A vehicle behavior abnormal pattern recognition method for vehicle-mounted terminals

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a vehicle behavior abnormal pattern recognition method for an on-board terminal, comprising acquiring sensor data of the on-board terminal; extracting features of the acquired data using a fully connected neural network based on tree ring bionic optimization to obtain a first feature; reducing the feature dimension of the first feature using an autoencoder neural network based on mask enhancement to obtain a second feature; constructing a classifier model using a fractional order neural network based on quantum chaos theory, and training the classifier model using the second feature; and recognizing vehicle behavior abnormal patterns using the trained classifier model. The present invention improves the recognition accuracy of abnormal driving patterns in practical applications.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a vehicle behavior abnormal pattern recognition method for a vehicle-mounted terminal. Background Art

[0002] The Chinese invention patent with application number CN202410552900.0 proposes a device and method for extracting and identifying features of an IoT terminal, which belongs to the field of IoT security protection technology. Without customizing and modifying the IoT terminal and ensuring that the terminal identity is not tampered with or forged, the physical characteristics of the IoT low-speed wired interface signal level are extracted to first distinguish the terminal connection line, and then identify the terminal characteristics, which are used as the terminal device fingerprint for terminal identity authentication.

[0003] The Chinese invention patent with application number CN202410345435.3 proposes an attack detection method, device, storage medium and electronic device for the Internet of Things. It relates to the field of artificial intelligence technology. Among them, the method includes: capturing a data packet in a first router to obtain a first data packet, wherein the first router includes: a router for exchanging data between the Internet of Things subnet and the external network; parsing the first data packet to obtain a first parsing result; inputting the first parsing result into a target detection model, and outputting a detection result, wherein the detection result is used to indicate whether the first data packet has a network attack behavior on the Internet of Things subnet, and the target detection model includes: a model obtained by training a machine learning model by simulating the network environment of the Internet of Things subnet.

[0004] The Chinese invention patent with application number CN202410281063.2 proposes an intelligent operation and maintenance system platform based on machine learning and industrial Internet of Things, involving the field of intelligent industrial control technology, including equipment management module, fault management module, report management module, system management module, monitoring module and data processing module; the monitoring module includes sensors and network equipment, which are used to collect and display the operating data of each device in the production process; the data processing module pre-processes the operating data, builds an LSTM model, and predicts the future data of each device based on the operating data to obtain the prediction of faults and the optimization plan of production methods, and adjusts the production equipment according to the predicted data to realize automatic optimization and intelligent scheduling of the production process.

[0005] The existing technology has the following problems that still need to be further solved:

[0006] (1) Traditional generative adversarial networks are prone to mode collapse during the generation process, resulting in unstable quality of generated data and unable to meet the needs of high-quality data expansion.

[0007] (2) Existing neural network structures are usually fixed and lack the ability to adaptively adjust according to data characteristics. When faced with complex and changeable IoT data, it is difficult to effectively extract features.

[0008] (3) Traditional autoencoders do not utilize random masking technology and may not be able to fully learn the implicit relationships in the data during the dimensionality reduction process, limiting the model’s ability to process complex data.

[0009] (4) Conventional neural network classifiers lack sensitivity to small changes, resulting in insufficient classification accuracy and sensitivity when dealing with abnormal patterns with small differences. Summary of the invention

[0010] In view of the above-mentioned deficiencies in the prior art, the present invention provides a vehicle behavior abnormal pattern recognition method for a vehicle-mounted terminal.

[0011] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0012] A vehicle behavior abnormal pattern recognition method for a vehicle terminal comprises the following steps:

[0013] Obtain sensor data from the vehicle terminal;

[0014] A fully connected neural network based on tree ring bionic optimization is used to extract features from the acquired data to obtain the first feature;

[0015] The first feature is reduced in dimension by using a mask-enhanced autoencoder neural network to obtain a second feature;

[0016] A fractional-order neural network based on quantum chaos theory is used to construct a classifier model, and the second feature is used to train the classifier model;

[0017] The trained classifier model is used to identify abnormal vehicle behavior patterns.

[0018] Preferably, a fully connected neural network based on tree ring bionic optimization is used to extract features from the acquired data to obtain the first feature, including:

[0019] Initialize the neural network and set the number of neurons in the primary layer;

[0020] Adjust the number of neurons and weights in each layer according to the distribution of training data;

[0021] The loss function is calculated after each training cycle, and the network structure is dynamically adjusted according to the loss function;

[0022] Repeat the above steps until the preset stop iteration condition is met.

[0023] Preferably, the number and weight of neurons in each layer are adjusted according to the distribution of training data as follows:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] Among them, η p (l) is the learning rate of the neurons in the lth layer, η base is the basic learning rate, δ p is the adjustment factor, Sig() is the Sigmoid activation function, Te is the accumulated gradient, is the number of neurons in the l+1th layer, is the number of neurons in the lth layer, is the weight of the neurons in the lth layer, is the loss function L p In Te iterations, the weight The gradient of is the loss function L p At the tth iteration, the weight The gradient of and are the increment of the number of neurons in the lth layer and the adjustment of the weights; α p is the first expansion coefficient; β p is the second expansion coefficient, lr is the current layer number; is the training error of the lth layer of neurons; is the input data of the lth layer of neurons; is the entropy of the data distribution in the lth layer; γ p To control the rate; represents normal distribution, and ~ represents obedience to a specific distribution.

[0030] Preferably, a mask-enhanced autoencoder neural network is used to perform feature dimension reduction on the first feature to obtain the second feature, including:

[0031] Initialize the parameters of the autoencoder network;

[0032] According to the framework of the greedy algorithm, set the independent training cycle and learning rate for each layer;

[0033] At the beginning of each training cycle, for each input vector, a mask matrix is ​​randomly generated;

[0034] The mask matrix is ​​fed into the autoencoder, which compresses the data into a low-dimensional feature representation.

[0035] The low-dimensional feature representation is sent to the decoder, and the decoder is used to reconstruct the low-dimensional feature representation to obtain a reconstructed vector;

[0036] Use greedy algorithm to optimize the parameters of each layer;

[0037] After all layers are trained, the models of each layer are integrated to form a complete autoencoder model, and the autoencoder model is fine-tuned;

[0038] Repeat the above steps until the preset stop iteration condition is met.

[0039] Preferably, for each input vector, the mask matrix is ​​randomly generated as follows:

[0040] m j ~Ber(p)

[0041]

[0042] Among them, m j is the jth element of the mask vector m, p is the mask ratio; Ber() is the Bernoulli distribution discrete function, Var(X TH ) is the input vector X TH The variance of tu is the preset variance threshold, and min() is the minimum value function.

[0043] Preferably, fine-tuning the autoencoder model is specifically as follows:

[0044]

[0045] Among them, Θ fine-tune is the fine-tuned model parameter, N tu is the number of training samples, is the i-th training sample, The parameters are and Θ loss function; argmin() represents the parameter value at which a function achieves the minimum value in its domain.

[0046] Preferably, a fractional-order neural network based on quantum chaos theory is used to construct a classifier model, and the second feature is used to train the classifier model, including:

[0047] Initialize the parameters of the fractional neural network;

[0048] The output vector of each layer is transformed through the quantum chaos activation function;

[0049] Calculate the loss function based on the difference between the predicted value of the output layer and the true label;

[0050] Calculate the gradient of each weight according to the loss function and update the weight using fractional calculus rules;

[0051] After each training cycle, the network status is checked by quantum state tomography, and the activation function parameters are adjusted by quantum state tomography.

[0052] Repeat the above steps until the preset stop iteration condition is met.

[0053] Preferably, the activation function parameters are adjusted as follows:

[0054]

[0055]

[0056] in, is the first modulation parameter, is the second modulation parameter, γ u is the first adjustment coefficient, δ u is the second adjustment coefficient, ∈ u is the third adjustment factor; is the variance of the output of the l-th layer fractional-order neural network before activation, is the skewness of the output of the l-th layer fractional-order neural network before activation, It is the median of the output of the l-th layer fractional-order neural network before activation.

[0057] Preferably, acquiring the sensor data of the vehicle-mounted terminal further includes:

[0058] A generative adversarial network based on a self-repair strategy is used to augment the acquired data.

[0059] Preferably, using a generative adversarial network based on a self-repair strategy to expand the acquired data includes:

[0060] Initialize the parameters of the generative adversarial network;

[0061] A self-repair module is used in the generator to monitor the quality of generated samples in real time and make self-adjustments;

[0062] In the training process of the generative adversarial network, an adaptive feature distillation strategy is used to optimize the feature generation process by analyzing the feedback at the feature level of the generated samples;

[0063] In the sample generation phase, the generator is used to generate new data samples based on the current network parameters;

[0064] Adopting adaptive threshold dynamic adjustment mechanism to optimize the discrimination process of new data samples generated by the generator;

[0065] Adaptive feature distillation loss and adversarial loss are used to optimize the generator and discriminator;

[0066] Repeat the above steps until the preset stop iteration condition is met.

[0067] The present invention has the following beneficial effects:

[0068] (1) The data quality expanded by the generative adversarial network algorithm with a self-repairing strategy is higher, which effectively avoids the quality degradation of the generated samples and enhances the generalization ability and stability of the model.

[0069] (2) The neural network based on the tree ring bionic strategy of the present invention can adaptively adjust its structure according to the complexity of the data, making the network more flexible and effectively improving the accuracy and efficiency of feature extraction.

[0070] (3) The mask-enhanced autoencoder of the present invention can maintain the structural characteristics of the data during feature dimensionality reduction, thereby enhancing the model's ability to interpret and process complex data.

[0071] (4) The fractional-order neural network of the present invention enhances the ability to capture subtle changes through quantum state tomography, thereby improving the system's accuracy in identifying abnormal driving patterns in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The present invention is a flow chart of a method for identifying abnormal vehicle behavior patterns for vehicle-mounted terminals. DETAILED DESCRIPTION

[0073] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0074] like Figure 1 As shown, the embodiment of the present invention provides a vehicle behavior abnormal pattern recognition method for a vehicle terminal, comprising the following steps S1 to S5:

[0075] S1. Obtain sensor data from the vehicle terminal;

[0076] In an optional embodiment of the present invention, the data of the present invention is collected from the internal sensors of the vehicle terminal, including temperature sensors, acceleration sensors, position sensors, etc., and the collected data is stored in JSON format. In one embodiment, the collected data attributes include: a1 is the vehicle speed, in km / h; R a2 is the acceleration in m / s 2 ; R a3 is the engine temperature, in °C; R a4 is the vehicle position, expressed as longitude and latitude (R a4x , R a4y );R a5 is the fuel efficiency, in km / L; R a6 R is the engine working status, such as start, stop; a7 is the battery power in percentage; R a8 is the external environment temperature, in °C; R a9 is the air quality index; R a10 Provides vehicle maintenance reminders, such as oil change, tire wear, etc.

[0077] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual use, the attributes of the data are usually more than 10 attributes, and the number of attributes of the data may reach dozens or even hundreds.

[0078] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include: normal driving mode, sudden acceleration, sudden braking, and abnormal driving, a total of 4 categories.

[0079] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention adopts a generative adversarial network algorithm based on a self-repairing strategy to generate samples, thereby achieving data expansion. The generative adversarial network algorithm based on a self-repairing strategy includes two parts: the generator G c and the discriminator D c The generator is responsible for generating new data that is close to the real data distribution, and the discriminator tries to distinguish the generated new data from the real data.

[0080] Aiming at the problems of mode collapse and training instability in traditional generative adversarial networks, this paper adopts a self-repair strategy to embed a self-repair module R in the generator. c , this module can automatically adjust the generation strategy according to the feedback provided by the discriminator to optimize the quality of the generated samples.

[0081] Specifically, the training process of the generative adversarial network algorithm based on the self-repair strategy is as follows:

[0082] S101, generator G c and the discriminator D c Perform parameter initialization. In one embodiment, the initialization method is expressed as follows:

[0083]

[0084]

[0085] In the formula, and are the parameters of the generator and the discriminator respectively; represents normal distribution; and are the number of neurons in the input layer of the generator and discriminator respectively; ~ indicates that it obeys a specific distribution.

[0086] S102, using self-repair module R in the generator c Monitor the quality of generated samples in real time and make self-adjustments. The parameter adjustment of the self-repair strategy follows the following rules:

[0087]

[0088] In the formula, Represents the update amount of the generator parameters; is the partial derivative of the discriminator output with respect to its input; x cr The input of the discriminator is the fake IoT data generated by the generator; is the partial derivative of the generator output with respect to its parameters; η xf is the self-repair learning rate; Z c is the random noise of the input; D c ( ) and G c ( ) represent the network functions of the discriminator and the generator respectively. Preferably, η xf Set to 0.01.

[0089] S103. In the training process of the generative adversarial network, an adaptive feature distillation strategy is used to optimize the feature generation process by analyzing the feedback at the feature level of the generated samples, so that the generated samples are more diverse and have higher quality. Specifically, the present invention uses the feature distillation loss L afd To optimize the key feature representation of the generated samples, the calculation method is:

[0090]

[0091] Where, L afdis the feature distillation loss, Ny is the number of generated sample features, My is the number of network layers of the generator, Represents the generator G c The i-th feature extracted from represents the target feature; σ c，j is the adaptive weight; W c，j is the weight of the jth layer; λ is the regularization parameter; || ||1 is the L1 norm, and || ||2 is the L2 norm. Preferably, My is set to 5, λ afd Set to 0.1.

[0092] In one embodiment, the adaptive weight σ c，j Adaptively adjust according to the feature importance of the data, and the calculation method is expressed as:

[0093]

[0094] In the formula, γ fd is a hyperparameter that controls sensitivity; L prev is the loss function of the previous round of training; Denotes the loss function L prev About weight W c，j The norm of the gradient of . Preferably, γ fd Set to 3.

[0095] S104, in the sample generation stage, the generator G c Generate new data samples based on the current network parameters, expressed as:

[0096]

[0097] In the formula, C c，new The new data generated by the generator according to the current parameters.

[0098] In the discrimination stage, the discriminator D c Evaluate the difference between the generated new data and the real data and provide feedback to the generator and self-repair module, expressed as:

[0099]

[0100] Furthermore, the self-repair module R c According to the feedback from the discriminator, the parameters are adaptively adjusted to correct the deviations that may occur in the generation process to optimize the quality of the generated samples, which can be expressed as:

[0101]

[0102] In the formula, α zr is the adjustment coefficient. Preferably, α zr Set to 0.05.

[0103] S105, using adaptive threshold dynamic adjustment mechanism for generator G c The discriminant process of the generated new data samples is optimized, especially when facing highly heterogeneous vehicle terminal data, it can flexibly adapt to the characteristics of different data. Specifically, the discriminator D c To generate data X c，new and the real data X real When evaluating, the threshold used to judge the authenticity of the data is dynamically adjusted through the adaptive threshold dynamic adjustment mechanism, and the baseline threshold is set as θ r0 , the threshold is based on the average discriminator output value calculated based on the initial training data set. In the initial stage, the threshold adjustment coefficient β is set r0 , which is initially 0.5. In each training iteration, the threshold is dynamically adjusted according to the discriminator output of the current data sample and the preset baseline threshold. The adjustment method is expressed as:

[0104]

[0105] In the formula, is the average value of the discriminator output of the current batch of generated data, θ c，new is the adjusted threshold.

[0106] Furthermore, an update mechanism is used to calculate the correct classification ratio r of the generated data. c Adjust Beta r , the adjustment method is expressed as:

[0107]

[0108] In the formula, is the updated threshold adjustment coefficient, λ r is a hyperparameter that controls the adjustment speed, r c is the proportion of generated data correctly identified as “generated” or “real” in the current iteration. Its ideal value is 0.5, indicating that the discriminator and the generator are balanced. Preferably, λ r Set to 3.

[0109] Furthermore, in the discrimination stage, the updated threshold θ is used c，new Judging the authenticity of the generated data is expressed as:

[0110]

[0111] Where D final Indicates the final judgment result.

[0112] S106, using adaptive feature distillation loss L afdThe performance of the generator and discriminator is optimized with adversarial loss, and the calculation method of the total loss function is expressed as:

[0113] L total =logD c (X real )+log(1-D c (G c (Z c )))+L afd

[0114] In the formula, X real represents the real data sample.

[0115] Furthermore, D c (G c (Z c The term )) reflects the discriminator's ability to evaluate the generated samples, and the calculation method is expressed as:

[0116]

[0117] Where Sig( ) is the Sigmoid activation function; and are the weight and bias of the discriminator respectively.

[0118] Furthermore, the parameters are updated according to the total loss function, and the updating method is expressed as:

[0119]

[0120]

[0121] In the formula, are the updated generator parameters, is the updated discriminator parameter, is the gradient of the total loss function with respect to the discriminator parameters, is the gradient of the total loss function with respect to the generator parameters, η gan is the updated learning rate. Preferably, η gan Set to 0.03.

[0122] S107, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0123] This embodiment uses a generative adversarial network algorithm based on a self-repairing strategy for data expansion. The generator can automatically adjust the generation strategy according to the feedback of the discriminator, which solves the problems of mode collapse and instability that may occur in traditional generative adversarial networks during training. The data expanded by the generative adversarial network algorithm with a self-repairing strategy has higher quality, effectively avoiding the quality degradation of the generated samples and enhancing the generalization ability and stability of the model.

[0124] S2, using a fully connected neural network based on tree ring bionic optimization to extract features from the acquired data to obtain the first feature;

[0125] In an optional embodiment of the present invention, the expanded data is input into a feature extraction model to train the feature extraction model. The present invention uses a 3-layer fully connected neural network for feature extraction. In the prior art, some schemes use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient exploding, or falling into a local optimal solution may be encountered, affecting the stability of training and the performance of the model.

[0126] The feature extraction model training phase of the present invention adopts a tree ring-based bionic optimization strategy to simulate the growth pattern of tree rings, and gradually expands the depth and width of the network in a hierarchical manner to adapt to complex IoT data feature extraction tasks.

[0127] In traditional neural network training, the network structure is usually static, that is, the depth of the network and the number of neurons in each layer are determined before training. The present invention adopts a dynamic structure adjustment mechanism based on the tree ring pattern, which enables the network to adaptively adjust its structure according to the complexity of the data, so that the network can respond more flexibly to various IoT data characteristics.

[0128] Specifically, the training process of the neural network algorithm based on tree ring bionic optimization is as follows:

[0129] S201. Initialize the neural network. The neural network starts with only one layer, which is equivalent to the primary growth rings of a tree. It contains the most basic neuron settings and sets the number of neurons in the primary layer. and weight It is expressed as:

[0130]

[0131]

[0132] Where N base The number of neurons based on p is the standard deviation of the initialization weight. Preferably, N base Set to 1.5 times the number of data features, σ pSet to 0.01.

[0133] S202. Decide whether to add a new layer based on the complexity of the training data and the current performance of the model. The addition of a new layer simulates the growth of tree rings. With each new layer, the capacity and complexity of the network increase to adapt to more complex data features. The number of neurons and weights of each layer are adjusted as follows:

[0134]

[0135]

[0136]

[0137]

[0138]

[0139] Where η p (l) is the learning rate of the lth layer, η base is the base learning rate, δ p is the adjustment factor, Sig() is the Sigmoid activation function, and Te represents the cumulative gradient considering the past Te iterations. is the number of neurons in the l+1th layer, is the number of neurons in the lth layer, is the weight of the lth layer, is the loss function L p In Te iterations, The gradient of is the loss function L p At the tth iteration, The gradient of and They represent the increment of the number of neurons in the lth layer and the adjustment of the weights respectively; α p is the first expansion coefficient; β p is the second expansion coefficient, lr is the current number of layers; is the training error of the lth layer; is the input data of the lth layer; represents the entropy of the data distribution in the lth layer; γ p Controls the rate at which weight increment decreases with layer depth. Preferably, α p Set to 0.2, β p Set the value to 3, γ p Set the value to 5.

[0140] Further, It is used to evaluate the complexity and uncertainty of the data, so as to make the expansion of the network structure more refined and adaptable. The calculation method is expressed as:

[0141]

[0142] In the formula, p k is the probability of the k-th feature appearing in the output data of the l-th layer.

[0143] S203. After each training cycle, the performance of the newly added layer is evaluated, and the network structure and parameters are adjusted according to the performance feedback. The performance evaluation is constrained by the loss function, that is, the evaluation is performed by comparing the actual output with the expected output. The method of calculating the loss function is expressed as:

[0144]

[0145] In the formula, y i is the actual output; is the network prediction output, obtained by the preset Softmax function; Mu is the total number of samples.

[0146] Furthermore, the network structure is dynamically adjusted according to the result of the loss function. Specifically, when L p >θ p When θ is θ, the network is expanded to improve the model performance. p is the performance threshold, preferably, θ p Set the value to 0.2.

[0147] S204, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0148] In this embodiment, feature extraction adopts a three-layer fully connected neural network optimized based on the tree ring bionic strategy. By simulating the growth pattern of tree rings to dynamically adjust the network structure, it solves the problem that the traditional neural network structure is fixed and difficult to adapt to the variability of data characteristics when processing complex Internet of Things data. It can adaptively adjust its structure according to the complexity of the data, making the network more flexible and effectively improving the accuracy and efficiency of feature extraction.

[0149] S3, using a mask-enhanced autoencoder neural network to reduce the dimension of the first feature to obtain a second feature;

[0150] In an optional embodiment of the present invention, the data after feature extraction is input into a feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention adopts a mask-enhanced autoencoder neural network as a feature dimensionality reduction algorithm. The mask-enhanced autoencoder neural network includes an encoder and a decoder, and realizes data compression and reconstruction by learning the effective expression of input data. On the basis of the traditional autoencoder, the present invention adopts a random mask in the input layer to selectively hide part of the features of the input data, forcing the network to reconstruct the complete output from the partially visible input data, thereby enhancing the model's ability to learn the implicit relationship in the data, so as to ensure that the data after dimensionality reduction can still maintain the original data structure characteristics.

[0151] Specifically, the training process of the mask-enhanced autoencoder neural network algorithm is as follows:

[0152] S301, initialize the parameters of the autoencoder network, including the weight and bias parameters of the autoencoder network, set Θ (l) is the parameter set of the lth layer, including weights and biases, that is, Θ (l) = {W tu (l) , b tu (l)}, and W tu (l) is the weight matrix of the lth layer, b tu (l) is the bias vector of the lth layer. In one embodiment, the initialization method is random initialization.

[0153] S302: According to the framework of the greedy algorithm, set the independent training cycle and learning rate of each layer to ensure that each layer can reach the local optimum before being added to the overall network. The number of independent training cycle iterations T for the lth layer is l Set to:

[0154]

[0155] In the formula, Ep is the total number of iterations, Ler is the total number of layers, δ l，Ler is the Kronecker function, ensuring that the last layer can have more training cycles.

[0156] S303, at the beginning of each training cycle, for each input vector X th , randomly generate a mask matrix m, which is used to randomly mask part of the input data. The way to generate the mask is expressed as:

[0157] m j ~Ber(p)

[0158] In the formula, m jis the jth element of the mask vector m, p is the mask ratio; Ber() is the Bernoulli distribution discrete function, which is used to generate the mask matrix, and each element m j Independently drawn from the Ber(p) distribution, the probability that the element has p is set to 1 (i.e., selected or activated), and the probability of 1-p is set to O (i.e., covered or inactivated).

[0159] Furthermore, the calculation method of the mask ratio p is expressed as:

[0160]

[0161] In the formula, Var(X TH ) represents the input data X TH The variance of tu is the preset variance threshold, and min() represents the minimum value function.

[0162] S304, after the input data is processed by masking, it is sent to the autoencoder, and the autoencoder compresses the data into a low-dimensional feature representation after processing by the activation function. The forward propagation method of the encoder is expressed as:

[0163] z tu =Sig(W tu (enc) ⊙X th +b tu (enc) )

[0164] In the formula, z tu is the encoded feature vector, Sig() is the Sigmoid activation function, W tu (enc) and b tu (enc) are the weights and biases of the encoder, and ⊙ represents the Hadamard product.

[0165] S305, the low-dimensional features are sent to the decoder, and the decoder attempts to reconstruct the original input data. By calculating the loss function, the back propagation algorithm updates the parameters in the network. The output of the decoder is expressed as:

[0166] X th ^=Sig(W tu (dec) ·z tu +b tu (dec) )

[0167] Where, X th ^ is the output of the decoder, that is, the reconstructed input data.

[0168] Furthermore, the calculation method of the loss function is expressed as:

[0169]

[0170] Where, L au is the total loss, is the reconstruction error, λ ru is the weight to control the isolation loss; ||z tu ||0 is isolated loss, indicating that z tu The number of non-zero elements of is used to measure the sparsity of the code. Preferably, λ ru Set to 0.3.

[0171] Furthermore, the isolated loss ||z tu The calculation method of ||0 is expressed as:

[0172]

[0173] Where n z is the vector z tu Dimensions; is the encoded feature vector of the i-th dimension; 1() is the indicator function, when The value is 1 when , otherwise it is 0.

[0174] S306. Optimize the parameters of each layer using a greedy algorithm. In one embodiment, the parameter update method uses an Adam optimizer, and the update method is expressed as:

[0175]

[0176] In the formula, and denote the parameters of the tth iteration and t+1th iteration of the lth layer, Adam() is the Adam optimizer function, η vu is the learning rate, L au is the calculated loss, is the gradient of the parameters of the t-th iteration of the first layer with respect to the loss function. Preferably, ηη vu Set the dynamic adjustment strategy, and its initial value is set to 0.01.

[0177] In one embodiment, let the learning rate of the tth iteration be The dynamic adjustment method is expressed as:

[0178]

[0179] In the formula, is the initial learning rate, is the initial reconstruction weight, is the reconstruction weight of the tth iteration, δ vuis the adjustment influence coefficient. Preferably, δ vu Set to 3, Set to 5.

[0180] Furthermore, the reconstruction weight of the tth iteration The adjustment of is based on the current effect of the reconstruction error and the sparsity weight, and is calculated as:

[0181]

[0182] In the formula, γ yc is the adjustment rate parameter, is the current reconstruction error, β yc is the sparsity weight, X th is the input data, X th ^ t is the reconstructed output of the tth iteration. Preferably, γ yc Set to 0.3, β yc Set to 5.

[0183] Furthermore, the gradient adjustment method in the Adam optimizer is expressed as:

[0184]

[0185] In the formula, m^ t and v^ t are the bias-corrected first-order and second-order moment estimates of the tth iteration, respectively, and ∈ is a small constant. Preferably, ∈ is set to 0.0001.

[0186] Furthermore, the bias-corrected first-order and second-order moment estimates m^ t and v^ t The calculation method is:

[0187]

[0188]

[0189] In the formula, m t and v t is the exponential moving average of the gradient and the exponential moving average of the square of the gradient at the tth iteration, and is the decay rate of the tth iteration. Preferably, and Set to 0.9 and 0.999 respectively.

[0190] S307. After the training of all layers is completed, the models of each layer are integrated to form a complete autoencoder model, and the overall fine-tuning is performed to optimize the global performance of the model. The fine-tuning method is expressed as:

[0191]

[0192] In the formula, Θ fine-tune is the fine-tuned parameter, N tu is the number of training samples, is the i-th training sample, Indicates that the parameter is and the loss function of Θ.

[0193] S308, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0194] The feature dimensionality reduction of this embodiment uses a mask-enhanced autoencoder neural network, which forces the network to learn the complete data structure from part of the data through random masking technology, thereby improving the model's ability to learn implicit relationships in the data. The mask-enhanced autoencoder can maintain the structural characteristics of the data during feature dimensionality reduction, enhancing the model's ability to interpret and process complex data.

[0195] S4, using a fractional-order neural network based on quantum chaos theory to build a classifier model, and using the second feature to train the classifier model;

[0196] In an optional embodiment of the present invention, the reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts a fractional-order neural network based on quantum chaos theory as the classifier model, and adopts a quantum state tomography method in the activation function to simulate the nonlinear behavior of the activation function through the state transformation of quantum bits to capture subtle changes in the data, enhance the model's sensitivity to abnormal patterns, and improve classification accuracy.

[0197] Specifically, the training process of the fractional-order neural network algorithm based on quantum chaos theory is as follows:

[0198] S401, initialize the parameters of the fractional-order neural network, taking the lth layer as an example, the weight of the lth layer and bias Set the initial value, expressed as:

[0199]

[0200]

[0201] In the formula, σ u is the standard deviation of the weights and biases at initialization. Preferably, σ u Set the value to 0.01.

[0202] S402, input data passes through the network. Taking the lth layer as an example, the output of the lth layer is converted by the quantum chaos activation function and transmitted layer by layer until the output layer, which is expressed as:

[0203]

[0204]

[0205]

[0206] Where, X u is the input vector of the fractional-order neural network, is the input vector of the fractional-order neural network at layer l-1, is the linear transformation result of the lth layer, is the output of the previous layer, is an activation function modulated by quantum chaos theory, is the first modulation parameter, is the second modulation parameter.

[0207] S403, using the loss function to evaluate the difference between the predicted value of the output layer and the true label, the calculation method is:

[0208]

[0209] Where, L u is the loss function of the fractional-order neural network, y i is the true label of the i-th sample, is the model's predicted value for the i-th sample, Mu is the total number of samples, and λ u is the dynamic regularization factor, is the regularization term.

[0210] In one embodiment, the dynamic regularization factor is based on the loss L of the current model. u and weight size |W u |Calculated, expressed as:

[0211]

[0212] In the formula, ρ bv is the baseline regularization strength, L cx is the cross entropy loss, κ bv is the loss sensitivity coefficient, Sig(|W u |) is the Sigmoid activation function. Preferably, ρ bv Let the value be 3, κ bv Set the value to 5.

[0213] S404. According to the result of the loss function, the gradient of each weight is calculated and updated using the fractional calculus rule. Specifically, taking the first layer as an example, the weight update rule of the first layer is expressed as:

[0214]

[0215]

[0216] In the formula, is the update amount of the weight of the lth layer of the fractional-order neural network, η u is the learning rate, is a weight update function based on fractional derivatives, L u is the loss function of the fractional-order neural network.

[0217] In one embodiment, the calculation method of the weight update function based on fractional derivatives is expressed as:

[0218]

[0219] In the formula, represents the Caputo-type fractional differential; αy is the fractional adjustment factor, and αy∈(0,1), preferably, αy is set to

[0220] Furthermore, the fractional derivative For non-integer derivatives, the calculation method is expressed as:

[0221]

[0222] Where Γ is the gamma function.

[0223] S405. After each training cycle, the network state is checked by quantum state tomography. The activation function parameters are adjusted to evaluate the performance by performing quantum state tomography. The calculation method is expressed as:

[0224]

[0225]

[0226] In the formula, γ u is the first adjustment coefficient, δ u is the second adjustment coefficient, ∈ u is the third adjustment factor; is the variance of the output of the l-th layer fractional-order neural network before activation, is the skewness of the output of the first layer of fractional-order neural network before activation, is the median of the output of the l-th layer of the fractional-order neural network before activation; the tan-1() function is used to convert the median into an angle between -π / 2 and π / 2. Preferably, γ u Set to 0.05, δ u Set to 0.03, ∈ u Set to 0.02.

[0227] S406, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0228] The classifier model of this embodiment adopts a fractional-order neural network based on quantum chaos theory, and uses quantum state tomography to simulate the activation function to capture small data changes and improve the sensitivity to abnormal patterns and classification accuracy. The fractional-order neural network enhances the ability to capture subtle changes through quantum state tomography, and improves the system's recognition accuracy of abnormal driving patterns in practical applications.

[0229] S5. Use the trained classifier model to identify abnormal vehicle behavior patterns.

[0230] In an optional embodiment of the present invention, the trained model will be used to process new samples to identify and classify the behavior patterns of vehicles. In one embodiment, the collected raw data is input into the trained feature extraction and feature dimension reduction model for feature processing, and further, the processed features are input into the classifier model for classifier training to obtain classification results. In this embodiment, the classification categories include: normal driving mode, sudden acceleration, sudden braking, abnormal driving, a total of 4 categories.

[0231] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0232] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0233] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0234] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0235] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A vehicle behavior abnormal pattern recognition method for a vehicle terminal, characterized in that: The following steps are involved: Obtain sensor data from the vehicle terminal; A fully connected neural network based on tree ring bionic optimization is used to extract features from the acquired data to obtain the first feature; The obtaining of the first feature comprises: Initialize the neural network and set the number of neurons in the primary layer; Adjust the number of neurons and weights in each layer according to the distribution of training data; The loss function is calculated after each training cycle, and the network structure is dynamically adjusted according to the loss function; Repeat the above steps until the preset stop iteration condition is met; The first feature is reduced in dimension by using a mask-enhanced autoencoder neural network to obtain a second feature; A fractional-order neural network based on quantum chaos theory is used to construct a classifier model, and the second feature is used to train the classifier model; Wherein, the classifier model training includes: Initialize the parameters of the fractional neural network; The output vectors of each layer are converted by a quantum chaos activation function; wherein the quantum chaos activation function is an activation function modulated by quantum chaos theory; Calculate the loss function based on the difference between the predicted value of the output layer and the true label; Calculate the gradient of each weight according to the loss function and update the weight using fractional calculus rules; After each training cycle, the network status is checked through quantum state tomography, and the activation function parameters are adjusted by quantum state tomography. The above steps are repeated until the preset stop iteration condition is met. Use the trained classifier model to identify abnormal vehicle behavior patterns; The activation function parameters are adjusted as follows: in, is the first modulation parameter, is the second modulation parameter, γ u is the first adjustment coefficient, δ u is the second adjustment coefficient, ∈ u is the third adjustment factor; is the variance of the output of the l-th layer fractional-order neural network before activation, is the skewness of the output of the l-th layer fractional-order neural network before activation, It is the median of the output of the l-th layer fractional-order neural network before activation.

2. The vehicle behavior abnormal pattern recognition method for a vehicle terminal according to claim 1 is characterized in that: The first feature is reduced in dimension using a mask-enhanced autoencoder neural network to obtain the second feature, including: Initialize the parameters of the autoencoder network; According to the framework of the greedy algorithm, set the independent training cycle and learning rate for each layer; At the beginning of each training cycle, for each input vector, a mask matrix is ​​randomly generated; The mask matrix is ​​fed into the autoencoder, which compresses the data into a low-dimensional feature representation. The low-dimensional feature representation is sent to the decoder, and the decoder is used to reconstruct the low-dimensional feature representation to obtain a reconstructed vector; Use greedy algorithm to optimize the parameters of each layer; After all layers are trained, the models of each layer are integrated to form a complete autoencoder model, and the autoencoder model is fine-tuned; Repeat the above steps until the preset stop iteration condition is met.

3. The vehicle behavior abnormal pattern recognition method for a vehicle terminal according to claim 2 is characterized in that: For each input vector, a random mask matrix is ​​generated: j ~Ber(p) Among them, m j is the jth element of the mask vector m, p is the mask ratio; Ber() is the Bernoulli distribution discrete function, Var(X TH ) is the input vector X TH The variance of tu is the preset variance threshold, and min() is the minimum value function.

4. The vehicle behavior abnormal pattern recognition method for a vehicle terminal according to claim 2 is characterized in that: Fine-tuning the autoencoder model is as follows: Among them, Θ fine-tune is the fine-tuned model parameter, N tu is the number of training samples, is the i-th training sample, For parameters and Θ loss function; argmin() represents the parameter value at which a function achieves the minimum value in its domain.

5. The vehicle behavior abnormal pattern recognition method for a vehicle terminal according to claim 1 is characterized in that: Acquiring sensor data from the vehicle terminal also includes: The acquired data is expanded using a generative adversarial network based on a self-repair strategy; The method of using a generative adversarial network based on a self-repair strategy to expand the acquired data includes: Initialize the parameters of the generative adversarial network; A self-repair module is used in the generator to monitor the quality of generated samples in real time and make self-adjustments; In the training process of the generative adversarial network, an adaptive feature distillation strategy is used to optimize the feature generation process by analyzing the feedback at the feature level of the generated samples; In the sample generation phase, the generator is used to generate new data samples based on the current network parameters; Adopting adaptive threshold dynamic adjustment mechanism to optimize the discrimination process of new data samples generated by the generator; Adaptive feature distillation loss and adversarial loss are used to optimize the generator and discriminator; Repeat the above steps until the preset stop iteration condition is met.

Images