New energy automobile intelligent cabin fault mode identification method

Abstract

The invention discloses a new energy automobile intelligent cabin fault mode identification method, and belongs to the field of machine learning model training, and the method comprises the steps: obtaining cabin data based on a plurality of sensor systems of a new energy automobile, marking the cabin data, and obtaining a marked data set; constructing a generative adversarial network for interactive exploration of quantum state information, and expanding the labeled data set based on the generative adversarial network to obtain a training data set; and performing feature extraction on the original data of the new sample by using the trained feature extraction model. And the obtained feature vectors are processed through the trained data dimension reduction model. And the feature vectors after dimension reduction are sent to a classifier for final fault type judgment. According to the method, the parameter selection of the support vector machine is optimized by adopting the Gaussian Newton method, the optimal solution is quickly converged through the minimization of the sum of squares of errors, and the classification performance is optimized.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning model training, and in particular relates to a method for identifying fault patterns in an intelligent cockpit of a new energy vehicle. Background Art

[0002] In the current industrial and technological context, new energy vehicles are developing rapidly, and the requirements for their safety, convenience and intelligence are also continuously increasing. Especially in the field of smart cockpit technology, how to accurately and quickly identify and prevent potential faults has become a research focus. The smart cockpit integrates a large number of sensors and electronic devices to monitor the status of the vehicle and provide the best driving and riding experience. However, the numerous electronic components and complex system interactions also greatly increase the possibility of failures. If these failures are not discovered and handled in time, they may cause safety accidents and affect the safety of passengers.

[0003] However, the existing technologies have the following defects: (1) Data generation in the existing technologies mostly relies on traditional neural networks, which lack the ability to accurately simulate complex data distributions; (2) Traditional parameter optimization techniques such as gradient descent may have difficulty finding the global optimal solution when faced with a much more complex parameter system and are prone to falling into local optimality; (3) Existing feature extraction methods may fail to effectively identify and retain key features in the data, resulting in information loss; (4) Under complex or changing environmental conditions, the fault diagnosis system of the existing technologies may not be robust enough, which may easily lead to a decrease in the accuracy of intelligent cockpit fault mode recognition. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for identifying fault modes of a new energy vehicle intelligent cockpit, comprising:

[0005] Acquire cockpit data based on multiple sensor systems of the new energy vehicle, annotate the cockpit data, and obtain annotated data sets;

[0006] Constructing a generative adversarial network for interactive exploration of quantum state information, and expanding the labeled data set based on the adversarial network to obtain a training data set;

[0007] The neural network model optimized based on the bird formation algorithm is trained by the training data set to obtain a feature extraction model;

[0008] Constructing a feature dimensionality reduction model based on a network flow optimized autoencoder neural network, inputting the data acquired by the feature extraction model into the feature dimensionality reduction model to obtain the reduced-dimensional data;

[0009] Constructing a support vector machine classifier based on Gauss-Newton optimization, inputting the dimension-reduced data into the classifier for training, and obtaining a trained classifier;

[0010] The fault mode of the smart cockpit of a new energy vehicle is identified based on the feature extraction model, the feature dimension reduction model and the trained classifier.

[0011] Preferably, the cabin data includes but is not limited to: battery voltage value, battery current value, battery temperature, battery charging state, battery health state, vehicle speed, acceleration, GPS location data, external environment temperature and cabin internal environment parameters;

[0012] In the labeled data set, the labeled categories include no fault, cockpit drive device fault, cockpit rotation device fault, cockpit gear fault and other faults.

[0013] Preferably, the process of obtaining the training data set includes:

[0014] Initialize the parameters of the quantum state generator in the quantum state space, and initialize the parameters of the discriminator in the classical neural network parameter space;

[0015] Simulating the quantum state of cockpit fault data based on the quantum state generator and current parameters;

[0016] Convert the quantum state into classical data, distinguish the classical data from the actually collected cockpit fault data through the discriminator, calculate the total loss function including the quantum state geometric loss, update the parameters of the quantum state generator and the discriminator using the back propagation algorithm and the parameter gradient descent method, and obtain a generative adversarial network for interactive exploration of quantum state information;

[0017] The labeled data set is expanded based on a generative adversarial network for interactive exploration of quantum state information to obtain a training data set.

[0018] Preferably, the process of obtaining the feature extraction model includes:

[0019] Constructing a neural network model, optimizing a combination of network parameters of the neural network model by using a bird formation algorithm, and obtaining an optimized network model;

[0020] The optimized network model is trained using the training data set to obtain the feature extraction model.

[0021] Preferably, the process of obtaining the optimized network model includes:

[0022] Initialize a group of birds in the bird formation algorithm, each bird corresponds to a set of network parameters;

[0023] A fitness function is set and a bird formation optimization is performed. The bird formation optimization includes an exploration phase and an utilization phase. In the exploration phase, the birds explore according to their own positions and the position of the leading bird, and explore the surrounding area through a small amount of random disturbance to find a possible better position. In the utilization phase, the birds adjust the flight direction and speed according to the known optimal position, and optimize their own position to approach the global optimal point.

[0024] After each round of optimization, the fitness function is used to evaluate the positions of all birds, that is, the corresponding network parameter combinations. The positions of the birds are adjusted according to the evaluation results to select the best performing parameters for the next round of iteration.

[0025] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed and the optimized network model is obtained.

[0026] Preferably, the process of obtaining the data after dimension reduction includes:

[0027] An encoder part and a decoder part are constructed based on a multi-layer neural network structure, and data is mapped into a low-dimensional feature space based on the encoder part;

[0028] Introducing a network flow optimization mechanism into the low-dimensional feature space to simulate the data flow process to optimize feature representation and generate a low-dimensional feature vector;

[0029] Mapping the low-dimensional feature vector back to the original data space through the decoder part to obtain a reconstructed data vector;

[0030] Comparing the difference between the reconstructed data vector and the original data, optimizing the parameters of the encoder part and the decoder part based on the difference, and obtaining a feature dimension reduction model;

[0031] The data acquired by the feature extraction model is input into the feature dimension reduction model to obtain the reduced dimension data.

[0032] Preferably, the process of obtaining the trained classifier includes:

[0033] Initialize all relevant parameters according to the requirements of the support vector machine, wherein the relevant parameters include weights and biases;

[0034] Construct a loss function and define rules based on data characteristics;

[0035] The support vector machine is trained using a sequential minimum optimization algorithm, and parameters are adjusted using the data-based rules;

[0036] The parameters adjusted by the rule-based method are optimized by using the Gauss-Newton method to obtain the trained classifier.

[0037] Preferably, the process of identifying the fault mode of the new energy vehicle smart cockpit includes:

[0038] Acquire new cockpit operation data, and input the new cockpit operation data into the feature extraction model to obtain new data features;

[0039] Inputting the new data features into the feature dimensionality reduction model to perform dimensionality reduction to obtain new dimensionality reduction data;

[0040] The newly reduced-dimensional data is input into the trained classifier to determine the fault type, thereby completing the identification of the fault mode points of the smart cockpit of the new energy vehicle.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] (1) The generative adversarial network based on quantum state information uses a quantum neural network as a generator to generate simulated data through quantum circuits. At the same time, it introduces a quantum state geometric loss function and uses quantum information geometry to evaluate the differences between quantum states, thereby improving the quality and diversity of data generation.

[0043] (2) The bird formation algorithm is used to dynamically adjust the weights and structure of the neural network and simulate the flock behavior of birds to optimize the network parameters, providing better adaptability and robustness in parameter optimization.

[0044] (3) The network flow optimized autoencoder neural network combines network flow optimization theory and autoencoder structure to optimize the data flow path through the "minimum cost maximum flow" model, effectively retaining key feature information while removing redundant and unimportant information.

[0045] (4) The Gauss-Newton method is used to optimize the parameter selection of the support vector machine, and the classification performance is optimized by quickly converging to the optimal solution by minimizing the sum of squared errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0047] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0048] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0049] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0050] Embodiment 1

[0051] like Figure 1 As shown, this embodiment provides a method for identifying fault modes of a new energy vehicle intelligent cockpit, including:

[0052] Constructing a generative adversarial network for interactive exploration of quantum state information, and expanding the labeled data set based on the adversarial network to obtain a training data set;

[0053] The neural network model optimized based on the bird formation algorithm is trained by the training data set to obtain a feature extraction model;

[0054] Constructing a feature dimensionality reduction model based on a network flow optimized autoencoder neural network, inputting the data acquired by the feature extraction model into the feature dimensionality reduction model to obtain the reduced-dimensional data;

[0055] Constructing a support vector machine classifier based on Gauss-Newton optimization, inputting the dimension-reduced data into the classifier for training, and obtaining a trained classifier;

[0056] The fault mode of the smart cockpit of a new energy vehicle is identified based on the feature extraction model, the feature dimension reduction model and the trained classifier.

[0057] Furthermore, the cabin data includes, but is not limited to: battery voltage value, battery current value, battery temperature, battery charging state, battery health state, vehicle speed, acceleration, GPS location data, external environment temperature and cabin internal environment parameters;

[0058] In the labeled data set, the labeled categories include no fault, cockpit drive device fault, cockpit rotation device fault, cockpit gear fault and other faults.

[0059] The data in this embodiment comes from multiple sensor systems of new energy vehicles, including engine status sensors, battery health monitoring sensors, and cabin internal environment sensors. The sensors can monitor and record various operating parameters of the vehicle in real time, and the data storage format is JSON.

[0060] In one embodiment, the collected data includes the following attributes:

[0061] ax: battery voltage value; ay: battery current value; az: battery temperature; aα: battery charging status; aβ: battery health status; aγ: vehicle speed; aδ: acceleration; a∈: GPS location data; aζ: external ambient temperature; aη: cabin internal environmental parameters.

[0062] It should be emphasized that this embodiment is only used to illustrate a data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0063] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include five categories: no fault, cabin drive device fault, cabin rotation device fault, cabin gear fault, and other faults.

[0064] Furthermore, the process of obtaining the training data set includes:

[0065] Initialize the parameters of the quantum state generator in the quantum state space, and initialize the parameters of the discriminator in the classical neural network parameter space;

[0066] Simulating the quantum state of cockpit fault data based on the quantum state generator and current parameters;

[0067] Convert the quantum state into classical data, distinguish the classical data from the actually collected cockpit fault data through the discriminator, calculate the total loss function including the quantum state geometric loss, update the parameters of the quantum state generator and the discriminator using the back propagation algorithm and the parameter gradient descent method, and obtain a generative adversarial network for interactive exploration of quantum state information;

[0068] The labeled data set is expanded based on a generative adversarial network for interactive exploration of quantum state information to obtain a training data set.

[0069] It is understandable that in the task of the present invention, the collection, acquisition, annotation and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability, while affecting the accuracy of the model. The present invention proposes a generative adversarial network based on interactive exploration of quantum state information for data expansion, and the generative adversarial network based on interactive exploration of quantum state information includes two components: a quantum state generator (G) and a discriminator (D). In traditional generative adversarial networks, the discriminator and the generator are constructed using classical neural networks. In the present invention, the quantum state generator uses data represented by quantum states and uses quantum neural networks to process these data to simulate complex data distributions. In addition, the present invention adds quantum state geometric loss to the loss function, and a new loss function that evaluates the differences between quantum states based on quantum information geometry helps to guide the quantum state generator to generate more accurate and diverse data, thereby improving the generalization ability of the model.

[0070] Specifically, the training process of the generative adversarial network algorithm based on interactive exploration of quantum state information is as follows:

[0071] Initialize the parameters of the quantum state generator and the discriminator D. In one embodiment, the parameters of the quantum state generator are randomly initialized in the quantum state space, and the parameters of D are initialized in the classical neural network parameter space, which can be expressed as:

[0072]

[0073] Where, QSG θ is the initial parameter of the quantum state generator, which obeys a normal distribution with a mean of 0; D φ is the initial parameter of the discriminator, which obeys a normal distribution with a mean of 0; represents a normal distribution with mean 0 and covariance matrix as the identity matrix.

[0074] Furthermore, the quantum state generator generates a quantum state simulating cockpit fault data using a quantum circuit according to current parameters, wherein the circuit parameters are determined by QSG θ Control, the output quantum state vector, can be expressed as:

[0075] |ψ(θ)>=U(θ)|0>;

[0076] Where |ψ(θ)> is the quantum state controlled by parameter θ, U(θ) is the parameterized quantum logic gate, and |0> is the initial state of the quantum bit.

[0077] Furthermore, the specific construction of U(θ) can be expressed as:

[0078]

[0079] In the formula, is the quantum Hamiltonian, θ is the control parameter, and i is the imaginary unit.

[0080] Furthermore, the generated quantum state needs to be converted into classical data through a specially designed quantum measurement process so that these data can be processed by the discriminator D. The way in which the quantum measurement process converts the quantum state into classical data can be expressed as:

[0081]

[0082] Where x is the measured classical data, ρ(θ) is the density matrix generated by the quantum state, Tr is the trace operation, is some observed quantity.

[0083] Furthermore, the observation The calculation method can be expressed as:

[0084]

[0085] In the formula, λ k is the eigenvalue of the observed quantity, and |k> is the corresponding eigenstate.

[0086] Furthermore, the calculation method of ρ(θ) can be expressed as:

[0087] ρ(θ)=|ψ(θ)><ψ(θ)|;

[0088] Furthermore, the discriminator D attempts to distinguish between the data generated from the quantum state generator and the actual cockpit fault data collected. Its training goal is to maximize its ability to identify generated data and real data, and the loss function is used to constrain its training process:

[0089]

[0090] In the formula, x real is the real data, x fake is the data generated by the quantum state generator, D φ is the discriminator network, and φ is the discriminator network parameter.

[0091] Furthermore, a total loss function including a quantum state geometric loss is calculated, wherein the quantum state geometric loss evaluates the geometric distance between the generated quantum state and the real data state to guide the quantum state generator to generate a quantum state closer to the real data. Then the total loss function It can be expressed as:

[0092]

[0093] Where QGL(θ) is the quantum state geometric loss, which is the distance between the generated quantum state and the target quantum state; ρtarget is the density matrix of the target quantum state, λ tr is the balancing factor. Preferably, λ tr Set to 0.03.

[0094] Furthermore, the calculation method of the quantum state geometric loss QGL(θ) can be expressed as:

[0095]

[0096] In the formula, ∥∥ F is the Frobenius norm, indicating that the quantum state geometric loss is evaluated by calculating the Frobenius norm of the difference between two density matrices.

[0097] Furthermore, the back-propagation algorithm and parameter gradient descent method are used to update the parameters of the quantum state generator and D. Both the quantum network parameters of the quantum state generator and the classical network parameters of D need to be adjusted according to the loss function. The update method can be expressed as:

[0098]

[0099] Where θ' and φ' are the updated control parameters and discriminator network parameters, respectively, and α as is the learning rate, is the total loss function, and is the gradient of the loss function with respect to the parameters θ and φ. Preferably, α as Set to 0.01.

[0100] Further, For example, the calculation method can be expressed as:

[0101]

[0102] Further, the above steps are repeated until the preset stop iteration condition is met, indicating 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.

[0103] Furthermore, the process of obtaining the feature extraction model includes:

[0104] Constructing a neural network model, optimizing a combination of network parameters of the neural network model by using a bird formation algorithm, and obtaining an optimized network model;

[0105] The optimized network model is trained using the training data set to obtain the feature extraction model.

[0106] Furthermore, the process of obtaining the optimized network model includes:

[0107] Initialize a group of birds in the bird formation algorithm, each bird corresponds to a set of network parameters;

[0108] A fitness function is set and a bird formation optimization is performed. The bird formation optimization includes an exploration phase and an utilization phase. In the exploration phase, the birds explore according to their own positions and the position of the leading bird, and explore the surrounding area through a small amount of random disturbance to find a possible better position. In the utilization phase, the birds adjust the flight direction and speed according to the known optimal position, and optimize their own position to approach the global optimal point.

[0109] After each round of optimization, the fitness function is used to evaluate the positions of all birds, that is, the corresponding network parameter combinations. The positions of the birds are adjusted according to the evaluation results to select the best performing parameters for the next round of iteration.

[0110] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed and the optimized network model is obtained.

[0111] The generated data is input into the feature extraction model to train the feature extraction model. The present invention uses a neural network model optimized based on the bird formation algorithm for feature extraction. The neural network based on the bird formation algorithm is dynamically adjusted by the bird formation algorithm. The optimizer of the neural network weights and structure. Specifically, the present invention optimizes the weights of the neural network by simulating the group behavior of flying birds, and each flying bird represents a set of network parameters. The interactive behavior between the flying birds helps to find the optimal or approximately optimal combination of network parameters, which enhances the adaptability and robustness of the model in a complex data environment.

[0112] Specifically, the training process of the neural network model optimized based on the bird formation algorithm is as follows:

[0113] Initialize a group of birds in the bird formation algorithm, each bird corresponds to a set of network parameters. In one embodiment, the initialization method can be expressed as:

[0114]

[0115] In the formula, Represents the initialized network parameter vector, n ci represents the dimension of the parameter vector; Represents a normal distribution with mean 0 and standard deviation 1.

[0116] Furthermore, by setting the fitness function, the network parameter combination corresponding to the bird with high fitness is more likely to be retained and copied. Specifically, the fitness function The calculation method for evaluating the parameter vector can be expressed as:

[0117]

[0118] In the formula, is the fitness function, is the input feature vector, It is a neural network parameter. The function form is the Sigmoid function in the preset logistic regression, which is used to map the linear output to the (0, 1) interval, indicating the confidence of the classification.

[0119] Furthermore, the bird formation optimization is performed, including the exploration phase and the utilization phase. In the exploration phase, the bird explores according to its own position and the position of the leading bird, explores the surrounding area through a small amount of random perturbations, and looks for a possible better position; in the utilization phase, the bird adjusts its flight direction and speed according to the known optimal position, and optimizes its position to approach the global optimal point. Specifically, in the exploration phase and the utilization phase, the position update method of the bird can be expressed as:

[0120]

[0121] In the formula, and are the parameter vectors at time t and time t+1, respectively, where the time is specifically the number of iterations; α ya is the learning rate, is the direction vector synthesized by individual learning and social learning, η ya is the learning rate that approaches the global optimal solution, is the currently known optimal parameter vector. Preferably, α ya Set to 0.01, η ya Set to 0.005.

[0122] Further, The calculation method can be expressed as:

[0123]

[0124] In the formula, β ta and γ ta is a weight factor that controls the influence of the local best and global best positions on the current position; It is the best position that the bird has found in history. is the optimal position found in the entire population. Preferably, β ta and γ ta Set to 0.3 and 0.7 respectively.

[0125] Furthermore, after each round of optimization, the fitness function is used to evaluate the positions of all birds, that is, the corresponding network parameter combinations, and the positions of the birds are adjusted according to the evaluation results to select the best performing parameters for the next round of iteration. Specifically, the condition for the updated fitness function to decide whether to accept the new position can be expressed as:

[0126]

[0127] Where f() is the fitness function.

[0128] Further, the above steps are repeated until the preset stop iteration condition is met, indicating 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.

[0129] Furthermore, the process of obtaining the data after dimensionality reduction includes:

[0130] An encoder part and a decoder part are constructed based on a multi-layer neural network structure, and data is mapped into a low-dimensional feature space based on the encoder part;

[0131] Introducing a network flow optimization mechanism into the low-dimensional feature space to simulate the data flow process to optimize feature representation and generate a low-dimensional feature vector;

[0132] Mapping the low-dimensional feature vector back to the original data space through the decoder part to obtain a reconstructed data vector;

[0133] Comparing the difference between the reconstructed data vector and the original data, optimizing the parameters of the encoder part and the decoder part based on the difference, and obtaining a feature dimension reduction model;

[0134] The data acquired by the feature extraction model is input into the feature dimension reduction model to obtain the reduced dimension data.

[0135] The data after feature extraction is further input into the feature dimension reduction module for dimension reduction operation. The present invention adopts the autoencoder neural network algorithm based on network flow optimization to effectively reduce the data dimension after feature extraction by encoding first and then decoding, while retaining key information to improve the accuracy of intelligent cockpit fault mode recognition. The autoencoder neural network based on network flow optimization combines network flow optimization theory and autoencoder structure, optimizes the data flow path to retain key feature information, and removes redundant and unimportant information, which can significantly improve the data representation efficiency and provide more accurate input for subsequent classification and monitoring tasks.

[0136] Specifically, the training process of the autoencoder neural network algorithm based on network flow optimization is as follows:

[0137] The encoder part is designed. This part uses a multi-layer neural network structure to map the high-dimensional input data into a low-dimensional feature space. The encoder is designed to capture the intrinsic structure and pattern in the data. The encoder part is designed with a multi-layer neural network, whose input is the high-dimensional data vector e after feature extraction. i , the output is the feature vector h after dimensionality reduction i , the encoding process can be expressed as:

[0138] h i =f enc (e i θ enc );

[0139] Among them, f enc is the encoder function, θ enc are the parameters of the encoder.

[0140] Furthermore, the encoder function f enc It consists of multiple layers of neural network, and the calculation of each layer can be expressed as:

[0141]

[0142] in, is the output of layer l (for the first layer, ), W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively, and σ is the ReLU activation function.

[0143] In one embodiment, the encoder and decoder are both designed as 3-layer fully connected neural networks. The number of neurons in the first layer is set to 128, the second layer is set to 64, and the last layer (i.e., the encoding layer and the decoding start layer) is set to 32.

[0144] Furthermore, in the encoded low-dimensional feature space, a network flow optimization mechanism is introduced to optimize the feature representation by simulating the "minimum cost maximum flow" problem in the data flow process, and by adjusting the connection weights between nodes in the network, key information is ensured to be transmitted first in the data flow process. Specifically, is the feature flow graph, where and Represent the nodes and edges in the graph respectively, and each edge The flow rate on uv and fees uv Calculated by the following formula:

[0145] f uv =min(cap uv ,demand v-supply u );

[0146] c uv =α uv ·(f uv ) 2 +β uv ·f uv ;

[0147] Among them, cap uv is the capacity of edge (u,v), demand v and supply u are the demand and supply of nodes v and u respectively, α uv and β uv It is a parameter that controls traffic costs.

[0148] In one embodiment, the edge capacity cap uv Dynamically set according to the average strength of the data flow, the initial value is set to 1.5 times the average strength of the data flow. Traffic cost parameter α uv and β uv The initial values ​​are set to 0.01 and 0.001, respectively, to balance the distribution of traffic and optimize the cost.

[0149] Furthermore, the flow rate f uv The calculation of needs to consider the supply and demand balance of the node, and its calculation method can be expressed as:

[0150] f uv =min(cap uv ,max(0,demand v -Σ k∈in(v) f kv ));

[0151] Among them, in(v) represents the set of all edges flowing into node v, ensuring that the flow does not exceed the capacity of the edge while meeting the demand of the node.

[0152] Furthermore, the decoder part is designed. This part also uses a multi-layer neural network structure to restore the data in the low-dimensional feature space to the original data space. The goal of the decoder is to reconstruct the original data as accurately as possible to verify the retention of key information during the encoding process. Specifically, the low-dimensional feature vector h i Mapping back to the original data space, we get the reconstructed data vector r i , the decoding process can be expressed as:

[0153] r i =f dec (h i θ dec );

[0154] Among them, f dec is the decoder function, θ dec are the decoder parameters.

[0155] Furthermore, the decoder function f dec Similar to the encoder structure, it is used to restore the original data space from the low-dimensional feature space. The calculation of each layer can be expressed as:

[0156]

[0157] in, is the output of the decoder layer l (for the first layer, ), and are the weights and biases of the decoder layer l, and φ is the activation function used by the decoder.

[0158] Furthermore, by comparing the differences between the original data and the reconstructed data, the parameters in the encoder and decoder are dynamically adjusted, as well as the weight setting in the network flow optimization process, to ensure that the feature dimension reduction process can effectively reduce the data dimension while retaining the information that is critical to the monitoring task. Specifically, the present invention uses an adaptive mapping adjustment method to compare the original data vector e i and reconstruct the data vector r i The difference between them can achieve adaptive adjustment, and the difference metric D(e i ,r i ) can be expressed by mean square error (MSE), which can be expressed as:

[0159]

[0160] According to the difference metric D, the parameters of the encoder and decoder, as well as the traffic and cost parameters in the network flow optimization, are dynamically adjusted to ensure that the feature dimensionality reduction process can effectively retain key information.

[0161] Furthermore, after training and optimization, the feature vector after dimensionality reduction is output, and further, dynamic feature reconstruction is performed. The present invention uses a preset dynamic feature reconstruction network to further refine and reconstruct the features to better meet the needs of the classifier. The network dynamically adjusts the parameters in the feature reconstruction process to achieve adaptive reconstruction of the features to enhance the description ability of the features for the device status. Specifically, the feature vector obtained after feature dimensionality reduction is denoted as h i , the goal of the dynamic feature reconstruction network is to use the network f DFRN h i Converted to the reconstructed feature vector h′ i , the conversion process can be expressed as:

[0162] h′ i=f DFRN (h i ; Θ DFRN );

[0163] Among them, Θ DFRN Represents the parameters of the dynamic feature reconstruction network. The adaptive adjustment of the dynamic feature reconstruction network is achieved through the loss function L DFRN Implementation, this loss function aims to minimize the difference between the reconstructed features and the original features, while optimizing the contribution of the reconstructed features to the classification results, which can be expressed as:

[0164]

[0165] Among them, ∥∥2 represents the L2 norm, which is used to calculate the difference between the original feature vector and the reconstructed feature vector; L class is the classification loss based on the reconstructed features, which is used to guide the feature reconstruction process; DFRN is a hyperparameter used to balance the two parts of loss.

[0166] Furthermore, the process of obtaining the trained classifier includes:

[0167] Initialize all relevant parameters according to the requirements of the support vector machine, wherein the relevant parameters include weights and biases;

[0168] Construct a loss function and define rules based on data characteristics;

[0169] The support vector machine is trained using a sequential minimum optimization algorithm, and parameters are adjusted using the data-based rules;

[0170] The parameters adjusted by the rule-based method are optimized by using the Gauss-Newton method to obtain the trained classifier.

[0171] The reduced-dimensional data is input into a classifier for classifier training. The present invention adopts a support vector machine classification algorithm based on Gauss-Newton optimization. The support vector machine distinguishes different categories by maximizing the edge of the decision boundary. In order to optimize the parameter selection in the support vector machine, the present invention adopts the Gauss-Newton method based on the minimization technology of the sum of squared errors to converge to the optimal solution more quickly.

[0172] Specifically, the training process of the support vector machine classification algorithm based on Gauss-Newton optimization is as follows:

[0173] All relevant parameters are initialized according to the requirements of the support vector machine. The relevant parameters include weights and biases. The initialization method can be expressed as:

[0174]

[0175] In the formula, Represents the initialized parameter vector, including the weights and biases of the support vector machine; Indicates that the mean is 0 and the variance is σ 2 The normal distribution of is, I is the identity matrix.

[0176] Furthermore, a loss function is constructed to evaluate the performance of the model. The loss function used in the present invention is combined with a regularization term and can be expressed as:

[0177]

[0178] In the formula, is the loss function, is the parameter vector, is the feature vector of the i-th data point, y i is the corresponding label, and λ is the regularization coefficient, which controls the strength of regularization. Preferably, λ is set to 0.5.

[0179] Furthermore, rules based on data characteristics are defined to guide the adaptive adjustment of parameters during the training process. Specifically, the defined rule control update method can be expressed as:

[0180]

[0181] In the formula, is the parameter vector adjusted by the rules. The adjust() function adjusts the parameters according to the defined rule set RuleSet, and α is the hyperparameter of the adjustment intensity.

[0182] Furthermore, the calculation method of the adjust() function can be expressed as:

[0183]

[0184] In the formula, β rule is the weight assigned according to the importance of the rule, is the parameter vector of the rule Preferably, the impact is determined by adjusting specific parameters according to the fault type.

[0185] Furthermore, the support vector machine is trained using the sequential minimum optimization algorithm, and the parameters are fine-tuned based on the rule-based optimization strategy, which depends on the rules defined above, such as adjusting the learning rate when encountering a specific type of fault data, which can be expressed as:

[0186] α t+1 =α t de_rate;

[0187] In the formula, α t and αt+1 is the learning rate in two consecutive iterations, and de_rate is the learning rate decay ratio. Preferably, the adaptive adjustment algorithm can be specifically implemented in the following manner:

[0188] α t+1 =α t ·exp(-γ th t);

[0189] In the formula, γ th is the decay constant, and t is the number of iterations. Preferably, γ th Set to 0.3.

[0190] Furthermore, the Gauss-Newton method is used to refine and optimize the parameters that have been adjusted by the rule-based method to ensure that the parameter settings can achieve the mathematically optimal solution and reduce the deviation caused by manual rule setting. Specifically, the Gauss-Newton iterative update method can be expressed as:

[0191]

[0192] In the formula, and are the parameter vectors before and after the update, J is the Jacobian matrix, r is the residual vector, and λ is the regularization parameter, which affects the size of the update step.

[0193] Furthermore, the calculation method of the Jacobian matrix J can be expressed as:

[0194]

[0195] In the formula, r i is the residual of the ith data point, q j is the jth parameter, Represents the partial derivatives of the residual with respect to the parameters and is used to construct the Jacobian matrix.

[0196] And, the calculation method of the residual vector r can be expressed as:

[0197]

[0198] In the formula, is the output predicted by the model, y i is the actual label value.

[0199] Further, the above steps are repeated until the preset stop iteration condition is met, indicating 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.

[0200] Furthermore, the process of identifying the fault mode of the new energy vehicle intelligent cockpit includes:

[0201] Acquire new cockpit operation data, and input the new cockpit operation data into the feature extraction model to obtain new data features;

[0202] Inputting the new data features into the feature dimensionality reduction model to perform dimensionality reduction to obtain new dimensionality reduction data;

[0203] The newly reduced-dimensional data is input into the trained classifier to determine the fault type, thereby completing the identification of the fault mode points of the smart cockpit of the new energy vehicle.

[0204] Use the trained model to process new samples and obtain the final fault diagnosis results. In one embodiment, new cockpit operation data is collected and input into the model. Further, the trained feature extraction model is used to extract features from the raw data of the new samples. Further, the obtained feature vector is processed by the trained data dimension reduction model. Further, the reduced feature vector is sent to the classifier for final fault type judgment. In this embodiment, the fault categories include five categories: no fault, cockpit drive device fault, cockpit rotation device fault, cockpit gear fault, and other faults.

[0205] The innovation of this embodiment is:

[0206] (1) The generative adversarial network based on quantum state information uses a quantum neural network as a generator to generate simulated data through quantum circuits. At the same time, it introduces a quantum state geometric loss function and uses quantum information geometry to evaluate the differences between quantum states, thereby improving the quality and diversity of data generation.

[0207] (2) The bird formation algorithm is used to dynamically adjust the weights and structure of the neural network and simulate the flock behavior of birds to optimize the network parameters, providing better adaptability and robustness in parameter optimization.

[0208] (3) The network flow optimized autoencoder neural network combines network flow optimization theory and autoencoder structure to optimize the data flow path through the "minimum cost maximum flow" model, effectively retaining key feature information while removing redundant and unimportant information.

[0209] (4) The Gauss-Newton method is used to optimize the parameter selection of the support vector machine, and the classification performance is optimized by quickly converging to the optimal solution by minimizing the sum of squared errors.

[0210] Based on the above innovations, the following technical effects are achieved:

[0211] (1) The use of quantum state generators and geometric loss functions ensures data diversity and high quality, thereby improving the accuracy of model training.

[0212] (2) The bird formation algorithm and network flow optimization enhance the adaptability and robustness of the model in complex data environments.

[0213] (3) Through the application of Gauss-Newton method, the speed of model training is accelerated and the efficiency of fault diagnosis is improved.

[0214] (4) The autoencoder structure combined with network flow theory effectively removes redundant parts of the data, retains key information, and reduces processing time and resource consumption.

[0215] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for identifying fault patterns of intelligent cockpits of new energy vehicles, characterized in that: include: Acquire cockpit data based on multiple sensor systems of the new energy vehicle, annotate the cockpit data, and obtain annotated data sets; Constructing a generative adversarial network for interactive exploration of quantum state information, and expanding the labeled data set based on the adversarial network to obtain a training data set; The neural network model optimized based on the bird formation algorithm is trained by the training data set to obtain a feature extraction model; Constructing a feature dimensionality reduction model based on a network flow optimized autoencoder neural network, inputting the data acquired by the feature extraction model into the feature dimensionality reduction model to obtain the reduced-dimensional data; Constructing a support vector machine classifier based on Gauss-Newton optimization, inputting the dimension-reduced data into the classifier for training, and obtaining a trained classifier; The fault mode of the smart cockpit of a new energy vehicle is identified based on the feature extraction model, the feature dimension reduction model and the trained classifier.

2. The method according to claim 1, characterized in that: The cabin data includes but is not limited to: battery voltage value, battery current value, battery temperature, battery charging status, battery health status, vehicle speed, acceleration, GPS location data, external environment temperature and cabin internal environment parameters; In the labeled data set, the labeled categories include no fault, cockpit drive device fault, cockpit rotation device fault, cockpit gear fault and other faults.

3. The method according to claim 1, characterized in that The process of obtaining the training data set includes: Initialize the parameters of the quantum state generator in the quantum state space, and initialize the parameters of the discriminator in the classical neural network parameter space; Simulating the quantum state of cockpit fault data based on the quantum state generator and current parameters; Convert the quantum state into classical data, distinguish the classical data from the actually collected cockpit fault data through the discriminator, calculate the total loss function including the quantum state geometric loss, update the parameters of the quantum state generator and the discriminator using the back propagation algorithm and the parameter gradient descent method, and obtain a generative adversarial network for interactive exploration of quantum state information; The labeled data set is expanded based on a generative adversarial network for interactive exploration of quantum state information to obtain a training data set.

4. The method according to claim 1, characterized in that: The process of obtaining the feature extraction model includes: Constructing a neural network model, optimizing a combination of network parameters of the neural network model by using a bird formation algorithm, and obtaining an optimized network model; The optimized network model is trained using the training data set to obtain the feature extraction model.

5. The method according to claim 4, characterized in that The process of obtaining the optimized network model includes: Initialize a group of birds in the bird formation algorithm, each bird corresponds to a set of network parameters; A fitness function is set and a bird formation optimization is performed. The bird formation optimization includes an exploration phase and an utilization phase. In the exploration phase, the birds explore according to their own positions and the position of the leading bird, and explore the surrounding area through a small amount of random disturbance to find a possible better position. In the utilization phase, the birds adjust the flight direction and speed according to the known optimal position, and optimize their own position to approach the global optimal point. After each round of optimization, the fitness function is used to evaluate the positions of all birds, that is, the corresponding network parameter combinations. The positions of the birds are adjusted according to the evaluation results to select the best performing parameters for the next round of iteration. Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed and the optimized network model is obtained.

6. The method according to claim 1, characterized in that The process of obtaining the data after dimensionality reduction includes: An encoder part and a decoder part are constructed based on a multi-layer neural network structure, and data is mapped into a low-dimensional feature space based on the encoder part; Introducing a network flow optimization mechanism into the low-dimensional feature space to simulate the data flow process to optimize feature representation and generate a low-dimensional feature vector; Mapping the low-dimensional feature vector back to the original data space through the decoder part to obtain a reconstructed data vector; Comparing the difference between the reconstructed data vector and the original data, optimizing the parameters of the encoder part and the decoder part based on the difference, and obtaining a feature dimension reduction model; The data acquired by the feature extraction model is input into the feature dimension reduction model to obtain the reduced dimension data.

7. The method according to claim 1, characterized in that The process of obtaining the trained classifier includes: Initialize all relevant parameters according to the requirements of the support vector machine, wherein the relevant parameters include weights and biases; Construct a loss function and define rules based on data characteristics; The support vector machine is trained using a sequential minimum optimization algorithm, and parameters are adjusted using the data-based rules; The parameters adjusted by the rule-based method are optimized by using the Gauss-Newton method to obtain the trained classifier.

8. The method according to claim 1, characterized in that The process of identifying the fault mode of the smart cockpit of a new energy vehicle includes: Acquire new cockpit operation data, and input the new cockpit operation data into the feature extraction model to obtain new data features; Inputting the new data features into the feature dimensionality reduction model to perform dimensionality reduction to obtain new dimensionality reduction data; The newly reduced-dimensional data is input into the trained classifier to determine the fault type, thereby completing the identification of the fault mode points of the smart cockpit of the new energy vehicle.

Images