An intelligent filtering method and system based on deep learning

Abstract

The present invention discloses an intelligent filtering method and system based on deep learning, which relates to the field of data filtering technology. It is to solve the technical problem that when using an intelligent optimization algorithm to optimize filter parameters, the optimization performance and computational efficiency conflict with each other. Obtain interference control instruction data; extract the individuals that each data passes through in the late optimization stage; after classifying the data, combine the late optimization individuals corresponding to the same type of data to obtain various sub-search spaces; after the filtered control instruction data is classified, call this type of sub-search space to optimize the filter parameters. The present invention extracts the late optimization individuals to focus on some filtering parameters that are close to the optimal filtering effect on the interference control instruction data; combines the late optimization individuals to obtain sub-search spaces unique to each type of data, and when processing the filtered control instruction data, only searches in the corresponding sub-search spaces, which reduces the search range and improves the search performance.

Description

Technical Field

[0001] The present invention relates to the field of data filtering technology, and in particular to an intelligent filtering method and system based on deep learning. Background Art

[0002] During actual operation, unmanned systems are inevitably subject to interference from various internal and external factors, such as sensor noise, electromagnetic interference, and environmental disturbances. These interferences can affect the transmission and execution of control commands. To ensure that unmanned systems can accurately control as expected, control commands need to be filtered to suppress noise and interference, thereby effectively extracting control command information.

[0003] The process of filtering control commands often requires setting parameters in the filtering algorithm. To address the difficulty of adapting to the traditional problem of manually presetting filter parameters, researchers use intelligent optimization algorithms to adaptively find a set of filter parameters that significantly improves the filtering effect. In the intelligent optimization process of filter parameters, a search space containing a large number of alternative filter parameters is first defined. Each individual in the search space represents a set of possible filter parameters. The intelligent optimization algorithm then simulates a natural phenomenon (birds foraging or ants foraging) to allow each individual to move in the search space and updates the individual's position and velocity according to the fitness function. The optimal filter parameters are approached through continuous iteration and updating of the position and velocity.

[0004] However, when filtering each new control instruction data to be processed, the intelligent optimization algorithm needs to perform extensive parameter searches and multiple rounds of calculations in the entire search space, which will cause computing resource consumption and time delays, extend the response time of the unmanned system, and thus affect its real-time decision-making ability and stability. Summary of the Invention

[0005] Based on this, in order to solve the technical problem in the prior art of the conflict between optimization performance and computational efficiency when using intelligent optimization algorithms to optimize filter parameters, the present invention provides an intelligent filtering method and system based on deep learning.

[0006] The present invention provides an intelligent filtering method based on deep learning, comprising:

[0007] Adding noise signals of different types and intensities to the collected real control instruction data to obtain a plurality of interference control instruction data;

[0008] A global search space for filter parameters is constructed, which contains several individuals. Each individual represents a set of filter parameters. The optimization goal is to minimize the gap between the denoised interference control command data obtained after filtering the interference control command data and the true control command data:

[0009] ;

[0010] in, is interference control instruction data; Interference control instruction data De-noised interference control instruction data obtained after filtering; is the filter, are the filter parameters; is the global search space; Indicates that in the global search space Find a set of optimal filter parameters in , which makes the gap between the denoised interference control command data and the real control command data smallest;

[0011] By adaptively adjusting the inertia weight used to control the particle search range in the original particle swarm optimization algorithm, the individual fluctuations experienced by the original particle swarm optimization algorithm in the later stage of optimization are reduced, and an improved particle swarm optimization algorithm is obtained. The improved particle swarm optimization algorithm is used in the global search space. The improved particle swarm optimization algorithm specifically includes:

[0012] I: Initialize the position and velocity of each individual in the particle swarm optimization algorithm;

[0013] II: is the fitness function, and the fitness value of each individual is calculated;

[0014] III: For each individual, its current fitness value and its best historical position Compare the corresponding fitness values. If the current individual's fitness value is high, the current individual position will be updated to the historical best position. ;

[0015] IV: For each individual, compare its current fitness value with the global best position Compare the corresponding fitness values. If the current fitness value is high, the current individual position will be updated to the global best position. ;

[0016] V: Update the speed and position of each individual:

[0017] ;

[0018] ;

[0019] ;

[0020] in, and In the t Second and t -1 iteration individual i In the d Dimensional speed; and In the t Second and t -1 iteration individual i In the d Dimensional location; and are the first learning factor and the second learning factor used to adjust the maximum learning step size; and are the first random number and the first random number used to increase the randomness of the search, and All obey uniform distribution in the interval [0,1]; It is in t The individual i In the d The optimal position of the dimension; is an individual i The global optimal position of It is in t Adjust the inertia weight of the search range at each iteration; is the maximum number of iterations; is the adjustment coefficient;

[0021] The meaning of the above formula is: inertia weight As the number of iterations t In the later stage of optimization, the speed of driving individuals also decreases with the number of iterations. t decreases with the increase of , so as to reduce the individual fluctuations that the original particle swarm optimization algorithm goes through in the later stage of optimization;

[0022] VI: Determine whether the maximum number of iterations has been reached. If so, stop the iteration; otherwise, return to II.

[0023] Extract the individuals that each interference control instruction data passes through in the later stage of optimization, including:

[0024] Get the first t -1 iteration, t The iteration and t +1 iteration fitness value fittness t-1 、 fittness t and fittness t+1 ;

[0025] Calculate the difference in the rate of change of fitness values ​​between adjacent iterations :

[0026] ;

[0027] When the difference in the rate of change of fitness values ​​in adjacent iterations When it is less than the set threshold, t The number of iterations is taken as the starting point of the late optimization stage, and the maximum number of iterations is taken as the end point of the late optimization stage, and the individuals passed from the starting point of the late optimization stage to the end point of the late optimization stage are obtained;

[0028] According to the similarity of the interference control instruction data, similar interference control instruction data are clustered into one category through a clustering algorithm, and several interference control instruction data are divided into different categories. The individuals corresponding to the interference control instruction data of the same category that have passed through the optimization stage are combined to obtain the sub-search space corresponding to each type of interference control instruction data;

[0029] Obtain the control instruction data to be filtered, classify the control instruction data to be filtered according to the similarity between the control instruction data to be filtered and various types of interference control instruction data, retrieve the sub-search space corresponding to the type of interference control instruction data, and use the optimization algorithm to search the filter parameters in the sub-search space.

[0030] The present invention provides an intelligent filtering system based on deep learning, comprising:

[0031] The data acquisition module is used to add noise signals of different types and intensities to the collected real control instruction data to obtain a plurality of interference control instruction data;

[0032] The sub-search space individual acquisition module is used to construct a global search space for filter parameters containing several individuals, where each individual represents a set of filter parameters. The optimization goal is to minimize the gap between the denoised interference control command data obtained after filtering the interference control command data and the true control command data. An optimization algorithm is used to search in the global search space to extract the individuals that each interference control command data passes through in the later stage of the optimization.

[0033] A sub-search space construction module is used to classify a number of interference control instruction data into different categories according to their similarity, and to combine the individuals corresponding to the interference control instruction data of the same category that have passed through the optimization stage to obtain sub-search spaces corresponding to each category of interference control instruction data;

[0034] The filtering module is used to obtain the control instruction data to be filtered, classify the control instruction data to be filtered according to the similarity between the control instruction data to be filtered and various types of interference control instruction data, call the sub-search space corresponding to the interference control instruction data of this type, and use the optimization algorithm to search the filter parameters in the sub-search space.

[0035] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0036] In the intelligent filtering method based on deep learning provided by the present invention, by extracting the individuals that the optimization algorithm passes through in the later stage of optimization, the focus is on the part of filtering parameters whose filtering effect on interference control instruction data gradually approaches the optimal one; the later stage individuals corresponding to similar interference control instruction data are combined to obtain sub-search spaces unique to each type of interference control instruction data. When processing the filtered control instruction data, only the corresponding sub-search space is used for search, which reduces the search range while increasing the probability of obtaining the optimal filtering parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0038] Figure 1 A schematic diagram of a flow chart of an intelligent filtering method based on deep learning provided by the present invention;

[0039] Figure 2 A schematic diagram of the process of constructing a global search space for filter parameters provided by the present invention;

[0040] Figure 3 A schematic diagram of the process of the improved particle swarm optimization algorithm provided by the present invention;

[0041] Figure 4 A schematic diagram of a computer device for implementing an intelligent filtering method based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0043] Unmanned systems are becoming increasingly widespread, with their autonomy and intelligence levels continuously improving. Control commands in unmanned systems are signals or commands used to guide and adjust the system's behavior and state. These commands are typically issued via a remote control or ground control station and transmitted via wireless communication links to the system's flight control system or corresponding actuators. However, during actual operation, unmanned systems are inevitably subject to interference from various internal and external factors, such as sensor noise, electromagnetic interference, and environmental disturbances. These interferences can affect the transmission and execution of control commands. To ensure that unmanned systems can accurately control as intended, filtering control commands is particularly important.

[0044] However, when filtering and optimizing based on the intelligent optimization algorithm, when filtering each new control instruction data to be processed, the intelligent optimization algorithm needs to perform extensive parameter searches and multiple rounds of calculations in the entire search space, which will cause computing resource consumption and time delays, extend the response time of the unmanned system, and thus affect its real-time decision-making ability and stability.

[0045] Based on this, the present invention provides an intelligent filtering method based on deep learning. The method obtains the optimal filtering parameter range corresponding to each interference control instruction by extracting the individuals (filtering parameters) passed by the interference control instruction in the later stage of optimization, and combines the individuals passed by similar interference control instructions in the later stage of optimization to obtain a sub-search space of similar interference control instructions. The subspace contains fewer individuals than the global search space, and the individuals contained in the subspace are more suitable for this type of interference control instruction. Searching within the search space can not only improve the quality of the optimal solution obtained, but also significantly reduce the search range and reduce the response time.

[0046] Example 1

[0047] Figure 1 The flow chart of the intelligent filtering method based on deep learning in this embodiment is shown. Figure 1 The method is described in detail and specifically comprises the following steps:

[0048] S1: adding noise signals of different types and intensities to a plurality of collected real control instruction data to obtain a plurality of interference control instruction data.

[0049] S101: Collecting real control instruction data.

[0050] Control command data in unmanned systems typically refers to control information transmitted from a control end (such as a remote control, autonomous driving algorithm, or ground control station) to a drone, unmanned vehicle, or other unmanned device. This control command data may include control commands, status data, and sensor input. The control end generates control commands and encapsulates them into data packets. These data packets are then transmitted to the corresponding unmanned device's receiving module via wireless communications such as Wi-Fi, 4G / 5G, or LoRa. Upon receiving the data packets, the corresponding unmanned device verifies them. If verification passes, the drone interprets the control commands and performs corresponding actions, such as adjusting flight altitude or heading. Each data packet contains a timestamp, source address, destination address, protocol, length, and packet information. Table 1 shows the contents of a data packet corresponding to a set of control commands sent from a ground control end to a drone to adjust the drone's flight altitude and heading. It should be understood that the information in Table 1 is merely an example, and the collected data is not limited to that in Table 1.

[0051] Table 1 Data packet contents corresponding to the control instructions for adjusting the flight altitude and heading of the UAV

[0052]

[0053] S102: Add noise to the real control instruction data to obtain interference control instruction data.

[0054] Noise is added to the actual control command data during unmanned system data transmission to simulate interference in the communication channel. First, different noise types and intensities are selected, including but not limited to Gaussian noise, salt and pepper noise, Poisson noise, and impulse noise. The noise is then added to the collected actual control commands. For example, after extracting the data packet list data, different Gaussian noise means and variances are set to simulate noise of varying intensities. Based on the set Gaussian noise mean and variance, a normal distribution is used to generate Gaussian noise of the same length as the data packet list. This generated noise is then added to the original control command data to obtain the interference control command data.

[0055] In step S1 , noises of different types and intensities are added to real control instruction data to simulate various real noises, thereby obtaining a plurality of interference control instruction data.

[0056] S2: Construct a global search space for filter parameters containing several individuals, where each individual represents a set of filter parameters. The optimization goal is to minimize the gap between the denoised interference control command data obtained after the interference control command data is filtered by the filter and the true control command data. Use the optimization algorithm to search in the global search space and extract the individuals that each interference control command data passes through in the later stage of optimization.

[0057] Traditional filtering methods often rely on manually preset parameter settings to achieve effective signal filtering. However, this manual approach is time-consuming and labor-intensive, and it is difficult to ensure that parameters can be adaptively adjusted to adapt to changing signal characteristics. To address this challenge, researchers have begun exploring the use of intelligent optimization algorithms, such as particle swarm optimization, to automatically search and determine filter parameters. Optimization algorithms, as a swarm intelligence-based optimization method, mimic natural phenomena or processes by transforming the filter parameter search process into an iterative search in a solution space. Each individual in the search space represents a set of possible filter parameters. They move through the solution space based on their position and velocity, and update their position and velocity according to a fitness function. To solve the filter parameter optimization problem, researchers first define a search space containing a large number of candidate parameters. Then, they use intelligent optimization algorithms to conduct a comprehensive search within this space to find the parameter combination that achieves the optimal filtering effect. This process is automatic and requires no human intervention, significantly improving the adaptive capabilities of the filtering method. Through continuous iteration and updates, the intelligent optimization algorithm gradually approaches the optimal solution, resulting in a set of parameters that significantly improves filtering performance.

[0058] S201: Construct a global search space for filter parameters containing several individuals.

[0059] In the field of signal processing, filters can filter and process signals according to specific frequency characteristics or statistical characteristics. Different filters have their own unique preset parameters, which determine the performance and effect of the filter. The preset parameters of the Gaussian filter include mean μ and variance σ The mean determines the center position of the Gaussian function, while the variance determines the width of the Gaussian function, that is, the degree of smoothness. The preset parameters of the Kalman filter include the process noise covariance Q and the observation noise covariance R. The process noise covariance describes the uncertainty in the system state changes, while the observation noise covariance describes the uncertainty in the observed data. The preset parameters of frequency domain filters (including low-pass filters, high-pass filters, and band-pass filters) include the cutoff frequency fc and the filter order N. The cutoff frequency determines the frequency range that the filter allows or blocks. The filter order affects the filter's performance, such as the transition bandwidth and stopband attenuation.

[0060] After clarifying the filter and its corresponding preset parameters, define the value range of each preset parameter. This value range needs to ensure that it can cover most actual application scenarios and filtering requirements. Combining these value ranges defines the search space of the entire parameter of the corresponding filter. Taking the low-pass filter as an example, the process of constructing the global search space of the filter parameters is as follows: Figure 2 Shown, including:

[0061] S2011: Determine preset parameters of the filter. The preset parameters of the low-pass filter include cutoff frequency fc, filter order N, passband ripple, and stopband attenuation.

[0062] S2012: Determine the value range of each preset parameter, such as setting the cutoff frequency fc between 1kHz and 10kHz, the filter order N between 1 and 25, the passband ripple between 0.1dB and 4dB, and the stopband attenuation between 10dB and 100dB.

[0063] S2031: Discretize the value range using uniform discretization or logarithmic discretization. Taking uniform discretization as an example: discretize the value range of the cutoff frequency fc with a step size of 100 Hz, discretize the value range of the filter order N with a step size of 1, discretize the value range of the passband ripple with a step size of 0.1 dB, and discretize the value range of the stopband attenuation with a step size of 2 dB.

[0064] S2014: Combine the discretized preset parameter values ​​to obtain the global search space. This operation can be achieved through Python's "itertools.product" function. The size of the global search space is N It is the product of the number of discretization points of all preset parameter ranges:

[0065] N=m 1 × ... × m i × ... × m n ;

[0066] in, n is the number of preset parameters in the filter, m i Indicates that the i-th preset parameter has m i A value.

[0067] S202: An optimization algorithm is used to search for an optimal solution in a global search space, with the optimization goal being to minimize the gap between the denoised interference control instruction data obtained after filtering the interference control instruction data and the true control instruction data.

[0068] In order to find the optimal filter configuration in the global search space, the particle swarm optimization algorithm is used to adjust the filter parameters so that the gap between the denoised interference control command data processed by the filter and the real interference-free control command data is minimized, thereby achieving the best filtering effect. The optimization goal is constructed as follows:

[0069] ;

[0070] in, is interference control instruction data; Interference control instruction data De-noised interference control instruction data obtained after filtering; is the filter, are the filter parameters; is the global search space; Indicates that in the global search space Find a set of optimal filter parameters in , which makes the gap between the denoised interference control command data and the real control command data Minimum.

[0071] Then the improved particle swarm optimization algorithm is used to search the global The improved particle swarm optimization algorithm adaptively adjusts the inertia weight used to control the particle search range in the original particle swarm optimization algorithm to reduce the individual fluctuations experienced by the original particle swarm optimization algorithm in the later stage of optimization. The process of the improved particle swarm optimization algorithm is as follows Figure 3 As shown, the specific steps include:

[0072] I: Initialize the position and velocity of each individual in the particle swarm optimization algorithm.

[0073] II: is the fitness function, and the fitness value of each individual is calculated.

[0074] III: For each individual, its current fitness value and its best historical position Compare the corresponding fitness values. If the current individual's fitness value is high, the current individual position will be updated to the historical best position. .

[0075] IV: For each individual, compare its current fitness value with the global best position Compare the corresponding fitness values. If the current fitness value is high, the current individual position will be updated to the global best position. .

[0076] V: Update the speed and position of each individual:

[0077] ;

[0078] ;

[0079] ;

[0080] in, and In the t Second and t -1 iteration individual i In the d Dimensional speed; and In the t Second and t -1 iteration individual i In the d Dimensional location; and are the first learning factor and the second learning factor used to adjust the maximum learning step size; and are the first random number and the first random number used to increase the randomness of the search, and All obey uniform distribution in the interval [0,1]; It is in t The individual i In the d The optimal position of the dimension; is an individual i The global optimal position of It is in t Adjust the inertia weight of the search range at each iteration; is the maximum number of iterations; is the adjustment coefficient; the meaning of the above formula is: inertia weight As the number of iterations t In the later stage of optimization, the speed of driving individuals also decreases with the number of iterations. t , so as to reduce the individual fluctuations that the original particle swarm optimization algorithm goes through in the later stage of optimization.

[0081] VI: Determine whether the maximum number of iterations has been reached. If so, stop the iteration; otherwise, return to II.

[0082] S203: Extracting the individuals that each interference control instruction data passes through in the later stage of optimization.

[0083] Get the first t -1 iteration, t The iteration and t +1 iteration fitness value fittness t-1 、 fittness t and fittness t+1 ; Calculate the difference in the rate of change of fitness values ​​between adjacent iterations :

[0084] .

[0085] When the difference in the rate of change of fitness values ​​in adjacent iterations When it is less than the set threshold, that is, the individual changes of adjacent iterations are relatively stable, so that the corresponding fitness value fluctuation frequency changes are based on the first t The number of iterations is taken as the starting point of the late optimization stage, and the maximum number of iterations is taken as the end point of the late optimization stage. The individual from the starting point of the late optimization stage to the end point of the late optimization stage is obtained, that is, the first t All individual positions at the iteration t+ All individual positions at iteration 1, ..., T The positions of all individuals at the iteration.

[0086] Step S2 constructs a global search space for filter parameters, which covers multiple filter parameter combinations with a larger value range. In order to find the optimal filter parameters in this wide search space, an improved particle swarm algorithm is used for optimization. As the optimization proceeds, the algorithm can gradually focus on a small number of parameter combinations in the later stage of the search, and these combinations gradually approach the optimal solution in terms of the filtering effect of the interference control instruction data. Specifically, the fitness value of the filter parameters tends to the optimal solution, while the corresponding parameter combination range is significantly reduced, which indicates that the convergence of the search process is getting stronger and stronger, and the optimization process gradually focuses on the optimal solution space area.

[0087] S3: Divide several interference control instruction data into different categories according to their similarity, combine the individuals corresponding to the same category of interference control instruction data that have passed through in the later stage of optimization, and obtain the sub-search space corresponding to each category of interference control instruction data.

[0088] S301: Classify a plurality of interference control instruction data into different categories.

[0089] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is used to automatically group similar interference control command data into clusters. DBSCAN can classify similar interference control command data into a single cluster without requiring labels, while minimizing the differences between different clusters. The interference control command data input to the DBSCAN algorithm is the feature vector obtained after feature extraction using a variational autoencoder (VAE). The VAE model consists of an encoder and a decoder. The encoder maps the input control command data to a probability distribution in a latent space, while the decoder maps samples in the latent space back to the original data space, reconstructing the input data. The loss function used in VAE training typically consists of two parts: reconstruction loss, which measures the difference between the reconstructed data and the original data, and KL divergence, which measures the difference between the latent representation and the prior distribution. After model training is complete, the encoder can be used to map new control command data into the latent space to obtain the corresponding feature vector.

[0090] S302: Combining the individuals corresponding to the same type of interference control instruction data that have passed through in the later stage of optimization.

[0091] Since each interference control instruction data corresponds to a filtering parameter optimization process, that is, each interference control instruction data has a "first t All individual positions at the iteration t+ All individual positions at iteration 1, ..., T All individual positions at the time of the iteration”; after classifying the interference control instruction data, the corresponding “ t All individual positions at the iteration t+ All individual positions at iteration 1, ..., T All individual positions at the iteration also have category attributes. By combining the individuals passed through the later stages of optimization with the same category attributes, the individuals corresponding to this type of noise can be obtained.

[0092] In order to improve the efficiency of interference control instruction data processing, step S3 constructs a smaller sub-search space by combining the solution space regions corresponding to interference control instruction data of specific similar properties. Compared with the global search space, the scope of this sub-search space is significantly reduced. And because the sub-search space is composed of multiple solution space regions focused on the optimal solution, the probability of obtaining the optimal filter parameters can be greatly improved. Specifically, by focusing on the area most likely to contain the optimal solution, the search range can be narrowed more efficiently, thereby accelerating the search process and improving the quality of the final solution. Compared with the traditional global search method, the construction of the sub-search space significantly improves the convergence speed and performance of the algorithm by reducing redundant calculations and exploring irrelevant solution areas. In addition, the design of the sub-search space not only takes into account the focus of the optimal solution area, but also combines the similarity characteristics of the interference control instruction data, so that the optimization process can better adapt to different types of data distributions, thereby improving the effect of filter parameter optimization.

[0093] S4: Obtain the control instruction data to be filtered, classify the control instruction data to be filtered according to the similarity between the control instruction data to be filtered and various types of interference control instruction data, retrieve the sub-search space corresponding to the type of interference control instruction data, and use the optimization algorithm to search the filter parameters in the sub-search space.

[0094] First, the control command data to be filtered is preprocessed. After data preprocessing is complete, the gap or similarity between the control command data to be filtered and the various types of interference control command data is calculated. The gap can be measured by calculating the Euclidean distance or Manhattan distance between the two sets of data. The similarity can be calculated using similarity metrics such as cosine similarity or the Pearson correlation coefficient. After calculating the gap or similarity, a set of values ​​is obtained, representing the difference or similarity between the data to be filtered and the various types of interference control command data. By comparing these results, the control command data to be filtered is classified into the category with the smallest calculated gap or the greatest similarity. In other words, the data to be filtered is assigned to the interference data category that is most similar to it. The sub-search space corresponding to this type of interference control command data is retrieved, and the filter parameters are searched within this sub-search space using an optimization algorithm.

[0095] based on Figure 1 The deep learning-based intelligent filtering method shown in the figure extracts the individuals that the optimization algorithm passes through in the later stage of optimization, focusing on the filtering parameters whose filtering effect on the interference control instruction data gradually approaches the optimal one; the individuals in the later stage of optimization corresponding to similar interference control instruction data are combined to obtain a sub-search space unique to each type of interference control instruction data. When processing the filtered control instruction data, only the corresponding sub-search space is used for search, which reduces the search range while increasing the probability of obtaining the optimal filtering parameters.

[0096] The above is an intelligent filtering method based on deep learning provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding intelligent filtering system based on deep learning, such as Figure 4 Shown, including:

[0097] The data acquisition module is used to add noise signals of different types and intensities to the collected real control instruction data to obtain a plurality of interference control instruction data.

[0098] The sub-search space individual acquisition module is used to construct a global search space for filter parameters containing several individuals, where each individual represents a set of filter parameters. The optimization goal is to minimize the gap between the denoised interference control instruction data obtained after the filter filters the interference control instruction data and the true control instruction data. The optimization algorithm is used to search in the global search space to extract the individuals that each interference control instruction data passes through in the later stage of the optimization.

[0099] The sub-search space construction module is used to classify a number of interference control instruction data into different categories according to the similarity of the interference control instruction data, and to combine the individuals corresponding to the interference control instruction data of the same category that have passed through the late optimization stage to obtain the sub-search space corresponding to each type of interference control instruction data.

[0100] The filtering module is used to obtain the control instruction data to be filtered, classify the control instruction data to be filtered according to the similarity between the control instruction data to be filtered and various types of interference control instruction data, call the sub-search space corresponding to the interference control instruction data of this type, and use the optimization algorithm to search the filter parameters in the sub-search space.

[0101] For the specific definition of the intelligent filtering system based on deep learning, please refer to the definition of the intelligent filtering method based on deep learning above, which will not be repeated here. The various modules in the above-mentioned intelligent filtering system based on deep learning can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0102] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 Provided is an intelligent filtering method based on deep learning.

[0103] The present invention also provides a computer device structure. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 Provided is an intelligent filtering method based on deep learning.

[0104] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0105] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. An intelligent filtering method based on deep learning, characterized in that: include: Adding noise signals of different types and intensities to the collected multiple real control command data to obtain multiple interference control command data; the real control command data is the control command data in the unmanned system; A global search space for filter parameters is constructed, containing several individuals, where each individual represents a set of filter parameters. The optimization objective is to minimize the gap between the denoised interference control command data obtained after filtering the interference control command data and the true control command data. An optimization algorithm is used to search for the best fit in the global search space, extracting the individuals that each interference control command data passes through in the later stages of the search. According to the similarity of the interference control instruction data, several interference control instruction data are divided into different categories, and the individuals corresponding to the interference control instruction data of the same category that have passed through the optimization stage are combined to obtain the sub-search space corresponding to each type of interference control instruction data; Obtaining control instruction data to be filtered, classifying the control instruction data to be filtered according to similarity between the control instruction data to be filtered and various types of interference control instruction data, retrieving a sub-search space corresponding to the type of interference control instruction data, and using an optimization algorithm to search for filter parameters in the sub-search space; The optimization goal is to minimize the gap between the denoised interference control instruction data obtained after filtering the interference control instruction data and the real control instruction data, and to use an optimization algorithm to search for the best in the global search space, specifically including: Build optimization goals: in, is interference control instruction data; Interference control instruction data The denoised interference control command data obtained after filtering; filter(·) is the filter, θ is the filter parameter; Ω is the global search space; Represents finding a set of optimal filter parameters θ in the global search space Ω optimal , which makes the gap between the denoised interference control command data and the real control command data smallest; By adaptively adjusting the inertia weight used to control the particle search range in the original particle swarm optimization algorithm, the individual fluctuations experienced by the original particle swarm optimization algorithm in the later stage of optimization are reduced, and an improved particle swarm optimization algorithm is obtained. The improved particle swarm optimization algorithm is used to search for the best solution in the global search space Ω. The extracting of the individual that each interference control instruction data passes through in the later stage of optimization specifically includes: Get the fitness value of the t-1th iteration, the tth iteration and the t+1th iteration t-1 、fitness t and fitness t+1 ; Calculate the difference Δt of the rate of change of fitness values ​​between adjacent iterations: When the difference Δt between the rates of change of fitness values ​​of adjacent iterations is less than the set threshold, the tth iteration is taken as the starting point of the late optimization stage and the maximum number of iterations is taken as the end point of the late optimization stage, and the individuals passed from the starting point of the late optimization stage to the end point of the late optimization stage are obtained.

2. The intelligent filtering method based on deep learning according to claim 1, characterized in that: The improved particle swarm optimization algorithm specifically includes: I: Initialize the position and velocity of each individual in the particle swarm optimization algorithm; II: is the fitness function, and the fitness value of each individual is calculated; III: For each individual, its current fitness value and its individual historical best position pbest i Compare the corresponding fitness values. If the current individual's fitness value is high, the current individual position will be updated to the historical best position pbest. i ; IV: For each individual, compare its current fitness value with the fitness value corresponding to the global best position gbest. If the current fitness value is higher, update the current individual position to the global best position gbest. V: Update the speed and position of each individual: in, and are the velocities of individual i in the dth dimension at the tth and t-1th iterations, respectively; and are the positions of individual i in the dth dimension at the tth and t-1th iterations respectively; c1 and c2 are the first learning factor and the second learning factor used to adjust the maximum learning step size respectively; r1 and r2 are the first random number and the first random number used to increase the randomness of the search respectively, and both r1 and r2 obey uniform distribution in the interval [0,1]; pbest id is the optimal position of individual i in the dth dimension at the tth iteration; gbest d is the global optimal position of individual i; w(t) is the inertia weight for adjusting the search range at the tth iteration; T is the maximum number of iterations; γ is the adjustment coefficient; The meaning of the above formula is: the inertia weight w(t) decreases with the increase of the number of iterations t, and the moving speed of the driving individual in the later stage of optimization also decreases with the increase of the number of iterations t, so as to reduce the individual fluctuations experienced by the original particle swarm optimization algorithm in the later stage of optimization; VI: Determine whether the maximum number of iterations has been reached. If so, stop the iteration; otherwise, return to II.

3. The intelligent filtering method based on deep learning according to claim 1, characterized in that: The method of classifying a plurality of interference control instruction data into different categories according to the similarity of the interference control instruction data is to cluster similar interference control instruction data into one category through a clustering algorithm.

4. An intelligent filtering system based on deep learning, characterized in that: include: The data acquisition module is used to add noise signals of different types and intensities to the collected real control instruction data to obtain a plurality of interference control instruction data; The sub-search space individual acquisition module is used to construct a global search space for filter parameters containing several individuals, where each individual represents a set of filter parameters. The optimization goal is to minimize the gap between the denoised interference control command data obtained after filtering the interference control command data and the true control command data. An optimization algorithm is used to search in the global search space to extract the individuals that each interference control command data passes through in the later stage of the optimization. A sub-search space construction module is used to classify a number of interference control instruction data into different categories according to their similarity, and to combine the individuals corresponding to the interference control instruction data of the same category that have passed through the optimization stage to obtain sub-search spaces corresponding to each category of interference control instruction data; A filtering module is used to obtain control instruction data to be filtered, classify the control instruction data to be filtered according to the similarity between the control instruction data to be filtered and various types of interference control instruction data, retrieve the sub-search space corresponding to the interference control instruction data of this type, and use an optimization algorithm to search for filter parameters in the sub-search space; The optimization goal is to minimize the gap between the denoised interference control instruction data obtained after filtering the interference control instruction data and the real control instruction data, and to use an optimization algorithm to search for the best in the global search space, specifically including: Build optimization goals: in, is interference control instruction data; Interference control instruction data The denoised interference control command data obtained after filtering; filter(·) is the filter, θ is the filter parameter; Ω is the global search space; Represents finding a set of optimal filter parameters θ in the global search space Ω optimal , which makes the gap between the denoised interference control command data and the real control command data smallest; By adaptively adjusting the inertia weight used to control the particle search range in the original particle swarm optimization algorithm, the individual fluctuations experienced by the original particle swarm optimization algorithm in the later stage of optimization are reduced, and an improved particle swarm optimization algorithm is obtained. The improved particle swarm optimization algorithm is used to search for the best solution in the global search space Ω. The extracting of the individual that each interference control instruction data passes through in the later stage of optimization specifically includes: Get the fitness value of the t-1th iteration, the tth iteration and the t+1th iteration t-1 、fitness t and fitness t+1 ; Calculate the difference Δt of the rate of change of fitness values ​​between adjacent iterations: When the difference Δt between the rates of change of fitness values ​​of adjacent iterations is less than the set threshold, the tth iteration is taken as the starting point of the late optimization stage and the maximum number of iterations is taken as the end point of the late optimization stage, and the individuals passed from the starting point of the late optimization stage to the end point of the late optimization stage are obtained.

Images