram of the monitoring results is shown in the figure below. Figure 4 As shown; 63: Feedback the analysis results to the control system or operators to make necessary adjustments to machining parameters or tool replacements.

[0039] This invention significantly improves the accuracy of tool wear monitoring, reduces the need for large amounts of labeled data, enhances the model's generalization ability, and is more suitable for application needs in actual production environments.

[0040] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A multi-loss joint optimization method for tool wear monitoring, characterized in that, Includes the following steps: Step S1: Data acquisition and preprocessing; acquire tool vibration signals and obtain vibration datasets after preprocessing. Divide the vibration datasets into training and testing sets, and label the tool wear level of each sample as a label. Step S2: Construct quadruplet sample pairs based on the training set to form a quadruplet training set; Step S3: Train the CNN network using the quadruples training set to obtain the feature extraction model; Step S4: Construct a tool wear monitoring model based on the feature extraction model and the softmax classification layer, and train the tool wear monitoring model using the training set; Step S5: Input the collected tool vibration signal into the trained tool wear monitoring model and output the tool wear state category.

2. The tool wear monitoring method based on multi-loss joint optimization according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Collect the raw vibration signal of the tool and perform signal slicing to obtain the vibration data of the tool; Step S12: Perform bandpass filter processing on the vibration data to filter out low-frequency mechanical interference and high-frequency noise in the processing environment; Step S13: Use the Min-Max normalization method to normalize the filtered vibration data, and map the data uniformly to the [0,1] interval to obtain the vibration dataset; Step S14: Divide the vibration dataset into a training set and a test set, and label each sample with the tool wear level as a label.

3. The tool wear monitoring method based on multi-loss joint optimization according to claim 1, characterized in that, Step S2 includes the following steps: S21: Randomly select a sample from the training set as the anchor sample. ; S22: Randomly select one sample from other samples of the same class as the anchor sample as the positive sample. ; S23: Randomly select two samples from samples of a different class than the anchor sample as negative samples. and negative samples To form a quadruple sample ( , , , ); S24: Repeat steps S21 to S23 until the required number of quadruplet samples are generated to obtain the quadruplet training set.

4. The tool wear monitoring method based on multi-loss joint optimization according to claim 3, characterized in that, Step S3 includes the following steps: Step S31: Input the quadruple samples into the CNN network for forward propagation and extract anchor samples respectively. Positive samples negative samples and negative samples Embedding features , , and ; Step S32: Embedded features , , and Input to the quadruple loss function and classification loss function ; Step S33: Based on the quadruplet loss function Reduce the distance between anchor samples and positive samples, and increase the distance between anchor samples and negative samples; Step S34: Classification Loss Function Cross-entropy loss is used to provide a supervisory signal for the overall classification of samples; Step S35: Using the joint loss function Backpropagation of network parameters is performed to optimize the model until convergence, thereby achieving compression of intra-class features and differentiation of inter-class features; ； in: This is the weighting factor.

5. The tool wear monitoring method based on multi-loss joint optimization according to claim 4, characterized in that, In step S33, the quadruple loss function for: ； in: and This is the interval parameter.

6. The tool wear monitoring method based on multi-loss joint optimization according to claim 4, characterized in that, In step S34, the classification loss function for: ； in, This is a real label; The probability predicted by the model output; C represents the total number of samples.

7. A multi-loss joint optimization tool wear monitoring system, based on the multi-loss joint optimization tool wear monitoring method according to any one of claims 1-6, characterized in that, It includes a data acquisition and preprocessing module, a quadruple construction module, a feature extraction module, a monitoring and training module, and a tool wear monitoring module; The data acquisition and preprocessing module is used to acquire data and form a vibration dataset; The quadruplet construction module is used to form a quadruplet training set based on the vibration dataset; The feature extraction module is used to train a CNN network using a quadruplet training set to obtain a feature extraction model. The monitoring and training module is used to build a tool wear monitoring model based on a feature extraction model and a softmax classification layer, and to train the tool wear monitoring model using a vibration dataset. The tool wear monitoring module is used to predict the wear state category of the output tool based on the trained tool wear monitoring model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a multi-loss joint optimization tool wear monitoring method as described in any one of claims 1-6.

Images