Machine tool component fault early warning and diagnosis method based on open set adversarial learning

Abstract

A machine tool component fault early warning and diagnosis method based on open set adversarial learning, after preprocessing and field construction of collected machine tool component monitoring data, an open set adversarial fitting network capable of self-adaptive fitting of classification decision boundary and hypothesis space boundary is constructed, dynamic adversarial learning strategy is used to ensure network training stability, Nash equilibrium of deep feature mapping module and boundary integrated fitting module is realized, open set fusion decision technology is used to fuse the output labels of multiple samples on the open set adversarial fitting network, accurate machine tool component fault early warning and diagnosis is realized. The present application includes both labeled data and unlabeled data in model training, which can diagnose both healthy operation type and fault type within the hypothesis space, and identify abnormal operation type outside the hypothesis space, breaking through the limitation of the prior art that can only identify a few fault types, which is of great significance to ensure the operation reliability of machine tool components.

Description

Technical Field

[0001] This invention relates to a technology in the field of machine tool fault diagnosis, specifically a method for early warning and diagnosis of machine tool component faults based on open set adversarial learning. Background Technology

[0002] Most existing machine tool component fault early warning and diagnosis methods are based on the closed-set identification assumption, which can only identify fault types within the model assumption space and requires the collection of a large amount of labeled data. However, machine tool component monitoring data collection is difficult, and the number of fault types within the model assumption space is limited, resulting in existing methods being able to diagnose a limited number of fault types and unable to identify abnormal operating states outside the assumption space. Furthermore, a large amount of unlabeled data exists in industrial settings, and existing methods have limited utilization of this data. Therefore, there is an urgent need to develop a machine tool component fault early warning and diagnosis method that can comprehensively utilize both labeled and unlabeled data to achieve timely identification of machine tool component anomalies and accurate fault diagnosis. Summary of the Invention

[0003] This invention addresses the problem that existing machine tool component fault early warning and diagnosis methods do not fully utilize unlabeled data, can only diagnose a limited number of fault types within the hypothesis space, and cannot identify abnormal operating states outside the hypothesis space. It proposes a machine tool component fault early warning and diagnosis method based on open-set adversarial learning. Through data preprocessing and domain construction techniques, both labeled and unlabeled data collected during the model training phase are incorporated into the model training. An open-set adversarial fitting network is designed to adaptively fit the classification decision boundary and hypothesis space boundary. A dynamic adversarial learning strategy is developed to effectively train the open-set adversarial fitting network by comprehensively utilizing data from different domains. Open-set fusion decision technology is used to calculate the early warning and diagnosis results, accurately identifying abnormal operating states of machine tool components and diagnosing specific fault types. This method overcomes the limitation of existing technologies that can only identify a limited number of fault types, and is of great significance for ensuring the operational reliability of machine tool components.

[0004] This invention is achieved through the following technical solution:

[0005] like Figure 1 As shown, this invention relates to a machine tool component fault early warning and diagnosis method based on open set adversarial learning. In the model training stage, the collected machine tool component monitoring data is first preprocessed and the domain is constructed to obtain source domain training samples and target domain training samples. Then, an open set adversarial fitting network that can simultaneously fit the classification decision boundary and the hypothesis space boundary is constructed. A dynamic adversarial learning strategy is used to ensure that the classification decision boundary is fully fitted while achieving Nash balance of different modules within the network. In the model application stage, open set fusion decision technology is used to integrate the output results of multiple samples on the open set adversarial fitting network, thereby improving the accuracy of machine tool component fault early warning and diagnosis.

[0006] The preprocessing involves filling in missing values ​​and removing outliers from the machine tool component monitoring data, followed by feature engineering to extract fault-sensitive features of the machine tool components, and then performing label verification and label encoding.

[0007] The label verification process involves comparing and analyzing the features of labeled training samples to identify samples with potentially abnormal labels, converting them into unlabeled training samples, thereby improving the label quality of the labeled training samples.

[0008] The aforementioned label encoding assumes that some labeled training samples contain... For each type of machine tool component failure, set... The 1st label uses one-hot encoding. The first label indicates whether the sample is in a healthy operating state, and the second to third labels... The bit label indicates whether the sample corresponds to a fault type. The bit label indicates whether the sample belongs to the abnormal operation type. The abnormal operation type is the machine tool component operation status type other than the healthy operation type and the fault type contained in the labeled training samples. It may be the early fault type or the rare fault type.

[0009] The aforementioned domain construction involves constructing a source domain containing labeled samples and a target domain containing unlabeled samples, wherein: the source domain training sample types include healthy operation type and... The types of faults and the types of training samples in the target domain include both healthy operation types and The types of faults include abnormal operation types, and the source domain training sample set is denoted as . The training samples in the target domain are denoted as ,in: The number of training samples in the source domain. The number of training samples for the target domain.

[0010] like Figure 2 As shown, the open-set adversarial fitting network includes a deep feature mapping module and a boundary ensemble fitting module. The deep feature mapping module is used to learn deep features that can separate healthy operation types, fault types, and abnormal operation types, providing discriminative features for machine tool component fault early warning and diagnosis. The boundary ensemble fitting module is used to fit the classification decision boundary and the hypothesis space boundary, and output the open-set fusion decision basis for machine tool component fault early warning and diagnosis.

[0011] The deep feature mapping module includes several convolutional layers, fully connected layers, and a gradient inversion layer. The target domain training samples pass through the convolutional layers, fully connected layers, and gradient inversion layer in sequence, while the source domain training samples pass through only the convolutional layers and fully connected layers. The gradient inversion layer does not change the data during forward propagation, but multiplies the gradient by a negative constant during backward propagation to achieve adversarial learning between the deep feature mapping module and the boundary integration fitting module.

[0012] The boundary integration fitting module includes: several fully connected layers, wherein: the number of neurons in the last fully connected layer is... ,forward Each neuron corresponds to a classification decision boundary, specifically: a mapping learned by the network from the source domain training samples to calculate the probability of whether a sample belongs to the healthy operation type and the fault type; the first neuron... Each neuron corresponds to the hypothesis space boundary, specifically: a mapping learned by the network from the source domain training samples and the target domain training samples for calculating the probability of whether a sample belongs to an abnormal operation type;

[0013] The first The training objective of each neuron determines the direction of adversarial learning between the deep feature mapping module and the boundary ensemble fitting module, specifically: , ,in: For the source domain The true labels of each training sample For the target domain Pseudo-labels for each training sample Assuming a spatial boundary sensitivity factor, This makes the training direction of the boundary integration fitting module set to let Approaching The training direction of the deep feature mapping module is to let keep away This enables the deep feature mapping module to learn features that can distinguish whether a sample belongs to an abnormal operation type.

[0014] The described dynamic adversarial learning strategy dynamically adjusts the adversarial learning process to ensure that the deep feature mapping module and the boundary ensemble fitting module stably reach Nash equilibrium. Specifically, the overall training objective of the deep feature mapping module is set as follows: The overall training objective of the boundary ensemble fitting module is set as follows: ,in: These are the trainable parameters for the deep feature mapping module. Trainable parameters for the boundary integration fitting module. For classification loss function, To counteract the learning loss function, For dynamic weighting functions, For training rounds.

[0015] The classification loss function mentioned is the cross-entropy function, specifically: ,in: For the source domain The training sample of the th training sample The actual value of the bit tag, For the source domain The training sample of the th training sample Predicted values ​​for bit labels.

[0016] The adversarial learning loss function is a binary cross-entropy function, specifically: ,in: For the target domain The training sample of the th training sample Predicted values ​​for bit labels.

[0017] The aforementioned dynamic weight function is used to control the influence of the adversarial learning loss function on the overall training objective, increasing with each training iteration. This allows the early stages of network training to focus on optimizing the classification loss, thereby improving the stability of network training. Specifically: ,in: To counteract the steady-state weights of the learning loss function, As a dynamic weighting growth factor, This represents the total number of rounds of network training.

[0018] The aforementioned open-set fusion decision-making technique fuses multiple samples on an open-set adversarial fitting network. The decision-making process utilizes bit labels, comprehensively considers the classification decision boundary and hypothesis space boundary output, and implements an integrated hard voting process to ensure the robustness of machine tool component fault early warning and diagnosis decisions. Specifically: ,in: For the first The decision type of each test sample. For the first The predicted label values ​​for each test sample; ,in: This is the set of decision types for this test. To integrate the number of samples; ,in: This is the majority type set for this test. For counting functions, ; ,in: This refers to the final decision type for this test.

[0019] Technical effect

[0020] This invention utilizes designed data preprocessing and domain construction techniques to clean, verify, encode, and classify machine tool component monitoring data, improving the quality of training data and enabling both labeled and unlabeled data to be included in model training. This ensures the quantity and diversity of training data. An open-set adversarial fitting network is used to achieve adaptive fitting between the classification decision boundary and the hypothesis space boundary. A dynamic adversarial learning strategy is employed to ensure network training stability, achieving Nash equilibrium between the deep feature mapping module and the boundary integration fitting module. Open-set fusion decision technology is used to fuse multiple samples with multi-bit labels on the open-set adversarial fitting network, improving the accuracy of machine tool component fault warning and diagnosis. Compared to existing technologies, this invention integrates machine tool component fault warning and fault diagnosis into a single framework. It can diagnose both healthy operation types and fault types within the hypothesis space, as well as identify abnormal operation types outside the hypothesis space, overcoming the limitation of existing technologies that can only identify a few fault types. By comprehensively utilizing labeled and unlabeled training data for model training, it avoids the insufficient utilization of unlabeled training data in existing technologies, thereby improving the accuracy of machine tool component fault warning and diagnosis. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 This is a schematic diagram of an open-set adversarial fitting network structure;

[0023] Figure 3 A comparison matrix of confusion between fault warning and diagnosis results under different methods. Detailed Implementation

[0024] This embodiment relates to a machine tool component fault early warning and diagnosis method based on open set adversarial learning. Taking machine tool spindle bearing fault early warning and diagnosis as an example, it specifically includes:

[0025] Step 1: Collect machine tool spindle bearing vibration monitoring data and data type labels, including tagged and untagged samples. Tagged samples include four sample types: healthy operation, outer race fault, inner race fault, and rolling element fault. Untagged samples include these four types as well as samples with abnormal operation. Perform missing value imputation and outlier removal, and extract spindle bearing fault-sensitive features, specifically: 10 time-domain features extracted from the original vibration waveform, 3 frequency-domain features extracted based on Hilbert and Fourier transforms, and 8 time-frequency-domain features extracted based on three-level wavelet packet decomposition.

[0026] Step 2: For labeled samples, label verification is performed. Based on the extracted 21 features, comparative analysis is conducted, and samples with potentially abnormal labels are converted into unlabeled training samples. The training samples that pass label verification are then labeled using one-hot encoding. Specifically, the first label indicates whether the sample belongs to the healthy operation type; the second to fourth labels indicate whether the sample belongs to the outer race fault, inner race fault, and rolling element fault, respectively; and the fifth label indicates whether the sample belongs to the abnormal operation type.

[0027] Step 3: Construct a source domain containing labeled samples and a target domain containing unlabeled samples. The source domain training samples include healthy operation types and three fault types. The target domain training samples include healthy operation types, the three fault types, and abnormal operation types. The set of training samples in the source domain is denoted as […]. The training samples in the target domain are denoted as ,in: The number of training samples in the source domain. The number of training samples for the target domain.

[0028] Step 4: Construct an open-set adversarial fitting network, specifically including:

[0029] 1) Deep Feature Mapping Module: This module is used to learn deep features that can separate healthy operation types, fault types, and abnormal operation types, providing discriminative features for machine tool component fault early warning and diagnosis. Specifically, it includes: 2 convolutional layers, 1 fully connected layer, and 1 gradient inversion layer. Each convolutional layer is followed by a max pooling layer, and the activation function is the ReLU function. Training samples in the target domain pass through the convolutional layer, fully connected layer, and gradient inversion layer in sequence, while training samples in the source domain pass through only the convolutional layer and fully connected layer. The gradient inversion layer does not change the data during forward propagation, but multiplies the gradient by -1 during backward propagation.

[0030] 2) Boundary Integration Fitting Module: This module fits the classification decision boundary to the hypothesis space boundary, outputting the open-set fusion decision basis for machine tool component fault early warning and diagnosis. Specifically, it includes two fully connected layers with ReLU activation functions. The last fully connected layer has 5 neurons. The first 4 neurons correspond to the classification decision boundary, specifically: a mapping learned by the network from the source domain training samples to calculate the probability of whether a sample belongs to the healthy operation type and the fault type. The 5th neuron corresponds to the hypothesis space boundary, specifically: a mapping learned by the network from the source domain training samples and the target domain training samples to calculate the probability of whether a sample belongs to the abnormal operation type.

[0031] The training direction for the fifth neuron is: , ,in: For the source domain The true labels of each training sample For the target domain The pseudo-labels of the training samples make the training direction of the boundary ensemble fitting module oriented as follows: Approaching The training direction of the deep feature mapping module is to let keep away This enables the deep feature mapping module to learn features that can distinguish whether a sample belongs to an abnormal operation type.

[0032] Step 5: Train the open-set adversarial fitting network using a dynamic adversarial learning strategy, ensuring that the deep feature mapping module and the boundary ensemble fitting module stably reach Nash equilibrium, thus obtaining the trained open-set adversarial fitting network. Specifically, set the overall training objective of the deep feature mapping module as follows: The overall training objective of the boundary ensemble fitting module is set as follows: ,in: These are the trainable parameters for the deep feature mapping module. Trainable parameters for the boundary integration fitting module. For classification loss function, To counteract the learning loss function, For dynamic weighting functions, For training rounds.

[0033] The classification loss function mentioned above is specifically the cross-entropy function: ,in: For the source domain The training sample of the th training sample The actual value of the bit tag, For the source domain The training sample of the th training sample Predicted values ​​for bit labels.

[0034] The adversarial learning loss function mentioned above is specifically the binary cross-entropy function: ,in: For the target domain The predicted value of the 5th label of each training sample.

[0035] The dynamic weighting function is specifically as follows: .

[0036] Step 6: For the test samples, based on the trained open-set adversarial fitting network, open-set fusion decision-making technology is used to obtain fault warning and diagnosis results. To improve the robustness of fault warning and diagnosis, each test outputs the final result by integrating the predicted labels of 10 samples on the open-set adversarial fitting network, specifically as follows: ,in: For the first The decision type of each test sample. For the first The predicted label values ​​for each test sample; ,in: This is the set of decision types for this test; ,in: This is the majority type set for this test. For counting functions, ; ,in: This refers to the final decision type for this test.

[0037] Table 1 shows the performance of our proposed method compared to existing fault warning and diagnosis methods. The methods compared include random forest based on prediction probability thresholds, support vector machines, and convolutional neural networks. The evaluation metrics used include fault diagnosis accuracy, abnormal operation type precision, abnormal operation type recall, and overall fault warning and diagnosis accuracy. The fault diagnosis accuracy is the recognition accuracy for four types of faults: healthy machine tool spindle operation, outer race faults, inner race faults, and rolling element faults. The overall fault warning and diagnosis accuracy is the recognition accuracy for all five types. Our proposed method outperforms the compared methods in all four evaluation metrics.

[0038] Table 1

[0039] method Fault diagnosis accuracy Precision rate of abnormal operation type detection Abnormal operation type recall rate Overall accuracy of fault warning and diagnosis Random Forest 0.832 1.000 0.880 0.842 Support Vector Machine 0.832 1.000 1.000 0.867 Convolutional Neural Networks 0.958 0.000 0.000 0.758 This embodiment 1.000 1.000 1.000 1.000

[0040] like Figure 3 As shown, a comparison of the confusion matrix of fault warning and diagnosis results between this embodiment and random forest, support vector machine, and convolutional neural network is provided. When using whether the predicted probability exceeds a set threshold to identify abnormal operating states, random forest and support vector machine can identify abnormal operating states relatively accurately, but the nonlinear fitting ability of these two methods is weak, resulting in low fault diagnosis accuracy. Convolutional neural network can have a high fault diagnosis accuracy, but it is prone to overfitting, which makes it unable to identify abnormal operating types. The method based on open set adversarial learning in this embodiment achieves accurate identification of five types of machine tool spindle bearings.

[0041] Compared with existing machine tool component fault early warning and diagnosis technologies, this invention considers both the identification of abnormal operation types and the diagnosis of specific fault types, resulting in higher accuracy in fault early warning and diagnosis. Specifically, it includes: data preprocessing and domain construction techniques that incorporate both labeled and unlabeled data into model training, increasing the quantity and diversity of training data; an open-set adversarial fitting network trained by a dynamic adversarial learning strategy that balances the fitting of classification decision boundaries and hypothesis space boundaries, enabling the network to diagnose both healthy operation types and fault types within the hypothesis space, as well as identify abnormal operation types outside the hypothesis space; and an open-set fusion decision technology that integrates the label prediction values ​​of multiple samples, further improving the accuracy of machine tool component fault early warning and diagnosis.

[0042] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for early warning and diagnosis of machine tool component faults based on open set adversarial learning, characterized in that, In the model training phase, the collected machine tool component monitoring data is first preprocessed and the domain is constructed to obtain source domain training samples and target domain training samples. Then, an open-set adversarial fitting network that can simultaneously fit the classification decision boundary and the hypothesis space boundary is constructed. A dynamic adversarial learning strategy is used to ensure that the classification decision boundary is fully fitted while achieving Nash balance of different modules within the network. In the model application phase, open-set fusion decision technology is used to integrate the output results of multiple samples on the open-set adversarial fitting network to improve the accuracy of machine tool component fault early warning and diagnosis. The preprocessing refers to: after imputing missing values ​​and removing outliers from the monitoring data of machine tool components, performing feature engineering to extract fault-sensitive features of machine tool components; Label verification is performed by comparing and analyzing the features of labeled training samples to identify samples with potentially abnormal labels. These samples are then converted into unlabeled training samples, thus improving the label quality of the labeled training samples. Label encoding is then performed, assuming that the labeled training samples contain... For each type of machine tool component failure, set... The 1st label uses one-hot encoding. The first label indicates whether the sample is in a healthy operating state, and the second to third labels... The bit label indicates whether the sample corresponds to a fault type. The bit label indicates whether the sample belongs to the abnormal operation type. The abnormal operation type specifically refers to the machine tool component operation status type other than the healthy operation type and the fault type contained in the labeled training samples. The aforementioned open-set adversarial fitting network refers to a network composed of a deep feature mapping module and a boundary ensemble fitting module. The deep feature mapping module is used to learn deep features that can separate healthy operation types, fault types, and abnormal operation types, providing discriminative features for machine tool component fault early warning and diagnosis. The boundary ensemble fitting module is used to fit the classification decision boundary and the hypothesis space boundary, and output the open-set fusion decision basis for machine tool component fault early warning and diagnosis. The dynamic adversarial learning strategy refers to dynamically adjusting the adversarial learning process to ensure that the deep feature mapping module and the boundary ensemble fitting module stably reach Nash equilibrium. Specifically, the overall training objective of the deep feature mapping module is set as follows: The overall training objective of the boundary ensemble fitting module is set as follows: ,in: These are the trainable parameters for the deep feature mapping module. Trainable parameters for the boundary integration fitting module. For classification loss function, To counteract the learning loss function, For dynamic weighting functions, For training rounds; The classification loss function mentioned is the cross-entropy function, specifically: ,in: For the source domain The training sample of the th training sample The actual value of the bit tag, For the source domain The training sample of the th training sample Predicted values ​​for bit tags; The adversarial learning loss function is a binary cross-entropy function, specifically: ,in: For the target domain The training sample of the th training sample Predicted values ​​for bit tags; The aforementioned dynamic weight function is used to control the influence of the adversarial learning loss function on the overall training objective, increasing with each training iteration. This allows the early stages of network training to focus on optimizing the classification loss, thereby improving the stability of network training. Specifically: ,in: To counteract the steady-state weights of the learning loss function, As a dynamic weighting growth factor, This represents the total number of rounds of network training.

2. The method for early warning and diagnosis of machine tool component faults based on open set adversarial learning according to claim 1, characterized in that, The aforementioned domain construction refers to: constructing a source domain containing labeled samples and a target domain containing unlabeled samples, wherein: the source domain training sample types include healthy operation type and... The types of faults and the types of training samples in the target domain include both healthy operation types and The types of faults include abnormal operation types, and the source domain training sample set is denoted as . The training samples in the target domain are denoted as ,in: The number of training samples in the source domain. The number of training samples for the target domain.

3. The method for early warning and diagnosis of machine tool component faults based on open set adversarial learning according to claim 1, characterized in that, The deep feature mapping module includes several convolutional layers, fully connected layers, and a gradient inversion layer. The target domain training samples pass through the convolutional layer, fully connected layer, and gradient inversion layer in sequence, while the source domain training samples pass through only the convolutional layer and fully connected layer. The gradient inversion layer does not change the data during forward propagation, but multiplies the gradient by a negative constant during backward propagation to achieve adversarial learning between the deep feature mapping module and the boundary integration fitting module.

4. The method for early warning and diagnosis of machine tool component faults based on open set adversarial learning according to claim 1, characterized in that, The boundary integration fitting module includes several fully connected layers, with the last fully connected layer having the following number of neurons: ,forward Each neuron corresponds to a classification decision boundary, specifically: a mapping learned by the network from the source domain training samples to calculate the probability of whether a sample belongs to the healthy operation type and the fault type; the first neuron... Each neuron corresponds to the hypothesis space boundary, specifically: a mapping learned by the network from the source domain training samples and the target domain training samples for calculating the probability of whether a sample belongs to an abnormal operation type; The first The training objective of each neuron determines the direction of adversarial learning between the deep feature mapping module and the boundary ensemble fitting module, specifically: , ,in: For the source domain The true labels of each training sample For the target domain Pseudo-labels for each training sample Assuming a spatial boundary sensitivity factor, This makes the training direction of the boundary integration fitting module set to let Approaching The training direction of the deep feature mapping module is to let keep away This enables the deep feature mapping module to learn features that can distinguish whether a sample belongs to an abnormal operation type.

5. The method for early warning and diagnosis of machine tool component faults based on open set adversarial learning according to claim 1, characterized in that, The aforementioned open-set fusion decision-making technique refers to fusing multiple samples on an open-set adversarial fitting network. The decision-making process utilizes bit labels, comprehensively considers the classification decision boundary and hypothesis space boundary output, and implements an integrated hard voting process to ensure the robustness of machine tool component fault early warning and diagnosis decisions. Specifically: ,in: For the first The decision type of each test sample. For the first The predicted label values ​​for each test sample; ,in: This is the set of decision types for this test. To integrate the number of samples; ,in: This is the majority type set for this test. For counting functions, ; ,in: This refers to the final decision type for this test.

Images