Ship ladder traveling cable motion online modeling method based on dynamic neural network model

Abstract

The invention relates to the technical field of neural network modeling, in particular to a ship ladder traveling cable motion online modeling method based on a dynamic neural network model, and the method comprises the steps: obtaining a second-order mixed partial derivative of a fusion feature function between any two physical tensors under all time steps before the current moment; all weights in the initial neural network model are updated by combining the interaction condition between the heat tensor and the electromagnetic tensor in all kinds of physical tensors, and an optimized neural network is obtained; and analyzing and optimizing the average distribution condition and dispersion degree of all weights between each pair of nodes of the neural network under all time steps before the current time step, screening out dynamic node pairs, and updating the weights between all the dynamic node pairs to obtain the ship ladder dynamic neural network. The invention aims at improving the efficiency and precision of ship ladder traveling cable motion online modeling and solving the key technical problem that the cable traveling motion situation lacks efficient and intelligent prediction.

Description

Technical Field

[0001] This application relates to the technical field of neural network modeling, and specifically to an online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model. Background Art

[0002] A marine elevator (referred to as a ship ladder) is a special electromechanical equipment fixedly installed on a ship to provide vertical or inclined transportation services for passengers, crew, or cargo. A ship ladder for ocean voyages needs to reliably cope with complex sea conditions such as typhoons, huge waves, and local abnormal weather. In particular, the trailing cable of the ship ladder is an important infrastructure for ensuring the power supply and information transmission between the elevator car and the control cabinet. Due to the multi-dimensional dynamic responses such as swaying and heaving generated by the ship during navigation under the interference of wind and waves, the trailing cable is forced to swing, resulting in interference damage or even breakage of the trailing cable with the shaft facilities such as the shaft wall, guide rail, and car, thus triggering safety accidents.

[0003] By online modeling, the movement state of the cable can be monitored in real time, potential safety hazards can be detected in time, safety accidents caused by cable failures can be avoided, and the safety of the ship ladder and personnel can be ensured. Based on the data and model analysis obtained from online modeling, it is also possible to predict the future operating state of the cable. For example, by analyzing the movement trend and force condition of the cable, the possible fault points and fault times of the cable in subsequent operation can be predicted, and maintenance measures can be taken in advance to reduce the probability of faults and improve the reliability of the system. However, the existing technologies often model mechanics, heat, and electromagnetic fields independently, ignoring the coupling effect. Traditional neural networks need to train and update all weights, which cannot meet the real-time requirements, reducing the efficiency and accuracy of the online modeling of the movement of the ship ladder trailing cable and unable to dynamically predict the movement trend of the trailing cable in real time. Summary of the Invention

[0004] In order to solve the above technical problems, this application provides an online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model to solve the existing problems.

[0005] The online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model of this application adopts the following technical solutions: An embodiment of this application provides an online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model, and this method includes the following steps: Obtain various physical field data of the cable during the operation of the ship ladder follow-up cable in real time. The various physical field data include: mechanical field, heat field, and electromagnetic field data. Use the various physical field data as the inputs of a multi-branch neural network respectively, extract various physical tensors from the neural network branches and transmit them to the fully connected layer for fusion to obtain an initial neural network model. Among them, the fusion process includes: obtaining the coupling coefficient matrix between any two physical tensors at the current time step, and fusing the any two physical tensors based on the coupling coefficient matrix to obtain the fusion characteristic function between any two physical tensors at the current time step; According to the weights between the nodes in the initial neural network, perform dynamic topology update on the initial neural network model to obtain an optimized neural network, and train the optimized neural network to obtain a ship ladder neural network model. The specific process is as follows: Based on the second-order mixed partial derivatives of the fusion characteristic function between any two physical tensors at all time steps before the current time step, and combining the interaction situation between the heat tensor and the electromagnetic tensor among all physical tensors, determine the coupling loss of the initial neural network model at the current time step, update all the weights in the initial neural network model to adjust the topological structure of the initial neural network model to obtain an optimized neural network; Analyze the average distribution and dispersion of the weights between each pair of nodes in the optimized neural network at all time steps before the current time step, screen out the dynamic node pairs from all pairs of nodes in the optimized neural network, and update the weights between all dynamic node pairs to obtain a ship ladder dynamic neural network.

[0006] Preferably, use the mechanical field data, heat field data, and electromagnetic field data as the inputs of the convolutional layer in the multi-branch neural network respectively, and output the neural network branches of the mechanical field, the neural network branches of the heat field, and the neural network branches of the electromagnetic field.

[0007] Preferably, the expression of the coupling coefficient matrix between any two physical tensors at the current time step is: ; In the formula, represents the coupling coefficient matrix between the p-th physical tensor and the q-th physical tensor at the current time step; represents the preset weight matrix at the current time step; represents the result of the concatenation operation on the p-th physical tensor and the q-th physical tensor at the current time step; represents the normalization function.

[0008] Preferably, the expression of the fusion characteristic function between any two physical tensors at the current time step is: ; represents the fusion characteristic function between the p-th physical tensor and the q-th physical tensor at the current time step; Represents the result of the fusion operation of the p-th physical tensor and the q-th physical tensor at the current time step.

[0009] Preferably, the expression for the coupling loss of the initial neural network model at the current time step is: ; where Represents the coupling loss of the initial neural network model at the current time step; Represents the cumulative sum of the second-order mixed partial derivatives of the fusion feature function between all physical tensors at time step t before the current time step; Represents the gradient of the heat field data at time step t before the current time step; Represents the gradient of the electromagnetic field intensity at time step t; k represents a preset value; N represents the number of all time steps before the current time step during the training of the initial neural network model; Represents Norm.

[0010] Preferably, updating all weights in the initial neural network model includes: The formula for updating the weight between node i and node j in the initial neural network model is: ; where , Represent the weights between node i and node j in the neural network at time steps t + 1 and t respectively; Represents the learning rate at time step t, where the initial learning rate is a preset first value; Represents the first-order partial derivative of the coupling loss with respect to the weight between node i and node j; Represents the error between the output result of the initial neural network model at time step t and the actually measured cable state, where the cable state includes: the position, speed, and acceleration of the cable; Represents Norm; Represents the adaptive learning rate function based on the error norm.

[0011] Preferably, adjusting the topological structure of the initial neural network model includes: After the initial neural network training is completed, if the weight between node i and node j in the initial neural network is less than the preset threshold, then the connection between node i and node j is deleted; otherwise, the connection between node i and node j is retained. Traverse the weights between all nodes to obtain the optimized neural network topological structure.

[0012] Preferably, screening out dynamic node pairs from all pairs of nodes in the optimized neural network includes: Calculate the mean and standard deviation of the weights of each pair of nodes in the optimized neural network over all iterations before the current time step. Denote the pairs of nodes whose mean is greater than a preset first value and standard deviation is greater than a preset second value as the dynamic node pairs in the optimized neural network at the current time step.

[0013] Preferably, updating the weights between all dynamic node pairs includes: The weight adjustment value of dynamic node pair d at the next moment of the current time step The expression is: ; where represents the weight between dynamic node pair d at the current time step; represents the learning rate at the current time step, where the initial learning rate in the process of updating the weights between dynamic node pairs is the preset second value; represents the gradient of the weight between dynamic node pair d at the current moment on the coupling loss L; represents the constrained projection operator.

[0014] Preferably, obtaining the ship ladder dynamic neural network includes: In the optimized neural network, obtain all dynamic node pairs at different time steps, update the weights of all dynamic node pairs, and use the neural network model after the optimized neural network training is completed as the ship ladder dynamic neural network.

[0015] This application has at least the following beneficial effects: This application dynamically adjusts the connection weights of the network layer according to real-time physical field data, realizes the co-evolution of the model structure and physical state, and significantly improves the adaptability of the model to the dynamic environment; further, introduces the cross-coupling term of mechanical-thermal-electromagnetic fields, quantifies the interaction of multiple physical fields through tensor fusion, reduces the prediction deviation, and improves the prediction accuracy of the model; further, adopts the local weight update strategy. When the environment changes violently, only the sub-network affected by the environmental change is locally trained, without the need to retrain the entire model, reducing the computational complexity and improving the efficiency and accuracy of online modeling of the ship ladder trailing cable movement, and solving the key technical problem of the lack of efficient and intelligent prediction of the trailing cable movement trend. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0017] Figure 1The flowchart of steps of an online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model provided by an embodiment of the present application; Figure 2 The schematic diagram of the process for obtaining the dynamic neural network model of the ship ladder provided by an embodiment of the present application. Detailed implementation manners

[0018] In order to further elaborate on the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the following, in combination with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features and effects of the online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model proposed according to the present application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present application belongs.

[0020] The following specifically describes the specific solution of the online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model provided by the present application with reference to the accompanying drawings.

[0021] An online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model provided by an embodiment of the present application. Specifically, the following online modeling method for the movement of a ship ladder trailing cable based on a dynamic neural network model is provided. Please refer to Figure 1 , the method includes the following steps: Step S1: Real-time obtain various physical field data of the cable during the operation of the ship ladder trailing cable. The various physical field data include: mechanical field, heat field and electromagnetic field data. Respectively use the various physical field data as the inputs of a multi-branch neural network, extract various physical tensors from the neural network branches and transmit them to the fully connected layer for fusion to obtain an initial neural network model. The fusion process includes: obtaining the coupling coefficient matrix between any two physical tensors at the current time step, and based on the coupling coefficient matrix, fusing the any two physical tensors to obtain the fusion characteristic function between any two physical tensors at the current time step.

[0022] In this embodiment, the neural network used is a convolutional neural network. In the actual application process, as other implementation manners, the implementer can also use other neural networks, and this embodiment does not make special restrictions. Among them, the convolutional neural network is a well-known technology, and its specific principle will not be elaborated here.

[0023] The input layer group of the neural network is responsible for receiving and processing data from multi-source sensors, providing a basis for subsequent model analysis and processing. During the actual operation of the ship's trailing cable, it will be affected by various factors. In order to comprehensively obtain the cable's status information, multi-source sensors are required for data acquisition. In this embodiment, these sensors include: Mechanical sensors: used to measure the mechanical stress on the cable and obtain relevant data of the cable under mechanical action in real time.

[0024] Temperature sensors: distributed at different positions of the cable, used to monitor the temperature change of the cable in real time. Since the cable will generate heat due to factors such as current passing through and friction during operation, the temperature change has an important impact on the cable's performance and lifespan. Therefore, real-time collection of the cable's temperature data provides a basis for analyzing the cable's thermal state.

[0025] Electromagnetic field sensors: used to detect the distribution and change of the electromagnetic field around the cable, obtain electromagnetic field intensity data, and provide data support for analyzing the cable's electromagnetic characteristics.

[0026] Among them, in this embodiment, the acquisition frequency of the physical field data is 1Hz. Therefore, during the following neural network training process, each time step is also set to 1s. In the actual application process, the implementer can also set it according to the specific situation by himself / herself, and this embodiment does not make special restrictions.

[0027] In order to eliminate the influence of data dimensions on the subsequent analysis, all the collected data is normalized to obtain the normalized values. In this embodiment, the maximum-minimum normalization method is used to normalize the data. In the actual application process, as other implementation methods, the implementer can also use other normalization methods such as the z-score standardization method. Regarding the selection of the normalization method, this embodiment does not make special restrictions. Among them, the maximum-minimum normalization method is a well-known technology, and its specific principle will not be elaborated here.

[0028] Furthermore, all kinds of physical field data are used as the input of the multi-branch neural network to obtain the neural network branches of various physical fields. Specifically: the mechanical field data, heat field data, and electromagnetic field data are respectively used as the input of the convolutional layer in the multi-branch neural network, and the neural network branches of the mechanical field, heat field, and electromagnetic field are output; furthermore, tensors of various physical field data are extracted from the neural network branches and transmitted to the fully connected layer for fusion to obtain the initial neural network model. Among them, for the convenience of description, the tensors of the physical field data are denoted as physical tensors, that is, the tensors of the mechanical field data, heat field data, and electromagnetic field data layers are respectively denoted as mechanical tensors, heat tensors, and electromagnetic tensors.

[0029] Among them, the process of extracting physical tensors from all neural networks of the neural network is a well-known technology, and the specific principle process will not be elaborated here.

[0030] It should be supplemented that the specific content of various neural network branches is as follows: (1) Neural network branch of the mechanical field: A 1D convolutional layer is adopted. By sliding the convolutional kernel over 1D data for convolutional operations, it can effectively extract local strain features. Convolutional operations can capture local information. In the mechanical stress analysis of the ship ladder trailing cable, this means that it can comprehensively consider the strain conditions of the cable within a certain range. For example, when detecting cable bending or stretching, it can integrate the stress change information of adjacent positions, thereby more comprehensively capturing the characteristic patterns of stress distribution.

[0031] After the mechanical tensor transmitted from the input layer enters the 1D convolutional layer, the convolutional kernel slides point by point over the data and performs convolutional calculations on each local area. In this way, the original mechanical field data is converted into more representative local strain features, which can highlight the deformation characteristics and laws of the cable under mechanical stress, providing key information for subsequent fusion with other physical field feature points and model analysis. In this embodiment, the convolutional kernel size is set to 5 and the stride is set to 1.

[0032] (2) Neural network branch of the heat field: A fully connected layer is used to map the temperature gradient distribution. In this embodiment, the number of neurons is set to 32, which effectively performs feature transformation on the input temperature data and learns the gradient change features of temperature at different positions or times of the cable, so as to accurately reflect the distribution of the heat field.

[0033] The input heat tensor enters the fully connected layer. Each neuron connects with all elements of the input data and performs weighted summation, and then undergoes a non-linear transformation through an activation function. In this way, the fully connected layer can map the original temperature data into a 32-dimensional feature space. Each dimension represents a certain feature of the temperature length, such as the trend of temperature change, the temperature difference at different positions, etc. These features are crucial for understanding the thermal state of the cable and its interaction with other physical fields.

[0034] (3) Neural network branch of the electromagnetic field: The fast Fourier transform is used as the frequency domain transformation layer to extract the spectral features of noise. In the complex electromagnetic environment where the ship ladder trailing cable is located, the electromagnetic field data contains rich frequency information. The fast Fourier transform can efficiently convert the time-domain electromagnetic signal to the frequency domain and reveal the distribution of different frequency components in the signal. By analyzing the noise spectral feature values, the frequency characteristics, intensity distribution, etc. of electromagnetic interference can be understood, which is very crucial for evaluating the operating state and performance of the cable in the electromagnetic environment.

[0035] After the input electromagnetic tensor enters the frequency-domain transformation layer, it is processed by the fast Fourier transform algorithm to convert the time-domain signal into a frequency-domain representation. The obtained frequency-domain data contains the amplitude and phase information at different frequencies. These information constitute the noise spectrum characteristics of the electromagnetic field, which can intuitively display the characteristics and changes of the electromagnetic environment around the cable, and provide important electromagnetic information for subsequent multi-physics field coupling analysis.

[0036] Furthermore, different types of tensors are fused. The specific process is as follows: In the actual operating environment of the ship ladder trailing cable, mechanical stress, temperature changes, and electromagnetic field interference often affect each other. The bending or stretching of the cable may cause changes in its internal resistance, thereby affecting the current distribution and electromagnetic field characteristics. At the same time, heat will also be generated due to friction and other reasons, changing the temperature of the cable. Through the above calculation method, the coupling tensor fusion layer can fully consider the complex coupling relationships between these physical fields and organically fuse the characteristics of each physical field. Compared with the traditional method of independently modeling each physical field, it can more realistically reflect the actual operating state of the cable, reduce the prediction deviation caused by ignoring the multi-physics field coupling effect, and provide more accurate and effective characteristic information for subsequent dynamic topology adjustment and cable state prediction.

[0037] To fuse the characteristics of different physical fields, a coupling coefficient matrix is defined , where represents the coupling coefficient matrix between the p-th physical tensor and the q-th physical tensor at the current time step; represents the preset weight matrix at the current time step; represents the result of concatenating the p-th physical tensor and the q-th physical tensor at the current time step; represents the normalization function.

[0038] It should be noted that the setting process of the preset weight matrix includes: when the neural network is not started for training iteration at the initial moment, the value of the weight matrix is randomly set. Subsequently, during the neural network training process, the weight matrix will be updated according to backpropagation. Among them, backpropagation is a well-known technology in the neural network training process, and its specific principle will not be elaborated here.

[0039] In addition, the process of concatenating different types of physical tensors is a well-known technology, and the specific process will not be elaborated here.

[0040] Furthermore, based on the coupling coefficient matrix, the above-mentioned any two physical tensors are fused to obtain the fusion characteristic function between any two physical tensors. The specific calculation formula is: The fusion characteristic function between the p-th physical tensor and the q-th physical tensor at the current time step is expressed as: ; It represents the result of the fusion operation on the p-th physical tensor and the q-th physical tensor at the current time step.

[0041] Among them, in this embodiment, the process of fusing different physical tensors needs to be fused according to the characteristics of different physical tensors. The fusion operation is a well-known technology, and its specific principle will not be elaborated here.

[0042] So far, the physical tensors have been fused to obtain a fusion feature function.

[0043] Step S2: Dynamically update the topology of the initial neural network model according to the relationship between different physical tensors to obtain an optimized neural network, and train the optimized neural network to obtain a ship ladder neural network model.

[0044] S201: Based on the second-order mixed partial derivatives of the fusion feature functions between any two physical tensors at all time steps before the current time step, and combined with the interaction between the heat tensor and the electromagnetic tensor among all physical tensors, determine the coupling loss of the initial neural network model at the current time step, and update all the weights in the initial neural network model to adjust the topological structure of the initial neural network model to obtain an optimized neural network.

[0045] The environment of the cable is changeable, and the fusion features of the multi-physical field data collected in real time are also constantly changing. By dynamically adjusting the topological structure according to the fusion features, the model can adapt to these changes in real time, timely adjust its own structure to better fit the new data, so as to maintain a high prediction accuracy under different working conditions, such as different tidal conditions, different load states of the ship, etc.

[0046] Therefore, based on the second-order mixed partial derivatives of the fusion feature functions between any two physical tensors, and combined with the interaction between the heat tensor and the electromagnetic tensor among all physical tensors, determine the coupling loss of the initial neural network model. According to the coupling loss, combined with backpropagation, dynamically update the weights in the initial neural network model. Specifically: The coupling loss of the initial neural network model at the current time step The expression is: ; where represents the cumulative sum of the second-order mixed partial derivatives of the fusion feature functions between all physical tensors at time step t before the current time step; represents the gradient of the heat field data at time step t before the current time step; represents the gradient of the electromagnetic field strength at time step t; k represents a preset value; N represents the number of all time steps before the current time step during the training of the initial neural network model; represents the norm.

[0047] It should be noted that the value of k is set artificially and is a material constant. In this embodiment, the value of k is 500 MPa. In actual application, the implementer can also set it according to specific circumstances, and this embodiment does not make special restrictions.

[0048] Among them, the calculation processes of the gradient of the heat field data and the gradient of the electromagnetic field intensity at time step t are both well-known technologies, and the specific calculation steps will not be elaborated here; and The norm is a well-known technology, and the specific principle will not be elaborated here.

[0049] Furthermore, update all the weights in the initial neural network model, including: The formula for updating the weight between node i and node j in the initial neural network model is: ; in the formula, and respectively represent the weights between node i and node j in the neural network at time steps t + 1 and t; represents the learning rate at time step t, where the initial learning rate is a preset first value; represents the first-order partial derivative of the coupling loss with respect to the weight between node i and node j; represents the error between the output result of the initial neural network model at time step t and the actually measured cable state, where the cable state includes: the position, speed, and acceleration of the cable; represents the norm; represents the adaptive learning rate function based on the error norm.

[0050] It should be noted that in this embodiment, the initial learning rate is set to have a value of 0.001, that is, the value of the preset first value is 0.001. In the subsequent iteration process, the learning rate will adaptively change during the iteration. The process of adaptive update of the learning rate is a well-known technology, and the specific principle will not be elaborated here.

[0051] It is further explained that the actually measured cable state includes the position, speed, and acceleration of the cable, all of which can be measured by various sensors.

[0052] Among them, the adaptive learning rate function based on the error norm is a well-known technology, and the specific principle will not be elaborated here.

[0053] Further, after the initial neural network training is completed, if the weight between node i and node j in the initial neural network is less than the preset threshold, it is considered that the connection between node i and j is too weak and contributes less to the information transmission and feature learning of the model, and the connection between node i and node j is deleted. Otherwise, the connection between node i and node j is retained. Traverse the weights between all nodes to obtain the updated topological structure of the initial neural network, which is used as the optimized neural network.

[0054] It should be noted that the value of the preset threshold is set manually. In this embodiment, the value of the preset threshold is 0.6. In the actual application process, the implementer can also set it according to the specific situation, and this embodiment does not make special restrictions.

[0055] S202: Analyze the average distribution and dispersion degree of all weights between each pair of nodes in the optimized neural network at all time steps before the current time step, screen out the dynamic node pairs from all pairs of nodes in the optimized neural network, and update the weights between all dynamic node pairs to obtain the ship ladder dynamic neural network.

[0056] The network weights are divided into static blocks and dynamic blocks. The static blocks contain those weight parameters that are relatively stable during the model training process and have important influences on the overall structure and basic features. These parameters do not need to be updated frequently during most of the running time of the model because they represent the inherent characteristics and basic laws of the cable under normal circumstances. The dynamic blocks contain those weight parameters that are easily affected by environmental changes, and these parameters need to be adjusted in a timely manner according to new data.

[0057] Therefore, analyzing the average distribution and dispersion degree of all weights between each pair of nodes in the optimized neural network at all time steps before the current time step, and screening out the dynamic node pairs from all pairs of nodes in the optimized neural network, specifically: Calculate the mean and standard deviation of the weights of each pair of nodes in the optimized neural network before the current time step during all iterations. Record the pair of nodes whose mean is greater than the preset first value and the standard deviation is greater than the preset second value as the dynamic node pairs in the optimized neural network at the current time step.

[0058] It should be noted that the values of the preset first value and the preset second value are both set manually. In this embodiment, the value of the preset first value is 0.1, and the value of the preset second value is 0.05. The implementer can also set it according to the specific situation, and this embodiment does not make special restrictions.

[0059] Further, update the weights between all dynamic node pairs to obtain the ship ladder dynamic neural network. Specifically: The weight adjustment value of the dynamic node pair d at the next moment of the current time step The expression of is: ; where, represents the weight between dynamic node pairs d at the current time step; represents the learning rate at the current time step, where the initial learning rate during the weight update process between dynamic node pairs is a preset second value; represents the gradient of the weight between dynamic node pairs d at the current moment on the coupling loss L; represents a constrained projection operator; in this embodiment, the feasible region has a range of .

[0060] It should be noted that the value of the initial learning rate during the weight update process between dynamic node pairs is set manually. In this embodiment, the value of the initial learning rate during the weight update process between dynamic node pairs is set to 0.001, that is, the value of the preset second value is 0.001. Implementers can also set it according to specific situations. There is no special limitation in this embodiment. During subsequent iterations, the learning rate will adaptively change during the iteration process. The process of adaptive learning rate update is a well-known technology, and the specific principle will not be elaborated here.

[0061] Among them, the process of calculating the gradient of the weight on the coupling loss, that is, the process of calculating the gradient of the weight on the loss function, and the constrained projection operator are both well-known technologies, and their specific principles will not be elaborated here.

[0062] In optimizing the neural network, all dynamic node pairs are obtained at different time steps, and the weights of all dynamic node pairs are updated. The neural network model after the training of the optimized neural network is used as the ship ladder dynamic neural network.

[0063] Preferably, the schematic diagram of the process for obtaining the ship ladder dynamic neural network model provided in this embodiment is as Figure 2 shown.

[0064] It should be noted that: the above sequence of embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above describes specific embodiments of this specification. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0065] Each embodiment in this specification is described in a progressive manner. The same or similar parts between each embodiment can be referred to each other, and the differences between each embodiment are emphasized.

[0066] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; modifying the technical solutions described in the foregoing embodiments, or equivalently replacing some of the technical features therein, does not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present application, and should all be included within the protection scope of the present application.

Claims

1. An online modeling method for the movement of the ship ladder trailing cable based on a dynamic neural network model, characterized in that The method includes the following steps: Obtain various physical field data of the cable during the operation of the ship ladder follow-up cable in real time. The various physical field data include: mechanical field, heat field, and electromagnetic field data. Use the various physical field data as the inputs of a multi-branch neural network respectively. Extract various physical tensors from the neural network branches and transmit them to the fully connected layer for fusion to obtain an initial neural network model. Among them, the fusion process includes: obtaining the coupling coefficient matrix between any two physical tensors at the current time step, and based on the coupling coefficient matrix, fusing the any two physical tensors to obtain the fusion characteristic function between any two physical tensors at the current time step; According to the weights between the nodes in the initial neural network, perform dynamic topology update on the initial neural network model to obtain an optimized neural network, and train the optimized neural network to obtain a ship ladder neural network model. The specific process is as follows: Based on the second-order mixed partial derivatives of the fusion characteristic functions between any two physical tensors at all time steps before the current moment, and combined with the interaction situation between the heat tensor and the electromagnetic tensor among all physical tensors, determine the coupling loss of the initial neural network model at the current time step, update all the weights in the initial neural network model to adjust the topological structure of the initial neural network model to obtain an optimized neural network; Analyze the average distribution and dispersion degree of the weights between each pair of nodes in the optimized neural network at all time steps before the current time step. Screen out the dynamic node pairs from all pairs of nodes in the optimized neural network, and update the weights between all dynamic node pairs to obtain a ship ladder dynamic neural network.

2. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, characterized in that, Respectively use the mechanical field data, heat field data, and electromagnetic field data as the inputs of the convolutional layer in the multi-branch neural network, and output the neural network branches of the mechanical field, the heat field, and the electromagnetic field.

3. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, characterized in that The expression for the coupling coefficient matrix between any two physical tensors at the current time step is as follows: ; where represents the coupling coefficient matrix between the p-th physical tensor and the q-th physical tensor at the current time step; represents the preset weight matrix at the current time step; represents the result of concatenating the p-th physical tensor and the q-th physical tensor at the current time step; represents the normalization function.

4. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 3, characterized in that, The expression of the fusion feature function between any two physical tensors at the current time step is as follows: ; represents the fusion feature function between the p-th physical tensor and the q-th physical tensor at the current time step; represents the result of the fusion operation on the p-th physical tensor and the q-th physical tensor at the current time step.

5. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, wherein The expression for the coupling loss of the initial neural network model at the current time step is as follows: ; where represents the coupling loss of the initial neural network model at the current time step; represents the cumulative sum of the second-order mixed partial derivatives of the fusion feature function between all physical tensors at time step t before the current time step; represents the gradient of the heat field data at time step t before the current time step; represents the gradient of the electromagnetic field strength at time step t; k represents a preset value; N represents the number of all time steps before the current time step during the training of the initial neural network model; represents norm.

6. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, wherein The update of all the weights in the initial neural network model includes: The formula for updating the weight between node i and node j in the initial neural network model is as follows: ; In the formula, , respectively represent the weights between node i and node j in the neural network at time step t + 1 and time step t; represents the learning rate at time step t, where the initial learning rate is a preset first value; represents the first-order partial derivative of the coupling loss with respect to the weight between node i and node j; represents the error between the output result of the initial neural network model at time step t and the actually measured cable state, where the cable state includes: the position, speed, and acceleration of the cable; represents norm; represents an adaptive learning rate function based on the error norm.

7. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, characterized in that, The adjustment of the topological structure of the initial neural network model includes: After the initial neural network training is completed, if the weight between node i and node j in the initial neural network is less than the preset threshold, then delete the connection between node i and node j; otherwise, retain the connection between node i and node j. Traverse the weights between all nodes to obtain the optimized neural network topological structure.

8. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, characterized in that, The screening of the dynamic node pairs from all pairs of nodes in the optimized neural network includes: Calculate the mean and standard deviation of the weights of each pair of nodes in the optimized neural network during all iterations before the current time step. Mark the pair of nodes whose mean is greater than the preset first value and the standard deviation is greater than the preset second value as the dynamic node pairs in the optimized neural network at the current time step.

9. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, wherein, The update of the weights between all dynamic node pairs includes: The weight adjustment value of the dynamic node pair d at the next moment of the current time step The expression is as follows: ; In the formula, represents the weight between the dynamic node pairs d at the current time step; represents the learning rate at the current time step, where the initial learning rate in the process of updating the weight between the dynamic node pairs is a preset second value; represents the gradient of the weight between the dynamic node pairs d at the current moment on the coupling loss L; represents the constraint projection operator.

10. The online modeling method for the movement of the ship ladder trailing cable based on the dynamic neural network model according to claim 1, characterized in that, The obtaining of the ship ladder dynamic neural network includes: In the optimized neural network, obtain all dynamic node pairs at different time steps, update the weights of all dynamic node pairs, and use the neural network model after the optimized neural network training is completed as the ship ladder dynamic neural network.

Images