Mobile target recognition method, device and equipment based on evolutionary neural network

Abstract

The application relates to a mobile target recognition method, device and equipment based on an evolutionary neural network, which comprises the following steps: collecting sensing signals of vehicles and pedestrians in motion through an unattended sensor system and performing signal preprocessing; the sensing signals comprise sound signals and ground vibration signals; a training set and a test set are constructed according to the sound signals and the ground vibration signals; the signal data in the training set and the test set are different; an evolutionary neural network is used to train the training set, so that an optimal feature extraction method and a neural network model are obtained; the environment noise, pedestrians and moving vehicles in the test set are recognized and classified according to the optimal feature extraction method and the neural network model, so that the recognized mobile targets are obtained; the mobile targets comprise pedestrians and moving vehicles. The evolutionary neural network can realize a high recognition accuracy with the smallest feature vector and the lowest network complexity.

Description

Technical Field

[0001] This invention belongs to the field of target recognition technology, and relates to a method, apparatus and device for recognizing moving targets based on evolutionary neural networks. Background Technology

[0002] Unattended ground sensor networks are wireless networks composed of a large number of stationary, randomly distributed unattended sensors in a self-organizing and multi-hop manner, used to detect abnormal events within the work area. The nodes are homogeneous, low-cost, small in size, and most can operate for extended periods. Unattended sensor networks are widely used in border control, protection of key facilities, and intrusion prevention in nature reserves. The detection methods of unattended sensors mainly include visible light, passive infrared, radar, magnetic fields, ground vibration, and sound. Among these methods, sound sensors and seismic sensors are a major research focus in the field of unattended sensor systems due to their long detection range, low cost, light weight, and low power consumption. Targets identified by unattended sensors based on sound and ground vibration signals typically include people, various vehicles, and ultra-low-altitude aircraft. Considering that vibration and sound signals generated by moving targets on the ground in urban areas are easily affected by environmental noise and the low power consumption requirement of unattended ground sensors, the current ability of unattended sensor systems to identify moving targets on the ground remains insufficient. Summary of the Invention

[0003] To address the problems existing in the above-mentioned traditional methods, this invention proposes a moving target recognition method based on an evolutionary neural network, a moving target recognition device based on an evolutionary neural network, a computer device, and a computer-readable storage medium. It can achieve a high recognition accuracy with the smallest feature vector and the lowest network complexity, and significantly improve the ability to recognize ground moving targets.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0005] On the one hand, a moving target recognition method based on evolutionary neural networks is provided, including:

[0006] The unattended sensor system collects and preprocesses the sensor signals of vehicles and pedestrians as they move; the sensor signals include sound signals and ground vibration signals.

[0007] Based on the sound signal and the ground vibration signal, a training set and a test set are constructed; the signal data in the training set and the test set are different.

[0008] An evolutionary neural network was trained on the training set to obtain the optimal feature extraction method and neural network model.

[0009] Based on the optimal feature extraction method and neural network model, environmental noise, pedestrians and moving vehicles in the test set are identified and classified to obtain the identified moving targets; the moving targets include pedestrians and moving vehicles.

[0010] In one embodiment, signal preprocessing includes normalizing signal units and removing DC components based on Fast Fourier Transform.

[0011] In one embodiment, the training set includes sampled signals of sound signals and ground vibration signals collected at multiple different times, and the test set includes sampled signals of sound signals and ground vibration signals collected at multiple different times.

[0012] In one embodiment, the process of training an evolutionary neural network on a training set includes:

[0013] Initialize the evolutionary neural network and generate the initial population;

[0014] Based on the fitness function of the evolutionary neural network, the fitness of each individual in the initial population is calculated using the training set, and the optimal solution is searched using the fitness as the criterion.

[0015] Evolutionary neural networks are subjected to evolutionary operations until the training termination condition is met, at which point the optimal feature extraction method and neural network model are output; evolutionary operations include selection, crossover, and mutation.

[0016] In one embodiment, the fitness function is:

[0017]

[0018] Wherein, acc(x i ) represents the i-th individual x in a certain iteration. i The classification performance, AN represents the maximum number of nodes in the optimized fully connected layer network, SN(x) i ) represents the i-th individual x i The total number of nodes in the neural network.

[0019] In one embodiment, the crossover operation includes:

[0020] Iterate through each parent individual and randomly select another parent individual from the parent population to pair up, forming N crossover combinations; N crossover combinations are used to generate offspring, where N is the number of individuals in the parent population;

[0021] Each crossover combination randomly generates a mask of the same length as the individual; a mask of 1 indicates that the offspring will inherit the genetic information of parent 1 from the crossover combination, and a mask of 1 indicates that the offspring will inherit the genetic information of parent 0 from the crossover combination.

[0022] In one embodiment, the training set is used to train the evolutionary neural network for 1500 epochs and for 200 generations.

[0023] On the other hand, a moving target recognition device based on an evolutionary neural network is also provided, comprising:

[0024] The signal acquisition module is used to collect and preprocess the sensor signals of vehicles and pedestrians moving through the unattended sensor system; the sensor signals include sound signals and ground vibration signals.

[0025] The dataset construction module is used to build training and test sets based on sound signals and ground vibration signals; the signal data in the training and test sets are different.

[0026] The model training module is used to train an evolutionary neural network on a training set to obtain the optimal feature extraction method and neural network model.

[0027] The identification and classification module is used to identify and classify environmental noise, pedestrians, and moving vehicles in the test set based on the optimal feature extraction method and neural network model, and to obtain the identified moving targets; the moving targets include pedestrians and moving vehicles.

[0028] In another aspect, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described moving target recognition method based on an evolutionary neural network.

[0029] Furthermore, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the aforementioned moving target recognition method based on an evolutionary neural network.

[0030] One of the above technical solutions has the following advantages and beneficial effects:

[0031] The aforementioned moving target recognition method, apparatus, and equipment based on evolutionary neural networks collect sound signals and ground vibration signals from moving vehicles and pedestrians using an unattended sensor system. Based on the collected signal data, training and test sets are constructed. The training set is then used to train the evolutionary neural network to output the optimal feature extraction method and neural network model. Finally, based on the optimal feature extraction method and neural network model, environmental noise, pedestrians, and moving vehicles in the test set are identified and classified to obtain the identified moving targets. By using evolutionary neural networks, the optimal feature extraction method and neural network structure are selected and designed. Since the obtained optimal feature extraction method and neural network structure have the smallest feature vector and the lowest network complexity, and experimental results show that using the obtained optimal feature extraction method and neural network with the smallest feature vector to identify moving targets on the test set significantly improves the accuracy of target recognition, thereby achieving the goal of significantly improving the ability to identify moving targets on the ground. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating a moving target recognition method based on an evolutionary neural network in one embodiment;

[0034] Figure 2 This is a schematic diagram of the training process of an evolutionary neural network in one embodiment;

[0035] Figure 3 This is a schematic diagram illustrating the specific process of training an evolutionary neural network in one embodiment.

[0036] Figure 4 This is a schematic diagram of the population composition in one embodiment;

[0037] Figure 5 This is a schematic diagram illustrating the relationship between parent and child generations in one embodiment;

[0038] Figure 6 This is a schematic diagram of genetic information encoding in one embodiment;

[0039] Figure 7 This is a schematic diagram of the genetic information encoding and conversion method in one embodiment;

[0040] Figure 8 This is a schematic diagram of a genetic code example with a length of 9 in one embodiment;

[0041] Figure 9 This is a schematic diagram of an optimized fully connected neural network structure in one embodiment;

[0042] Figure 10 This is a schematic diagram of the composition of a signal acquisition device in one embodiment;

[0043] Figure 11 This is a schematic diagram of the accuracy curves and loss curves of the training set and test set during the training process in one embodiment, where (a) is the accuracy curve and (b) is the loss curve;

[0044] Figure 12 This is a schematic diagram of the confusion matrix of a model on the test set in one embodiment;

[0045] Figure 13 This is a schematic diagram of the module structure of a moving target recognition device based on an evolutionary neural network in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0048] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various locations throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments.

[0049] Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items, and all possible combinations thereof.

[0050] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0051] In one embodiment, such as Figure 1 As shown, a moving target recognition method based on an evolutionary neural network according to the present invention includes the following steps S12 to S18:

[0052] S12 collects and preprocesses the sensor signals of vehicles and pedestrians moving through an unattended sensor system; the sensor signals include sound signals and ground vibration signals.

[0053] Understandably, in practical applications, data sources can be real-time or historical collections of sound signals and ground vibration signals from unattended sensor systems during vehicle and pedestrian movement. Data can be directly obtained by connecting to the unattended sensor system or indirectly through intermediate devices. The specific acquisition method can be selected based on the conditions of the actual application scenario.

[0054] S14. Based on the sound signal and the ground vibration signal, construct a training set and a test set; the signal data in the training set and the test set are different.

[0055] It is understood that the dataset used in this embodiment consists of two parts: a training set and a test set. The training set is used for training and iterative evolution of the neural network to find the required feature extraction method and neural network structure. The test set is used for the identification of moving targets. The signal data included in both the training set and the test set are signal data collected in batches at different times within the monitoring area.

[0056] S16 uses an evolutionary neural network to train on the training set to obtain the optimal feature extraction method and neural network model.

[0057] Evolutionary neural networks (ENNs) are a novel network model that organically integrates evolutionary computation and neural networks. They utilize the principles of biological evolution to search for neural networks that perform well for a given task within a feasible domain. In this embodiment, the ENN uses a genetic algorithm for optimization, focusing on the selection of feature extraction methods and the fully connected neural network structure, with the optimization objectives being the highest network classification accuracy and the lowest network complexity. The obtained training set is input into the ENN for training and iterative evolution, ultimately outputting the optimal feature extraction method and neural network model.

[0058] S18. Based on the optimal feature extraction method and neural network model, the environmental noise, pedestrians and moving vehicles in the test set are identified and classified to obtain the identified moving targets; the moving targets include pedestrians and moving vehicles.

[0059] It is understandable that after obtaining the optimal feature extraction method and neural network model, the test set is input into the obtained neural network model and the obtained optimal feature extraction method is used for feature extraction processing. The model extracts image features from the test set, identifies and classifies noise, vehicles and pedestrians, and finally outputs the recognition results, such as vehicles and pedestrians in the monitoring area.

[0060] The aforementioned moving target recognition method based on evolutionary neural networks collects sound signals and ground vibration signals from moving vehicles and pedestrians using an unattended sensor system. Training and test sets are constructed based on the collected signal data. The training set is then used to train the evolutionary neural network to output the optimal feature extraction method and neural network model. Finally, based on the optimal feature extraction method and neural network model, environmental noise, pedestrians, and moving vehicles in the test set are identified and classified to obtain the identified moving targets. By using evolutionary neural networks, the optimal feature extraction method and neural network structure are selected and designed. Since the obtained optimal feature extraction method and neural network structure have the smallest feature vector and the lowest network complexity, and experimental results show that using the obtained optimal feature extraction method and neural network with the smallest feature vector to identify moving targets on the test set significantly improves the accuracy of target recognition, thereby achieving the goal of significantly improving the ability to identify moving targets on the ground.

[0061] In one embodiment, signal preprocessing includes normalizing signal units and removing DC components based on Fast Fourier Transform.

[0062] It is understandable that signal preprocessing includes two parts: normalizing the signal units and removing the DC component based on Fast Fourier Transform. For the normalization of the signal units: the ground vibration signal (denoted as Raw_Signal) acquired by the seismic detector in the unattended sensor system after being acquired by the AD acquisition card is in V (volts). The goal is to convert the ground vibration signal (denoted as velocity after conversion) to mm / s. According to the sensitivity of the seismic detector (e.g., 100mv / m / s), the velocity signal conversion formula for ground vibration can be obtained (1). The sound signal (denoted as Raw Signal) acquired by the sound sensor in the unattended sensor system after being acquired by the AD acquisition card is in V. The goal is to convert the sound signal into the voltage signal (denoted as Sensitivity) acquired by the standard microphone. According to the microphone sensitivity formula (2) and the amplification factor of the amplifier circuit (e.g., 110dB), the calculation formula for the standard microphone output signal (Standard signals) can be obtained (3).

[0063] velocity mm / s =Raw_Signal V / 100 (1)

[0064]

[0065]

[0066] The removal of the DC component involves eliminating the DC bias generated by the acquisition instrument. First, a Fast Fourier Transform (FFT) is performed on the normalized sensor signal to remove the zero-frequency signal. Then, an Inverse Fast Fourier Transform (IFFT) is used to obtain the sound signal and ground vibration signal after removing the DC component.

[0067] Through the above preprocessing, the sound signal and ground vibration signal data can be standardized, which can be used to quickly build a standardized dataset and improve the efficiency of network training and evolution.

[0068] In one embodiment, the training set includes sampled signals obtained by subsampling sound signals and ground vibration signals collected at multiple different times, and the test set includes sampled signals obtained by subsampling sound signals and ground vibration signals collected at multiple different times.

[0069] As is understood, Table 1 shows the VI (vibration intensity) and SPL (sound pressure level) of ambient noise, the temperature and humidity of the environment at the time of data collection, and the date of data collection for each data acquisition. Table 1 is only one example of data acquisition; in practical applications, more batches of data can be collected as needed for identification. The dataset includes two parts: a training set and a test set. The training set is the data sample used for model fitting, while the test set is used to evaluate the generalization ability of the final model and for target recognition. In some implementations, the optimal feature extraction method and neural network model obtained through training evolution can be directly utilized. After preprocessing the sound signals and ground vibration signals collected on-site, the data can be directly input into the aforementioned neural network model to complete on-site moving target recognition or real-time target recognition.

[0070] In some implementations, to ensure a more effective evaluation of the generalization ability of the obtained neural network model, the training and test sets are each composed of sound and vibration signals collected at multiple different times. For example, as shown in Table 1, the sound and vibration signals of various targets collected in the first to fourth collection experiments constitute the training set of the dataset, and the sound and vibration signals of various targets collected in the fifth collection experiment constitute the test set of the dataset. The dataset can be formed by subsampling the collected sound signals and ground vibration signals in 4-second intervals with a 1-second step. Table 2 shows the size of the processed dataset in seconds.

[0071] Table 1

[0072] VI / dB SPL / dB temperature date First collection experiment 36.88 25.71 26℃ 05.30.2022 Second collection experiment 38.82 26.32 25℃ 05.31.2022 Third collection experiment 39.03 24.49 23℃ 06.01.2022 Fourth collection experiment 41.99 28.44 23℃ 06.11.2022 Fifth collection experiment 40.83 26.69 27℃ 06.15.2022

[0073] Table 2

[0074]

[0075] In one embodiment, such as Figure 2 As shown, the process of training an evolutionary neural network on a training set includes the following steps S04 to S08:

[0076] S04, Initialize the evolutionary neural network and generate the initial population;

[0077] S06. Based on the fitness function of the evolutionary neural network, the fitness of each individual in the initial population is calculated using the training set, and the optimal solution is searched using the fitness as the criterion.

[0078] S08 performs evolutionary operations on the evolutionary neural network until the training termination condition is met, at which point the optimal feature extraction method and neural network model are output; the evolutionary operations include selection, crossover, and mutation operations.

[0079] It is understandable that target recognition methods deployed in unattended sensor systems should simultaneously possess both high recognition accuracy and low processing complexity. This presents a significant challenge to the structural design of neural networks and the optimal selection of feature extraction methods. To overcome this challenge, this application proposes the aforementioned moving target recognition method based on evolutionary neural networks.

[0080] Method Principle: Evolutionary neural networks consist of four parts: initialization, evaluation, evolution, and termination condition determination. Figure 3 As shown. Evolutionary neural networks encompass several concepts, including population, individual, genotype, genetic information, parent, and offspring. Genetic information refers to the binary encoding of the feature extraction method and the neural network structure; genotype is a combination of genetic information encodings; an individual is a neural network carrying the genotype; and a population is a collection of multiple individuals, such as... Figure 4 As shown; the parent individual refers to the previous generation individual that passes on genetic information to the next generation during the evolutionary process; the offspring individual refers to the next generation individual that receives genetic information from the previous generation during the evolutionary process; the parent generation refers to the previous generation population during the evolutionary process; the offspring generation refers to the next generation population during the inheritance process, such as... Figure 5 As shown.

[0081] Specifically, the initialization process involves first setting the population size, selection probability, crossover probability, mutation probability, and number of generations based on the computer's performance and the size of the dataset used in the experiment. Then, the optimization problem is encoded by converting the optimizable objects into binary genetic codes. Each solution to this optimization problem has a unique corresponding binary code. In this embodiment, the genetic information encoding includes three parts: feature selection encoding, the number of fully connected network layers, and the number of nodes in each layer. Figure 6As shown, the genetic information encoding of an evolutionary neural network consists of a binary code of length d + M1 + M2·M3, where M1 and M2 are parameters to be set, corresponding to the maximum number of layers in a fully connected neural network. and the maximum number of nodes per layer Related. The parameters M1 and M2 depend specifically on the complexity of the training data for the fully connected neural network. The higher the complexity, the greater the maximum number of layers and the maximum number of nodes per layer should be. M3 is the number of layers in the fully connected network.

[0082] The first to the d-th codes represent the feature selection results, where d is the number of candidate features. Each value in the feature selection code corresponds to a feature; a value of 1 indicates that the feature is selected, and a value of 0 indicates that the feature is discarded. The (d+1)-M1+d-th codes represent the number of layers in the fully connected neural network. Where x i This represents the i-th value in this binary code. The codes from the (d+M1+(n-1)×M2)-th encoding to the (d+M1+n×M2)-th encoding represent the number of nodes in the n-th fully connected layer. Where x i This is the i-th value in this binary code.

[0083] For example, with parameters d=5, M1=4, and M2=4, and a total encoding length of 69, the genetic information encoding conversion method is as follows: Figure 7 As shown, the first to fifth codes are feature selection codes, the sixth to ninth codes are codes for the number of fully connected network layers, and the codes from 10+(n-1)*4 to 10+n*4 are codes for the number of nodes in the nth layer. "Feature" represents a feature, "Enable" indicates it is selected, and "Disable" indicates it is not selected. The feature selection code 01001 can be translated as the second and fifth features being selected, and the remaining features being discarded. The number of fully connected network layers 0011 can be translated as a 3-layer network. The code for the number of nodes in the first layer 1011 can be translated as the first layer having 11 nodes.

[0084] Finally, a series of binary codes are randomly generated as the initial population. Starting from this initial population, the best and worst are selected to approach the optimal solution.

[0085] Evaluation Process: The evaluation process involves calculating the fitness of each individual in the population based on the fitness function, and using this fitness as the criterion to search for the optimal solution. The fitness function is usually set according to the objective of finding the optimal solution. Evolutionary neural networks aim for the highest network classification accuracy and the lowest network complexity, with a focus on classification accuracy. Therefore, the fitness function includes two parts: a network classification accuracy score and a network complexity score, each assigned a different weight coefficient.

[0086] Furthermore, the fitness function is as follows:

[0087]

[0088] In the formula, acc(x i ) is the i-th individual x in a certain cycle. i The classification performance is represented by the classification performance of a neural network classifier. AN represents the maximum number of nodes in the optimized fully connected layer network. SN(x) i ) represents the i-th individual x i The total number of nodes in the neural network.

[0089] The evolutionary process mainly includes three parts: selection, crossover, and mutation. Figure 3 As shown, the selection process involves eliminating individuals with poor fitness and retaining those with good fitness. During selection, the population undergoes N random, repeatable samplings to determine heritable genotypes for the next generation (the same genotype may be sampled multiple times), where N is the number of individuals in the population. The probability of an individual being selected is determined by the proportion of its fitness value to the total fitness value of the population. Let the population size be M, and the fitness of an individual i be f. i The probability of it being selected is expressed as:

[0090]

[0091] Clearly, the higher the fitness of an individual in a population, the greater the probability of passing on its genes to the next generation. The selected genotype serves as the parent individual for the next step of crossover and mutation operations.

[0092] Furthermore, crossover operations can specifically include the following processing:

[0093] Iterate through each parent individual and randomly select another parent individual from the parent population to pair up, forming N crossover combinations; N crossover combinations are used to generate offspring, where N is the number of individuals in the parent population;

[0094] Each crossover combination randomly generates a mask of the same length as the individual; a mask of 1 indicates that the offspring will inherit the genetic information of parent 1 from the crossover combination, and a mask of 1 indicates that the offspring will inherit the genetic information of parent 0 from the crossover combination.

[0095] Specifically, during the crossover process, each parent individual is traversed, and an individual is randomly selected from the parent population to pair with the offspring, forming N crossover combinations to generate offspring, where N is the number of individuals in the parent population. Each crossover combination randomly generates a 0-1 mask of the same length as the individual. If the mask is 1, the offspring will inherit the genetic information of parent 1 in the crossover combination; if the mask is 0, the offspring will inherit the genetic information of parent 0 in the crossover combination.

[0096] During mutation, the genetic information of offspring is encoded with a small probability at a specific encoding position in an individual, undergoing variable transformation (e.g., 0 becomes 1, 1 becomes 0). Offspring that have completed these three genetic evolutionary processes constitute the next generation population. This process repeats until the set optimization criteria are met. Figure 8 As shown, taking a 9-bit genetic code as an example, assuming the mask is 000101010, positions 4, 6, and 8 from parent individual 1, and positions 1, 2, 3, 6, and 8 from parent individual 2 will be selected to form the next generation's genes. Here, Mask Code represents the code, Parent Generation represents the parent generation, Children Generation represents the offspring generation, Mutate represents mutation, and Crossover represents crossover. Assuming position 5 is selected as the mutation position, the final offspring's genetic code should be 111011110. Figure 8 In the offspring code, positions 4, 6, and 8 are inherited from parent individual 1, positions 1, 2, 3, 7, and 9 are inherited from parent individual 2, and position 5 is the variant portion.

[0097] Optimization Results: The optimization results of the evolutionary neural network include the selection of feature extraction methods and the structure of the fully connected neural network. Through the above processing, the evolutionary neural network can be effectively trained and evolved.

[0098] In one embodiment, the number of training epochs for training the evolutionary neural network using the training set is 1500 epochs, and the number of evolutionary generations is 200.

[0099] Specifically, the dataset mentioned above is used as the metric for fitness calculation in the evaluation process of evolutionary neural networks. The implementation process is as follows:

[0100] 1) Calculate 120 features for the training and test sets respectively, and form them into feature vectors.

[0101] 2) Design the parameters of the evolutionary neural network, where the population size is 50, the crossover probability is 0.8, the mutation probability is 0.03, and the number of generations is 200. Randomly generate an initial population.

[0102] 3) For the genetic information in the population, generate a feature vector and a neural network corresponding to each individual, and calculate the fitness of each individual. In the fitness calculation process, the neural network is first trained on the training set for 1500 epochs. The classification score of the trained neural network on the test set will be the classification performance score in the fitness calculation. One epoch represents all the data fed into the network, completing one forward computation + backpropagation process.

[0103] 4) Select, crossover, and mutate the parent population to generate the offspring population.

[0104] 5) Determine if the requirement of 200 generations of evolution is met. If it is met, output the individual with the highest fitness; otherwise, return to step 3.

[0105] After 200 generations of evolution, the individual with the optimal fitness adopted one ground vibration signal feature and one sound signal feature, and its neural network had two layers. The selected features were kurtosis (as shown in Equation (6)) and vibration energy (as shown in Equation (7)). The optimized fully connected neural network structure is as follows: Figure 9 As shown.

[0106]

[0107]

[0108] Where, x i Let be the i-th data point in x, and n be the length of signal x. This fully connected neural network uses cross-entropy loss as its objective function; therefore, the classifier loss function of this fully connected neural network is:

[0109]

[0110] Where N is the number of samples, C is the number of target types, and t ij Let θ represent the probability that the i-th sample belongs to class j, and let Θ represent the set of all parameters.

[0111] Through the above training and evolution, the optimal feature extraction method and neural network structure can be obtained with minimal computational cost, further saving computational resources and time costs.

[0112] In one example, to better illustrate and facilitate understanding of the methods described above, an experimental test example is provided. It should be noted that the experimental test examples in this application are for illustrative and illustrative purposes only and are not intended to limit the implementation of the methods described above.

[0113] Evaluation Metrics: In unattended sensor systems, accuracy, false alarm rate (FAR), and false alarm rate (FAR) are commonly used to evaluate the performance of a model. Accuracy refers to the proportion of correctly predicted samples out of the total number of samples, calculated as ACC = (TP + TN) / (TP + TN + FP + FN). Correctly classified intrusion events are called true positives (TP), correctly classified noise samples are called true negatives (TN), missed intrusion events are called false negatives (FN), and misclassified noise as intrusion events are called false positives (FP). False alarm rate (FAR) is the proportion of samples misclassified as noise out of the total number of identified intrusion events. It is calculated as FAR = FP / (TP + TN). False alarm rate (FAR) is the proportion of samples misclassified as noise out of the total number of identified intrusion events. It is calculated as UR = FN / (TP + TN).

[0114] Dataset: The dataset used in this example was collected during a sound and vibration signal target recognition experiment. The experiment took place on campus and various experiments were conducted over two weeks. The signal acquisition device was placed on one side of a road on campus. Pedestrians and vehicles moved along the road; pedestrians moved within a 15m range of the unattended sensor, and vehicles moved within a 30m range. The signal acquisition device is shown below. Figure 10 As shown, the system consists of a sensor module, an AD signal acquisition module, and a laptop computer. The sensor module includes a seismic detector and a sound sensor. The seismic detector has a natural frequency of 4.5Hz and a sensitivity of 100mV / m / s. The sound sensor includes an analog MEMS microphone with a sensitivity of -42dBv and an amplifier circuit with a gain of 110dB. The AD signal acquisition module is a four-channel AD acquisition card with a 24-bit acquisition accuracy and a sampling frequency set to 10000Hz. The acquired data types include pedestrians, small vehicles, and background noise (no target).

[0115] The detection range of unattended sensors is closely related to the noise intensity of their deployment area. In noisy urban areas, the detection range of unattended sensors is shorter, while in quiet rural areas, the detection range is longer. For vibration signals, vibration intensity is usually used to describe their noise level. For sound signals, sound pressure level (SPL) is used as the unit of measurement for noise level. The formulas for vibration intensity (VI) and sound pressure level (SPL) are as follows:

[0116]

[0117]

[0118] Among them, S RMS S is the root mean square of the vibration signal. ref The relative reference value for vibration intensity is taken as 10e-5. P RMS P is the root mean square of the sound signal. ref This is a relative reference value for the sound pressure level, taken as 10e-5.

[0119] Experimental Results: Using the training set from the aforementioned dataset as input to the optimized model, the model was trained for 500 epochs. The model reached a relatively stable state after the 100th epoch. Then, the classification ability of the trained model was evaluated using data from the test set. The median results from 11 repeated trials are discussed. Figure 11 The model's accuracy curves (a) and loss curves (b) on the training and test sets during the training process are shown. Figure 12 This is the confusion matrix of the model on the test set.

[0120] Statistical calculations show that the overall recognition accuracy for the three categories is 98.04%. For single target types, the classification accuracy for small vehicles and pedestrians is very high, both exceeding 99%. For noise, the classification accuracy is 94.79%, with a false alarm rate of 1.71% and a false negative rate of 0.24%.

[0121] Table 3 compares the classification performance of other advanced machine learning algorithms proposed by scholars in the field of ground moving target recognition based on vibration signals with the classification performance of the method proposed in this application. These advanced machine learning algorithms include Genetic Algorithm-Optimized Support Vector Machine (GA-SVM), improved BP neural network, and Vib-CNN. Among them, GA-SVM is a traditional machine learning method, and Vib-CNN is an end-to-end deep learning method. GA-SVM uses wavelet decomposition energy ratio, zero-crossing rate, mean, peak value, and waveform factor as feature extraction methods. The improved BP neural network uses peak value, zero-crossing rate, and the ratio of high-frequency energy to low-frequency energy as feature extraction methods, and LFCC (linear frequency cepstral coefficients) as the network's feature extraction method. To fairly evaluate the performance of these methods, each model will undergo 11 experiments, with each experiment consisting of 200 epochs of training. The median classification accuracy on the test set from the 11 experiments will be used as the classification performance of the model for discussion, and the computation time of the algorithm on the test set will be recorded.

[0122] Table 3

[0123] Model accuracy False alarm rate underreporting rate computation time GA-SVM 98.56％ 0.96％ 0.48％ 20.27s Improved BP Neural Network 97.47％ 0.49％ 2.04％ 18.37S Vib-CNN 98.16％ 0.43％ 1.41％ 20083ms This application method 98.21％ 1.31％ 0.49％ 4.54S

[0124] The classification performance of each model is shown in Table 3. The classification performance of each model on the test set is similar, but their computation time varies significantly. Vib-CNN has a much longer computation time than other models, but its classification accuracy is not significantly different. GA-SVM and the improved BP neural network have similar classification performance and computation time on the test set, with GA-SVM having slightly higher classification accuracy and computation time than the improved BP neural network. Compared to traditional machine learning and other deep learning methods, the proposed moving target recognition method based on evolutionary neural networks has lower algorithmic complexity and higher recognition accuracy, and can be deployed on lower-cost and lower-power hardware devices.

[0125] It should be understood that, although Figures 1 to 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order requirement for the execution of these steps; they can be executed in other orders. Figures 1 to 3 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0126] In one embodiment, such as Figure 13As shown, a moving target recognition device 100 based on an evolutionary neural network is provided, including a signal acquisition module 11, a dataset construction module 13, a model training module 15, and a recognition and classification module 17. The signal acquisition module 11 is used to acquire and preprocess sensor signals from moving vehicles and pedestrians through an unattended sensor system; the sensor signals include sound signals and ground vibration signals. The dataset construction module 13 is used to construct training and testing sets based on the sound signals and ground vibration signals; the signal data in the training and testing sets are different. The model training module 15 is used to train an evolutionary neural network on the training set to obtain the optimal feature extraction method and neural network model. The recognition and classification module 17 is used to identify and classify environmental noise, pedestrians, and moving vehicles in the testing set according to the optimal feature extraction method and neural network model to obtain the identified moving targets; the moving targets include pedestrians and moving vehicles.

[0127] The aforementioned moving target recognition device 100 based on an evolutionary neural network collects sound signals and ground vibration signals from moving vehicles and pedestrians through an unattended sensor system. Based on the collected signal data, it constructs training and testing sets. Then, it uses the training set to train the evolutionary neural network to output the optimal feature extraction method and neural network model. Finally, based on the optimal feature extraction method and neural network model, it identifies and classifies environmental noise, pedestrians, and moving vehicles in the testing set to obtain the identified moving targets. By using an evolutionary neural network, the optimal feature extraction method and neural network structure are selected and designed. Since the obtained optimal feature extraction method and neural network structure have the smallest feature vector and the lowest network complexity, and experimental results show that using the obtained optimal feature extraction method and neural network with the smallest feature vector to identify moving targets on the testing set significantly improves the accuracy of target recognition, thereby achieving the goal of significantly improving the ability to identify moving targets on the ground.

[0128] In one embodiment, signal preprocessing includes normalizing signal units and removing DC components based on Fast Fourier Transform.

[0129] In one embodiment, the training set includes sampled signals obtained by subsampling sound signals and ground vibration signals collected at multiple different times, and the test set includes sampled signals obtained by subsampling sound signals and ground vibration signals collected at multiple different times.

[0130] In one embodiment, during the training of the evolutionary neural network on the training set, the model training module 15 is specifically used to initialize the evolutionary neural network and generate an initial population; calculate the fitness of each individual in the initial population using the training set according to the fitness function of the evolutionary neural network, and search for the optimal solution using fitness as the criterion; perform evolutionary operations on the evolutionary neural network until the training termination condition is reached, and output the optimal feature extraction method and neural network model; the evolutionary operations include selection, crossover and mutation operations.

[0131] In one embodiment, the fitness function is:

[0132]

[0133] Wherein, acc(x i ) represents the i-th individual x in a certain iteration. i The classification performance, AN represents the maximum number of nodes in the optimized fully connected layer network, SN(x) i ) represents the i-th individual x i The total number of nodes in the neural network.

[0134] In one embodiment, the crossover operation includes:

[0135] Iterate through each parent individual and randomly select another parent individual from the parent population to pair up, forming N crossover combinations; N crossover combinations are used to generate offspring, where N is the number of individuals in the parent population;

[0136] Each crossover combination randomly generates a mask of the same length as the individual; a mask of 1 indicates that the offspring will inherit the genetic information of parent 1 from the crossover combination, and a mask of 1 indicates that the offspring will inherit the genetic information of parent 0 from the crossover combination.

[0137] In one embodiment, the number of training epochs for training the evolutionary neural network using the training set is 1500 epochs, and the number of evolutionary generations is 200.

[0138] For specific limitations regarding the moving target recognition device 100 based on evolutionary neural networks, please refer to the corresponding limitations of the moving target recognition method based on evolutionary neural networks mentioned above, which will not be repeated here. Each module in the aforementioned moving target recognition device 100 based on evolutionary neural networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of a device with specific data processing functions, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of computer devices already existing in the art.

[0139] In one embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program, can perform the following steps: collecting and preprocessing sensing signals of moving vehicles and pedestrians through an unattended sensor system; the sensing signals include sound signals and ground vibration signals; constructing a training set and a test set based on the sound signals and ground vibration signals; the signal data in the training set and the test set are different; training an evolutionary neural network on the training set to obtain an optimal feature extraction method and a neural network model; and identifying and classifying environmental noise, pedestrians, and moving vehicles in the test set based on the optimal feature extraction method and the neural network model to obtain identified moving targets; the moving targets include pedestrians and moving vehicles.

[0140] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the computer equipment in different application scenarios, and will not be listed and described in detail in this specification.

[0141] In one embodiment, when the processor executes the computer program, it can also implement the steps or sub-steps added in the various embodiments of the moving target recognition method based on the evolutionary neural network.

[0142] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When executed by a processor, the computer program can perform the following steps: collecting and preprocessing sensing signals of moving vehicles and pedestrians through an unattended sensor system; the sensing signals include sound signals and ground vibration signals; constructing a training set and a test set based on the sound signals and ground vibration signals; the signal data in the training set and the test set are different; training an evolutionary neural network on the training set to obtain an optimal feature extraction method and a neural network model; and identifying and classifying environmental noise, pedestrians, and moving vehicles in the test set based on the optimal feature extraction method and the neural network model to obtain identified moving targets; the moving targets include pedestrians and moving vehicles.

[0143] In one embodiment, when the computer program is executed by a processor, it can also implement the steps or sub-steps added in the various embodiments of the moving target recognition method based on evolutionary neural networks.

[0144] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus DRAM (RDRAM), and interface DRAM (DRDRAM), etc.

[0145] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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 this specification.

[0146] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, all of which fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An evolutionary neural network-based moving target recognition method, characterized in that, The method comprises the following steps: Collecting sensing signals of vehicles and pedestrians in motion through an unattended sensor system and performing signal preprocessing; the sensing signals include sound signals and ground vibration signals; According to the sound signals and the ground vibration signals, a training set and a test set are constructed; the signal data in the training set and the test set are different; Using an evolutionary neural network to train in the training set to obtain an optimal feature extraction method and a neural network model; According to the optimal feature extraction method and the neural network model, the environmental noise, pedestrians and moving vehicles in the test set are identified and classified to obtain the identified moving targets; The moving targets include pedestrians and moving vehicles; The process of training in the training set using an evolutionary neural network comprises the following steps: Initializing an evolutionary neural network and generating an initial population; According to the fitness function of the evolutionary neural network, the fitness of each individual in the initial population is calculated using the training set, and the optimal solution is searched according to the fitness as the criterion; Performing evolutionary operation on the evolutionary neural network until the training termination condition is reached to output the optimal feature extraction method and the neural network model; the evolutionary operation includes selection, crossover and mutation operations; The fitness function is as follows: in, Indicates the first iteration of a certain loop. i individual Classification performance, AN This represents the maximum number of nodes in the optimized fully connected layer network. Indicates the first i individual The total number of nodes in the neural network; The crossover operation comprises the following steps: Traverse each parent individual and randomly select another parent individual in the parent population to form N crossover combinations; the N crossover combinations are used to generate offspring, and N is the number of individuals in the parent population; Each crossover combination randomly generates a mask with the same length as the individual; the mask is 1, indicating that the offspring will inherit the gene information of parent 1 in the crossover combination, and the mask is 0, indicating that the offspring will inherit the gene information of parent 0 in the crossover combination. 2.The mobile target recognition method based on evolutionary neural network according to claim 1, characterized in that, The signal preprocessing includes normalizing the signal unit and removing the direct current component based on the fast Fourier transform.

3. The mobile target recognition method based on evolutionary neural network according to claim 1 or 2, characterized in that, The training set includes subsampled signals of the sound signals and the ground vibration signals collected at different times, and the test set includes subsampled signals of the sound signals and the ground vibration signals collected at different times. 4.The mobile target recognition method based on evolutionary neural network according to claim 1, characterized in that, The number of training rounds of training the evolutionary neural network using the training set is 1500 epochs, and the number of evolutionary generations is 200 generations.

5. An apparatus for mobile target recognition based on evolutionary neural networks, characterized in that, The method comprises the following steps: A signal collection module is configured to collect sensing signals of vehicles and pedestrians in motion through an unattended sensor system and perform signal preprocessing; the sensing signals include sound signals and ground vibration signals; A data set construction module is configured to construct a training set and a test set according to the sound signals and the ground vibration signals; the signal data in the training set and the test set are different; A model training module is configured to use an evolutionary neural network to train in the training set to obtain an optimal feature extraction method and a neural network model; An identification and classification module is configured to identify and classify the environmental noise, pedestrians and moving vehicles in the test set according to the optimal feature extraction method and the neural network model to obtain the identified moving targets; The moving targets include pedestrians and moving vehicles; In the process of training the evolutionary neural network in the training set, comprising: initializing the evolutionary neural network and generating an initial population; calculating the fitness of each individual in the initial population using the training set according to the fitness function of the evolutionary neural network and searching for the optimal solution as the criterion; evolving the evolutionary neural network until the training termination condition is reached to output the optimal feature extraction method and the neural network model; the evolution operation includes selection, crossover and mutation operations; the fitness function is: wherein, denotes the classification performance of the i-th individual in the j-th cycle, i AN denotes the optimized maximum number of nodes of the fully connected layer network, denotes the total number of nodes of the neural network of the i-th individual, i the i-th individual in the j-th cycle.​​​ the crossover operation includes: traversing each parent individual and randomly selecting another parent individual in the parent population for pairing to form N crossover combinations; the N crossover combinations are used to generate offspring, and N is the number of individuals in the parent population; each crossover combination randomly generates a mask equal in length to the individual; the mask is 1, indicating that the offspring will inherit the genetic information of the parent 1 in the crossover combination, and the mask is 0, indicating that the offspring will inherit the genetic information of the parent 0 in the crossover combination. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-5 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 4.

Images