An aero-engine abnormality detection method based on coded time domain convolution

Abstract

The present disclosure discloses an aero-engine anomaly detection method based on encoding time domain convolution, comprising the following steps: S100; collecting and preprocessing multi-sensor time domain data of an aero-engine; S200: encoding and reconstructing the preprocessed multi-sensor time domain data to obtain reconstructed multi-sensor time domain data; S300: predicting the reconstructed multi-sensor time domain data by using a time domain convolution network to obtain prediction data; S400: calculating the dynamic cumulative error of the reconstructed multi-sensor time domain data and the prediction data; S500: comparing the dynamic cumulative error with a detection threshold, and performing anomaly detection on the aero-engine according to the comparison result.

Description

Technical Field

[0001] This disclosure belongs to the field of aero-engine fault diagnosis technology, and specifically relates to a multi-sensor temporal anomaly detection method based on coded temporal convolution. Background Technology

[0002] Aero engines are a concentrated reflection of a nation's comprehensive national strength, industrial base, and technological level. Aero engines are highly complex thermodynamic machines that must operate reliably under extreme conditions of high temperature, high pressure, high speed, and alternating loads. They are a crucial strategic guarantee for national defense and a nation's status as a powerful power. A failure in any part of the engine can spread to the entire system, causing enormous economic losses and casualties. Therefore, online monitoring of the aero engine's operating status and real-time anomaly detection methods are essential for improving engine operational safety.

[0003] Generally, methods for detecting anomalies in aero-engines fall into two main categories: physics-based and data-based. Physics-based methods utilize failure mechanism models or other system phenomenon descriptions to determine the possible operating states of components or systems, allowing for accurate judgment of the engine's operating state with relatively little data. However, this requires sufficient information about failure modes, and aero-engines have extremely complex structures; physical models cannot completely reproduce their internal structures. Therefore, physics-based methods are difficult to model and have low accuracy. Data-based methods use sensor observation data to determine the characteristics of degradation processes and predict future states, requiring almost no knowledge of physical failure models. Data-based methods are further divided into signal processing methods and deep learning-based time-series outlier detection methods. Due to the limited locations for sensor installation on aero-engines, inappropriate sensor placement and selection will result in lower accuracy for signal processing methods. Deep learning-based time-series outlier detection methods, on the other hand, can overcome expert experience by learning the deep nonlinear characteristics of aero-engines, improving anomaly detection accuracy. Traditional time-series outlier detection methods, such as support vector machines, k-nearest neighbors, and isolated forests, are based on classification and use the distribution of normal data to set decision boundaries. Such methods are inefficient, inaccurate, and difficult to integrate multi-sensor information when processing high-dimensional nonlinear data. In recent years, many deep learning algorithms with automatic feature extraction capabilities have been applied to outlier detection in time series data, such as Long Short-Term Memory networks, autoencoders, and recurrent neural networks. However, due to the complex and variable operating conditions of aero-engines, automatic feature extraction methods often fail to learn the complex transient patterns of time series signals. Furthermore, these deep learning algorithms suffer from poor interpretability and parallel computing capabilities, and they also fail when processing long-term data series due to gradient vanishing. Therefore, researching multi-sensor time-series anomaly detection methods with operating condition adaptability is of great significance for solving the aforementioned practical problems in aero-engine operational anomaly detection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this disclosure is to provide a multi-sensor temporal anomaly detection method based on coded temporal convolution. This method can achieve accurate anomaly detection even without fault samples, greatly improving the robustness of the detection method in practical applications and effectively solving the problem of low detection efficiency of data-driven methods caused by complex changes in engine operating conditions.

[0005] To achieve the above objectives, this disclosure provides the following technical solutions:

[0006] An aero-engine anomaly detection method based on encoding temporal convolution includes the following steps:

[0007] S100; Acquires and preprocesses multi-sensor time-domain data of aero-engines;

[0008] S200: Encode and reconstruct the preprocessed multi-sensor time-domain data to obtain the reconstructed multi-sensor time-domain data;

[0009] S300: Uses a temporal convolutional network to predict the reconstructed multi-sensor temporal data to obtain prediction data;

[0010] S400: Calculates the dynamic cumulative error between the reconstructed multi-sensor time-domain data and the predicted data;

[0011] S500: Compare the dynamic cumulative error with the detection threshold, and perform anomaly detection on the aero-engine based on the comparison result.

[0012] Preferably, in step S 100, the preprocessing of the multi-sensor time-domain data includes the following steps:

[0013] S101: Perform a sliding window operation on multi-sensor time-domain data to obtain a multi-source training dataset;

[0014] S102: Normalize the multi-source training dataset to obtain a normalized multi-source training dataset.

[0015] Preferably, in step S200, the time-domain data of the multi-sensor is encoded and reconstructed by an autoencoder, which includes an encoding layer, an intermediate layer and a decoding layer.

[0016] Preferably, in step S300, the temporal convolutional network includes a dilated dilated convolution module, a random deactivation module, a weight normalization module, a residual connection module, and a ReLU activation function module.

[0017] Preferably, in step S300, the training steps of the temporal convolutional network are as follows:

[0018] S301: Divide the reconstructed multi-sensor time-domain data into training and testing sets;

[0019] S302: Train the temporal convolutional network using the training set. When the loss function of the temporal convolutional network converges, the training is complete, and the optimal temporal convolutional network is obtained. If the loss function does not converge, tune the parameters and train again until convergence.

[0020] S303: Use the optimal temporal convolutional network to make predictions on the test set.

[0021] This disclosure also provides an aero-engine anomaly detection device based on coded temporal convolution, comprising:

[0022] The acquisition module is used to acquire time-domain data from multiple sensors of the aero-engine.

[0023] The reconstruction module is used to encode and reconstruct multi-sensor time-domain data to obtain reconstructed multi-sensor time-domain data.

[0024] The prediction module is used to predict the encoded and reconstructed multi-sensor temporal data using a temporal convolutional network to obtain prediction data.

[0025] The calculation module is used to calculate the dynamic cumulative error of the multi-sensor time-domain data and prediction data;

[0026] The detection module is used to compare the dynamic cumulative error with the detection threshold and detect the anomaly of the aero-engine based on the comparison result.

[0027] Preferably, the device further includes a preprocessing module for preprocessing the multi-sensor time-domain data.

[0028] Preferably, the device further includes a threshold acquisition module for acquiring a detection threshold.

[0029] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0030] 1. An autoencoder is used to process multi-sensor time-series data. The original multi-sensor time-series data is input into the autoencoder for reconstruction, and the time-series data of multiple operating conditions is mapped into a sequence of variable length. This facilitates the storage of large-scale time-series data of aero-engines and enables the learning of the normal operating conditions of aero-engines under multiple operating conditions. It captures the features of normal samples at multiple levels and obtains multiple models with good reconstruction and recognition capabilities for normal samples.

[0031] 2. Using a temporal convolutional network for time series sample prediction, and a dilated and dilated convolutional network for parallel computation, the degree of dilation is determined according to the number of sensors, which can better solve the gradient explosion problem in long-term series prediction of aero-engines.

[0032] 3. A method is proposed to use the design threshold of reconstruction error and the dynamic changes of error distribution for anomaly detection. This reduces the possibility of missed or false alarms due to changes in operating conditions and improves the robustness of anomaly detection. This is because when operating conditions change, the distribution of sampled data will change temporarily, but after the change, it will still be consistent with the distribution before the change. However, once the mechanical equipment fails, its distribution will change irreversibly. Therefore, the robustness of anomaly detection can be improved based on the dynamic changes of error distribution. Attached Figure Description

[0033] Figure 1This is a flowchart of a multi-sensor temporal anomaly detection method based on coded temporal convolution provided in one embodiment of the present disclosure;

[0034] Figure 2 A schematic diagram of the structure of a self-encoder provided in another embodiment of this disclosure;

[0035] Figure 3 This is a schematic diagram of the structure of a temporal convolutional network provided in another embodiment of the present disclosure;

[0036] Figure 4 This is a frequency distribution diagram of the loss function value and the reconstructed value provided in another embodiment of this disclosure;

[0037] Figures 5(a) to 5(c) for Figure 1 The method shown is applied to the statistical chart of anomaly detection results in the embodiment. In the figure, Figure 5(a) is the original vibration signal of the same device with a single sensor; Figure 5(b) is the data after encoding and reconstructing the vibration signal of the same device with a single sensor; and Figure 5(c) is the predicted signal of the original vibration signal.

[0038] Figure 6(a) shows the anomaly detection results of a single sensor;

[0039] Figure 6(b) shows the anomaly detection results of multiple sensors on the same device. Detailed Implementation

[0040] The following will refer to the appendix. Figures 1 to 6(b) Specific embodiments of this disclosure are described in detail. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0041] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions in the specification describe preferred embodiments of this disclosure; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of this disclosure. The scope of protection of this disclosure is determined by the appended claims.

[0042] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of this disclosure.

[0043] In one embodiment, such as Figure 1 As shown, this disclosure proposes a multi-sensor temporal anomaly detection method based on encoding temporal convolution, including the following steps:

[0044] S100; Acquires and preprocesses multi-sensor time-domain data of aero-engines;

[0045] S200: Encode and reconstruct the preprocessed multi-sensor time-domain data to obtain the reconstructed multi-sensor time-domain data;

[0046] S300: Uses a temporal convolutional network to predict the reconstructed multi-sensor temporal data to obtain prediction data;

[0047] S400: Calculates the dynamic cumulative error between the reconstructed multi-sensor time-domain data and the predicted data;

[0048] S500: Compare the dynamic cumulative error with the detection threshold, and perform anomaly detection on the aero-engine based on the comparison result.

[0049] The above embodiments constitute the complete technical solution of this disclosure. This embodiment, by encoding and reconstructing multi-sensor time-series data, can map multi-condition time-series data into sequences of variable length, facilitating the storage of large-scale time-series data of aero-engines. Furthermore, it can learn the normal operating conditions of aero-engines under multiple operating conditions, capturing the features of normal samples at multiple levels, and obtaining multiple models with good reconstruction and recognition capabilities for normal samples. In summary, this embodiment utilizes data from aero-engines in a healthy operating state to establish an unsupervised intelligent anomaly detection model for aero-engines. This effectively addresses the problem of aero-engines being unable to detect faults due to the difficulty in collecting fault data, and can be largely extended to other scenarios lacking fault data.

[0050] In another embodiment, step S100, preprocessing the multi-sensor time-domain data, includes the following steps:

[0051] S101: Perform a sliding window operation on multi-sensor time-domain data to obtain a multi-source training dataset;

[0052] In this step, L segments of data are extracted from the multi-sensor time-domain data using a sliding time window of fixed length. These L segments are then concatenated to form a multi-source training dataset. The panes used for data extraction should be no less than 1 / 10 of the sampling frequency of the sensor corresponding to each segment of time-series data, and the panes slide at fixed intervals with no overlap between them to ensure that the extracted data contains information about the complete operating cycle of the aero-engine. For example, if a sensor has a sampling frequency of 20,000 and a sampling interval of 10 minutes, sampling a total of 984 data segments, then 20,480 data points are extracted from each data segment each time data is extracted. These 984 × 20,480 data points are then concatenated. Since the sampling frequency of each sensor is different, the length of the time-series data collected at the same time is inconsistent, so the time-series length of each training dataset can be inconsistent.

[0053] S102: Normalize the multi-source training dataset to obtain the normalized multi-source training dataset.

[0054] In this step, the multi-source training dataset is normalized using the following formula:

[0055]

[0056] Where, x i x represents all sampled samples from the i-th sensor in the multi-source training dataset before normalization. i (1) represents the first sampling point of the sensor sample, and min and max represent the sampling points with the minimum and maximum sampling values, respectively. i ′ represents the sampled data of the i-th sensor after normalization.

[0057] After normalization, the value of each original sampling point in the multi-source training dataset falls within [-1, 1]. For example, if a certain sampling data is [1 2 3 4 5 4 3 2 1], the result after normalization is: [-1.0000 -0.500000.50001.00000.50000 -0.5000 -1.0000].

[0058] Normalization can prevent excessive differences in data between different sensors due to malfunctions in the target monitoring equipment, which could lead to errors in subsequent model inference. It can also prevent excessive differences in data between sensor groups due to different measurement ranges, which could hinder network synchronization operations.

[0059] In another embodiment, in step S200, the preprocessed multi-sensor time-domain data is encoded and reconstructed by an autoencoder, which includes an encoding layer, an intermediate layer, and a decoding layer.

[0060] In this step, the encoding reconstruction is mainly accomplished by an autoencoder, which includes an encoding layer, an intermediate layer, and a decoding layer. The corresponding encoding formula is as follows:

[0061]

[0062] In the formula, the encoder transforms the variable-length input sequence into a fixed-shape context variable c, and encodes the information of the input sequence in this context variable. At time step t, the recurrent neural network converts the input feature vector x of the word xt into a variable c. t and h t-1 (i.e., the hidden state at the previous time step) is transformed into h. t (i.e., the current hidden state). The encoder transforms the hidden states of all time steps into context variables using a selected function q. At any time step t′ on the output sequence, the recurrent neural network will convert the output y from the previous time step... t′-1 The context variable c is used as its input, and then at the current time step, they are combined with the previous hidden state S. t′-1 Transform into hidden state s t .

[0063] The more layers an autoencoder has, the more complex the network becomes, which may lead to phenomena such as gradient explosion; while the fewer layers, the weaker the model's learning ability. Therefore, after experiments, the autoencoder structure in this embodiment is set to an encoding layer with 16×8 neurons, an intermediate layer with 3 neurons, and a decoding layer with 16×8 neurons.

[0064] After obtaining the hidden state of the decoder, the output layer and softmax operation can be used to compute the output y at time step t′. t′ The conditional probability distribution P(y) t′ |y1,...,y t′-1 c), the softmax formula is as follows:

[0065]

[0066] in, This represents the output after the softmax transformation, where e represents the base of the natural logarithm function. i This indicates that the i-th reconstructed value is transformed. It represents the sum of a set of encoded data after transformation.

[0067] The softmax operation allows all encoded values ​​to be mapped to the range [0, 1].

[0068] In another embodiment, in step S300, as Figure 3As shown, the temporal convolutional network includes a dilated and dilated convolution module, a random deactivation module, a weight normalization module, a residual connection module, and a ReLU activation function module.

[0069] In this embodiment, dilated dilated convolution injects holes into standard convolution, increasing the receptive field by expanding the kernel size. For aero-engine sampling signals, this increased receptive field includes more signal fragment information, allowing the network to better learn the distribution of data under normal conditions. The random deactivation module is a structure used to reduce overfitting in neural networks. During each training iteration, a portion of neurons are removed based on a predetermined probability before training continues, preventing overfitting. The weight normalization module is a reparameterization process that decouples the magnitude and direction of the weight tensor, aiming for faster convergence and stronger learning rate robustness. The residual connection module primarily avoids gradient explosion during network training. The calculation methods for each module are as follows:

[0070]

[0071]

[0072]

[0073] Where F(s) represents the dilated and dilated convolution function operation, and k and d represent the kernel size and dilation coefficient, respectively. This represents the encoded data, where sd·i represents the number of weights selected in the current layer. This represents the output of each layer of the temporal convolutional network and the input of the next layer. Activation represents the ReLU activation function module, 'o' represents the result of the activation function operation, and 'y' represents the result after the ReLU activation function operation. Weight normalization is used in each layer of the temporal convolutional network. During temporal convolution, the encoded data after each truncation is... Simultaneously input into a temporal convolutional network for training.

[0074] Furthermore, the training process for the temporal convolutional network is as follows:

[0075] 1. Divide the encoded multi-sensor time-domain data into training and testing sets in a 6:4 ratio;

[0076] 2. Train the temporal convolutional network using the training set. The loss function is the cross-entropy loss function. The size of the temporal convolutional kernel is set to 1*7*25, the time step is set to 30, the value of the random deactivated module is 0.5, the number of training iterations is set to 50, and the batch size is set to 16. When the loss function converges, the training is complete, and the optimal temporal convolutional network is obtained. If the loss function does not converge, the relevant parameters are adjusted and the training is performed again.

[0077] 3. Deploy the trained temporal convolutional network using the test set.

[0078] In another embodiment, the difference diff is calculated by comparing the encoded and reconstructed multi-sensor temporal data with the predicted value output by the temporal convolutional network, i.e., diff = y. true -y pred The difference between each comparison is recorded as the error value e. Afterward, a new error value e is obtained for each comparison at time step t. t After obtaining the error value e for each segment of data. t When, if the error value e in the previous moment t The error value e at the next time step is greater than the error value at the next time step. t+1 Then et and e t+1 The values ​​remain unchanged; if the error value e from the previous moment remains unchanged. t Less than the next time e t+1 The error value, then e t The value of e remains unchanged. t+1 Set the value to -1* et+1 Meanwhile, the cumulative error and cutmsum(et) for all time periods are calculated. For example, if the difference over a certain time period is diff = [0.2, 0.1, 0.05, 0.1], then according to the method described in this disclosure, the difference over that time period is set to diff = [0.2, -0.1, -0.05, 0.1].

[0079] In another embodiment, in step S500, the detection threshold is determined by calculating the error distribution of the multi-sensor time-domain data and the reconstructed multi-sensor time-domain data, and using the extreme value of the error distribution as the detection threshold.

[0080] In this embodiment, at each time step, the decoder predicts the probability distribution of the output sampled data. The distribution is obtained by using a softmax transform and optimized by calculating the cross-entropy loss function as shown below:

[0081]

[0082] in, This represents the value after encoding and after softmax transformation, y i This represents the normalized encoding value of the i-th multi-sensor sampled data, and n represents the number of multi-sensor sampled data.

[0083] Then, set the epochs of the autoencoder to 100 and the batch size to 20, train the autoencoder network to minimize the cross-entropy loss function, save the model, and compare the loss function value with the encoded value. Statistical analysis was performed and frequency distribution plots of both were calculated. The maximum value of the error in the frequency distribution plot was used as the threshold Γ for anomaly detection.

[0084] After determining the threshold Γ, at each time step, the accumulated dynamic error and cumsum(e) are calculated. t The cumulative dynamic error is compared with a threshold Γ. If the cumulative dynamic error exceeds the threshold Γ, the aero-engine is considered to have malfunctioned. For example, in the subsequent verification, the frequency distribution of the loss function and the encoded value is shown in the following figure. Figure 4 As shown, it can be seen that the error is mainly distributed in the range of 0 to 0.1. Therefore, the detection threshold is set to 0.1. If the calculated value of the cumulative dynamic error is greater than 0.1, as shown by the black curve in Figure 6(a), then the engine is considered to be abnormal.

[0085] This embodiment utilizes the design threshold of reconstruction error and the dynamic changes in error distribution for anomaly detection. In reality, due to the highly complex and variable operating conditions of aero-engines, conventional anomaly detection methods often result in false alarms and missed alarms. This embodiment reconstructs multi-sensor time-domain data using an autoencoder, learning the operating patterns of the data itself. The final learning result ensures that the error range of data reconstruction under different operating conditions is less than the minimum error value (i.e., the threshold). In actual deployment, since the operating conditions of each aero-engine are not exactly the same, if an aero-engine malfunctions, the error in the data distribution will gradually increase, and vice versa. Based on this, this embodiment compares the dynamic cumulative error with the threshold to detect aero-engine anomalies, reducing the possibility of missed or false alarms due to changes in operating conditions, and improving the robustness of anomaly detection.

[0086] To better illustrate the technical effects of this disclosure, two anomaly detection scenarios were sampled for experimental verification in this embodiment: 1) single-sensor data from the same device; 2) multi-sensor data from the same device. Scenario 1) consists of 984 data segments, each containing 20,480 sampling points collected by the sensor. Scenario 2) consists of two sensors, comprising 491 data segments, each containing 32,768 sampling points collected by the sensor. The samples were preprocessed, temporal convolutional network predicted, and dynamic threshold calculated according to the method provided in this disclosure. Training and testing were conducted in both scenarios. The original data and the encoded reconstructed data in Scenario 1) are shown in Figures 5(a) and 5(b), and the predicted data is shown in Figure 5(c). The anomaly detection results obtained from the two scenarios are shown in Table 1 and Figures 6(a) and 6(b).

[0087] Table 1 Results of Multi-Sensor Temporal Anomaly Detection Using Encoding Temporal Convolution

[0088]

[0089] Figure 6(a) shows the bearing life dataset from NSAS-IMS, where the bearing has been confirmed to have experienced initial failure in data collection file 533; Figure 6(b) shows the bearing life dataset from XJTU-SY, where the bearing has been confirmed to have experienced initial failure in file 479. The fixed curve in the figures represents a defined threshold, while the continuously changing curve represents the dynamic error value. Once the dynamic error value exceeds the threshold, it indicates that the bearing has experienced initial failure.

[0090] As can be seen, the method provided in this disclosure can achieve accurate anomaly detection in two different scenarios, with a false positive rate of no more than 5% and an anomaly detection rate of over 99%. This indicates that the method described in this disclosure has high sensitivity to anomaly data. Furthermore, the detection results are compared with those of commonly used outlier detection networks (GANs), and the comparison results are shown in Table 2.

[0091] Table 2 Results of GAN Network Multi-Sensor Temporal Anomaly Detection

[0092]

[0093] As shown in Table 2, the method proposed in this disclosure outperforms commonly used anomaly detection GAN networks in both accuracy and the comprehensive index F1-score.

[0094] The foregoing general description of the invention and its specific embodiments should not be construed as a limitation on the technical solution of the invention. Those skilled in the art, based on the disclosure of this application, can add, reduce, or combine the disclosed technical features in the foregoing general description and / or specific embodiments (including examples) without departing from the constituent elements of the invention, to form other technical solutions within the scope of protection of this application.

Claims

1. An aero-engine anomaly detection method based on encoding temporal convolution, comprising the following steps: S100; Acquires and preprocesses multi-sensor time-domain data of aero-engines; S200: Encode and reconstruct the preprocessed multi-sensor time-domain data to obtain the reconstructed multi-sensor time-domain data; S300: Uses a temporal convolutional network to predict the reconstructed multi-sensor temporal data to obtain prediction data; S400: Calculates the dynamic cumulative error between the reconstructed multi-sensor time-domain data and the predicted data; S500: Compare the dynamic cumulative error with the detection threshold, and perform anomaly detection on the aero-engine based on the comparison result; in, The difference (diff) is calculated by comparing the encoded and reconstructed multi-sensor temporal data with the predicted values ​​output by the temporal convolutional network. The difference between each comparison is recorded as the error value e. A new error value is obtained after each comparison at time step t. After obtaining the error value of each segment of data. If the error value of the previous moment Greater than the error value at the next time step ,but and The values ​​remain unchanged; if the error value at the previous moment remains unchanged... Less than the next time step The error value, then The value remains unchanged, and... Set the value to Calculate the cumulative error at all times. ; In step S500, the detection threshold is determined by calculating the error distribution of the multi-sensor time-domain data and the reconstructed multi-sensor time-domain data, and using the extreme value of the error distribution as the detection threshold. At each time step, the decoder predicts the probability distribution of the output sampled data. The distribution is obtained by using a softmax transform and optimized by calculating the cross-entropy loss function as shown below: ， in, This represents the value after encoding and after softmax transformation. This represents the normalized encoding value of the i-th multi-sensor sampled data, and n represents the number of multi-sensor sampled data. Then, set the epoch of the autoencoder to 100 and the batch size to 20, train the autoencoder network to minimize the cross-entropy loss function, save the autoencoder network, and compare the loss function value with the encoded value. Statistical analysis was performed, and the frequency distribution plots of both were calculated. The maximum value of the error in the frequency distribution plot was used as the threshold for anomaly detection. .

2. The method according to claim 1, wherein, Step S100, the preprocessing of the multi-sensor time-domain data includes the following steps: S101: Perform a sliding window operation on multi-sensor time-domain data to obtain a multi-source training dataset; S102: Normalize the multi-source training dataset to obtain a normalized multi-source training dataset.

3. The method according to claim 2, wherein, In step S200, the preprocessed multi-sensor time-domain data is encoded and reconstructed using an autoencoder, which includes an encoding layer, an intermediate layer, and a decoding layer.

4. The method according to claim 1, wherein, In step S300, the temporal convolutional network includes a dilated dilated convolution module, a random deactivation module, a weight normalization module, a residual connection module, and a ReLU activation function module.

5. The method according to claim 1, wherein, In step S300, the training steps of the temporal convolutional network are as follows: S301: Divide the reconstructed multi-sensor time-domain data into training and testing sets; S302: Train the temporal convolutional network using the training set. When the loss function of the temporal convolutional network converges, the training is complete, and the optimal temporal convolutional network is obtained. If the loss function does not converge, tune the parameters and train again until convergence. S303: Use the optimal temporal convolutional network to make predictions on the test set.

6. An apparatus for an aero-engine anomaly detection method based on encoding temporal convolution according to any one of claims 1 to 5, comprising: The acquisition module is used to acquire time-domain data from multiple sensors of the aero-engine. The reconstruction module is used to encode and reconstruct multi-sensor time-domain data to obtain reconstructed multi-sensor time-domain data. The prediction module is used to make predictions on the reconstructed multi-sensor temporal data using a temporal convolutional network to obtain prediction data. The calculation module is used to calculate the dynamic cumulative error of the multi-sensor time-domain data and prediction data; The detection module is used to compare the dynamic cumulative error with the detection threshold and detect the anomaly of the aero-engine based on the comparison result.

7. The apparatus according to claim 6, wherein, The device also includes a preprocessing module for preprocessing multi-sensor time-domain data.

8. The apparatus according to claim 6, wherein, The device also includes a threshold acquisition module for acquiring a detection threshold.

Images