A monitoring method for a smart track vehicle articulation system

Abstract

A kind of monitoring method of smart rail car hinging device system, comprising the following steps: first, data acquisition is carried out by test;Second, data is preprocessed;Third, based on the optimization of generative adversarial network, the generator in generative adversarial network is optimized using evolutionary algorithm, and the data after preprocessing is expanded;Fourth, based on adaptive residual capsule network, the expanded data is subjected to deep learning, and the parameters of the model used for monitoring are obtained, specifically referring to: first, the expanded data is divided, then the prominent features in data signal are extracted using adaptive residual, and finally output by capsule;Fifth, install the model after determining parameters to smart rail car for real-time monitoring.The present application realizes real-time monitoring of smart rail car corner and damping arm pressure through angle sensor and pressure sensor of hinging device, can find out problems in time and even avoid component damage caused by long-term vibration, improve the safety of stable operation of hinging device.

Description

Technical Field

[0001] This invention relates to a system monitoring method, specifically to a monitoring method for an intelligent rail transit vehicle articulated device system. Background Technology

[0002] In the field of urban transportation, with the development of the social economy and the improvement of people's living standards, the number of motor vehicles has increased dramatically, leading to a deterioration of the urban traffic environment, a decrease in road network efficiency, and slow traffic on main and auxiliary roads. Intelligent rail vehicles (IRVs) combine the advantages of modern trams and buses. IRVs employ virtual trajectory following control technology, using onboard sensors to identify virtual track lines on the road surface, achieving track-like operation. The articulated joint device of the IRV allows it to maintain a small turning radius while keeping the train length consistent, solving the turning difficulties caused by its extra-long body. However, prolonged use of the articulated joint device can lead to damping arm failure. If not repaired or replaced in time, this can cause excessive turning, potentially resulting in serious traffic accidents at high speeds. Current technologies do not address monitoring techniques for the articulated joint device of IRVs, nor do they consider real-time fault warning and diagnostic solutions for this component. Summary of the Invention

[0003] This invention addresses the lack of monitoring technology for current intelligent rail transit (IRT) articulation device systems by proposing a monitoring method for IRT articulation device systems. This method enables real-time monitoring of the operational status of the IRT articulation device during service, thereby improving the safety and stability of the articulation device.

[0004] The technical means adopted by this invention to solve the above problems is as follows: a monitoring method for an intelligent rail vehicle articulated device system, comprising the following steps: First, data acquisition through experiments; second, data preprocessing; third, augmenting the preprocessed data by using an evolutionary algorithm to optimize the generator in the optimized generative adversarial network; fourth, performing deep learning on the augmented data based on an adaptive residual capsule network to obtain the parameters of the model used for monitoring, specifically: first dividing the augmented data, then using adaptive residuals to extract prominent features from the data signal, and finally outputting through capsules; fifth, installing the model with determined parameters onto the intelligent rail vehicle for real-time monitoring.

[0005] Furthermore, the first step of the test includes bench testing and vehicle testing. Bench testing involves collecting pressure signals from the sensors of the articulation device and converting them into digital pressure signals. Vehicle testing involves installing the articulation device on a vehicle and conducting multiple rounds of real-world driving tests under the upper and lower load limits during the service life of the articulation device, collecting pressure signals from the sensors of the articulation device and converting them into digital pressure signals.

[0006] Furthermore, in the second step, during data preprocessing, abnormal abrupt signals are detected and removed to obtain the first hydraulic digital signal and the first vibration signal.

[0007] Furthermore, the specific preprocessing method in the second step is as follows: First, the digital pressure signal and angle signal are normalized to map the data to the range of 0 to 1; then, abnormal mutation values ​​are detected and removed from the normalized data to reduce computational complexity and reduce the impact of abnormal mutation values ​​on fault warning and diagnosis; finally, the data are oversampled to enhance the data and highlight its features.

[0008] Furthermore, in the third step, the generative adversarial network consists of a generator G and a discriminator D. Real data and generated data are input into the generator G to obtain the probability of the data being real or fake. A random noise vector z is input to obtain fake data, ultimately generating highly realistic fake data. The loss function is:

[0009]

[0010] In the formula: E represents the expected value; P d With P z Let x be the distribution of the real data and z be the random noise.

[0011] During network training, the model parameters are updated using the Adam algorithm, iterating until the distribution of fake samples equals the distribution of real samples, reaching Nash equilibrium and achieving optimal training. The loss function of the optimization algorithm is:

[0012]

[0013] After optimization, the authenticity of the generated data is first evaluated. The data generated each time is input into the discriminator D to calculate the sample diversity.

[0014]

[0015] in, Representing the gradient operator, the specific loss function is:

[0016]

[0017] Where l is the loss function for true and false labels; l gc The loss function for class labels; L max This is the generator loss function.

[0018] Furthermore, in the fourth step, the adaptive residuals calculate the optimal coefficients for each channel, thereby adaptively amplifying sensitive features and suppressing irrelevant features. The Sigmoid function effectively maintains the range of output values, as shown below:

[0019]

[0020] x i σ represents the feature of the i-th channel; i Represents the adaptive coefficient of the i-th channel;

[0021] Then the obtained adaptive coefficient σ i Multiply by the output feature α of another branch in the adaptive module:

[0022] y = σ i ×α.

[0023] In this way, the network can retain effective information by identifying the optimal coefficients of each channel, using adaptive methods and suppressing features, based on the importance of the features.

[0024] Furthermore, in the fourth step, the capsule network includes a primary capsule layer and a digital capsule layer. The digital capsule layer includes a dynamic routing algorithm. The main idea of ​​the capsule network is divided into three stages. The first stage is completed in the primary capsule layer, where the prediction vector U is calculated. j|i The calculation formula is as follows:

[0025] U j|i =W ij U i ,

[0026] U i W is the output of the i-th neuron in the previous convolutional network layer. ij To be related to the neuron's output U i Multiply the weight matrix.

[0027] Furthermore, the second stage modifies the prediction vector U obtained in the first stage. j|i Multiply by c ij After performing a weighted summation, we finally obtain an output vector S. j The calculation formula is as follows:

[0028] S j =∑ i c ij *U j|i

[0029] c ij It is a coupling coefficient, determined by a dynamic routing algorithm. Its function is to allow the input neuron to autonomously choose the best path to transmit to the next layer of neurons.

[0030] Furthermore, the coupling coefficient c ij By using the initial vector b ij The initial vector b is obtained by performing Softmax normalization. ijFrom the prediction vector U j|i and output vector V j The iterative update is performed, and the calculation formula is as follows:

[0031]

[0032] b ij =b ij +V j *U j|i .

[0033] Furthermore, the third stage will use the intermediate vector S obtained in the second stage. j The output vector V of the routing capsule layer is obtained by performing a nonlinear transformation using the nonlinear activation vector compression function Squashing. j The advantage of this activation function is that it can both maintain the direction of the initial vector and compress the magnitude of the input vector to the range (0,1). The calculation formula is as follows:

[0034]

[0035] in, Represents the total input vector S j The reduction factor; Represents the total input vector S j The unit vector.

[0036] The beneficial effects of this invention are:

[0037] 1. This invention utilizes angle and pressure sensors in the articulation device to achieve real-time monitoring of the turning angle and damping arm pressure of the intelligent rail vehicle. This allows for timely detection of problems and even prevention of component damage caused by long-term vibration, improving the safety of the articulation device's stable operation. It also solves the problems of early warning of impending failure during the service life of the articulation device and fault diagnosis after failure, avoiding risks such as component damage caused by failure to repair or replace in a timely manner. Furthermore, the deep learning algorithm of this invention is simple, practical, and suitable for engineering applications.

[0038] 2. In the case of insufficient original data, this invention expands the dataset based on a small amount of experimental data, which can reduce experimental costs and improve efficiency. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the hinge device system structure in Embodiment 1;

[0040] Figure 2 This is a schematic diagram of the monitoring system for the articulated device in Example 1;

[0041] Figure 3 This is a flowchart of the monitoring process for Example 1;

[0042] Figure 4 This is a schematic diagram of the adaptive residual capsule network in Example 1;

[0043] Figure 5 This is a schematic diagram illustrating the construction principle of the reference database in Example 1;

[0044] Figure 6 This is a schematic diagram of the failure warning and diagnosis process in Example 1;

[0045] In the diagram: 1. First damping arm, 2. Second damping arm, 3. First pressure sensor, 4. Second pressure sensor, 5. First angle sensor, 6. Second angle sensor, 7. Third angle sensor, 8. Hinge device, 9. Controller, 10. Data acquisition module, 11. Data analysis module, 12. Fault monitoring, early warning and diagnosis module, 13. Reference data test platform, 14. Host computer. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings. The drawings are for illustrative purposes only, representing schematic diagrams rather than actual physical objects, and should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0047] Example 1

[0048] A monitoring method for an intelligent rail transit vehicle articulated device system, such as Figure 1 As shown, the hinge device 8 has a controller 9 inside, which is connected to a first angle sensor 5, a second angle sensor 6, and a third angle sensor 7. The hinge device 8 has a first damping arm 1 and a second damping arm 2 outside, with a first pressure sensor 3 and a second pressure sensor 4 respectively mounted on the first damping arm 1 and the second damping arm 2. When the intelligent rail vehicle turns, the first angle sensor 5, the second angle sensor 6, and the third angle sensor 7 monitor the angle signal at this time, thereby limiting the turning radius of the intelligent rail vehicle through the damping of the first damping arm 1 and the second damping arm 2. The minimum number of pressure sensors is two, and the minimum number of angle sensors is three; of course, more can be used, as long as the installation space requirements are met. Figure 1 and Figure 2 As shown, the entire articulated device system also includes a data acquisition module 10, a data analysis module 11, a fault monitoring, early warning and diagnosis module 12, a reference data test platform 13 and a host computer 14.

[0049] The first pressure sensor 3, the second pressure sensor 4, the first angle sensor 5, the second angle sensor 6, and the third angle sensor 7 transmit the real-time measured pressure and angle signals to the data acquisition module 10, and then transmit the data to the data analysis module 11. The data analysis module 11 converts the signals to obtain the unprocessed original pressure and angle signals.

[0050] The data analysis module 11 transmits the converted signal to the fault monitoring, early warning, and diagnosis module 12. The fault monitoring, early warning, and diagnosis module 12 first normalizes the original signal to eliminate outliers, and then performs data augmentation on the original data using overlapping sampling. The augmented dataset is then fed into a trained adaptive residual capsule network for fault early warning and diagnosis. Finally, the real-time status of the articulated device system is transmitted to the host computer 14 for display. When the host computer 14 receives failure information, it provides early warning and fault diagnosis for the current failure phenomenon.

[0051] The reference data testing platform 13 is used to establish a reference database. The established reference database is then transmitted to the monitoring, early warning, and diagnosis module 12 to train the adaptive residual capsule network multiple times, saving the optimal training model parameters. When signals from the articulated device 8 during its service life are collected and transmitted to the monitoring, early warning, and diagnosis module 12, the trained model can be directly used for real-time status monitoring, early warning, and diagnosis. Figure 5 This is a schematic diagram illustrating the construction principle of a reference database.

[0052] Specifically, such as 1- Figure 4 As shown, the monitoring process of the articulated device system in this embodiment includes the following steps: First, data is acquired through experiments, including bench tests and vehicle tests. During the bench test, the data acquisition module 10 acquires the pressure electrical signals from the first pressure sensor 3 and the second pressure sensor 4, and the data analysis module 11 converts the pressure electrical signals into digital pressure signals through a certain formula; then, the signals are normalized and preprocessed, and data is enhanced before data storage. During the vehicle test, after the articulated device 8 is installed in the vehicle, multiple rounds of actual vehicle operation tests are conducted under the upper and lower load limits during the service life of the articulated device 8. The data acquisition module 10 acquires the pressure electrical signals from the first pressure sensor 3 and the second pressure sensor 4, as well as the angle signals from the first angle sensor 5, the second angle sensor 6, and the third angle sensor 7, and the data analysis module 11 converts the pressure electrical signals into digital pressure signals through a certain formula; then, the digital pressure signals and angle signals are normalized and preprocessed, and data is enhanced before data storage. Finally, the bench test data and the vehicle test data are used to construct a reference database to train an adaptive residual capsule network.

[0053] The second step is to preprocess the data, detect and remove abnormal mutation signals, and obtain the first hydraulic digital signal and the first vibration signal. The specific process is as follows: (1) Normalize the pressure digital signal and the angle signal to map the data to the range of 0 to 1; (2) Detect and remove abnormal mutation values ​​from the normalized data to reduce the computational complexity and reduce the impact of abnormal mutation values ​​on fault warning and diagnosis; (3) Perform overlay sampling on the normalized data to enhance the data and highlight the data features.

[0054] The third step involves optimizing the generator in the generative adversarial network (GAN) using an evolutionary algorithm. This optimizes the generator to generate new samples with a distribution similar to the original samples under different loss functions, thereby expanding the dataset.

[0055] A Generative Adversarial Network (GAN) consists of a generator G and a discriminator D. Real data and generated data are input into the generator G to obtain the probabilities of the data being real or fake. A random noise vector z is input to generate fake data. The network's ultimate goal is to generate deceptively realistic fake data. The loss function is:

[0056]

[0057] In the expression: E represents the expected value; P d With P z Let x be the distribution of the real data and z be the random noise.

[0058] During network training, the model parameters are updated using the Adam algorithm, iterating continuously until the distribution of fake samples equals the distribution of real samples, reaching Nash equilibrium, and the training reaches its optimal state.

[0059] The optimized generative adversarial network (GAN) incorporates a Softmax classifier to enable classification capabilities, and an evolutionary algorithm is used to optimize the network's generator. The loss function of the optimization algorithm is:

[0060]

[0061] After optimization, the authenticity of the generated data is first evaluated. The data generated each time is input into the discriminator D to calculate the sample diversity.

[0062]

[0063] in, This represents the gradient operator.

[0064] The diversity of generated data is calculated using the gradient of the discriminator D. A high diversity score corresponds to a small gradient of the discriminator D, effectively preventing problems such as pattern collapse. The specific loss function is:

[0065]

[0066] Where l is the loss function for true and false labels; l gc The loss function for class labels; L max This is the generator loss function.

[0067] The fourth step involves using an adaptive residual capsule network to calculate the optimal coefficients for each channel, adaptively amplifying sensitive features and suppressing irrelevant features. Specifically, this means first dividing the expanded data, then using adaptive residuals to extract prominent features from the data signal, and finally outputting the data through capsules.

[0068] In this step, the adaptive model obtains the optimal coefficients for each channel feature, thereby adaptively amplifying sensitive features and suppressing irrelevant features. In the adaptive model, the Sigmoid function effectively maintains the range of output values. The adaptive output can be expressed as:

[0069]

[0070] In the formula: x i σ represents the feature of the i-th channel; i Let represent the adaptive coefficient of the i-th channel. Then, the obtained adaptive coefficient σ i Multiply by the output feature 'a' of another branch in the adaptive module:

[0071] y = σ i ×α.

[0072] In this way, the network can further retain effective information by identifying the optimal coefficients for each channel, using adaptive methods and suppressing features, based on the importance of the features.

[0073] Capsule networks consist of a primary capsule layer and a digital capsule layer. The digital capsule layer includes a dynamic routing algorithm. The main idea of ​​capsule networks is that the operation is divided into three stages. The first stage is completed in the primary capsule layer, where the prediction vector U is calculated. j|i The calculation formula is as follows:

[0074]

[0075] Among them, U i W is the output of the i-th neuron in the previous convolutional network layer. ij To be related to the neuron's output U i Multiplying the weight matrix; U j|i This is the prediction vector.

[0076] The second stage involves the prediction vector U obtained from the previous stage. j|i Multiply by c ijAfter performing a weighted summation, we finally obtain an output vector S. j The calculation formula is as follows:

[0077] S j =∑ i c ij *U j|i

[0078] c ij It is a coupling coefficient, determined by a dynamic routing algorithm. Its function is to allow the input neuron to autonomously choose the best path to transmit to the next layer of neurons.

[0079] The core idea at this stage is the dynamic routing algorithm, and the core task of dynamic routing is to find the optimal coupling coefficient c. ij c ij The second operational stage of the capsule network plays a crucial role, determining the connection between the current capsule layer and the previous capsule layer. This is achieved by adjusting the initial vector b. ij The initial vector is obtained by performing Softmax normalization, and the initial vector is obtained by the prediction vector U. j|i and output vector V j The iterative update is performed, and the calculation formula is as follows:

[0080]

[0081] b ij =b ij +V j *U j|i

[0082] In each forward propagation, b ij Initialized to 0, the coupling coefficient c is calculated. ij The initial value is then used to calculate V through the forward propagation of the network. j Each forward propagation updates b. ij The value of c is then updated. ij The value of S is then updated further. j The value of V ultimately changed j The value of is determined through iterative updates step by step, finding an optimal set of coupling coefficients.

[0083] The third stage will use the intermediate vector S obtained in the second stage. j The output vector V of the routing capsule layer is obtained by nonlinear transformation using a novel nonlinear activation vector compression function, Squashing. j The advantage of this activation function is that it can both maintain the direction of the initial vector and compress the magnitude of the input vector to the range (0,1). Its calculation formula is as follows:

[0084]

[0085] in, Represents the total input vector S j The reduction factor; Represents the total input vector S j The unit vector.

[0086] In the above process, the original data is first processed into multiple samples of the same length. The processed samples are then divided into training and test sets in an 8:2 ratio. The training set is first used to train the network model, and then the test set is used for validation. The parameters of the model with the best fault warning and diagnosis effect are determined.

[0087] Fifth step, as Figure 5 As shown, the model with determined parameters is installed on the intelligent rail vehicle for real-time monitoring. In this embodiment, the different damping conditions of the articulated device damping arm are divided into: 1% deviation of the measured damping value from the lower limit (warning), 2% deviation of the measured damping value from the lower limit (warning), less than 3% deviation of the measured damping value from the lower limit (warning), less than 4% deviation of the measured damping value from the lower limit (warning), and greater than 5% deviation of the measured damping value from the lower limit (warning). The corresponding warning thresholds for vibration amplitude are: threshold 1, threshold 2, threshold 3, threshold 4, and threshold 5. Failure modes include: sensor failure warning and measured damping value deviation of 1% to 5% warning. Figure 6As shown, the data analysis module 10 uses the pressure signals from the first pressure sensor 3 and the second pressure sensor 4 of the damping arm, and the angle signals from the first angle sensor 5, the second angle sensor 6, and the third angle sensor 7 to monitor the working status of the articulation device 8 during its service life. When the real-time signals of the articulation device 8 during its service life are input into the network model, the module determines whether its working status is normal based on the signal characteristics of the input signals and highlights the working status through indicator lights. If all tensor matrix values ​​output by the network model are 0 or constant, the model outputs a sensor failure warning, and simultaneously displays its current status on the host computer 14, at which point a blue indicator light illuminates. Gradient thresholds are set in the network model; different thresholds represent different working states of the articulation device. If the RMS value of the tensor matrix output by the network model is greater than threshold 1 and less than threshold 2, the model outputs a damping failure warning 1 (measured damping value deviates from the lower limit by 1%), and simultaneously displays its current status on the host computer 14. At this time, a yellow indicator light flashes for 15 seconds. If the RMS value of the tensor matrix output by the network model is greater than threshold 2 and less than threshold 3, the model outputs a damping failure warning 2 (measured damping value deviates from the lower limit by 2%), and simultaneously displays its current status on the host computer 14. At this time, a yellow indicator light flashes for 10 seconds. If the RMS value of the tensor matrix output by the network model is greater than threshold 3 and less than threshold 4, the model outputs a damping failure warning. 3. If the measured damping value deviates from the lower limit by 3%, a yellow indicator light will flash for 7 seconds, simultaneously displaying the current status on the host computer's display (14). If the RMS value of the tensor matrix output by the network model is greater than threshold 4 but less than threshold 5, the model will output a damping failure warning 4 (measured damping value deviates from the lower limit by 4%), simultaneously displaying the current status on the host computer's display (14). A yellow indicator light will flash for 5 seconds. If the RMS value of the tensor matrix output by the network model is greater than threshold 5, the model will output a damping failure warning 5 (measured damping value deviates from the lower limit by 5%), simultaneously displaying the current status on the host computer's display (14). A red indicator light will flash for 1 second. These failure warnings can be performed simultaneously with fault diagnosis.

[0088] The articulated mechanism works by using a damping arm to limit the turning radius and angle of the intelligent rail vehicle. The angle and pressure signals measured by the sensors provide feedback to each other. If the sensors or the damping ratio fails, the turning radius of the intelligent rail vehicle cannot be effectively controlled, leading to safety hazards and potentially serious traffic accidents. This method, however, uses angle and pressure sensors in the articulated mechanism to monitor the turning angle and damping arm pressure of the intelligent rail vehicle in real time. This allows for timely detection of problems and even prevention of component damage caused by long-term vibration, ensuring the safety and reliability of the intelligent rail vehicle.

[0089] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes or modifications without departing from the spirit and scope of the invention. Therefore, all equivalent technical solutions should also fall within the protection scope of the invention, which should be defined by the claims.

Claims

1. A monitoring method for an articulated rail vehicle system, characterized in that: The process includes the following steps: First, data is collected through experiments; second, the data is preprocessed; third, based on an optimized generative adversarial network (GAN), the generator in the GAN is optimized using an evolutionary algorithm to augment the preprocessed data; fourth, deep learning is performed on the augmented data based on an adaptive residual capsule network to obtain the parameters of the monitoring model, specifically: first, the augmented data is divided, then adaptive residuals are used to extract prominent features from the data signals, and finally, the data is output through capsules; fifth, the model with determined parameters is installed in the intelligent rail vehicle for real-time monitoring. The first step of the test includes bench testing and vehicle testing. Bench testing collects the pressure electrical signals from the sensors of the articulation device and converts the pressure electrical signals into digital pressure signals. Vehicle testing involves installing the articulation device on a vehicle and conducting multiple rounds of actual vehicle operation tests under the upper and lower load limits during the service life of the articulation device, collecting the pressure electrical signals from the sensors of the articulation device and converting the pressure electrical signals into digital pressure signals. In the fourth step, the adaptive residuals calculate the optimal coefficients for each channel, thereby adaptively amplifying sensitive features and suppressing irrelevant features. The Sigmoid function effectively maintains the range of output values, as shown below: ， Representing the Characteristics of each channel; Representing the Adaptive coefficients for each channel; Then the obtained adaptive coefficients Multiply by the output feature α of another branch in the adaptive module: y=σ i ×a.

2. The monitoring method for the intelligent rail vehicle articulation device system as described in claim 1, characterized in that: In the second step, during data preprocessing, abnormal abrupt signals are detected and removed to obtain the first hydraulic digital signal and the first vibration signal.

3. The monitoring method for the intelligent rail vehicle articulation device system as described in claim 2, characterized in that: The specific preprocessing method in the second step is as follows: First, the digital pressure signal and angle signal are normalized to map the data to the range of 0~1; then, abnormal mutation values ​​are detected and removed from the normalized data to reduce computational complexity and reduce the impact of abnormal mutation values ​​on fault warning and diagnosis; finally, the data is oversampled and enhanced to highlight data features.

4. The monitoring method for the intelligent rail vehicle articulation device system as described in claim 1, characterized in that: In the third step, the Generative Adversarial Network (GAN) consists of a generator G and a discriminator D. Real data and generated data are input into the generator G to obtain the probabilities of the data being real or fake. A random noise vector z is input to generate fake data, ultimately producing indistinguishable fake data. The loss function is: ， In the formula: Represents mathematical expectation; and For real data With random noise Distribution; During network training, the model parameters are updated using the Adam algorithm, iterating until the distribution of fake samples equals the distribution of real samples, reaching Nash equilibrium and achieving optimal training. The loss function of the optimization algorithm is: ， After optimization, the authenticity of the generated data is first evaluated, and the data generated each time is input into the discriminator. In this process, the diversity of the sample is calculated: ， in, Representing the gradient operator, the specific loss function is: ， in, The loss function is for true and false labels; The loss function is for the category label; This is the generator loss function.

5. The monitoring method for the intelligent rail vehicle articulated device system as described in claim 1, characterized in that: Capsule networks consist of a primary capsule layer and a digital capsule layer. The digital capsule layer includes a dynamic routing algorithm. The main idea of ​​capsule networks is that the operation is divided into three stages. The first stage is completed in the primary capsule layer, where the prediction vector is calculated. The calculation formula is as follows: ， For the first convolutional layer of the previous layer The output of each neuron; To be related to the output of neurons Multiply the weight matrix.

6. The monitoring method for the intelligent rail vehicle articulated device system as described in claim 5, characterized in that: The second stage provides a prediction vector obtained from the first stage. Multiply The weighted summation of rows yields a final output vector. The formula is as follows: ， It is a coupling coefficient, determined by a dynamic routing algorithm. Its function is to allow the input neuron to autonomously choose the best path to transmit to the next layer of neurons.

7. The monitoring method for the intelligent rail vehicle articulated device system as described in claim 6, characterized in that: Coupling coefficient By using the initial vector The initial vector is obtained by performing Softmax normalization. From the prediction vector and output vector The iterative update is performed, and the calculation formula is as follows: ， 。 8. The monitoring method for the intelligent rail vehicle articulation device system as described in claim 6, characterized in that: The third stage will use the intermediate vector obtained in the second stage. The output vector of the routing capsule layer is obtained by performing a nonlinear transformation using the nonlinear activation vector compression function Squashing. The calculation formula is as follows: ， in, Represents the total input vector The reduction factor; Represents the total input vector The unit vector.

Images