Robot cable fault detection method based on neural network

Abstract

The invention discloses a robot cable fault detection method based on a neural network, and the method specifically comprises the following steps: S1, analyzing the soft fault and features of an industrial robot control cable, and extracting the isum and Pj features in the fault as output to detect the novelty; s2, proposing an anomaly detection method based on an auto-encoder (AE), training the AAE by using a healthy sample by adopting an adversarial auto-encoder (AAE), and projecting the input and the reconstructed input thereof into the trained AAE to obtain paired activation values; and S3, introducing an along-projection path reconstruction module (RaPP) comparison input and reconstruction input, and comparing hidden activation and corresponding reconstruction in a hidden space to provide support for a decision process. The method has cost effectiveness, does not need additional sensors and signal generators in application, does not need physical parameters and related information of cables, can effectively diagnose health and fault conditions, and is suitable for automatic production lines running in changing and severe environments.

Description

Technical Field

[0001] The present invention relates to the field of industrial robots, and in particular to a robot cable fault detection method based on a neural network. Background Art

[0002] Cables are essential for motor control in industrial applications such as vehicles and factory automation. Monitoring cable health and early fault diagnosis are critical to ensure system stability. Accurate diagnosis reduces downtime and improves system reliability and productivity. In modern manufacturing, a single machine failure can cause the entire production line to shut down because safety signals are managed by process control units. Studies have shown that artificial neural network diagnostic methods based on salient features are reliable, while modern robust control can overcome certain robot failures by compensating for disturbances.

[0003] Cable faults include hard faults (such as open circuits and short circuits) and soft faults (small impedance changes). Harsh environments may gradually transform soft faults (local modifications) into hard faults (local damage). In automated factories, it is very common that machines may continue to operate without signs of faults, while the cables are damaged. Therefore, it is critical to diagnose and handle potential soft faults in a timely manner to prevent further damage to the equipment. In addition, the frequency, amplitude, and phase sequence of industrial robot control cables change according to instructions and are affected by operating conditions. Ensuring the health of the cables is very important to maintain system stability and efficiency.

[0004] Disadvantages and problems of existing technology:

[0005] (1) Additional sensors, signal generators, and signal sensing devices are required: Additional equipment is required to generate a reference signal and inject the reference signal into the target cable. For example, an arbitrary waveform generator is used to create the designed reference signal, a network analyzer is used to inject a sine wave into the target cable, and the reflected waveform is measured. For signal injection, an inductive coupling device is used, while others use a three-way connector, which increases cost effectiveness.

[0006] (2) Physical cable parameters are required, such as cable length information, to design the event signal by adjusting the signal bandwidth, center frequency, and time duration; in addition, before establishing a diagnostic model, the surrounding noise conditions need to be analyzed to recover the reflected incident signal, which makes the process long and laborious.

[0007] (3) A large amount of fault data is required: Existing methods usually build diagnostic models based on fault data sets by analyzing and learning data sets, and use variable resistors and MATLAB simulation to artificially reproduce fault conditions, which is very restrictive and not suitable for industry. Summary of the invention

[0008] The present invention provides a neural network-based robot cable fault detection method, the application of which does not require additional sensors and signal generators, nor does it require physical parameters and related information of the cable, can effectively diagnose health and fault conditions, and is suitable for automated production lines operating in changing and harsh environments.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is: a robot cable fault detection method based on neural network, which specifically includes the following steps:

[0010] Step S1, analyze the soft faults and features of the industrial robot control cable and extract the i sum and P j Features as outputs to test novelty;

[0011] Step S2, an autoencoder (AE)-based anomaly detection method is proposed and an adversarial autoencoder (AAE) is adopted. The AAE is trained using healthy samples, and the input and its reconstructed input are projected into the trained AAE to obtain paired activation values;

[0012] Step S3, the Reconstruction Along Projection Path (RaPP) module is introduced to compare the input and the reconstructed input, and also compare the hidden activations and the corresponding reconstructions in the latent space to support the decision process;

[0013] Step S4, training and testing procedures.

[0014] Preferably, the specific content of step S1 is:

[0015] Step S1.1, establish the soft fault resistance R f and the cross-sectional area of ​​the fault conductor S f Mathematical model, soft fault refers to the fault with small impedance change caused by local change of cable, such as local change of conductor or insulation layer in harsh environment. The relationship can be expressed as (1)-(4):

[0016] S f =αS o , 0<α≤1 (1)

[0017]

[0018] In the formula, S o is the cross-sectional area of ​​the normal cable, S f is the remaining cross-sectional area of ​​the cable after damage; α indicates the degree of damage according to the cross-sectional area of ​​the damaged conductor in the cable; R o is the normal cable resistance, R f is the fault resistance of the faulty cable, where ρ is the resistivity of the conductor material and l is the length of the conductor;

[0019] Step S1.2, determine the value range of α to evaluate the degree of cable damage. The value of α is 1 for a healthy cable and less than 1 for a faulty cable; for an open circuit and a short circuit, α is zero and greater than 1, respectively.

[0020] Step S1.3, the motor type in the robot joint is a three-phase AC servo motor, and the three-phase current is used for fault diagnosis. The three-phase current flowing in the C&I cable can be expressed as equations (5)-(7):

[0021] i a (t) = I ma cos(ωt+φ a ) (5)

[0022] i b (t) = I mb cos(ωt+φ b -120°) (6)

[0023] i c (t) = I mb cos(ωt+φ c +120°) (7)

[0024] Where I ma ,I mb ,I mc is the three-phase current amplitude, φ a ,φ b ,φ c is the corresponding phase angle;

[0025] Step S1.4, determine the ratio of each single-phase current to the total current, which reflects the imbalance of the fault control network. The current ratio P j The definition is shown in formula (8):

[0026]

[0027] Where j represents any phase among a, b and c, P j Compare each single-phase current (SC j ) to the total three-phase current (SSC);

[0028] Step S1.5, taking the vector sum of the three-phase currents at the wire input stage as a feature, due to the leakage current caused by unhealthy wires, the vector sum of the three-phase currents is not zero, and its expression is shown in (9):

[0029] i sum =i a +i b +i c (9)

[0030] For healthy leads, isum The magnitude of is zero, and it is non-zero when a soft fault occurs on the conductor;

[0031] Step S1.6, use negative sequence current (NSC) to describe the imbalance of three-phase current, which is calculated by formula (10):

[0032]

[0033] where i NS is the three-phase current NSC, It is a complex number representing a phase shift of 120 degrees counterclockwise in the complex plane. This phase shift is used to analyze and identify unbalanced conditions in three-phase power systems, where the presence of negative sequence components usually indicates that the system is asymmetric.

[0034] Preferably, the specific content of step S2 is:

[0035] Step S2.1, using the adversarial autoencoder (AAE), A consists of an encoder g and a decoder f;

[0036] The encoder compresses the input and reduces the dimensionality of the input, and the decoder acts as an inverse mapping;

[0037] Step S2.2, the reduced dimensionality reconstruction input is Select the feature P a ,P b ,P c ,I sum As the output of the proposed network, Represents the composition of functions;

[0038] Step S2.3, the encoder outputs the features into a simplified representation, i.e., a latent code. The encoder acts as a latent code generator, trying to mislead the discriminator D z , leading it to believe that the latent codes come from a normal distribution;

[0039] Step S2.4, use D z Distinguish whether a given sample is generated by the latent code of the AAE or a random vector sample from a normal distribution, where the latent distribution is similar to the prior distribution during training.

[0040] Preferably, the specific content of step S3 is:

[0041] Step S3.1, since the corresponding hidden reconstruction is equivalent to the hidden activation of the reconstructed input, the reconstruction is fed into the trained AAE to indirectly calculate the hidden reconstruction; is fed into g to compute the hidden reconstruction, where is the reconstructed input of the trained AAE, whose input x is based on given;

[0042] Step S3.2, set the number of hidden layers in g and f, where and The calculation of g is shown in formula (11):

[0043]

[0044] Step S3.3, x and Input A, and then get their hidden representation pair As shown in equations (12) and (13):

[0045] h i (x) = g :i (x) (12)

[0046]

[0047] Step S3.4, using two hidden activations To obtain the novelty score, RaPP uses the normalized path-by-path aggregation (s NAP ) The calculation score of the indicator is shown in formula (14):

[0048]

[0049] In the formula X is the given training set, D is a matrix, for x i ∈x, whose i-th row corresponds to d(x i );

[0050] Step S3.5, in order to normalize the distance, calculate This metric takes into account the properties of the latent space and the distance distribution of the pairs in equations (12) and (13). This process helps capture the distribution characteristics of the data in the latent space, thereby more accurately evaluating the novelty score of the sample.

[0051] Preferably, the specific contents of the steps are:

[0052] Step S4.1, train the AAE on the normal data sample, then load the test sample x into the trained AAE to generate the reconstructed input A(x) so that the difference between the input x and A(x) is minimized; the novelty of the test sample x is measured by the reconstruction formula (15):

[0053] ε=||xA(x)||2 (15)

[0054] When the error ε(x) increases, the test sample x can be identified as novel, which means that x is farther away from the manifold of AAE;

[0055] Step S4.2, using RaPP technology to obtain novelty score s NAP , the area under the receiver operating characteristic line (AUROC) was used to assess diagnostic accuracy;

[0056] Step S4.3, reconstruct the error ε(x) and s NAP List them together and compare the results;

[0057] In step S4.4, the Adam optimizer stochastic gradient descent method is used for a mini-batch of size 20; binary cross entropy is used as the loss function, the sum is reduced, and the learning rate is set to 0.001;

[0058] In step S4.5, AAE adopts a symmetrical structure with three layers of encoder and decoder. The hidden state size of the encoder and decoder is 30, and each layer uses Leaky-ReLU activation.

[0059] The beneficial effects of adopting the above technical solution are:

[0060] 1. The present invention is cost-effective: its application does not require additional sensors and signal generators, nor does it require the physical parameters and related information of the cable. It can effectively diagnose health and fault conditions and is suitable for automated production lines operating in changing and harsh environments.

[0061] 2. The present invention has strong applicability and can solve the blind spot problem of too short cables: it only utilizes the three-phase current passing through the control cable and is applicable to various types of control cables without being limited by the cable conductor thickness, insulation material and other cable parameters. The blind spot problem is caused by the fact that too short cables may cause overlap of incident and reflected signals when the signal bandwidth and measuring instruments are incorrectly selected. The proposed fault diagnosis is applicable to very short cables because the transmitted signal is used for diagnosis.

[0062] 3. The present invention is versatile and does not require a large amount of fault data: it is based on an unsupervised learning model and only requires normal operation data by assuming unimodal normality; therefore, it is easily applicable to industry, while other methods are usually based on fault data sets to build diagnostic models by analyzing and learning data sets, and using variable resistors and MATLAB simulation to artificially reproduce fault conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is the logic control diagram of the present invention;

[0064] Figure 2 is an equivalent circuit model diagram of a faulty cable of the present invention;

[0065] Figure 3 The overall structure diagram of the diagnostic system of the present invention;

[0066] Figure 4 A graphical illustration of the AAE architecture of the present invention;

[0067] Figure 5 This is a data diagram of the degree of damage to the wire and cable of the present invention. DETAILED DESCRIPTION

[0068] The specific implementation methods of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, with the aim of helping those skilled in the art to have a more complete, accurate and in-depth understanding of the concept and technical solution of the present invention and facilitating its implementation.

[0069] like Figures 1 to 5 As shown, the present invention is a neural network-based robot cable fault detection method, the application of which does not require additional sensors and signal generators, nor does it require physical parameters and related information of the cable. It can effectively diagnose health and fault conditions and is suitable for automated production lines operating in changing and harsh environments.

[0070] Specifically, Figures 1 to 5 As shown, the specific steps include:

[0071] Step S1, analyze the soft faults and features of the industrial robot control cable and extract the i sum and P j Features as outputs to test novelty;

[0072] Step S2, an autoencoder (AE)-based anomaly detection method is proposed and an adversarial autoencoder (AAE) is adopted. The AAE is trained using healthy samples, and the input and its reconstructed input are projected into the trained AAE to obtain paired activation values;

[0073] Step S3, the Reconstruction Along Projection Path (RaPP) module is introduced to compare the input and the reconstructed input, and also compare the hidden activations and the corresponding reconstructions in the latent space to support the decision process;

[0074] Step S4, training and testing procedures.

[0075] The specific content of step S1 is:

[0076] Step S1.1, establish the soft fault resistance R f and the cross-sectional area of ​​the fault conductor S f Mathematical model, soft fault refers to the fault with small impedance change caused by local change of cable, such as local change of conductor or insulation layer in harsh environment. The relationship can be expressed as (1)-(4):

[0077] S f =αS o , 0<α≤1 (1)

[0078]

[0079] In the formula, S o is the cross-sectional area of ​​the normal cable, S f is the remaining cross-sectional area of ​​the cable after damage; α indicates the degree of damage according to the cross-sectional area of ​​the damaged conductor in the cable; R o is the normal cable resistance, R f is the fault resistance of the faulty cable, where ρ is the resistivity of the conductor material and l is the length of the conductor;

[0080] Step S1.2, determining the value range of α to evaluate the degree of cable damage, the value of α is 1 for a healthy cable and less than 1 for a faulty cable; for an open circuit and a short circuit, α is zero and greater than 1, respectively;

[0081] In this study, when the current flows through the fault area, it will be affected by the damaged conductor and insulation material, resulting in the existence of fault resistance (Rf) and leakage current through the damaged insulation layer, which in turn causes an imbalance in the power circuit. Given that the motor used in the robot joint is a three-phase AC servo motor, the connecting cable between its controller and the motor constitutes a three-phase power supply network. Therefore, this study will use a three-phase current analysis method to diagnose the fault in order to accurately identify and locate the fault point. Most industrial controllers measure and use current to control objects, so current information can be easily obtained in most cases without the need for additional equipment;

[0082] Step S1.3, the motor type in the robot joint is a three-phase AC servo motor, and the three-phase current is used for fault diagnosis. The three-phase current flowing in the C&I cable can be expressed as equations (5)-(7):

[0083] i a (t) = I ma cos(ωt+φ a ) (5)

[0084] i b (t) = I mb cos(ωt+φ b -120°) (6)

[0085] i c (t) = I mb cos(ωt+φ c +120°) (7)

[0086] Where I ma ,I mb ,I mc is the three-phase current amplitude, φ a ,φ b ,φ cis the corresponding phase angle;

[0087] Step S1.4, determine the ratio of each single-phase current to the total current, which reflects the imbalance of the fault control network. The current ratio P j The definition is shown in formula (8):

[0088]

[0089] Where j represents any phase among a, b and c, P j Compare each single-phase current (SC j ) to the total three-phase current (SSC);

[0090] Step S1.5, taking the vector sum of the three-phase currents at the wire input stage as a feature, due to the leakage current caused by unhealthy wires, the vector sum of the three-phase currents is not zero, and its expression is shown in (9):

[0091] i sum =i a +i b +i c (9)

[0092] For healthy leads, i sum The magnitude of is zero, and it is non-zero when a soft fault occurs on the conductor;

[0093] Step S1.6, use negative sequence current (NSC) to describe the imbalance of three-phase current, which is calculated by formula (10):

[0094]

[0095] where i NS is the three-phase current NSC, It is a complex number representing a phase shift of 120 degrees counterclockwise in the complex plane. This phase shift is used to analyze and identify unbalanced conditions in three-phase power systems, where the presence of negative sequence components usually indicates that the system is asymmetric.

[0096] The specific content of step S2 is:

[0097] Step S2.1, using the adversarial autoencoder (AAE), A consists of an encoder g and a decoder f;

[0098] The encoder compresses the input and reduces the dimensionality of the input, and the decoder acts as an inverse mapping;

[0099] Step S2.2, the reduced dimensionality reconstruction input is Select the feature P a ,P b ,P c ,I sumAs the output of the proposed network, Represents the composition of functions;

[0100] Step S2.3, the encoder outputs the features into a simplified representation, i.e., a latent code. The encoder acts as a latent code generator, trying to mislead the discriminator D z , leading it to believe that the latent codes come from a normal distribution;

[0101] Step S2.4, use D z Distinguish whether a given sample is generated by the latent code of the AAE or a random vector sample from a normal distribution, where the latent code distribution is similar to the prior distribution during training.

[0102] The specific content of step S3 is:

[0103] Step S3.1, since the corresponding hidden reconstruction is equivalent to the hidden activation of the reconstructed input, the reconstruction is fed into the trained AAE to indirectly calculate the hidden reconstruction; is fed into g to compute the hidden reconstruction, where is the reconstructed input of the trained AAE, whose input x is based on Given.

[0104] Step S3.2, set the number of hidden layers in g and f, where and The calculation of g is shown in formula (11):

[0105]

[0106] Step S3.3, x and Input A, and then get their hidden representation pair As shown in equations (12) and (13):

[0107] h i (x) = g :i (x) (12)

[0108]

[0109] Step S3.4, using two hidden activation values To obtain the novelty score, RaPP uses the normalized path-by-path aggregation (s NAP ) The calculation score of the indicator is shown in formula (14):

[0110]

[0111] In the formula X is the given training set, D is a matrix, for x i ∈x, whose i-th row corresponds to d(xi );

[0112] Step S3.5, in order to normalize the distance, calculate This metric takes into account the properties of the latent space and the distance distribution of the pairs in equations (12) and (13). This process helps capture the distribution characteristics of the data in the latent space, thereby more accurately evaluating the novelty score of the sample.

[0113] The specific content of step S4 is:

[0114] Step S4.1, train the AAE on the normal data sample, then load the test sample x into the trained AAE to generate the reconstructed input A(x) so that the difference between the input x and A(x) is minimized; the novelty of the test sample x is measured by the reconstruction formula (15):

[0115] ε=||xA(x)||2 (15)

[0116] When the error ε(x) increases, the test sample x can be identified as novel, which means that x is farther away from the manifold of AAE;

[0117] Step S4.2, using RaPP technology to obtain novelty score s NAP , the area under the receiver operating characteristic line (AUROC) was used to assess diagnostic accuracy;

[0118] Step S4.3, reconstruct the error ε(x) and s NAP List them together and compare the results;

[0119] Step S4.4, using Adam optimizer stochastic gradient descent for a mini-batch size of 20; using binary cross entropy as the loss function, reducing the sum, and setting the learning rate to 0.001;

[0120] In step S4.5, AAE adopts a symmetrical structure with three layers of encoder and decoder. The hidden state size of the encoder and decoder is 30, and each layer uses Leaky-ReLU activation.

[0121] The present invention is described above by way of example in conjunction with the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention; or the above-mentioned concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A robot cable fault detection method based on neural network, characterized in that: The specific steps include: Step S1, analyze the soft faults and features of the industrial robot control cable and extract the i sum and P j Features as outputs to test novelty; Step S2, an anomaly detection method based on autoencoder (AE) is proposed and an adversarial autoencoder (AAE) is adopted. The AAE is trained using healthy samples, and the input and its reconstructed input are projected into the trained AAE to obtain paired activation values; Step S3, the Reconstruction Along Projection Path (RaPP) module is introduced to compare the input and the reconstructed input, and also compare the hidden activations and the corresponding reconstructions in the latent space to support the decision process; Step S4, training and testing procedures.

2. The robot cable fault detection method based on neural network according to claim 1 is characterized in that: The specific content of step S1 is: Step S1.1, establish the soft fault resistance R f and the cross-sectional area of ​​the fault conductor S f Mathematical model, soft fault refers to the fault with small impedance change caused by local change of cable, such as local change of conductor or insulation layer in harsh environment. The relationship can be expressed as (1)-(4): S f =αS o ,0<α≤1 (1) In the formula, S o is the cross-sectional area of ​​the normal cable, S f is the remaining cross-sectional area of ​​the cable after damage; α indicates the degree of damage according to the cross-sectional area of ​​the damaged conductor in the cable; R o is the normal cable resistance, R f is the fault resistance of the faulty cable, where ρ is the resistivity of the conductor material and l is the length of the conductor; Step S1.2, determining the value range of α to evaluate the degree of cable damage, the value of α is 1 for a healthy cable and less than 1 for a faulty cable; for an open circuit and a short circuit, α is zero and greater than 1, respectively; Step S1.3, the motor type in the robot joint is a three-phase AC servo motor, and the three-phase current is used for fault diagnosis. The three-phase current flowing in the C&I cable can be expressed as formula (5)-(7): i a (t)=I ma cos(ωt+φ a ) (5) i b (t)=I mb cos(ωt+φ b -120°) (6) i c (t)=I mb cos(ωt+φ c +120°) (7) Where I ma ,I mb ,I mc is the three-phase current amplitude, φ a ,φ b ,φ c is the corresponding phase angle; Step S1.4, determine the ratio of each single-phase current to the total current, which reflects the imbalance of the fault control network. The current ratio P j The definition is shown in formula (8): Where j represents any phase among a, b and c, P j Compare each single-phase current (SC j ) to the total three-phase current (SSC); Step S1.5, taking the vector sum of the three-phase current of the wire input stage as a feature, due to the leakage current caused by unhealthy wires, the vector sum of the three-phase current is not zero, and its expression is shown in (9): i sum =i a +i b +i c (9) For healthy leads, i sum The magnitude of is zero, and it is non-zero when a soft fault occurs on the conductor; Step S1.6, use negative sequence current (NSC) to describe the imbalance of three-phase current, which is calculated by formula (10): where i NS is the three-phase current NSC, It is a complex number representing a phase shift of 120 degrees counterclockwise in the complex plane. This phase shift is used to analyze and identify unbalanced conditions in three-phase power systems, where the presence of negative sequence components usually indicates that the system is asymmetric.

3. The robot cable fault detection method based on neural network according to claim 1, characterized in that: The specific content of step S2 is: Step S2.1, using the adversarial autoencoder (AAE), A consists of an encoder g and a decoder f; The encoder compresses the input and reduces the dimensionality of the input, and the decoder acts as an inverse mapping; Step S2.2, the reduced dimensionality reconstruction input is Select the feature P a ,P b ,P c ,I sum As the output of the proposed network, Represents the composition of functions; Step S2.3, the encoder outputs the features into a simplified representation, i.e., a latent code. The encoder acts as a latent code generator, trying to mislead the discriminator D z , leading it to believe that the latent codes come from a normal distribution; Step S2.4, use D z Distinguish whether a given sample is generated by the latent code of the AAE or a random vector sample from a normal distribution, where the latent code distribution is similar to the prior distribution during training.

4. The robot cable fault detection method based on neural network according to claim 1, characterized in that: The specific content of step S3 is: Step S3.1, since the corresponding hidden reconstruction is equivalent to the hidden activation of the reconstructed input, the reconstruction is fed into the trained AAE to indirectly calculate the hidden reconstruction; is fed into g to compute the hidden reconstruction, where is the reconstructed input of the trained AAE, whose input x is based on given; Step S3.2, set the number of hidden layers in g and f, where and The calculation of g is shown in formula (11): Step S3.3, x and Input A, and then get their hidden representation pair As shown in equations (12) and (13): h i (x)=g :i (x) (12) Step S3.4, using two hidden activation values To obtain the novelty score, RaPP uses the normalized path-wise aggregation (s NAP ) The calculation score of the indicator is shown in formula (14): In the formula X is the given training set, D is a matrix, for x i ∈x, whose i-th row corresponds to d(x i ); Step S3.5, in order to normalize the distance, calculate This metric takes into account the properties of the latent space and the distance distribution of the pairs in equations (12) and (13). This process helps capture the distribution characteristics of the data in the latent space, thereby more accurately evaluating the novelty score of the sample.

5. The robot cable fault detection method based on neural network according to claim 1, characterized in that: The specific content of step S4 is: Step S4.1, train the AAE on the normal data sample, then load the test sample x into the trained AAE to generate the reconstructed input A(x) so that the difference between the input x and A(x) is minimized; the novelty of the test sample x is measured by the reconstruction formula (15): ε=||xA(x)||2 (15) When the error ε(x) increases, the test sample x can be identified as novel, which means that x is farther away from the manifold of AAE; Step S4.2, using RaPP technology to obtain novelty score s NAP , the area under the receiver operating characteristic line (AUROC) was used to assess diagnostic accuracy; Step S4.3, reconstruct the error ε(x) and s NAP List them together and compare the results; In step S4.4, the Adam optimizer stochastic gradient descent method is used for a mini-batch of size 20; binary cross entropy is used as the loss function, the sum is reduced, and the learning rate is set to 0.001; In step S4.5, AAE adopts a symmetrical structure with three layers of encoder and decoder. The hidden state size of the encoder and decoder is 30, and each layer uses Leaky-ReLU activation.

Images