A health baseline construction and fault diagnosis method based on twin residual network

Abstract

The application discloses a health baseline construction and fault diagnosis method based on a twin residual network, and comprises the following steps: dividing a diagnosis signal into three parts of a pre-training data set, a training data set and a test data set, and performing normalization processing; constructing a twin residual network, and training a residual network feature extraction part in the twin residual network by using the pre-training data set to form a health baseline model; taking out the residual network feature extraction part in the health baseline model to construct a residual network diagnosis model, and training the residual network diagnosis model by using the training data set; and performing fault diagnosis test on the trained residual network diagnosis model by using the test data set, and evaluating network performance.

Description

Technical Field

[0001] This invention relates to the field of health baseline construction and fault diagnosis technology, specifically to a method for health baseline construction and fault diagnosis based on twin residual networks. Background Technology

[0002] Currently, various mechanical and electrical equipment are widely used in all aspects of daily life and industrial production. Due to harsh operating conditions and component degradation, these devices inevitably malfunction, potentially leading to unacceptable property damage and personal injury. Therefore, establishing a health baseline for the equipment and diagnosing equipment malfunctions promptly and accurately is crucial.

[0003] Existing methods for establishing health baselines and diagnosing faults can be categorized into three types based on their principles: knowledge-based, model-based, and data-driven. While knowledge-based methods can establish fault diagnosis systems for equipment under challenging conditions, they often rely on expert knowledge and struggle to achieve high accuracy. Model-based methods, although offering high diagnostic accuracy, require precise mathematical models of the equipment, limiting their widespread use in fault diagnosis. Therefore, these two methods are now less commonly used. Data-driven fault diagnosis methods, while not requiring precise mathematical models of the system or any prior knowledge of its distribution, heavily rely on historical system data. However, obtaining complete, balanced, and correctly labeled historical system operational data is extremely difficult and expensive in reality. Therefore, data-driven methods based on deep neural networks, capable of effectively handling imbalanced, incomplete, and poorly labeled data, have become increasingly popular.

[0004] While fault diagnosis methods based on deep neural networks possess extremely strong learning capabilities, powerful generalization abilities, and high flexibility, as neural networks become increasingly deeper and the parameter scale expands, existing deep neural network models have begun to encounter various training challenges. On the one hand, from the perspective of the model itself, existing deep network models are difficult to train and prone to gradient problems and non-convergence issues. On the other hand, from the perspective of the training process, existing deep network models have excessively high training costs, long training times, large data requirements, and high computational resource consumption. Summary of the Invention

[0005] This invention provides a method for constructing a health baseline and diagnosing faults based on twin residual networks, in order to solve the technical problems of existing deep neural networks being difficult to train and having excessively high training costs.

[0006] This invention provides a method for constructing a health baseline and diagnosing faults based on twin residual networks, including: Acquire diagnostic signal data for multiple fault modes of the device, and divide the diagnostic signal data into a pre-training dataset, a training dataset, and a test dataset; Construct a twin residual network, and use the pre-trained dataset to train the residual network feature extraction part of the twin residual network to form a healthy baseline model; By extracting the residual network features from the health baseline model, a residual network diagnostic model is constructed, and the residual network diagnostic model is trained using the training dataset to obtain a trained residual network diagnostic model. The trained residual network diagnostic model is tested for fault diagnosis using the test dataset to obtain the fault diagnosis test results.

[0007] Preferably, constructing the twin residual network includes: Construct residual network and twin network frameworks respectively; A twin residual network is constructed using the residual network and the twin network framework.

[0008] Preferably, constructing a twin residual network using the residual network and the twin network framework includes: The twin residual network is constructed by embedding the feature extraction portion of the residual network into two parallel branches of the twin network framework.

[0009] Preferably, the training dataset and the test dataset consist of samples of different fault categories; the pre-training dataset consists of multiple sample pairs, each sample pair containing two samples; wherein, the sample is a segment of diagnostic signal data of a fixed length.

[0010] Preferably, the sample pairs in the pre-training dataset include positive sample pairs and negative sample pairs; wherein, the positive sample pair is formed by pairing two samples of the same fault category; and the negative sample pair is formed by pairing two samples of different fault categories.

[0011] Preferably, the step of training the residual network feature extraction part of the Siamese residual network using the pre-trained dataset to form a healthy baseline model includes: By inputting the positive and negative sample pairs from the pre-training dataset into the residual network feature extraction part of the Siamese residual network for training, a trained Siamese residual network is obtained. The trained twin residual network is used to extract data features of normal samples, and the standard deviation of the data feature distribution of the normal samples is calculated. The upper and lower limits of the healthy baseline model are constructed by adding or subtracting three times the standard deviation from the data features, and the data features of the normal samples and the upper and lower limits are used together to form the healthy baseline model.

[0012] Preferably, the step of extracting the residual network features from the healthy baseline model to construct the residual network diagnostic model includes: By extracting the residual network feature extraction part from the parallel branch of the Siamese network framework in the health baseline model, and then connecting it to a classification layer composed of fully connected layers and a classification cross-entropy loss function, a residual network diagnostic model is constructed.

[0013] The advantages and positive effects of this invention are as follows: (1) The health baseline construction method based on twin residual networks proposed in this invention has a higher health status identification degree compared with traditional methods; (2) Compared with traditional models, the twin residual network diagnostic model proposed in this invention has higher diagnostic accuracy and better training performance, which can significantly reduce the training time of the diagnostic model and reduce the training cost, thereby solving the problems of difficult training and high training cost of deep neural network models. (3) On imbalanced datasets, the fault diagnosis method based on twin residual networks proposed in this invention still has a certain degree of stable diagnostic capability. Attached Figure Description

[0014] Figure 1 This is a framework diagram of a method for constructing a health baseline and diagnosing faults based on twin residual networks provided by the present invention; Figure 2 This is a schematic diagram of the training curve during the construction process of the health baseline model provided by this invention; Figure 3 This is a schematic diagram of the training curve of the residual network diagnostic model provided by the present invention; Figure 4 This is the feature clustering diagram provided by the present invention; Figure 5 This is a schematic diagram illustrating the accuracy variation trend of the residual network based on the health baseline model provided by this invention on an imbalanced set; Figure 6 This is a schematic diagram of the confusion matrix on dataset E provided by the present invention; Figure 7 This is a feature clustering diagram of imbalanced data provided by the present invention; Figure 8 This is a schematic diagram illustrating the accuracy variation trend of the residual network without using a healthy baseline model provided by this invention on an imbalanced set; Figure 9This is a schematic diagram of the confusion matrix of the residual network without using a healthy baseline model provided by the present invention on dataset E; Figure 10 This is a schematic diagram comparing the trend of accuracy changes provided by the present invention; Figure 11 This is a flowchart of the health baseline construction and fault diagnosis method based on twin residual networks provided by the present invention. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0016] This invention addresses the problem of training deep neural networks by leveraging the characteristics of residual networks, which are less prone to degradation and easier to train. It also utilizes Siamese networks to construct a health baseline model, thereby reducing training costs, improving the training efficiency of the diagnostic model, and enhancing the model's diagnostic performance.

[0017] Figure 11 This is a flowchart of the health baseline construction and fault diagnosis method based on twin residual networks provided by the present invention, as shown below. Figure 11 As shown, it includes: Step S101: Acquire diagnostic signal data for multiple fault modes of the device, and divide the diagnostic signal data into a pre-training dataset, a training dataset, and a test dataset; Step S102: Construct a Siamese residual network and train the residual network feature extraction part of the Siamese residual network using the pre-trained dataset to form a healthy baseline model; Step S103: By extracting the residual network feature extraction part from the health baseline model, a residual network diagnostic model is constructed, and the residual network diagnostic model is trained using the training dataset to obtain a trained residual network diagnostic model. Step S104: Use the test dataset to perform fault diagnosis tests on the trained residual network diagnostic model to obtain fault diagnosis test results.

[0018] Furthermore, the construction of the twin residual network includes: constructing a residual network and a twin network framework respectively; and constructing a twin residual network using the residual network and the twin network framework.

[0019] Specifically, constructing a twin residual network using the residual network and the twin network framework includes: constructing the twin residual network by embedding the feature extraction part of the residual network into two parallel branches of the twin network framework.

[0020] The training dataset and the test dataset consist of samples from different fault categories; the pre-training dataset consists of multiple sample pairs, each containing two samples; wherein each sample is a segment of diagnostic signal data of a fixed length. Further, the sample pairs in the pre-training dataset include positive sample pairs and negative sample pairs; wherein a positive sample pair is formed by pairing two samples of the same fault category; and a negative sample pair is formed by pairing two samples of different fault categories.

[0021] Specifically, the step of training the residual network feature extraction part of the Siamese residual network using the pre-training dataset to form a healthy baseline model includes: inputting positive sample pairs and negative sample pairs from the pre-training dataset into the residual network feature extraction part of the Siamese residual network for training to obtain a trained Siamese residual network; extracting data features of normal samples using the trained Siamese residual network and calculating the standard deviation of the data feature distribution of the normal samples; constructing upper and lower limit thresholds of the healthy baseline model by adding or subtracting three times the standard deviation from the data features, and combining the data features of the normal samples and the upper and lower limit thresholds to form the healthy baseline model.

[0022] Furthermore, after training the residual network feature extraction part of the Siamese residual network using the pre-training set, the trained model is used to extract data features from normal samples (i.e., sample data collected by the equipment in a non-faulty state). For example, for normal samples... After a series of convolution and activation operations, the sample features are obtained. and statistical data characteristics y Distribution The standard deviation of the data is used to construct the upper and lower thresholds of the health baseline model by adding or subtracting three times the standard deviation from the data features. The specific formula for constructing the health baseline model is as follows:

[0023] in up This is the upper threshold of the health baseline model. down This is the lower limit threshold of the health baseline model. This represents the mean of the characteristic distribution of the normal samples. denoted as the standard deviation of the characteristic distribution of the normal sample.

[0024] The data characteristics of normal samples and their upper and lower thresholds together constitute a healthy baseline model.

[0025] Furthermore, the step of extracting the residual network feature extraction part from the healthy baseline model to construct the residual network diagnostic model includes: extracting the residual network feature extraction part from the parallel branch of the Siamese network framework, and then connecting it to a classification layer composed of fully connected layers and a classification cross-entropy loss function to form the residual network diagnostic model.

[0026] This invention extracts the residual network features from the healthy baseline model, connects two fully connected layers and a cross-entropy loss function layer at the end to construct a residual network diagnostic model, and trains the residual network diagnostic model using the training dataset. Specifically, the residual network diagnostic model trains the samples... Perform a series of convolution, activation, and classification operations to predict the sample Classification The loss function layer is used to predict and classify. The classification error is obtained by comparing it with the true classification C. Classification error The gradient is propagated back along the network to update the parameters of the residual network diagnostic model, and finally the trained residual network diagnostic model is obtained.

[0027] Figure 1 This is a framework diagram of a health baseline construction and fault diagnosis method based on twin residual networks provided by the present invention, as shown below. Figure 1 As shown, the method may include: Step 1: Divide the raw diagnostic signal data into three parts: a pre-training dataset, a training dataset, and a test dataset. The training and test datasets consist of samples from different fault categories, each sample being a segment of a diagnostic signal of a fixed length. The pre-training dataset consists of several sample pairs, each pair containing two samples, each a segment of a diagnostic signal of a fixed length. Sample pairing methods include intra-class and inter-class pairing. Intra-class pairing involves pairing two samples of the same fault category to form a positive sample pair, while inter-class pairing involves pairing two samples of different fault categories to form a negative sample pair. Positive and negative sample pairs together constitute the pre-training dataset.

[0028] Step 2: First, construct the feature extraction part of the residual network and the Siamese network framework. Then, embed the feature extraction part of the residual network into two parallel branches of the Siamese network framework to form a Siamese residual network. Train the Siamese residual network using the pre-training set, where positive samples are expected to output 0 and negative samples are expected to output infinity. In practice, a threshold margin is usually set, and the difference between the actual output and the expected output is calculated using a loss function. The network parameters are updated through backpropagation of the loss function gradient. This method of learning using fault sample pairs can learn the correlation and differences between different types of faults. Therefore, the Siamese residual network with updated parameters can accurately determine the health status of equipment and serves as a health baseline model for the equipment.

[0029] Step 3: Extract the residual network feature extraction part from the parallel branch of the Siamese network framework in the healthy baseline model, and then connect it to a classification layer composed of fully connected layers and a classification cross-entropy loss function to form a residual network diagnostic model. Then, train the residual network diagnostic model using the training set to further update the network parameters.

[0030] Step 4: Input the test set into the residual network diagnostic model to test it, observe the network output and statistically analyze the diagnostic performance.

[0031] Example

[0032] 1. Description of Experimental Data This invention presents a general method for establishing equipment health baselines and diagnosing faults. To verify the effectiveness of this invention, experiments were conducted using motor data. The experimental data came from a comprehensive mechanical fault simulation test bench (MFS), manufactured by Spectra Quest, USA. This test bench primarily consists of a drive motor, motor controller, magnetic brake, pulse-type optical speed sensor, transmission shaft system, and vibration sensor. In conjunction with the data acquisition system, this test bench enables technical research on motor fault diagnosis, load-side fault diagnosis, and motor health status assessment.

[0033] In the data acquisition process, X, Y, and Z axis vibration data of the motor vibration points were collected. A total of 6 samples were taken, each sampling representing one fault mode, including normal operation, rotor shaft bending, bearing inner ring failure, rotor imbalance, rotor bar breakage, and rotor misalignment. Each sampling used a frequency of 25600Hz, a sampling time of 192s, and 4,915,200 sampling points. The motor speed was 30 r / min, and the load was set to level 1. Details are shown in Table 1.

[0034] After dividing the collected data into samples, different numbers of fault samples were selected to form datasets A to F, as shown in Table 2. Among them, dataset A is a balanced dataset, and B to F are imbalanced datasets, with the imbalance ratio gradually increasing from B to F.

[0035] Table 1: Raw Data Information

[0036] Table 2: Number of samples for each fault state in the dataset

[0037] In the process of constructing the health baseline model, samples need to be paired to form sample pairs as input to the health baseline model. Taking dataset A as an example, 10 samples of each fault state are taken from dataset A, for a total of 60 fault samples. Intra-class and inter-class pairings are performed, resulting in 21 pairing methods and 1830 sample pairs. The sample pair size is (2, 4096, 3). All sample pairs constitute the pre-training set. The number of sample pairs is shown in Table 3. In the 0-0 fault intra-class pairing, 10 samples are paired with themselves to form 10 sample pairs, and paired with other samples to form 10 sample pairs. Therefore, 0-0 pairing yields 55 pairings; in 0-1 fault class pairing, 10 samples of type 0 are paired with 10 samples of type 1, resulting in 100 pairings. The number of pairings for the remaining 20 pairing methods follows the same pattern, resulting in a total of 1830 pairings, forming a pre-training dataset. The pre-training set is created using datasets B-F in the same way as dataset A, also with 21 pairing methods, resulting in 1830 pairings.

[0038] Table 3: Number of Pairs

[0039] In the formal training process, in order to prevent overfitting of the samples, it is necessary to remove the training set samples that have been used in the construction of the healthy baseline model. After removing the pre-training samples, the datasets A to F are named datasets A1 to F1, and their sample composition is shown in Table 4.

[0040] Table 4: Training set for model construction samples after removing healthy baselines

[0041] 2. Case study of health baseline construction technology based on twin residual networks First, a Siamese residual network model was constructed, and the pre-training set created using dataset A was input into the Siamese residual network to build a health baseline model. In this process, the batch size was set to 16, the training epochs to 300, the optimizer to Adam, and the learning rate to 0.00025. During this learning process, the fault sample pairing learning method allows the model to learn the correlations and differences between different faults. Therefore, the Siamese residual network model, after parameter updates, can accurately determine the health status of the equipment, possessing the characteristics of an equipment health baseline model.

[0042] The training curves obtained during the construction of the health baseline model are shown below. Figure 2 As shown in the training curve, the Siamese residual network begins to converge around the 250th batch, with a good overall convergence trend. The fluctuations during convergence are small, indicating that the network has good convergence stability. At the convergence point, the network's classification results for different pairings are relatively clear, and the clustering of 21 pairing methods can be clearly seen, indicating that the network has strong feature extraction capabilities and high classification accuracy.

[0043] 3. Validation of Motor Fault Diagnosis Case Based on Balanced Dataset The residual network feature extraction part of the healthy baseline model is extracted from the Siamese network framework, and then followed by a classification layer consisting of fully connected layers and a classification cross-entropy loss function to form a residual network diagnostic model. This model is then input into the training set A1 for further training. During training, the batch size is set to 32, the number of training epochs is 100, the optimizer is Adam, and the learning rate is set to 0.0001. The loss function and accuracy curves during training are shown below. Figure 3 It can be seen that the residual network diagnostic model begins to converge around the 50th batch. During the convergence process, the fluctuations in the loss function and accuracy are small, and the overall convergence trend is smooth, showing good convergence performance. After complete convergence, the loss function and accuracy show almost no fluctuations, showing good convergence point stability.

[0044] To evaluate the performance of the trained residual network diagnostic model, 20 samples were taken for each fault state, forming a test set of 120 samples. The trained residual network was then tested. To eliminate random factors, the model was trained five times and tested separately for each fault state. The fault diagnosis accuracy was statistically analyzed. The experimental data are shown in Table 5. It can be seen that the residual network diagnostic model, after the healthy baseline model construction process, achieved 100% diagnostic accuracy on both the training and test sets, demonstrating excellent diagnostic performance.

[0045] Table 5: Performance Indicators of Residual Network Diagnostic Model

[0046] To test the feature extraction capability, the data features extracted by the residual network were dimensionality-reduced using t-distributed stochastic neighbor embedding (t-SNE) and plotted as a clustering graph, as shown below. Figure 4 .from Figure 4 As can be seen, the data features extracted by the residual network are clearly classified, with reasonable inter-class distances, high intra-class clustering, and no misclassification points. This indicates that the data features extracted by the residual network can better reflect category features rather than individual features.

[0047] Compared to the training and diagnostic process of residual networks that do not use a healthy baseline model but are directly trained on dataset A, it can be seen that the healthy baseline model construction process significantly improves the training and diagnostic performance of the residual networks. The performance comparison statistics are shown in Table 6. It can be seen that the residual network built with the healthy baseline model begins to converge around the 50th batch, while the residual network without this process begins to converge around the 95th batch. This indicates that the construction process reduces the number of times training data is used, i.e., reduces the amount of training data. Furthermore, the residual network after this process exhibits better convergence trend, convergence smoothness, and convergence stability, demonstrating better training performance. The diagnostic accuracy of the residual network after this process is improved by 0.7% and 0.5% on the training and test sets, respectively, indicating better diagnostic performance. In terms of convergence time, the residual network after this process reduces the average time from 19.22 seconds to 9.08 seconds, a reduction of 52.8%, indicating that the healthy baseline model construction process can significantly reduce the training time of the residual network and lower training costs.

[0048] Table 6: Performance Comparison of Residual Networks with and without Healthy Baseline Construction Process on Balanced Datasets

[0049] 4. Case study validation of motor fault diagnosis based on unbalanced dataset To further explore the training and diagnostic performance of the Siamese residual network, datasets B-F were used as the training set. A certain number of samples were extracted and paired according to the aforementioned scheme to form a pre-training set. Samples used in the health baseline model construction process were removed from datasets B-F, forming datasets B1-F1, which served as the formal training dataset. To test the diagnostic performance of the trained network, 20 samples were taken for each fault state, totaling 120 samples, to test the trained residual network. To eliminate random factors, the network was trained five times and tested each time. The fault diagnosis accuracy was statistically analyzed. Experimental data are shown in Table 7.

[0050] Table 7: Diagnostic accuracy of twin residual networks on unbalanced sets

[0051] As shown in the table, the residual network constructed using the health baseline model exhibits high diagnostic accuracy when the imbalance ratio is 400:50 and 400:40. When the imbalance ratio increases to 400:30, the average diagnostic accuracy is 98.66%; when it increases to 400:20, the average diagnostic accuracy is 95.50%; and when the imbalance ratio is 400:15, the average diagnostic accuracy is 78.00%, with a minimum of 70.83%, still exceeding 50%. This indicates that the health baseline construction and fault diagnosis technology based on Siamese residual networks still possesses a certain degree of stable diagnostic capability when facing highly imbalanced data conditions. The accuracy trend curves for datasets B1 to F1 are shown below. Figure 5 .

[0052] The confusion matrix of the diagnostic results of the residual network trained using dataset E is as follows: Figure 6 It can be seen that under these data conditions, the highest probability of misclassifying fault 5 as fault 3 is 3.83%; followed by misclassifying fault 1 as fault 4, at 0.50%; and finally misclassifying fault 4 as fault 1, at 0.17%.

[0053] To test the feature extraction capability of the residual network after the healthy baseline model construction process, the data features extracted by the residual network were dimensionality-reduced using t-distributed random neighborhood embedding and plotted as a clustering graph, as shown below. Figure 7 The correspondence between the colors of the cluster points and the numbers of each fault category in the clustering diagram is shown in Table 8. It can be seen that the feature classification is relatively clear, the inter-class distance is large, the intra-class clustering is high, and there are very few misclassified points. This indicates that even under highly imbalanced data conditions, the feature extraction part of the residual network pre-trained with Siamese can still extract the categorical features of the data, rather than individual features.

[0054] Table 8: Correspondence between cluster point colors and fault numbers

[0055] To compare the performance of fault diagnosis techniques based on Siamese residual networks, the residual network was directly trained on datasets B to E without undergoing a health baseline model construction process. A test set of 120 samples (20 samples for each fault state) was formed to test the residual network without a health baseline model construction process. To eliminate random factors, the network was trained five times and tested separately. The fault diagnosis accuracy was statistically analyzed. Experimental data are shown in Table 9.

[0056] Table 9: Diagnostic accuracy of residual networks without using a healthy baseline model on imbalanced sets.

[0057] As shown in the table, the residual networks without the health baseline model construction process exhibit low diagnostic accuracy on datasets B through E, with significant fluctuations in accuracy. At imbalance ratios of 400:50 and 400:40, the residual networks without this process achieve only 87.17% and 85.00% diagnostic accuracy, respectively. When the imbalance ratio increases to 400:30, the average diagnostic accuracy drops to 73.67%, reaching an unacceptable level. When the imbalance ratio increases to 400:20, the average diagnostic accuracy reaches 69.00%, with a minimum accuracy of 50.00% and a maximum of 89.17%. The trend of diagnostic accuracy variation of the residual networks without the health baseline model construction process on datasets B through E is shown below. Figure 8 As shown.

[0058] The test confusion matrix of the residual network without the healthy baseline model building process on dataset E is as follows: Figure 9 As shown, the residual network without the healthy baseline model construction process loses its ability to make accurate diagnoses on dataset E. On highly imbalanced datasets, it can only make accurate diagnoses for the larger number of normal samples, and may not be able to make accurate diagnoses for the smaller number of faulty samples. The high-frequency categories for fault misclassification are: fault 1 misdiagnosed as fault 5, fault 3 misdiagnosed as fault 0, fault 4 misdiagnosed as fault 1, and fault 5 misdiagnosed as both fault 1 and fault 4; the categories with a relatively high occurrence of misdiagnosis are: fault 1 misdiagnosed as fault 3, fault 3 misdiagnosed as fault 4 or fault 5, and fault 4 misdiagnosed as fault 5.

[0059] Table 10: Comparison of Residual Network Diagnostic Accuracy in Model Construction Process with and without Healthy Baseline on Imbalanced Sets

[0060] Table 10 compares the performance of residual networks with and without a healthy baseline model construction process on imbalanced datasets, and compares the accuracy trend curves for both. Figure 10 As shown, when faced with highly imbalanced data, the residual network without a healthy baseline model construction process loses its ability to achieve stable convergence and accurate diagnosis. In contrast, the residual network after the healthy baseline model construction process can stably handle this highly imbalanced data situation. This indicates that the healthy baseline construction and fault diagnosis technology based on Siamese residual networks can reduce the training difficulty and cost of the diagnostic model and improve the diagnostic accuracy.

[0061] In summary, the proposed health baseline construction and fault diagnosis method based on Siamese residual networks can effectively reduce the training difficulty and cost of diagnostic models. On balanced datasets, the Siamese residual network-based health baseline construction and fault diagnosis technique significantly improves convergence performance compared to residual network diagnostic models, increasing diagnostic accuracy by 0.5% and reducing convergence time by 52.8%, thus significantly reducing the training cost of residual networks. On imbalanced datasets, the Siamese residual network-based health baseline construction and fault diagnosis technique exhibits better adaptability to extreme conditions, achieving higher diagnostic accuracy under highly imbalanced data conditions, further reducing the training difficulty and cost of residual networks.

[0062] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the invention should be considered within the scope of the invention.

Claims

1. A method for constructing a health baseline and diagnosing faults based on twin residual networks, characterized in that, include: Acquire diagnostic signal data for multiple fault modes of the device, and divide the diagnostic signal data into a pre-training dataset, a training dataset, and a test dataset; A residual network and a Siamese network framework are constructed respectively. The Siamese residual network is constructed by embedding the feature extraction part of the residual network into two parallel branches of the Siamese network framework. The healthy baseline model is formed by training the residual network feature extraction part of the Siamese residual network using the pre-training dataset. This includes: training the Siamese residual network by inputting positive and negative sample pairs from the pre-training dataset into the residual network feature extraction part; extracting data features of normal samples using the trained Siamese residual network and calculating the standard deviation of the data feature distribution of the normal samples; constructing upper and lower thresholds for the healthy baseline model by adding or subtracting three times the standard deviation from the data features; and combining the data features of the normal samples with the upper and lower thresholds to form the healthy baseline model. By extracting the residual network feature extraction part from the parallel branch of the Siamese network framework in the health baseline model, and connecting it with a classification layer composed of fully connected layers and a classification cross-entropy loss function, a residual network diagnostic model is constructed. The residual network diagnostic model is then trained using the training dataset to obtain a trained residual network diagnostic model. The trained residual network diagnostic model is tested for fault diagnosis using the test dataset to obtain the fault diagnosis test results.

2. The method according to claim 1, characterized in that, The training dataset and the test dataset consist of samples of different fault categories; the pre-training dataset consists of multiple sample pairs, each containing two samples; wherein, the sample is a segment of diagnostic signal data of a fixed length.

3. The method according to claim 2, characterized in that, The sample pairs in the pre-training dataset include positive sample pairs and negative sample pairs; wherein, the positive sample pair is formed by pairing two samples of the same fault category; and the negative sample pair is formed by pairing two samples of different fault categories.

Images