Sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method

Abstract

The invention discloses a sensor algorithm parameter tuning method based on non-convex optimization and an entropy weight method, and the method comprises the steps: S1, setting an optimization objective function of a sensor, initializing the parameters of the sensor, and carrying out the data collection; s2, optimizing the optimization objective function of the sensor through a stochastic gradient descent method; s3, introducing a SiLU activation function to enhance the performance of the optimization process; s4, after local parameter tuning is completed, an entropy weight method is introduced to dynamically adjust the weight of the sensor node; and S5, ensuring that the tuning result is shared in the large-scale sensor network by combining centralized and distributed global data synchronization mechanisms. The invention discloses a sensor algorithm parameter tuning method based on non-convex optimization and an entropy weight method, and aims to solve the global optimization problem of sensor parameter tuning in a large-scale and complex environment. The method aims at the defect that a traditional tuning algorithm in a large-scale sensor network is limited to local optimum.

Description

Technical Field

[0001] The present invention belongs to the technical field of sensor parameter tuning, and in particular relates to a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method. Background Art

[0002] In large-scale sensor networks, the tuning of sensor parameters is crucial to ensure the accuracy of data collection and the stability of the system. Traditional tuning methods are mostly based on classical optimization algorithms (such as particle swarm and genetic algorithm), which perform well on local adjustment of parameters, but are prone to fall into local optimality in global optimization and complex environments in high-dimensional space. Although particle swarm optimization (PSO) has shown good global search capabilities in many optimization problems, it often encounters the situation of falling into local optimality when dealing with high-dimensional non-convex optimization problems due to the limitations of group search strategies, especially in complex, multi-peak non-convex spaces. PSO relies on the collaborative relationship between individuals and groups, which easily leads to the stability of the search direction and the loss of the ability to explore the global optimal. Genetic algorithm (GA), as an algorithm based on evolutionary mechanism, realizes the exploration of solution space through operations such as selection, crossover and mutation. However, GA usually requires longer iteration cycles to converge in high-dimensional complex environments, and is easily affected by local solutions in non-convex spaces, resulting in low search efficiency.

[0003] In addition, when faced with nonlinear and non-convex optimization problems, existing methods are insufficient when dealing with large-scale sensor networks. Therefore, a new tuning method is needed that can effectively cope with large-scale, multi-dimensional sensor network tuning in complex environments. Summary of the invention

[0004] The main purpose of the present invention is to provide a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method, aiming to solve the global optimization problem of sensor parameter tuning in large-scale and complex environments. In view of the shortcomings of traditional tuning algorithms in large-scale sensor networks that are limited to local optimality, the tuning efficiency and accuracy of the overall network are improved by optimizing the architecture and data synchronization mechanism. This method mainly solves the non-convex optimization problem through stochastic gradient descent (SGD) and non-convex projection gradient method, and introduces SiLU (Sigmoid-Weighted Linear Unit) activation function to improve the nonlinear performance and stability of the optimization process. At the same time, the entropy weight method dynamically adjusts the weights of sensor nodes, thereby improving the system's adaptability and global optimization capabilities.

[0005] To achieve the above objectives, the present invention provides a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method, comprising the following steps:

[0006] Step S1: Set the optimization objective function of the sensor and initialize sensor parameters and data collection;

[0007] Step S2: (to prevent falling into a local optimal solution) optimizing the optimization objective function of the sensor by using a stochastic gradient descent method;

[0008] Step S3: Introduce SiLU activation function to enhance the performance of the optimization process;

[0009] Step S4: After completing the local parameter tuning, the entropy weight method is introduced to dynamically adjust the weights of the sensor nodes;

[0010] Step S5: By combining centralized and distributed global data synchronization mechanisms, the tuning results are ensured to be shared within the large-scale sensor network.

[0011] As a further preferred technical solution of the above technical solution, for step S1:

[0012] Assume that the number collected by the sensor is x i , the sensor parameter is θ i , the optimization goal is to minimize the non-convex objective function f(θ), which is in the form of:

[0013]

[0014] Among them, w i represents the weight of the i-th sensor, g(θ i ,x i ) is a non-convex function that represents the relationship between sensor performance and its parameters and data.

[0015] As a further preferred technical solution of the above technical solution, for step S2:

[0016] First, the parameters are randomly initialized (θ0), and then the gradient of the objective function is calculated To update the sensor parameters, specifically:

[0017]

[0018] Among them, η is the learning rate, which controls the step size of parameter update; after each parameter update, the parameters are projected to ensure that the parameters are within the feasible domain, specifically:

[0019]

[0020] Among them, Proj C () is the projection function, C is the feasible domain of the constraints, which is related to the physical or application constraints of the sensor (such as energy consumption limits or upper and lower limits of parameters).

[0021] As a further preferred technical solution of the above technical solution, for step S3:

[0022] The formula of SiLU activation function is:

[0023]

[0024] (The SiLU activation function improves the smoothness and stability of the gradient update process through nonlinear processing, which can avoid the problem of gradient disappearance or explosion.) During the gradient descent process, the SiLU activation function is applied to the gradient update formula to optimize the nonlinear performance of the tuning. The updated gradient formula is:

[0025]

[0026] The gradient update after the SiLU activation function can effectively avoid falling into the local optimum in non-convex optimization and improve the convergence speed of the global optimum.

[0027] As a further preferred technical solution of the above technical solution, for step S4, the entropy weight method determines the importance (in the system) by calculating the change rate of sensor data, and dynamically adjusts the weight of each node according to the size of information entropy. The specific calculation method is: for each sensor node data change rate p ij Calculate information entropy H i :

[0028]

[0029] Among them, p ij represents the data change rate of the i-th sensor at the j-th moment, and the information entropy H i The higher the value, the greater the data change of the node, and the higher the priority is given in the tuning; therefore, the node weight is dynamically adjusted to:

[0030]

[0031] Through the dynamic adjustment mechanism, weights are adaptively allocated in real time, allowing nodes with drastic changes to respond to environmental changes more quickly, thereby improving the tuning efficiency of the entire network.

[0032] As a further preferred technical solution of the above technical solution, for step S5:

[0033] (In order to ensure that the tuning results can be shared in a large-scale sensor network, the present invention designs a hierarchical data synchronization mechanism based on a central node) First, the central node collects information including delay and load (etc.) of all nodes in the network to dynamically adjust the data synchronization frequency, and flexibly sets the synchronization interval through an adaptive control algorithm, so as to respond in time when the data flow changes (avoiding network congestion caused by too frequent synchronization and reducing the data lag problem caused by low-frequency synchronization), so (this dynamic frequency control method can) achieve efficient synchronization of the entire network according to the node scale and data flow;

[0034] During data transmission, in order to ensure the integrity of synchronized data, a packet loss detection and compensation mechanism is introduced. The central node monitors the transmission order and missing status of data packets. Once packet loss is detected, a compensation request is triggered immediately, requiring the corresponding node to resend the lost data packet to ensure complete data transmission. In order to reduce the impact of packet loss and delay on system stability, a delay processing algorithm is used to dynamically adjust the delay of data packets, and delayed data is processed by priority sorting to ensure the timeliness of key information.

[0035] The synchronization mechanism combines centralized and distributed solutions: the central node is responsible for the synchronization control of the entire network and the aggregation of key data, while the distributed nodes can also directly transmit local data to reduce the dependence on the central node and improve the response speed and stability of synchronization (the data synchronization method combining centralized and distributed methods helps to achieve stable, two-way data sharing and efficient data synchronization in large-scale sensor networks); after each local parameter tuning is completed, each sensor node uploads the tuning results to the central node through wireless communication, and the central node summarizes the tuning results of all nodes and calculates the global optimal parameter θ global , the formula is:

[0036]

[0037] The central node feeds back the global optimal parameters to each sensor node to ensure coordinated optimization of parameters on a global scale (this distributed data synchronization mechanism can not only greatly reduce the computing pressure of the central node, but also improve the tuning efficiency of the entire sensor network, especially in large-scale sensor networks).

[0038] The beneficial effects of the present invention are:

[0039] The present invention can effectively solve the non-convex optimization problem of complex sensor networks and avoid local optimal solutions. By combining non-convex optimization algorithms, stochastic gradient descent, projection method and SiLU activation function, the optimization process is smoother and the effect is better. The entropy weight method realizes dynamic weight adjustment and enhances the system adaptability and efficiency. The distributed data synchronization mechanism ensures global parameter sharing, performs stably in large-scale networks, and is suitable for complex scenarios such as smart cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method of the present invention. DETAILED DESCRIPTION

[0041] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0042] In the preferred embodiments of the present invention, those skilled in the art should note that the sensors and the like involved in the present invention may be regarded as prior art.

[0043] Preferred embodiments.

[0044] The present invention discloses a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method, comprising the following steps:

[0045] Step S1: Set the optimization objective function of the sensor and initialize sensor parameters and data collection;

[0046] Step S2: (to prevent falling into a local optimal solution) optimizing the optimization objective function of the sensor by using a stochastic gradient descent method;

[0047] Step S3: Introduce SiLU activation function to enhance the performance of the optimization process;

[0048] Step S4: After completing the local parameter tuning, the entropy weight method is introduced to dynamically adjust the weights of the sensor nodes;

[0049] Step S5: By combining centralized and distributed global data synchronization mechanisms, the tuning results are ensured to be shared within the large-scale sensor network.

[0050] Specifically, for step S1:

[0051] Assume that the number collected by the sensor is x i , the sensor parameter is θ i, the optimization goal is to minimize the non-convex objective function f(θ), which is in the form of:

[0052]

[0053] Among them, w i represents the weight of the i-th sensor, g(θ i ,x i ) is a non-convex function that represents the relationship between sensor performance and its parameters and data.

[0054] More specifically, for step S2:

[0055] First, the parameters are randomly initialized (θ0), and then the gradient of the objective function is calculated To update the sensor parameters, specifically:

[0056]

[0057] Among them, η is the learning rate, which controls the step size of parameter update; after each parameter update, the parameters are projected to ensure that the parameters are within the feasible domain, specifically:

[0058]

[0059] Among them, Proj C () is the projection function, C is the feasible domain of the constraints, which is related to the physical or application constraints of the sensor (such as energy consumption limits or upper and lower limits of parameters).

[0060] Furthermore, for step S3:

[0061] The formula of SiLU activation function is:

[0062]

[0063] (The SiLU activation function improves the smoothness and stability of the gradient update process through nonlinear processing, which can avoid the problem of gradient disappearance or explosion.) During the gradient descent process, the SiLU activation function is applied to the gradient update formula to optimize the nonlinear performance of the tuning. The updated gradient formula is:

[0064]

[0065] The gradient update after the SiLU activation function can effectively avoid falling into the local optimum in non-convex optimization and improve the convergence speed of the global optimum.

[0066] Furthermore, for step S4, the entropy weight method determines the importance (in the system) by calculating the change rate of sensor data, and dynamically adjusts the weight of each node according to the size of information entropy. The specific calculation method is: for each sensor node data change rate p ij Calculate information entropy H i :

[0067]

[0068] Among them, p ij represents the data change rate of the i-th sensor at the j-th moment, and the information entropy H i The higher the value, the greater the data change of the node, and the higher the priority is given in the tuning; therefore, the node weight is dynamically adjusted to:

[0069]

[0070] Through the dynamic adjustment mechanism, weights are adaptively allocated in real time, allowing nodes with drastic changes to respond to environmental changes more quickly, thereby improving the tuning efficiency of the entire network.

[0071] Preferably, for step S5:

[0072] (In order to ensure that the tuning results can be shared in a large-scale sensor network, the present invention designs a hierarchical data synchronization mechanism based on a central node) First, the central node collects information including delay and load (etc.) of all nodes in the network to dynamically adjust the data synchronization frequency, and flexibly sets the synchronization interval through an adaptive control algorithm, so as to respond in time when the data flow changes (avoiding network congestion caused by too frequent synchronization and reducing the data lag problem caused by low-frequency synchronization), so (this dynamic frequency control method can) achieve efficient synchronization of the entire network according to the node scale and data flow;

[0073] During data transmission, in order to ensure the integrity of synchronized data, a packet loss detection and compensation mechanism is introduced. The central node monitors the transmission order and missing status of data packets. Once packet loss is detected, a compensation request is triggered immediately, requiring the corresponding node to resend the lost data packet to ensure complete data transmission. In order to reduce the impact of packet loss and delay on system stability, a delay processing algorithm is used to dynamically adjust the delay of data packets, and delayed data is processed by priority sorting to ensure the timeliness of key information.

[0074] The synchronization mechanism combines centralized and distributed solutions: the central node is responsible for the synchronization control of the entire network and the aggregation of key data, while the distributed nodes can also directly transmit local data to reduce the dependence on the central node and improve the response speed and stability of synchronization (the data synchronization method combining centralized and distributed methods helps to achieve stable, two-way data sharing and efficient data synchronization in large-scale sensor networks); after each local parameter tuning is completed, each sensor node uploads the tuning results to the central node through wireless communication, and the central node summarizes the tuning results of all nodes and calculates the global optimal parameter θ global , the formula is:

[0075]

[0076] The central node feeds back the global optimal parameters to each sensor node to ensure coordinated optimization of parameters on a global scale (this distributed data synchronization mechanism can not only greatly reduce the computing pressure of the central node, but also improve the tuning efficiency of the entire sensor network, especially in large-scale sensor networks).

[0077] For the present invention, the stochastic gradient descent method is first used to process large-scale data, and a small number of random samples are selected in each iteration to calculate the gradient, thereby improving the computational efficiency. In the implementation process of the stochastic gradient descent method, the specific steps include initializing the learning rate and randomly selecting a sample set. In each iteration, the method randomly selects a subset from the data to calculate the gradient, and uses an adaptive learning rate to adjust the update step size of the algorithm. In order to avoid falling into the local optimum, SGD usually adopts a momentum strategy, that is, superimposing the update directions of the previous few times during the update process, thereby helping the search to jump out of the local area. In addition, when facing noisy data, SGD ensures convergence by gradually reducing the learning rate, and regularizes the data when necessary, thereby further improving the robustness of the optimization process.

[0078] The non-convex projected gradient method is used to constrain the updated solution within the feasible region to ensure the validity of the solution. At the same time, the entropy weight method dynamically adjusts the node weight according to the data information entropy of each sensor node, so that the node with higher information has a greater impact in the system. First, by analyzing the rate of change of sensor data, the data fluctuation of each node over a period of time is calculated to obtain the information entropy of each node. The weight adjustment process dynamically adjusts the node weight according to the information entropy value. The higher the entropy value, the greater the amount of information contained in the node, and the weight increases accordingly; otherwise, it decreases. This can respond quickly to environmental changes and improve the system's adaptability. Through this dynamic adjustment, the entropy weight method effectively avoids the problem of insufficient sensitivity to environmental changes in traditional fixed weight methods and enhances the stability of the system in a changing environment. By calculating the node weights and adjusting them in time during each optimization process, the system can maintain adaptability and global optimization effects in a changing environment, thereby achieving higher accuracy and robustness in positioning and navigation.

[0079] It is worth mentioning that the technical features such as sensors involved in the patent application of this invention should be regarded as prior art. The specific structure, working principle and possible control method and spatial layout method of these technical features can be selected by conventional methods in the field, and should not be regarded as the inventive point of this patent. This patent will not be further elaborated.

[0080] For those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method, characterized in that: The following steps are involved: Step S1: Set the optimization objective function of the sensor and initialize sensor parameters and data collection; Step S2: Optimizing the optimization objective function of the sensor by stochastic gradient descent method; Step S3: Introduce SiLU activation function to enhance the performance of the optimization process; Step S4: After completing the local parameter tuning, the entropy weight method is introduced to dynamically adjust the weights of the sensor nodes; Step S5: By combining centralized and distributed global data synchronization mechanisms, the tuning results are ensured to be shared within the large-scale sensor network.

2. According to claim 1, a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method is characterized in that: For step S1: Assume that the number collected by the sensor is x i , the sensor parameter is θ i , the optimization goal is to minimize the non-convex objective function f(θ), which is in the form of: Among them, w i represents the weight of the i-th sensor, g(θ i ,x i ) is a non-convex function that represents the relationship between sensor performance and its parameters and data.

3. According to claim 2, a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method is characterized in that: For step S2: First, the parameters are randomly initialized, and then the gradient of the objective function is calculated To update the sensor parameters, specifically: Among them, η is the learning rate, which controls the step size of parameter update; after each parameter update, the parameters are projected to ensure that the parameters are within the feasible domain, specifically: Among them, proj C () is the projection function, C is the feasible domain of the constraint conditions, which is related to the physical or application constraints of the sensor.

4. According to claim 3, a sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method is characterized in that: For step S3: The formula of SiLU activation function is: During the gradient descent process, the SiLU activation function is applied to the gradient update formula to optimize the nonlinear performance of the tuning. The updated gradient formula is: The gradient update after the SiLU activation function can effectively avoid falling into the local optimum in non-convex optimization and improve the convergence speed of the global optimum.

5. The sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method according to claim 4 is characterized in that: For step S4, the entropy weight method determines the importance by calculating the change rate of sensor data and dynamically adjusts the weight of each node according to the size of information entropy. The specific calculation method is: for each sensor node data change rate p ij Calculate information entropy H i : Among them, p ij represents the data change rate of the i-th sensor at the j-th moment, and the information entropy H i The higher the value, the greater the data change of the node, and the higher the priority is given in the tuning; therefore, the node weight is dynamically adjusted to: Through the dynamic adjustment mechanism, weights are adaptively allocated in real time, allowing nodes with drastic changes to respond to environmental changes more quickly, thereby improving the tuning efficiency of the entire network.

6. The sensor algorithm parameter tuning method based on non-convex optimization and entropy weight method according to claim 5 is characterized in that: For step S5: First, the central node collects information about all nodes in the network, including delay and load, to dynamically adjust the frequency of data synchronization and flexibly set the synchronization interval through an adaptive control algorithm, so as to respond promptly when data traffic changes, thereby achieving efficient synchronization of the entire network according to the node scale and data traffic; During data transmission, in order to ensure the integrity of synchronized data, a packet loss detection and compensation mechanism is introduced. The central node monitors the transmission order and missing status of data packets. Once packet loss is detected, a compensation request is triggered immediately, requiring the corresponding node to resend the lost data packet to ensure complete data transmission. In order to reduce the impact of packet loss and delay on system stability, a delay processing algorithm is used to dynamically adjust the delay of data packets, and delayed data is processed by priority sorting to ensure the timeliness of key information. The synchronization mechanism combines centralized and distributed solutions: the central node is responsible for the synchronization control of the entire network and the aggregation of key data, while the distributed nodes directly transmit local data to reduce the dependence on the central node and improve the response speed and stability of synchronization; after each local parameter tuning is completed, each sensor node uploads the tuning results to the central node through wireless communication, and the central node summarizes the tuning results of all nodes and calculates the global optimal parameter θ global , the formula is: The central node feeds back the global optimal parameters to each sensor node to ensure the coordinated optimization of parameters on a global scale.

Images