Semi-supervised fatigue test condition monitoring method based on adaptive confidence active learning

Abstract

The present invention relates to a semi-supervised fatigue test state monitoring method based on adaptive confidence active learning, which belongs to the field of equipment state monitoring technology and solves the problems of high labeling cost and poor adaptability to working conditions in traditional monitoring methods. First, a data set consisting of fatigue test state monitoring data is constructed and divided into an initial training set, an active learning set and a validation set. Then, an LSTM is used to construct a state recognition model, and the initial training set is used for pre-training. After pre-training, an active learning model is constructed by adding a Softmax function, and the confidence of the active learning set samples is calculated. Samples below the threshold are screened out, and a retraining set is formed after manual labeling. The retraining set is used to train the state recognition model and adjust hyperparameters, and finally the model is evaluated on the validation set. By integrating active learning with adaptive confidence and a semi-supervised learning mechanism, the present invention significantly reduces the dependence on labeled data, has lower labeling costs, and has strong adaptability to complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment status monitoring, and in particular to a semi-supervised fatigue test status monitoring method based on adaptive confidence active learning. Background Art

[0002] Modern industrial production and scientific experiments place higher demands on the reliability of materials and structures. To ensure the accuracy and safety of fatigue tests and avoid data deviations or test interruptions caused by equipment failure, condition monitoring of fatigue testing equipment is necessary. Condition monitoring is a technique that assesses the operating status of equipment through periodic or continuous measurement of key parameters. This technology aims to detect anomalies or deterioration trends early, enabling predictive maintenance and avoiding unplanned downtime.

[0003] However, traditional monitoring methods may face some challenges. First, they require a large amount of labeled data for training. Effective machine learning is only possible with large amounts of labeled fault data and normal data. However, in practice, obtaining fault data is difficult or expensive. Second, they have poor adaptability to monitoring environments. Under different environments or working conditions, the operating status of equipment will change, making traditional monitoring methods unable to adapt.

[0004] Semi-supervised learning can address some of these issues. It leverages large amounts of readily available normal operation data. Even without sufficient fault labels, it can use machine learning algorithms to uncover system operating patterns and identify abnormal patterns. However, this approach is limited by the quality and representativeness of the small amount of labeled data. Summary of the Invention

[0005] In order to solve the problems of high labeling cost and poor adaptability to working conditions in traditional monitoring methods, as well as the problem that semi-supervised methods are limited by the quality and representativeness of a small amount of labeled data, the present invention provides a semi-supervised fatigue test state monitoring method based on adaptive confidence active learning, which can adapt to complex and changeable environments and working conditions, and realize accurate and effective adaptive evaluation of the operating status of fatigue test equipment.

[0006] In order to solve the above problems, the technical solutions adopted by the present invention are as follows:

[0007] A semi-supervised fatigue test condition monitoring method based on adaptive confidence active learning includes the following steps:

[0008] Step 1: Obtain fatigue test state monitoring data collected by multiple sensors and resample the low sampling frequency data to align it with the high sampling frequency data to obtain a data set;

[0009] Step 2: Divide the dataset into a training set and a validation set, then divide the training set into an initial training set and an active learning set, and manually label all samples in the validation set and the initial training set;

[0010] Step 3: Use long short-term memory network to build a state recognition model;

[0011] Step 4: Use the initial training set to pre-train the state recognition model, and The value is used as the evaluation indicator to evaluate the pre-trained model;

[0012] Step 5: Add a Softmax activation function to the output of the linear layer in the pre-trained model to make the model output the sample category probability and construct an active learning model;

[0013] Step 6: Input the active learning set into the active learning model, and the active learning model outputs the confidence of each sample;

[0014] Step 7: Filter out samples in the active learning set that are below the confidence threshold to form a retraining set, and manually label all samples in the retraining set;

[0015] Step 8: Input the retraining set into the state recognition model for retraining and hyperparameter adjustment;

[0016] Step 9: Input the validation set into the adjusted state recognition model to evaluate the model.

[0017] Beneficial effects of the present invention:

[0018] The proposed method significantly reduces fatigue test status monitoring's reliance on labeled data by integrating active learning with semi-supervised learning mechanisms based on adaptive confidence. This significantly reduces labeling costs while ensuring monitoring accuracy. Furthermore, the method offers strong adaptability to complex operating conditions, addressing the shortcomings of traditional monitoring methods in these conditions. This method enables adaptive machine assessment of equipment operating status, effectively determining the current fatigue test status while reducing labeling costs and expanding its application scope. This minimizes the economic losses and R&D interruptions caused by untimely fatigue test status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to an embodiment of the present invention;

[0020] Figure 2 The state recognition model of the present invention and the non-active learning model trained with full data on the validation set Value comparison chart. DETAILED DESCRIPTION

[0021] To make the technical problems, technical solutions, and advantages to be solved by the present invention more clearly apparent, the following detailed description will be given in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions of the present invention, not all of its components.

[0022] like Figure 1 As shown, the semi-supervised fatigue test state monitoring method based on adaptive confidence active learning provided by this embodiment specifically includes the following steps 1 to 9.

[0023] Step 1: Data acquisition and preprocessing: Obtain fatigue test state monitoring data collected by multiple sensors, resample the low sampling frequency data to align it with the high sampling frequency data, and then construct the data set K.

[0024] The present invention uses semi-supervised fatigue test state monitoring data. In order to construct a data set K for a state recognition model, this step collects fatigue test state monitoring data covering a variety of typical operating conditions of fatigue test equipment to ensure the comprehensiveness and representativeness of the data set K. The data comes from a variety of sensors and is used to collect and monitor key parameters of fatigue tests. Since the data sampling frequencies of different sensors are different, for example, the sampling frequencies of sensors such as voltage, current, and power are relatively low, in order to ensure the effectiveness of subsequent data fusion and analysis, it is necessary to resample the low sampling frequency data to align it with the high sampling frequency data. For different types of data, the present invention adopts different resampling methods.

[0025] For voltage, current, power and other data, linear interpolation method is used for resampling. The formula of linear interpolation method is as follows:

[0026]

[0027] in, is the starting point of linear interpolation; is the end point of linear interpolation; is the interpolated value, that is, The estimated value corresponding to the location.

[0028] For pressure data collected by pressure sensors, the Fourier resampling method is used to resample low sampling frequency data. The Fourier resampling method is mainly used on vibration signals. The original signal exists in the form of a time series, and each time point corresponds to a sampling value. The Fourier resampling method specifically includes the following steps: first, the time domain signal is converted to the frequency domain through Fourier transform to obtain the frequency domain spectrum of the signal. The spectrum represents the energy distribution of the signal at different frequencies; secondly, the signal is processed in the frequency domain to change the number of sampling points of the signal, such as using The function interpolates the spectrum to more accurately estimate the value of the spectrum at the new frequency point (i.e., the sampling point); finally, the processed frequency domain spectrum is converted back to the time domain through the inverse Fourier transform to obtain the resampled signal. The function formula is as follows:

[0029]

[0030]

[0031] in, Indicates frequency.

[0032] Step 2: Dataset division: Divide the dataset K into a training set L0 and a validation set U0, and then further divide the training set L0 into an initial training set L1 and an active learning set L2. Manually label all samples in the validation set U0 and the initial training set L1.

[0033] Dataset Partitioning and Labeling: The constructed dataset K is divided into a training set L0 and a validation set U0. To reliably evaluate the generalization ability of the model, all samples in the validation set U0 need to be manually labeled to ensure the accuracy of their labels.

[0034] Training set segmentation and labeling: Furthermore, a simple random partitioning method is used to divide the training set L0 into an initial training set L1 and an active learning set L2 according to a certain ratio. To ensure that the initial state recognition model can learn sufficient and representative features and thus achieve good initial performance, all samples in the initial training set L1 also need to be manually labeled.

[0035] Step 3: Build a state recognition model.

[0036] This paper uses the Long Short-Term Memory Network (LSTM) as a state recognition model, aiming to leverage its advantages in processing time series data. By introducing gating mechanisms (forget gate, input gate, and output gate), LSTM effectively addresses the vanishing or exploding gradient problems that often occur during the training of traditional recurrent neural networks (RNNs). This gating mechanism allows LSTM to learn and memorize important information in long sequences while ignoring irrelevant information, thereby capturing long-term dependencies in the data.

[0037] The state recognition model receives a shape of A three-dimensional tensor of as input, where is the batch size, which indicates the number of samples used in each training iteration; is the sequence length, indicating the time steps contained in each sample; is the number of features, indicating the feature dimensions contained in each time step.

[0038] The LSTM layer performs the following operations on each time step of the input sequence to update the cell state and hidden state.

[0039] 1. Calculate the forget gate, which determines which information is discarded from the cell state. Its formula is:

[0040]

[0041] in, As the output of the forget gate, its shape is ,in is the batch size, is the dimension of the hidden layer, and The value range is between 0 and 1. A value close to 1 indicates retention and a value close to 0 indicates discarding.

[0042] yes Activation function, which compresses the output value to between 0 and 1;

[0043] is the weight matrix of the forget gate, which controls the degree of information retention in the cell state by weighting the input information;

[0044] Represents the hidden state of the previous moment and the current input splicing;

[0045] is the bias vector of the forget gate, which adjusts the overall activation level of the forget gate by introducing an additional learnable parameter.

[0046] 2. The input gate determines what new information will be added to the cell state in order to learn long-term dependencies. The formula for the input gate is as follows:

[0047]

[0048] in, represents the output of the input gate;

[0049] is the weight matrix of the input gate;

[0050] is the bias vector of the input gate.

[0051] 3. Cell Status It is the long-term memory in the LSTM unit that can pass information between time steps, discard unimportant information through the forget gate, and add new information through the input gate, thereby modeling long-term dependencies. The formula is as follows:

[0052]

[0053] in, is the cell state in the previous step;

[0054] Indicates forgetting the previous step;

[0055] Indicates that new information Added to the cell state.

[0056] Step 4: Model pre-training: Use the initial training set L1 to pre-train the state recognition model, and The value is used as the evaluation indicator to evaluate the pre-trained model.

[0057] This step uses the initial training set L1 to pre-train the state recognition model built in step 3. At this stage, the model will extract key features from the input data and establish a mapping relationship between the input and the label.

[0058] After pre-training is complete, use The value is used as the evaluation indicator to evaluate the pre-trained model. The value comprehensively considers both precision and recall, effectively avoiding evaluation errors caused by class imbalance. This evaluation provides a preliminary understanding of the model's accuracy, providing an important basis for subsequent hyperparameter adjustments and model improvements.

[0059] The value is calculated as follows:

[0060]

[0061] in, stands for precision, which refers to the proportion of samples predicted by the model as positive examples that are actually positive examples, that is, the precision of positive examples for the model's prediction. The formula is as follows:

[0062]

[0063] in, True positive examples refer to the number of samples that are predicted by the model to be positive and are actually positive examples;

[0064] False positives refer to the number of samples that the model predicts to be positive but are actually negative.

[0065] Recall rate refers to the proportion of samples that are truly positive that are correctly predicted by the model, that is, the proportion of correct predictions made by the model for the entire training set. The formula is as follows:

[0066]

[0067] in, False negatives refer to the number of samples that are predicted by the model to be negative but are actually positive.

[0068] Step 5. Build an active learning model framework: Add a Softmax activation function to the output of the linear layer in the pre-trained model so that the model outputs the sample category probability and builds an active learning model.

[0069] Based on the state recognition model constructed in step 3, the present invention further constructs an active learning model framework. The core idea of ​​active learning is that not all data contributes equally to model learning. The model actively selects the most informative and representative samples for learning, which can more effectively improve performance, thereby reducing dependence on large amounts of labeled data. This significantly reduces labeling costs while maximizing model performance.

[0070] In order to achieve active learning and enable the model to effectively evaluate the amount of information in the sample, the present invention makes key modifications to the linear layer of the original state recognition model. Specifically, the Softmax activation function is added to the output of the linear layer. This change makes the model no longer directly output the predicted category label, but outputs two probability values, which respectively represent the probability of the sample belonging to different categories. Through the Softmax function, the model converts the output of the linear layer into a probability distribution, so that the probability information can be used to guide the active learning process. The calculation formula of the Softmax activation function is as follows:

[0071]

[0072] in, Represents the output of the model linear layer for category 1; Represents the output of the model linear layer for category 2; is the base of natural logarithms; Indicates the predicted category (category 1 here); is the model input (3D tensor).

[0073] Step 6. Confidence-based active learning sample selection: Input the active learning set L2 into the active learning model constructed in step 5, and the active learning model outputs the confidence of each sample.

[0074] Input the active learning set L2 into the active learning model constructed in step 5. The active learning model will output the probability value of each sample belonging to each category based on the Softmax activation function. For a given sample, assuming that the probabilities output by the model are and ,in and The value of is between 0 and 1, and In order to measure the degree of certainty of the model for the classification result, the highest value of the two probability values ​​is taken as the confidence score of the sample, which is recorded as For example, if and , then the confidence level of the sample is 0.95, indicating that the model is highly confident that the sample belongs to category 1. On the contrary, if and , the confidence level of the sample is 0.52, indicating that the classification result of the model for the sample is relatively vague and it is difficult to determine the category to which it belongs.

[0075] This confidence assessment method based on the highest probability is intuitive, easy to calculate, and effectively reflects the model's understanding of each sample. However, relying solely on confidence to select samples can result in the selected samples being concentrated in the model's "comfort zone"—that is, samples that the model can already distinguish well, while ignoring samples that are most valuable for improving the model. Therefore, in practical applications, in addition to confidence, other metrics are needed to assess the information content of samples.

[0076] To better utilize confidence information for active learning sample selection, we further introduce a confidence threshold and comprehensively consider sample diversity. Specifically, we first sort the samples in the active learning set L2 according to confidence. Then, we define a new strategy and select or adjust it based on the actual situation:

[0077] Set a lower confidence threshold , select confidence level lower than For example, the confidence threshold The probability of the samples being less than 0.627 is 0.627. According to this strategy, samples with probability less than 0.627 are selected from the active learning set L2. These samples are often difficult for the model to distinguish, usually located near the category boundary or containing noise, and are of great value to model learning.

[0078] Step 7: Screen and label high-value samples: Filter out samples below the confidence threshold in the active learning set L2 to form the retraining set L3, and manually label all samples in the retraining set L3.

[0079] After the confidence-based active learning sample selection in step 6, a group of samples that the model judged as "difficult to classify" were obtained. The common feature of these samples is that even after initial training, the model still cannot give prediction results with sufficient confidence. In other words, these samples represent the weak links in the model's knowledge system and contain high learning value. Screening out the samples to form the retraining set L3 is one of the key steps of the present invention.

[0080] The retraining set L3 is highly targeted and has high information density. It represents data that the model still cannot accurately judge after initial learning. It can focus on reflecting the model's current shortcomings and provide a clear direction for subsequent training. Compared with the initial training set L1, each sample in the retraining set L3 contains more information, which can effectively help improve the model's performance. To support subsequent model learning, the samples in the retraining set L3 need to be manually labeled. During the labeling process of the retraining set L3, strict specifications must be followed to ensure the labeling quality. Specifically, the following steps are included:

[0081] (1) Develop detailed annotation guidelines: The annotation guidelines should clearly define the meaning, characteristics, and criteria of each category to avoid ambiguity among annotators;

[0082] (2) Training of annotation personnel: Provide systematic training to annotation personnel to ensure that they fully understand the annotation guidelines and master the correct annotation methods;

[0083] (3) Conduct labeling quality control: Establish a complete labeling quality control system, conduct spot checks and reviews on labeling results, and promptly discover and correct errors.

[0084] The present invention adopts a confidence-based sample screening strategy to manually label high-value samples with high model uncertainty, which not only avoids the high cost of full labeling, but also breaks through the limitations of semi-supervised learning that are limited by the quality and representativeness of the initial labels.

[0085] Step 8: Iterative model training and adjustment: Input the retraining set L3 into the state recognition model for retraining and hyperparameter adjustment.

[0086] In this step, the retraining set L3 is input into the initial state model constructed in step 3 for iterative training and hyperparameter adjustment. Its core goal is to extract key information from the carefully screened and manually labeled retraining set L3, thereby improving the recognition ability of the model and enabling the model to better generalize to unseen data. However, since the retraining set L3 is not randomly sampled, but a "hard nut to crack" selected through active learning methods, its size is usually much smaller than the initial training set L1. This difference in data size makes the model very prone to overfitting on the retraining set L3. The overfitting phenomenon refers to the model over-learning specific patterns and noise in the retraining set L3, and losing the ability to capture broader and more general data patterns, resulting in a decrease in generalization performance in practical applications.

[0087] In order to effectively mitigate the risk of overfitting and ensure that the model can learn truly valuable knowledge from the retraining set L3, the present invention mainly includes the following steps in the process of fine-tuning hyperparameters:

[0088] Learning rate adjustment: Given the relatively small size of the retraining set L3, the model may not require as drastic adjustments as during the initial training phase. Therefore, fine-tuning the learning rate is necessary. For example, a fixed but lower learning rate can be used: the learning rate during the fine-tuning phase can be set to a fixed value lower than the initial training learning rate. The selection of this fixed learning rate requires experimental verification, typically trying multiple different values ​​and determining the optimal value based on the model's performance on the validation set. During the experiment, the training loss and validation set performance metrics for different learning rates should be recorded for comparison and selection.

[0089] Adjusting the number of iterations: Excessive iterations may cause the model to overfit the training data, reducing generalization ability. Therefore, it is necessary to reasonably control the number of iterations based on the size of the retraining set L3 and the convergence of the model during training.

[0090] In order to mitigate the risk of overfitting and ensure that the model can learn truly valuable knowledge from the retraining set L3, the learning rate is optionally adjusted to 1×10 during fine-tuning. -4 , the number of iterations is adjusted to 200.

[0091] Step 9. Validation set performance evaluation: Input the validation set U0 into the state recognition model adjusted in step 8 to evaluate the generalization ability and prediction accuracy of the model.

[0092] In order to comprehensively and objectively evaluate the performance of the active learning framework proposed in this invention, this step evaluates the performance of the state recognition model after retraining and hyperparameter adjustment on the validation set U0 to measure its generalization ability and prediction accuracy in the real world.

[0093] The focus of performance evaluation is to verify whether the recognition ability of the model has been significantly improved after fine-tuning the retraining set L3, as well as the degree of improvement in the model's generalization performance. The value is used as the evaluation indicator, and the adjusted state recognition model is tested on the validation set U0. The value is 0.91, which is the same as that obtained in step 4. The value is 0.76 compared to The value has increased. The improvement in the value shows that the adjusted state recognition model can not only identify positive examples more accurately, but also cover all true positive examples more comprehensively, thereby improving the practicality and reliability of the model.

[0094] Furthermore, full data training is used as a comparison. All data are labeled and put into the non-active learning model for training. It is found that the non-active learning model has a better performance on the validation set. The value is 0.92, such as Figure 2 As shown in the figure, under the premise of little difference in F1 value, 30% of data is saved, which shows that the present invention can effectively improve the performance of the model under resource-constrained conditions.

[0095] The proposed semi-supervised fatigue test state monitoring method based on adaptive confidence active learning significantly reduces the reliance of fatigue test state monitoring on labeled data by integrating adaptive confidence active learning with semi-supervised learning mechanisms. This significantly reduces labeling costs while ensuring monitoring accuracy. Furthermore, the method is highly adaptable to complex operating conditions, addressing the shortcomings of traditional monitoring methods in such conditions. This method enables adaptive machine assessment of the equipment's operating status, effectively determining the current fatigue test status while reducing labeling costs and expanding its application scope. This minimizes the economic losses and R&D interruptions caused by untimely fatigue test state monitoring.

[0096] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A semi-supervised fatigue test state monitoring method based on adaptive confidence active learning is characterized by: The following steps are involved: Step 1: Obtain fatigue test state monitoring data collected by multiple sensors and resample the low sampling frequency data to align it with the high sampling frequency data to obtain a data set; Step 2: Divide the dataset into a training set and a validation set, then divide the training set into an initial training set and an active learning set, and manually label all samples in the validation set and the initial training set; Step 3: Use a long short-term memory network to build a state recognition model, in which: The formula of the forget gate is: ； in, is the output of the forget gate, and its value range is between 0 and 1; yes Activation function; is the weight matrix of the forget gate; Represents the hidden state of the previous moment and the current input splicing; is the bias vector of the forget gate; The formula for the input gate is: ； in, represents the output of the input gate; is the weight matrix of the input gate; is the bias vector of the input gate; Cell state The formula is: ； in, is the cell state in the previous step; Indicates forgetting the previous step; Indicates that new information Added to the cell state; Step 4: Use the initial training set to pre-train the state recognition model, and The value is used as the evaluation indicator to evaluate the pre-trained model; Step 5: Add the Softmax activation function to the output of the linear layer in the pre-trained model to make the model output the sample category probability and construct an active learning model; Step 6: Input the active learning set into the active learning model, and the active learning model outputs the confidence of each sample; Step 7: Filter out samples in the active learning set that are below the confidence threshold to form a retraining set, and manually label all samples in the retraining set; Step 8: Input the retraining set into the state recognition model for retraining and hyperparameter adjustment; Step 9: Input the validation set into the adjusted state recognition model to evaluate the model.

2. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 1 is characterized in that: The input of the state recognition model is a three-dimensional tensor in is the batch size; is the sequence length; is the number of features.

3. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 1 or 2, characterized in that: Use linear interpolation or Fourier resampling to resample low sampling frequency data.

4. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 3 is characterized in that: The formula of the linear interpolation method is: ； in, is the starting point of linear interpolation; is the end point of linear interpolation; is The estimated value corresponding to the location.

5. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 3 is characterized in that: The Fourier resampling method comprises the following steps: Convert the time domain signal into the frequency domain spectrum through Fourier transform; use The function interpolates the frequency domain spectrum to change the number of sampling points; The processed frequency domain spectrum is converted back to the time domain through inverse Fourier transform to obtain the resampled signal.

6. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 1 or 2, characterized in that: The value is calculated as: ； in, is the accuracy, is the recall rate, For a real example, For a false positive example, is a false negative example.

7. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 1 or 2, characterized in that: The calculation formula of the Softmax activation function is: ； in, Represents the output of the model linear layer for category 1; Represents the output of the model linear layer for category 2; is the base of natural logarithms; Indicates the predicted category; Represents the model input.

8. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 1 or 2, characterized in that: In step 8, the hyperparameters of the state recognition model that are adjusted include a learning rate and the number of iterations.

9. The semi-supervised fatigue test state monitoring method based on adaptive confidence active learning according to claim 8 is characterized in that: The adjusted learning rate is lower than the learning rate during the pre-training phase in step 4.

Images