A method for detecting imbalance anomalies in complex equipment based on reconstruction similarity

Abstract

This application provides a method for detecting complex equipment imbalance anomalies based on reconstruction similarity. Using the CAE-SIAMESE model, it achieves the detection of aero-engine anomalies under imbalanced sample conditions. The CAE-SIAMESE model includes a convolutional autoencoder and a Siamese neural network. The convolutional autoencoder is used for representation learning, and the Siamese neural network is used to measure the reconstruction similarity of samples. This application offers high diagnostic accuracy and strong noise resistance. The introduction of a reconstruction similarity learning mechanism results in greater reconstruction similarity for normal samples and lower reconstruction similarity for abnormal samples obtained by CAE-SIAMESE, which is more conducive to the diagnosis of normal and abnormal samples. The convolutional layers in the network can better capture the complex relationships in multi-dimensional monitoring data and better represent the characteristics of the original samples. The Siamese neural network can resist the influence of noise when measuring similarity, effectively solving the problem of noise in engine monitoring data.

Description

Technical Field

[0001] This application belongs to the field of complex equipment imbalance anomaly detection technology, and more specifically, it relates to a method for complex equipment imbalance anomaly detection based on reconstruction similarity. Background Technology

[0002] Complex equipment is inextricably linked to modern industry, directly impacting a nation's comprehensive national strength and core competitiveness. This includes aerospace propulsion systems, large communication systems, large ships, power grid control systems, high-speed trains, and large weaponry. This application uses aero-engines as an example of complex equipment. As a key technology for aviation flight safety, aero-engine anomaly detection is becoming an increasingly important research area. If aero-engine anomalies are not detected and addressed appropriately in a timely manner, they will gradually evolve into serious malfunctions. These can range from minor unplanned flight delays or cancellations to catastrophic accidents such as engine failures in flight or even aircraft crashes, inevitably causing huge economic losses and significant casualties. Therefore, real-time status monitoring and anomaly detection are essential to improve the reliability and safety of aero-engines and reduce maintenance costs.

[0003] Anomaly detection methods for aero-engines are generally divided into two main categories: model-driven anomaly detection methods and data-driven anomaly detection methods. Model-driven anomaly detection methods require accurate modeling of the aero-engine performance degradation process, which necessitates extensive expertise in engine design and assembly. However, this core expertise is often kept confidential, limiting the widespread application of these methods. Data-driven methods, on the other hand, can discover potential correlations and causal relationships within data and infer corresponding engine states without knowing the underlying operational mechanisms. Firstly, airlines collect vast amounts of engine monitoring data, significantly enhancing data availability and providing more favorable conditions for implementing data-driven methods. Secondly, many data-driven methods have emerged in recent years, including support vector machines, isolated forests, and artificial neural networks, achieving good diagnostic results in anomaly detection. For example, some researchers have applied improved support vector machines to anomaly detection, demonstrating good diagnostic results even under imbalanced data conditions. Others have used stacked denoising autoencoders to extract features from raw data, achieving high diagnostic accuracy.

[0004] However, engine monitoring data exhibits a high degree of imbalance, meaning the number of normal samples far exceeds the number of abnormal samples. During training, supervised learning methods tend to focus more on normal samples, leading to poor diagnostic performance for abnormal samples. Furthermore, the abnormal samples collected by airlines do not fully cover all types of engine anomalies. If new types of anomalies emerge, supervised learning methods will fail. However, semi-supervised learning methods can effectively mitigate the shortcomings of supervised learning methods on imbalanced datasets. A popular approach is to train the model only on normal samples, learning their distribution, then evaluate them using abnormal samples and untrained normal samples, finally determining whether they are abnormal based on the degree to which they deviate from the normal sample distribution.

[0005] In recent years, numerous semi-supervised anomaly detection models have emerged applicable to various fields, including one-class support vector machines, Gaussian mixture models, and neural networks. One-class support vector machines and traditional autoencoders use one-dimensional input data, ignoring the two-dimensional structure of the original data, making it difficult to capture the co-variation relationships between aero-engine performance parameters. However, convolutional autoencoders, a significant variant of traditional autoencoders, containing convolutional layers, are well-suited for processing two-dimensional data. Nevertheless, directly using convolutional autoencoders for aero-engine anomaly detection still has two drawbacks. First, calculating the reconstructed similarity of samples—that is, the similarity between input and reconstructed samples—requires a meaningful and computable distance metric in the original space. Second, engine monitoring data contains noise, which adversely affects the calculation of reconstructed similarity. Summary of the Invention

[0006] To achieve the above objectives, the technical solution adopted in this application is: to provide a method for detecting anomalies in complex equipment based on reconstructed similarity, wherein the method for detecting anomalies in complex equipment based on reconstructed similarity uses the CAE-SIAMESE model to detect anomalies in aero-engines under conditions of sample imbalance.

[0007] The CAE-SIAMESE model includes a convolutional autoencoder and a Siamese neural network; the convolutional autoencoder is used for representation learning; and the Siamese neural network is used to measure the similarity of sample reconstructions.

[0008] Optionally, the complex equipment imbalance anomaly detection method based on reconstruction similarity includes a training step;

[0009] The training step, after data preprocessing, requires training the CAE-SIAMESE model. The training steps for the CAE-SIAMESE model are as follows:

[0010] (1) Only normal samples are used to train the convolutional autoencoder, that is, the loss function L is minimized through the backpropagation algorithm. cae ;

[0011] (2) Pair normal samples with their corresponding reconstructed samples and collect all paired samples as input to the Siamese neural network;

[0012] (3) Train the Siamese neural network using paired samples, i.e., minimize the loss function L. sia .

[0013] Alternatively, the Siamese neural network can be trained by directly training the Siamese neural network from scratch using paired samples.

[0014] Alternatively, the Siamese neural network can be trained by directly transferring the weights of the trained encoder to the Siamese neural network, and then using paired samples to train the Siamese neural network.

[0015] Optionally, the complex equipment imbalance anomaly detection method based on reconstruction similarity includes a testing step;

[0016] The testing steps, after training is complete, involve testing the CAE-SIAMESE model. The testing steps for the CAE-SIAMESE model are as follows:

[0017] (1) Input the test samples into the convolutional autoencoder to obtain the corresponding reconstructed representation;

[0018] (2) Input the sample pairs consisting of the input samples and the corresponding reconstructed representations into the Siamese neural network to obtain the corresponding embedded representations;

[0019] (3) The similarity between the input sample and the reconstructed representation is measured by calculating the squared Euclidean distance between the embedded representations, i.e. the reconstruction similarity of the sample;

[0020] (4) Compare the reconstruction similarity of the samples with the threshold to screen out abnormal samples.

[0021] Optionally, a convolutional autoencoder includes an encoder and a decoder;

[0022] The encoder consists of two convolutional layers, one batch layer, and a LeakyReLU activation function;

[0023] The decoder consists of two deconvolutional layers, a batch layer, a ReLU activation function, and a Tanh layer.

[0024] Optionally, for a single-channel input x, the latent representation of the i-th feature map in a convolutional layer is formally defined as follows:

[0025]

[0026] In the formula, f represents the activation function in the encoder (LeakyReLU is used in this application), b i Let represent the bias of the i-th feature map, and let the sign be . Represents a two-dimensional convolution operation;

[0027] A deconvolutional layer can be viewed as the inverse mapping of a convolutional layer, and the reconstructed representation of input x is defined as follows:

[0028]

[0029] In the formula, g represents the activation function in the decoder, c represents the bias of each input channel, and H represents a set of latent feature maps;

[0030] The reconstruction error of the sample, which is the loss function of the convolutional autoencoder, is expressed as:

[0031]

[0032] In the formula, ||·||1 represents the 1-norm;

[0033] The 1-norm is used to measure the reconstruction error between the input sample and the reconstructed sample.

[0034] Optionally, the Siamese neural network represents a sample x and its corresponding reconstruction. As input, and output the corresponding embedded representation F W (x) and

[0035] A simple distance metric in embedded representations can approximate the neighborhood relationships between input samples;

[0036] By calculating the Euclidean distance between the embedded representations:

[0037]

[0038] This distance is used to compare the input sample x and the reconstructed sample. The degree of similarity.

[0039] Optionally, in Siamese neural networks, the loss function used is triangular loss, which is an improvement on contrastive loss;

[0040] The formal definition of contrastive loss is as follows:

[0041]

[0042] In the formula, y represents the paired data. Tags;

[0043] when When they belong to the same category, y = 1; when y = 0 when they belong to different classes;

[0044] d represents a threshold that restricts the Euclidean distance between embedded representations belonging to different classes to the range [0, d]. That is, when the distance exceeds d, the loss is set to 0.

[0045] Since the label for all paired data is 1, the contrastive loss becomes:

[0046]

[0047] The learned mapping F is obtained by using contrastive loss. W (·) might be a constant mapping, i.e. At this point, the loss function is at its minimum, and its value is 0;

[0048] The formal definition of triangular loss is as follows:

[0049]

[0050] Optionally, after obtaining the reconstructed similarity of samples using a Siamese neural network, it is also necessary to set an appropriate threshold to distinguish between normal and abnormal sample regions. The threshold selection method adopts a maximum likelihood estimation-based method. Based on the six sigma criterion, the threshold β for sample deviation is set to the 3.4 quantile of the gamma distribution.

[0051]

[0052] In the formula, r β (s) represents the diagnostic result of the sample;

[0053] If the deviation of a sample is less than the decision threshold β, it will be diagnosed as a normal sample; otherwise, it will be diagnosed as an abnormal sample.

[0054] This application presents a complex equipment imbalance anomaly detection method based on reconstruction similarity, which boasts high diagnostic accuracy and strong noise resistance. It addresses the problem of poor diagnostic performance of existing semi-supervised anomaly detection methods on imbalanced and noisy high-dimensional monitoring data. This method effectively distinguishes between normal and abnormal anomaly scores, and then uses a threshold selection method based on maximum likelihood estimation to set thresholds and perform anomaly detection. The constructed CAE-SIAMESE consists of a convolutional autoencoder and a Siamese neural network. The convolutional autoencoder is optimized only for normal samples to calculate the reconstruction representation and depth features of normal samples. The Siamese neural network, acting as a reconstruction similarity learning mechanism, is optimized for sample pairs composed of normal samples and their corresponding reconstruction representations to calculate the reconstruction similarity of the samples, which is then used as the anomaly score for anomaly detection. The advantages of CAE-SIAMESE are: (1) The reconstruction similarity learning mechanism makes the reconstruction similarity of normal samples obtained by CAE-SIAMESE greater and the reconstruction similarity of abnormal samples smaller, which is more conducive to the diagnosis of normal and abnormal; (2) The convolutional layer in the network can capture the complex relationship in multidimensional monitoring data and better represent the characteristics of the original sample; (3) The Siamese neural network can resist the influence of noise when measuring similarity, effectively solving the problem of noise in engine monitoring data. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is the overall flowchart of the anomaly detection framework;

[0057] Figure 2 This is a diagram illustrating the triangular loss pattern.

[0058] Figure 3 The figure shows the anomaly detection effect of the method considered under noise interference. Detailed Implementation

[0059] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0060] This application now describes a method for detecting anomalies in complex equipment based on reconstructed similarity. This method, using the CAE-SIAMESE model, detects anomalies in aero-engines under conditions of sample imbalance.

[0061] The CAE-SIAMESE model includes a convolutional autoencoder and a Siamese neural network.

[0062] Convolutional autoencoders are used for representation learning.

[0063] Siamese neural networks are used to measure the similarity of sample reconstructions.

[0064] The CAE-SIAMESE model is a convolutional auto-encoder and deep Siamese convolutional neural network (CAE-SIAMESE).

[0065] The complex equipment imbalance anomaly detection method based on reconstruction similarity includes training and testing steps, such as... Figure 1 As shown.

[0066] The training step, following data preprocessing, requires training the CAE-SIAMESE model. The training steps are summarized below:

[0067] (1) Only normal samples are used to train the convolutional autoencoder, that is, the loss function L is minimized through the backpropagation algorithm. cae ;

[0068] (2) Pair normal samples with their corresponding reconstructed samples and collect all paired samples as input to the Siamese neural network;

[0069] (3) Train the Siamese neural network using paired samples, i.e., minimize the loss function L. sia .

[0070] There are two different training methods for Siamese neural networks. One is to train the Siamese neural network from scratch using paired samples directly, and the other is to directly transfer the weights of the trained encoder to the Siamese neural network, and then train the Siamese neural network using paired samples.

[0071] The testing steps, after training is complete, involve testing the CAE-SIAMESE model. The testing steps for the CAE-SIAMESE model are summarized as follows:

[0072] (1) Input the test samples into the convolutional autoencoder to obtain the corresponding reconstructed representation;

[0073] (2) Input the sample pairs consisting of the input samples and the corresponding reconstructed representations into the Siamese neural network to obtain the corresponding embedded representations;

[0074] (3) The similarity between the input sample and the reconstructed representation is measured by calculating the squared Euclidean distance between the embedded representations, i.e. the reconstruction similarity of the sample;

[0075] (4) Compare the reconstructed similarity of the samples with the threshold to screen out abnormal samples, thereby achieving the purpose of anomaly detection.

[0076] The specific structure of the convolutional autoencoder in this application is shown below. The convolutional autoencoder includes an encoder and a decoder.

[0077] The encoder comprises two convolutional layers, one batch layer, and one Leaky ReLU activation function, while the decoder comprises two deconvolutional layers, one batch layer, one ReLU activation function, and one Tanh layer. Compared to traditional autoencoders, the convolutional autoencoder of this application uses convolutional layers instead of fully connected layers in traditional autoencoders, and uses deconvolutional layers instead of fully connected layers in traditional autoencoders.

[0078] Convolutional layers are a key component that distinguishes convolutional autoencoders from traditional autoencoders, significantly reducing the number of parameters in the network. This improves the network's robustness and reduces the risk of overfitting. For a single-channel input x, the latent representation of the i-th feature map is formally defined as follows:

[0079]

[0080] In the formula, f represents the activation function in the encoder (LeakyReLU is used in this application), b i Let represent the bias of the i-th feature map, and let the sign be . This represents a two-dimensional convolution operation.

[0081] A deconvolutional layer can be viewed as the inverse mapping of a convolutional layer, and the reconstructed representation of input x is defined as follows:

[0082]

[0083] In the formula, g represents the activation function in the decoder (ReLU is used in this application), c represents the bias of each input channel, and H represents a set of latent feature maps.

[0084] The loss function of a convolutional autoencoder is typically the reconstruction error of the samples, which can be expressed as:

[0085]

[0086] In the formula, ||·||1 represents the 1-norm.

[0087] This application uses the 1-norm to measure the reconstruction error between the input sample and the reconstructed sample.

[0088] The network structure of the Siamese neural network in this application is consistent with the encoder in a convolutional autoencoder. The Siamese neural network includes two convolutional layers, one batch layer, and one LeakyReLU activation function. There are two ways to initialize the weights of the Siamese neural network: one is to directly train the Siamese neural network from scratch using paired samples, and the other is to directly transfer the weights of the trained encoder to the Siamese neural network, and then train the Siamese neural network using paired samples.

[0089] The twin neural network is a coupled architecture based on two artificial neural networks. This application consists of two convolutional neural networks with identical structures and shared weights.

[0090] Siamese neural networks are represented by a sample x and its corresponding reconstruction. As input, and output the corresponding embedded representation F W (x) and

[0091] A simple distance metric in an embedded representation can approximate the neighborhood relationships between input samples.

[0092] By calculating the Euclidean distance between the embedded representations:

[0093]

[0094] This distance is used to compare the input sample x and the reconstructed sample. The similarity between the input and reconstructed samples is determined by the distance metric used in the input space. Furthermore, regardless of the distance metric used, directly calculating the reconstructed similarity is susceptible to noise. However, Siamese neural networks effectively eliminate the influence of noise because the embedded representations of the input and reconstructed samples belong to the same class and are thus "pulled" together. For example, in object detection, although the target is affected by external factors such as lighting and angle, Siamese neural networks can still identify it, demonstrating that they are not easily affected by external noise.

[0095] In Siamese neural networks, the loss function used in this application is triangular loss, such as... Figure 2 As shown.

[0096] Triangular loss is an improvement on contrast loss.

[0097] The formal definition of contrastive loss is as follows:

[0098]

[0099] In the formula, y represents the paired data. The tag (when) When they belong to the same category, y = 1; when (y = 0 when they belong to different classes).

[0100] d represents a threshold that restricts the Euclidean distance between embedded representations belonging to different classes to the range [0, d]. That is, when the distance exceeds d, the loss is set to 0.

[0101] The label for all paired data is 1. Therefore, the contrastive loss becomes:

[0102]

[0103] The learned mapping F is obtained by using contrastive loss. W (·) might be a constant mapping, i.e. At this point, the loss function is at its minimum, with a value of 0.

[0104] The formal definition of triangular loss is as follows:

[0105]

[0106] Known If D F (F W (x), h)→0 and According to the limit theorem, we can obtain Furthermore, if So minimize This ensures that the Siamese neural network can "pull" together embedded representations belonging to the same category. Furthermore, given that h is a latent representation of the input sample x, if x1 ≠ x2, then h1 ≠ h2. This means that the learned mapping F... W (·) cannot be a constant mapping, thus solving the degradation problem of Siamese neural networks.

[0107] After obtaining the reconstructed similarity of samples using a Siamese neural network, a suitable threshold needs to be set to distinguish between normal and abnormal samples. This application also proposes a threshold selection method based on maximum likelihood estimation. For the convenience of the following analysis, a theorem on the gamma distribution and its corollaries are first given.

[0108] Theorem 1: Suppose random variables X1, X2, ..., Xn are given.n If the elements are independent and identically distributed according to a standard normal distribution N(0,1), then the statistic is... Follows a gamma distribution

[0109] Corollary 1: Assume random variables X1, X2, ..., Xn. n Independent and identically distributed in a normal distribution N(μ, σ) 2 ), where μ is a known constant, then the statistic is Follows a gamma distribution

[0110] In this application, the reconstructed similarity score of samples is defined as follows:

[0111]

[0112] In the formula,

[0113] and

[0114] Assumption and They all approximately follow a normal distribution but It approximately follows a normal distribution N(0, σ). 2 According to Corollary 1, the reconstructed similarity s of the samples approximately follows a gamma distribution Ga(k, θ), where... and For a gamma distribution Ga(k, θ), if k is a known constant, then the maximum likelihood estimate of θ is:

[0115]

[0116] In the formula, N1 represents the number of normal samples in the training set.

[0117] The threshold selection method based on maximum likelihood estimation, according to the six sigma criterion, sets the threshold β for sample deviation to the 3.4 quantile of the gamma distribution.

[0118]

[0119] In the formula, r β (s) represents the diagnostic result of the sample. In other words, if the deviation of the sample is less than the decision threshold β, it will be diagnosed as a normal sample. Otherwise, it will be diagnosed as an abnormal sample.

[0120] To demonstrate the effectiveness of the proposed imbalance anomaly detection method, this section uses real monitoring data from 36 aero engines of an airline to verify its effectiveness. In fact, aero engines are highly complex and precise thermodynamic machines with numerous performance parameters to be monitored. This application selects three performance parameters that best characterize the engine's state: Delta Exhaust Gas Temperature (DEGT), Delta Fuel Flow (DFF), and Delta Core Speed ​​(DN2). The sample construction process is shown in Table 1 below: First, research shows that the order of parameters affects the performance of the convolutional neural network. This application selects the DEGT-DN2-DFF order. Then, samples are constructed using a sliding window method. The window length is set to 10, and the window step size is set to 5. Table I shows the 6100 samples collected, including 6064 normal samples and 36 abnormal samples. Clearly, the dataset is severely imbalanced.

[0121] Table I: Engine Dataset

[0122]

[0123] This application was experimentally verified on a device with a six-core Intel(R) Core(TM) i7-9750H CPU @ 2.59GHz processor and 8GB DDR4 memory. To ensure that the latent representation of the samples and the embedded representation are consistent in dimensionality, the network structure of the Siamese neural network in this application is the same as the encoder in the convolutional autoencoder. The network structure of the convolutional autoencoder in this application mainly includes the number of convolution channels, the size of the convolution kernel, the stride of the convolution, and the padding method. The hyperparameter settings of the convolutional autoencoder are shown in Table II.

[0124] Table II. Hyperparameters of Convolutional Autoencoders:

[0125] hyperparameters Convolutional layer 1 Convolutional layer 2 Deconvolution layer 1 Deconvolution layer 2 Number of channels (1，4) (4，16) (16，4) (4，1) kernel size 2×2 3×3 3×3 2×2 Kernel stride 1 1 1 1 filling 1 0 0 1

[0126] If the number of abnormal samples is close to the number of normal samples, then overall accuracy can be used to effectively evaluate the diagnostic performance of the anomaly detection model. However, with improvements in manufacturing and assembly levels, as well as advancements in monitoring and repair technologies, aero-engines have become highly reliable and mature products. Therefore, for aero-engines, the number of normal samples far exceeds the number of abnormal samples, as detailed in Table III. In this case, overall accuracy is not very useful. For example, in the test set, there are 1820 normal samples, while there are only 36 abnormal samples. Even if the anomaly model diagnoses all test samples as normal, its overall accuracy is still as high as 98.06%. Clearly, overall accuracy is not suitable for evaluating the diagnostic performance of anomaly models under imbalanced sample conditions. Therefore, this application selects AUC and confusion matrix to evaluate the proposed anomaly detection model.

[0127] Table III Dataset Overview:

[0128] state training set test set total normal samples 4244 1820 6064 Abnormal samples 0 36 36 total 4244 1856 6100

[0129] The confusion matrix, as shown in Table IV, categorizes samples in the test set using two criteria: the true class and the predicted class. Comparing TP and FN reveals the diagnostic performance of normal samples, while comparing FP and TN reveals the diagnostic performance of abnormal samples. Since there is no overlap between the various metrics, the confusion matrix is ​​not affected by data imbalance.

[0130] The receiver operating characteristic (ROC) curve (AUC) shows the false positive rate (FPR) on the x-axis and the true positive rate (TPR) on the y-axis. The area enclosed by the AUC and the coordinate axis is defined as the AUC. Studies have shown that under imbalanced data conditions, the AUC is a better measure of classifier performance than overall accuracy.

[0131] Table IV Confusion Matrix

[0132]

[0133] To verify the anomaly detection performance of the CAE-SIAMESE model, this application conducted five sets of experiments. The first set used a single-class support vector machine (OC-SVM) to directly classify the original data. The second set used a traditional autoencoder (AE) to classify the original data. The third set used a convolutional autoencoder (CAE) to classify the original data. The fourth set first used a CAE to learn the distribution of normal samples, and then used a Siamese neural network (SIAMESE) to classify the samples. The network structure of the Siamese neural network was consistent with the encoder of the convolutional autoencoder, and the weights were randomly initialized. The fifth set of experiments was similar to the fourth set, except that the weights of the Siamese neural network were directly transferred from the trained encoder. To distinguish it from the fourth set of experiments, this experiment can be abbreviated as CAE-SIAMESE-WT. Furthermore, to eliminate the problems caused by the randomness of the neural network, each set of experiments in this application was repeated ten times, and the test and validation sets were re-divided for each experiment.

[0134] Table V shows the mean and standard deviation of the AUC for each experimental group. If AUC characterizes the diagnostic accuracy of each model, the models with the highest to lowest diagnostic accuracy are: CAE-SIAMESE-WT model, CAE-SIAMESE model, CAE model, AE model, and OC-SVM model. Compared to the OC-SVM and AE models, the CAE model achieved better diagnostic accuracy, which is undoubtedly directly attributable to the convolutional and deconvolutional layers in the CAE model. These layers can capture the 2D structure in engine monitoring data, thus making it easier to capture the co-variational relationships between performance parameters.

[0135] Furthermore, compared with other anomaly detection methods, the CAE-SIAMESE-WT model established in this application achieved the highest diagnostic accuracy, and the CAE-SIAMESE model also achieved the second-best diagnostic accuracy. This fully demonstrates that using Siamese neural networks can better measure the reconstruction similarity of test samples, thereby better distinguishing normal samples from abnormal samples. In addition, directly transferring the weights of the trained encoder to the Siamese neural network improves both the model's diagnostic accuracy and robustness. Furthermore, while the CAE-SIAMESE and CAE-SIAMESE-WT models exhibit better robustness than the CAE model...

[0136] The mean and standard deviation of the AUC for the methods considered in Table V.

[0137] serial number method mean Standard deviation 1 OC-SVM 0.8595 ---- 2 AE 0.8711 0.0002 3 CAE 0.9104 0.0036 4 CAE-SIAMESE(our) 0.9418 0.0029 5 CAE-SIAMESE-WT(our) 0.9494 0.0025

[0138] Table VI. Mean AUC of the methods considered under different noise conditions

[0139]

[0140] To verify the noise resistance of the Siamese neural network in measuring the similarity of reconstructed samples, this application conducted four sets of experiments. In the first set of experiments, additive white Gaussian noise with a signal-to-noise ratio (SNR) of 20 dB was added to the engine monitoring data. Subsequently, the SNR was increased by 5 dB each time. The SNR is defined as follows:

[0141]

[0142] In the formula, P signal and P noise Let represent the power of the original signal and the power of the noise, respectively. According to the above formula, if the signal-to-noise ratio (SNR) is too low, the original signal will be "overwhelmed" by the noise; if the SNR is too high, the original signal will be almost unaffected by the noise. Neither of these scenarios can satisfy the experimental objective.

[0143] Table VI shows the mean AUC values ​​of the CAE model, CAE-SIAMESE model, and CAE-SIAMESE-WT model under different noise intensities. Experimental results show that the diagnostic accuracy of all models decreases under noise interference, and the higher the noise intensity, the worse the diagnostic accuracy. However, under the same noise intensity, the CAE-SIAMESE model and CAE-SIAMESE-WT model have better diagnostic accuracy than the CAE model. This indicates that Siamese neural networks have better noise resistance when measuring sample reconstruction similarity. The CAE model directly uses the sample reconstruction error as the reconstruction similarity, and is affected by noise regardless of the distance metric used. However, the Siamese neural network "pulls" the embedded representations of the input sample and the reconstructed sample together, thereby reducing the impact of noise on the calculation of sample reconstruction similarity. Furthermore, Figure 3 This also more intuitively illustrates the changing trend of the diagnostic accuracy of the above model under interference from noise of different intensities.

[0144] Table VII shows the confusion matrix of the CAE-SIAMESE-WT diagnostic results. In the test set, there were 1820 normal samples, of which 1675 were correctly diagnosed, resulting in a true negative rate of approximately 0.9203. There were 36 abnormal samples, of which 31 were correctly diagnosed, resulting in a true positive rate of approximately 0.8611. The results indicate that the threshold selection method based on maximum likelihood estimation proposed in this application has high diagnostic accuracy and significant application potential in practical engineering.

[0145] Table VII shows the diagnostic results of CAE-SIAMESE-WT.

[0146]

[0147] This application provides a method for detecting imbalanced anomalies in complex equipment based on reconstruction similarity. Addressing the problem that existing semi-supervised anomaly detection methods perform poorly on imbalanced and noisy high-dimensional monitoring data, this application introduces a reconstruction similarity learning mechanism on top of a convolutional autoencoder network, constructing a convolutional-Siamese network model (CAE-SIAMESE) to learn anomaly scores that can effectively distinguish between normal and abnormal data. Then, a threshold selection method based on maximum likelihood estimation is used to set the threshold and perform anomaly detection. The constructed CAE-SIAMESE consists of a convolutional autoencoder and a Siamese neural network. The convolutional autoencoder is optimized only by normal samples to calculate the reconstruction representation and depth features of normal samples. The Siamese neural network, as the reconstruction similarity learning mechanism, is optimized by sample pairs composed of normal samples and their corresponding reconstruction representations to calculate the reconstruction similarity of the samples, which is then used as the anomaly score for anomaly detection. The advantages of CAE-SIAMESE are: (1) The reconstruction similarity learning mechanism makes the reconstruction similarity of normal samples obtained by CAE-SIAMESE greater and the reconstruction similarity of abnormal samples smaller, which is more conducive to the diagnosis of normal and abnormal; (2) The convolutional layer in the network can capture the complex relationship in multidimensional monitoring data and better represent the characteristics of the original sample; (3) The Siamese neural network can resist the influence of noise when measuring similarity, effectively solving the problem of noise in engine monitoring data.

[0148] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for detecting imbalance anomalies in complex equipment based on reconstruction similarity, characterized in that: The complex equipment imbalance anomaly detection method based on reconstruction similarity, through the CAE-SIAMESE model, realizes the detection of aero-engine anomalies under sample imbalance conditions; The CAE-SIAMESE model includes a convolutional autoencoder and a Siamese neural network; the convolutional autoencoder is used for representation learning; the Siamese neural network is used to measure the similarity of sample reconstruction. The training steps for the CAE-SIAMESE model are as follows: (1) Only normal samples are used to train the convolutional autoencoder, that is, the loss function is minimized through the backpropagation algorithm. ; (2) Pair normal samples with their corresponding reconstructed samples and collect all paired samples as input to the Siamese neural network; (3) Train the Siamese neural network using paired samples, i.e. minimize the loss function. ; The twin neural network uses samples and its corresponding reconstruction representation As input, and output the corresponding embedded representation. and ; The simple distance metric of embedded representation approximates the neighborhood relationship between input samples; By calculating the Euclidean distance between the embedded representations: This distance is used to compare input samples. and reconstructed samples The degree of similarity; In the aforementioned Siamese neural network, the loss function used is triangular loss. The formal definition of the triangular loss is as follows: in, Represents paired data The tag, Input sample The potential representation of; After obtaining the reconstructed similarity of samples using Siamese neural networks, it is necessary to set an appropriate threshold to distinguish between normal and abnormal sample regions. The threshold selection method adopts a maximum likelihood estimation-based threshold selection method, which, according to the six sigma criterion, sets the threshold value for sample deviation. If set to the 3.4 quantile of the gamma distribution, then In the formula, Indicates the diagnostic results of the sample; If the deviation of the sample is less than the decision threshold If it is normal, it will be diagnosed as a normal sample; otherwise, it will be diagnosed as an abnormal sample.

2. The method for detecting imbalance anomalies in complex equipment based on reconstruction similarity as described in claim 1, characterized in that: The Siamese neural network is trained by directly using paired samples to train the Siamese neural network from scratch.

3. The method for detecting imbalance anomalies in complex equipment based on reconstruction similarity as described in claim 1, characterized in that: The Siamese neural network is trained by directly transferring the weights of the trained encoder to the Siamese neural network, and then using paired samples to train the Siamese neural network.

4. The method for detecting imbalance anomalies in complex equipment based on reconstruction similarity as described in claim 1, characterized in that: The method for detecting imbalance anomalies in complex equipment based on reconstructed similarity includes testing steps; The testing steps, after training is complete, involve testing the CAE-SIAMESE model. The testing steps for the CAE-SIAMESE model are as follows: (1) Input the test samples into the convolutional autoencoder to obtain the corresponding reconstructed representation; (2) Input the sample pairs consisting of the input samples and the corresponding reconstructed representations into the Siamese neural network to obtain the corresponding embedded representations; (3) The similarity between the input sample and the reconstructed representation is measured by calculating the squared Euclidean distance between the embedded representations, i.e., the reconstruction similarity of the sample; (4) Compare the reconstruction similarity of the samples with the threshold to filter out abnormal samples.

5. The method for detecting imbalance anomalies in complex equipment based on reconstruction similarity as described in claim 1, characterized in that: A convolutional autoencoder consists of an encoder and a decoder; The encoder consists of two convolutional layers, one batch layer, and a LeakyReLU activation function; The decoder consists of two deconvolutional layers, a batch layer, a ReLU activation function, and a Tanh layer.

6. The method for detecting imbalance anomalies in complex equipment based on reconstruction similarity as described in claim 5, characterized in that: Convolutional layers for single-channel input , No. The formal definition of the latent representation of a feature map is as follows: In the formula, This represents the activation function in the encoder. Indicates the first The bias of each feature map, and the sign Represents a two-dimensional convolution operation; If a deconvolutional layer can be viewed as the inverse mapping of a convolutional layer, then the input... The reconstruction representation is defined as follows: In the formula, This represents the activation function in the decoder. This represents the bias of each input channel, and Represents a set of latent feature maps; The reconstruction error of the sample, which is the loss function of the convolutional autoencoder, is expressed as: In the formula, Represents the 1-norm; The 1-norm is used to measure the reconstruction error between the input sample and the reconstructed sample.

Images