Ship personnel positioning method based on distributed deep network model

Abstract

The invention relates to a ship personnel positioning method based on a distributed deep network model, and the method comprises the steps: constructing and training a distributed deep network model which comprises a plurality of wireless sensor network nodes which are divided into a target node, a coordinator node and a plurality of anchor nodes, the distributed deep network model carries out distance estimation between an anchor node and a target node based on the received broadcast signal; after distance estimation is obtained, a statistical redundancy deletion principle is introduced, a minimum error positioning algorithm is designed, a ship personnel positioning system based on received signal strength indication is constructed, and coordinates of ship personnel positioning points are output. Compared with the prior art, the method has the advantages of being accurate and real-time in distance measurement estimation, improving positioning accuracy and stability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of ship personnel positioning, and in particular to a ship personnel positioning method based on a distributed deep network model. Background Art

[0002] With the continuous development of the shipping economy, modern ships have become increasingly complex, intelligent and modern. The safety requirements for ship personnel are constantly increasing, and the demand for ship personnel positioning is increasing.

[0003] Research on wireless sensor networks is a hot topic in the field of wireless communications. Network nodes in wireless sensor networks can achieve self-positioning through mutual communication, which can be applied to practical projects such as environmental and engineering monitoring, and target tracking. Existing common node localization technologies in wireless sensor networks can be roughly divided into ranging-based and non-ranging positioning algorithms. Representative ranging-based positioning algorithms include time of arrival (TOA)-based positioning solutions, time difference of arrival (TDOA)-based positioning solutions, angle of arrival (AOA)-based positioning solutions, and received signal strength indicator (RSSI)-based positioning solutions. Common positioning frameworks include trilateration, extended Kalman filtering, and maximum likelihood estimation. These common algorithms have initially solved the node localization problem in wireless sensor networks.

[0004] However, due to the complex structure of ships, with the bulk of the hull constructed of metal, radio signals are subject to multipath effects and complex noise interference, resulting in low positioning accuracy in ship environments. A common method for locating personnel on board ships is to estimate distance using a path attenuation model, followed by position estimation using trilateration. The core concept of the path attenuation model is to estimate distance based on the attenuation of the received signal strength (RSS) between the anchor node and the target node as the distance between the nodes decreases. Position estimation is then performed using the least squares method using distance information greater than or equal to the three-sided distance. While the path attenuation model can estimate position to a certain extent and offers advantages such as low computational complexity and high operability, environmental changes during a ship's motion significantly impact the RSS. Path fading models for wireless channels cannot effectively represent signal strength in real space. Therefore, it is necessary to comprehensively consider the relationship between the RSS and the ship's ambient noise.

[0005] Deep networks excel at automatically sensing and reasoning about underlying relationships between variables, and are capable of exploring the relationship between signal characteristics, ambient noise, and node distance. However, building deep network models requires significant computing power, which is often insufficient for cost considerations, making them unsuitable for direct use in ship personnel positioning.

[0006] After searching, Chinese invention patent application publication number CN115052245A discloses a deep learning-based drone-assisted wireless sensor network node positioning method. This positioning method is applied in outdoor areas. Specifically, the drone moves periodically in a fixed trajectory within an area parallel to the positioning area and broadcasts a beacon signal at a fixed time period. Ground sensor nodes receive the drone beacon signal, calculate the RSSI value to form an RSSI vector, calculate the RSSI similarity between nodes, and then calculate the distance between the anchor node and the unknown node. A convolutional neural network positioning model is established, and the distance between the node and the anchor node is input to estimate the node position coordinates. This existing patent application has the problem of not using a distributed architecture and static scene positioning is not suitable for ship movement scenarios.

[0007] How to achieve real-time and accurate positioning of ship personnel has become a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a ship personnel positioning method based on a distributed deep network model.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] According to one aspect of the present invention, a method for positioning ship personnel based on a distributed deep network model is provided, the method comprising:

[0011] Constructing and training a distributed deep network model, the distributed deep network model includes multiple wireless sensor network nodes, divided into a target node, a coordinator node, and several anchor nodes, and the distributed deep network model estimates the distance between the anchor node and the target node based on the received broadcast signal;

[0012] After obtaining the distance estimation, the statistical redundancy reduction principle is introduced to design a minimum error positioning algorithm, and a ship and personnel positioning system based on the received signal strength indication is constructed to output the coordinates of the ship and personnel positioning points.

[0013] Preferably, the target node is the node that needs to be located; the target node is responsible for broadcasting signals to the coordinator node and the anchor node.

[0014] Preferably, each anchor node deploys a lightweight standard neural network, whose hidden layer is a fully connected structure; the input of the lightweight standard neural network includes temperature, humidity, its own ID tag, the received signal strength of the target node, and the signal connection quality of the target node; the output of the lightweight standard neural network is the distance between the anchor node and the target node.

[0015] Preferably, the coordinator node is responsible for the aggregation and distribution of parameters during the distributed deep network model training process: all anchor nodes send local network model parameters to the coordinator node. After the coordinator node receives the parameters of all anchor nodes, it aggregates the parameters to form global network model parameters, and then distributes the global network model parameters to each anchor node. Each anchor node continues to train using the local training set based on the global network model parameters.

[0016] More preferably, the optimization target of the local network model parameters of the anchor node is:

[0017]

[0018] Where f(w) is the sum of the sample mean square error and the regularization term, w is the local network model parameter of the anchor node, i is the index of the training data sample, d i is the distance label value corresponding to the i-th data sample, is the distance prediction value corresponding to the i-th data sample, and γ is the regularization weight coefficient.

[0019] More preferably, the distributed deep network model uses stochastic gradient descent to find the global optimal value of the loss function;

[0020] The anchor nodes continuously adjust the weight matrix through back propagation to minimize the loss function:

[0021]

[0022] Where, the subscript t represents the number of iterations, α(t) represents the learning rate, the superscript k represents the index of the anchor node, and w k is the weight matrix of the kth anchor node.

[0023] More preferably, after performing multiple iterations locally, the anchor node sends the trained local network parameter model to the coordinator node for aggregation, specifically:

[0024]

[0025] Among them, B represents the number of local iterations of each anchor node, n is the total number of anchor nodes participating in the aggregation, is the weight matrix after training t+B rounds for the kth anchor node, w t+BAggregated weight matrix for the coordinator node.

[0026] Preferably, the statistical redundancy reduction principle is specifically:

[0027] Use n anchor nodes to perform ranging and obtain the ranging information from each anchor node to the target node, where n is greater than 3;

[0028] Among all anchor nodes, some anchor nodes are retained as the final positioning output nodes.

[0029] The retained anchor nodes are used as nodes with effective observation distances, and the minimum error positioning algorithm is used to output the final positioning point coordinates.

[0030] Preferably, the retention principle of some anchor nodes is as follows:

[0031]

[0032] Among them, M represents the number of retained anchor nodes, N represents the total number of anchor nodes, P represents the probability that the ranging error is greater than the set threshold, and Z + Is a positive integer.

[0033] Preferably, in addition to having the function of an anchor node, the coordinator node is also configured with a temperature and humidity sensor to distribute temperature and humidity information to the anchor node.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1) In view of the limited computational and communication volume of wireless sensor network nodes, the present invention designs a distributed deep network model for distance estimation. The distributed architecture reduces the computational workload of the coordinator node through decentralization, and the number of communication rounds is limited. After obtaining the distance estimate, the statistical redundancy reduction principle is introduced. Based on the principle of minimizing the error between the positioning output point and the center of gravity points determined between each anchor node, a minimum error positioning algorithm is designed to locate the ship personnel system, thereby improving the ranging and positioning accuracy.

[0036] 2) The anchor node of the present invention deploys a lightweight standard neural network to output the distance between the anchor node and the target node; the coordinator node is responsible for the aggregation and distribution of parameters during the training of the distributed deep network model. The two coordinate iterative training, balancing the amount of computation and communication, making the distance estimation more real-time, stable and reliable. Compared with other algorithms, the distance measurement error of the present invention is minimized.

[0037] 3) The optimization target of the local network model parameters in the present invention adopts regularization technology and applies stochastic gradient descent to find the global optimal value of the loss function. The anchor node continuously adjusts the weight matrix through back propagation to minimize the loss function, which not only prevents overfitting but also speeds up the convergence speed during training.

[0038] 4) The present invention uses more than three anchor nodes for ranging. Due to the existence of redundant information, the principle of statistical redundancy reduction is applied. Based on the principle of minimizing the error between the positioning output point and the center of gravity points determined between each anchor node, a minimum error positioning algorithm is designed to output the final positioning point coordinates, making positioning more accurate and stable. Compared with other algorithms, the positioning accuracy and stability of the method of the present invention are the best. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the process of the ship personnel positioning method of the present invention;

[0040] Figure 2 Schematic diagram of the positional relationship between different equidistant circles;

[0041] Figure 3 This is a schematic diagram of deploying wireless sensor network nodes on board a ship in the present invention;

[0042] Figure 4 Schematic diagram of the training effect of different numbers of iterations in the present invention;

[0043] Figure 5 Schematic diagram of ranging accuracy with different numbers of iterations in the present invention;

[0044] Figure 6 Schematic diagram comparing the ranging accuracy of the method of the present invention and other algorithms;

[0045] Figure 7 Schematic diagram comparing the positioning point and path tracking effects of the method of the present invention and other algorithms. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0047] This embodiment provides a method for locating personnel on board a ship based on a distributed deep network model. This method overcomes the inaccuracies of traditional trilateration methods and aims to improve the accuracy of positioning personnel on board. This method utilizes a distributed design of the deep network model based on the computational performance of wireless sensor network nodes to adapt to their performance.

[0048] The process of the ship personnel positioning method is as follows: Figure 1 Shown, including:

[0049] S1, considering the characteristics of wireless sensor network nodes such as limited computing and communication volume, a distributed deep network model is designed for distance estimation. The distributed architecture is decentralized, which can reduce the computing volume of the coordinator node, and the number of communication rounds is limited.

[0050] S2, after obtaining a relatively accurate distance estimation, introduces the principle of statistical redundancy reduction, designs a minimum error positioning algorithm, and constructs a ship and personnel positioning system based on received signal strength indication to improve the ranging and positioning accuracy.

[0051] S3, deploying the method of the present invention on a real ship to verify the positioning accuracy of the ship personnel positioning system. The experimental results show that the ship personnel positioning method is feasible and can effectively locate ship personnel with higher positioning accuracy and stability.

[0052] In S1, the distributed deep network model can balance the computing and communication volume of each anchor node, realize the node ranging requirements, and is more suitable for scenarios with high real-time requirements, large data volume, and strong stability.

[0053] The distributed deep network model includes multiple wireless sensor network nodes, namely:

[0054] 1 target node: the node that needs to be located. It is only responsible for broadcasting signals so that other modules can receive the broadcast signals;

[0055] Several anchor nodes: Each of them deploys a lightweight standard neural network with a 3x8 fully connected hidden layer. The input is 5 bits, including temperature, humidity, its own ID tag, RSSI (received signal strength) and LQI (signal quality) of the signal received from the target node; the output is 1 bit, including the distance between the anchor node and the target node.

[0056] 1 coordinator node: In addition to having the function of an anchor node, the coordinator node has a temperature and humidity sensor that distributes temperature and humidity information to the anchor nodes. It is also responsible for aggregating and distributing parameters during the training process. That is, after a certain number of training rounds (for example, 5 times), all anchor nodes send local network model parameters (called local parameters) to the coordinator node. After the coordinator receives the parameters of all anchor nodes, it aggregates the parameters by averaging and other methods to form new parameters (called global network model parameters). The global network model parameters are then distributed to each anchor node. Each anchor node continues to train using the local training set based on the global network model parameters, and the above operations are repeated.

[0057] Generally speaking, at the beginning of a network iteration, the coordinator node sends the current global network model parameters to each anchor node. Each anchor node trains its local neural network based on the global state and its local training set. After training to a certain level (e.g., a fixed number of local training rounds), each anchor node sends its own network model parameters to the coordinator node. Finally, the coordinator node applies the updated network model parameters to the global state, and the process repeats. The algorithm details are as follows.

[0058] (1) Since the target of the distributed deep network model used for ranging is a non-convex neural network target, and when there is not enough training data or the model is overtrained, overfitting often occurs. As the training process progresses, the model complexity increases, the error on the training set gradually decreases, but the error on the validation set gradually increases. Regularization technology can prevent the deep network model from overfitting, but the model convergence speed slows down. Based on this, the optimization target of the local network model parameters of each anchor node is as follows (1):

[0059]

[0060] Where f(w) is the sum of the sample mean square error and the regularization term, w is the local network model parameter of the anchor node, i is the index of the training data sample in each module, and d i is the distance label value corresponding to the i-th data sample, is the distance prediction value corresponding to the i-th data sample, γ is the regularization weight coefficient, is the L2 regularization term.

[0061] (2) The present invention applies stochastic gradient descent to find the global optimum of the loss function (the same optimization objective as formula (1)). Gradient descent is an optimization technique that updates weights based on the aggregated error of the training data. Although gradient descent is fast, it can get stuck in a local optimum when the loss function is non-convex. Stochastic gradient descent is state-of-the-art in data center environments and is generally more accurate than gradient descent, although it can slow model convergence.

[0062] (3) The hidden layer adopts a fully connected structure and uses the ReLU activation function, which can increase the nonlinearity of the network, prevent the gradient from disappearing, and improve the calculation speed of the neural network. The anchor node continuously adjusts the weight matrix through back propagation to minimize the loss function, as shown in the following formula:

[0063]

[0064] Where, the subscript t represents the number of iterations, α(t) represents the learning rate, the superscript k represents the index of the anchor node, and w k is the weight matrix of the kth anchor node.

[0065] (4) After each anchor node has been trained to a certain extent locally (usually for a fixed number of local iterations B), it sends the trained local network parameter model to the coordinator for aggregation. In order to reduce the amount of communication between nodes, multiple iterations are performed locally before aggregation of the coordinator nodes, as shown in Equation (3):

[0066]

[0067] Among them, B represents the number of local iterations of each anchor node, n is the total number of anchor nodes participating in the aggregation, is the weight matrix after training t+B rounds for the kth anchor node, w t+B Aggregated weight matrix for the coordinator node.

[0068] Below is the complete pseudo code of the distributed deep network model:

[0069]

[0070] In S2, the present invention designs a minimum error positioning algorithm and constructs a ship personnel positioning system based on the received signal strength indication to improve positioning accuracy. In an ideal situation, if the distance information from the target node to the three anchor nodes is known, the positioning coordinates can be determined according to formula (4).

[0071]

[0072] Among them, (x k ,y k ) represents the position coordinates of the anchor node k, (x, y) represents the position coordinates of the target node, r k represents the distance from anchor node k to the target node, and N is the number of anchor nodes.

[0073] Due to the existence of errors, the equidistant lines of the three anchor nodes may not intersect at one point. The positional relationship between different equidistant circles is as follows: Figure 2 As shown in Figure 2 , the typical solution is to use the proportional radius calculation method to determine the center of gravity of two equally spaced circles, then average all the center of gravity points to obtain the final positioning point. To improve positioning accuracy, six anchor nodes are used for ranging. Due to the presence of redundant information, a minimum error positioning algorithm is designed based on the principle of minimizing the error between the positioning output point and the center of gravity points determined by each anchor node to output the final positioning point coordinates.

[0074] The specific implementation is as follows:

[0075] 201, obtaining the ranging information of each anchor node through a distributed deep network model.

[0076] 202. Due to the influence of the ship environment, the multipath effect and shadow effect are serious. In order to remove the large error interference, some anchor nodes are retained as the final positioning output nodes among all anchor nodes. The retention principle is as shown in formula (5).

[0077]

[0078] Among them, M represents the number of retained anchor nodes, N represents the total number of anchor nodes, P represents the probability that the ranging error is greater than the set threshold, and Z + Is a positive integer.

[0079] 203, using the retained anchor nodes as nodes with effective observation distances to output the final positioning point coordinates. This algorithm can ensure that the positioning output has the minimum mean square error value.

[0080] Example 2

[0081] This embodiment also relates to an application of a method for positioning ship personnel based on a distributed deep network model. The method is deployed on a real ship and compared with other methods in an experiment.

[0082] In order to collect training samples, the 8.2*7.5m 2 Six anchor nodes are fixed at known locations on the bridge deck, three of which are fixed at non-collinear positions on the ceiling (the first to third anchor nodes), and the remaining anchor nodes (the fourth to fifth anchor nodes) are fixed 1 meter above the ground. Figure 3 shown.

[0083] First, the present invention conducted comparative experiments on the key parameters of the distributed deep network model. Communication volume is one of the most important metrics in positioning system design. Under fixed communication volume conditions, a higher number of local iterations results in a higher number of overall training times, a lower loss function, and better ranging accuracy. However, given the characteristics of distributed deep network models, this rule may not necessarily hold true. Therefore, experiments are necessary to determine the optimal number of local iterations for the system established by the present invention.

[0084] Under the same communication volume, the impact of different local iteration rounds on the loss function and ranging accuracy of the coordinator node after aggregation is as follows: Figure 4 and Figure 5 . Figure 4In the example, when the number of local training rounds B is less than 20, the training effect improves as B increases. This is because, given the same communication volume, a larger B allows for more local iterations, which leads to better model convergence. As B continues to increase, although each model undergoes more local training, the training effect is better for the local model. However, after these local models are aggregated, different anchor nodes optimize the model in different directions. This results in more local training rounds and more significant differences between them, causing the aggregated model to become unstable. Figure 5 In the ,distance measurement accuracy results are similar to the training results.,When B = 5, the model has a smaller loss function, and the distance,estimation error is less than 1.1 meters in 80% of the cases,,achieving the highest accuracy.

[0085] Then, the above optimal configuration is used to build a positioning system based on the distributed deep network model (NN (DCOL), Neural Network (Distributed Cluster Optimization Learning)), and a comparative experiment is conducted with other related ranging models. Other related ranging models include the centralized deep network ranging model (NN (Centr.), Neural Network (Central)), the multivariate linear regression ranging model (MLR: Multiple Linear Regression), and the single linear regression ranging model (SLR: Single Linear Regression) based on the received signal strength indicator. The ranging accuracy results are as follows: Figure 6 . Figure 6 The position estimation error of the univariate linear regression ranging model is the largest, which is due to the interference of changes in climate conditions and noise factors on the received signal strength indication during the experiment. The distributed deep network model (NN(DCOL)) of the present invention is better than other models, and its ranging error median and average values are both the smallest. This shows that the introduction of environmental variables improves the ranging accuracy. Due to the high-order nonlinear relationship between the input variables and the distance, the multivariate linear regression cannot accurately fit this relationship, and the deep neural network has the ability to fit arbitrary functions. Therefore, the average value of the ranging error of the multivariate linear regression ranging model (MLR) is greater than that of the deep network ranging model.

[0086] Finally, through statistical judgment, 22% of the ranging results in the experimental data have an error higher than 1 meter. According to the retention principle of formula (5), it can be calculated that in the positioning system of the distributed deep network model (NN (DCOL)) of the present invention, the number of retained nodes of the minimum error positioning algorithm should be four. At the same time, a minimum node positioning algorithm is set, that is, only three anchor nodes with the smallest ranging error are retained, and a maximum node positioning algorithm is set, that is, all anchor nodes are retained, and the final positioning output results are compared and analyzed. In the experimental environment, a circular trajectory is constructed, and 86 positioning points are collected at equal distances. Different algorithms are combined with the minimum error positioning algorithm to compare the path tracking effects. Based on the analysis of the actual positioning points, the range of the third quartile of the path tracking error is drawn, as shown in FIG. Figure 7 .from Figure 7 As can be seen from the figure, the third quartile of the univariate linear regression model (SLR) is approximately 0.73 meters, with the largest path tracking error. The third quartile of the distributed deep network model (NN(DCOL)) is approximately 0.26 meters, and the tracking error range is close to that of the centralized deep network model (NN(Centr.)), which is consistent with the distance estimation accuracy results.

[0087] In summary, this method improves the positioning accuracy and stability of the positioning system based on distance estimation. The final positioning error of this method is less than 0.96 meters in 80% of cases, and the node density is 0.11 per square meter.

[0088] Example 3

[0089] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0090] Many components in a device are connected to the I / O interface, including: input units, such as a keyboard and mouse; output units, such as various types of displays and speakers; storage units, such as magnetic disks and optical disks; and communication units, such as network cards, modems, and wireless communication transceivers. The communication unit allows the device to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunication networks.

[0091] The processing unit performs the various methods and processes described above. For example, in some embodiments, the method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the CPU can be configured to execute the method in any other appropriate manner (e.g., by means of firmware).

[0092] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0093] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0094] In the context of the present invention, machine-readable medium can be a tangible medium that can contain or store a program for use with an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. More specific examples of machine-readable storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0095] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A ship personnel positioning method based on a distributed deep network model, characterized in that: The method includes: Constructing and training a distributed deep network model, the distributed deep network model includes multiple wireless sensor network nodes, divided into a target node, a coordinator node, and several anchor nodes, and the distributed deep network model estimates the distance between the anchor node and the target node based on the received broadcast signal; After obtaining the distance estimation, the statistical redundancy reduction principle is introduced to design a minimum error positioning algorithm, and a ship and personnel positioning system based on the received signal strength indication is constructed to output the coordinates of the ship and personnel positioning points.

2. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: The target node is the node that needs to be located; the target node is responsible for broadcasting signals to the coordinator node and the anchor node.

3. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: Each anchor node deploys a lightweight standard neural network, whose hidden layer is a fully connected structure; the input of the lightweight standard neural network includes temperature, humidity, its own ID tag, the received signal strength of the target node, and the signal connection quality of the target node; the output of the lightweight standard neural network is the distance between the anchor node and the target node.

4. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: The coordinator node is responsible for the aggregation and distribution of parameters during the distributed deep network model training process: all anchor nodes send local network model parameters to the coordinator node. After receiving the parameters of all anchor nodes, the coordinator node aggregates the parameters to form global network model parameters, and then distributes the global network model parameters to each anchor node. Each anchor node continues to train using the local training set based on the global network model parameters.

5. A method for positioning ship personnel based on a distributed deep network model according to claim 4, characterized in that: The optimization goal of the local network model parameters of the anchor node is: Where f(w) is the sum of the sample mean square error and the regularization term, w is the local network model parameter of the anchor node, i is the index of the training data sample, d i is the distance label value corresponding to the i-th data sample, is the distance prediction value corresponding to the i-th data sample, and γ is the regularization weight coefficient.

6. A method for positioning ship personnel based on a distributed deep network model according to claim 5, characterized in that: The distributed deep network model uses stochastic gradient descent to find the global optimal value of the loss function; The anchor nodes continuously adjust the weight matrix through back propagation to minimize the loss function: Where, the subscript t represents the number of iterations, α(t) represents the learning rate, the superscript k represents the index of the anchor node, and w k is the weight matrix of the kth anchor node.

7. A method for positioning ship personnel based on a distributed deep network model according to claim 6, characterized in that: After multiple iterations locally, the anchor node sends the trained local network parameter model to the coordinator node for aggregation, specifically: Among them, B represents the number of local iterations of each anchor node, n is the total number of anchor nodes participating in the aggregation, is the weight matrix after training t+B rounds for the kth anchor node, w t+B Aggregated weight matrix for the coordinator node.

8. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: The statistical redundancy reduction principle is specifically as follows: Use n anchor nodes to perform ranging and obtain the ranging information from each anchor node to the target node, where n is greater than 3; Among all anchor nodes, some anchor nodes are retained as the final positioning output nodes. The retained anchor nodes are used as nodes with effective observation distances, and the minimum error positioning algorithm is used to output the final positioning point coordinates.

9. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: The specific principles for retaining some anchor nodes are as follows: Among them, M represents the number of retained anchor nodes, N represents the total number of anchor nodes, P represents the probability that the ranging error is greater than the set threshold, and Z + Is a positive integer.

10. A method for positioning ship personnel based on a distributed deep network model according to claim 1, characterized in that: In addition to having the function of an anchor node, the coordinator node is also equipped with a temperature and humidity sensor to distribute temperature and humidity information to the anchor node.

Images