Novel unmanned vehicle fault detection method and system

Abstract

The invention relates to a novel unmanned vehicle fault detection method and system. The method comprises the following steps: collecting physical quantity information of an unmanned vehicle; wherein the physical quantity information comprises noise data, image data and surface shape feature data of each part of the vehicle; inputting the physical quantity information into a fault diagnosis model for fault diagnosis; wherein the fault diagnosis model is obtained by training a meta-learning framework based on historical physical quantity information of the unmanned vehicle; generating a fault report for the fault diagnosis result; wherein the fault report comprises a fault type, fault time, a fault position and an influence range. According to the method, the accuracy and the real-time performance of fault detection are improved by utilizing the sensor and the edge computing capability, meanwhile, the dependence on diagnosis hardware is reduced, the overall operation cost is reduced, and the requirement for large-scale deployment of unmanned vehicles in the future is met.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis, and particularly to a new method and system for fault detection of unmanned vehicles. Background Art

[0002] The fault detection technology of unmanned vehicles is an important part of ensuring the safe and reliable operation of autonomous driving vehicles. With the rapid development of unmanned driving technology and its wide application in multiple fields such as logistics distribution, park patrol, and public transportation, it has become a crucial issue to ensure that unmanned vehicles can operate stably and respond to various fault situations in a timely manner. Common fault diagnosis methods are as follows:

[0003] Sensor fault detection utilizes redundant sensor design, detects sensor faults through data consistency checks, adopts a sensor self - checking mechanism, and combines historical data and statistical methods to identify abnormal behaviors of sensors.

[0004] Actuator fault detection monitors the deviation between commands and feedback of steering motors, braking systems, etc., to determine whether there are faults in the actuators.

[0005] Software fault detection applies software fault injection technology, simulates various software faults during the development and testing stages to evaluate the robustness and fault - tolerance ability of the system, and integrates software monitoring tools to monitor abnormal situations during the code execution process in real - time.

[0006] Communication fault detection monitors the quality of in - vehicle network communication, including indicators such as packet loss rate and latency, to identify potential communication faults. Adopts a redundant communication architecture to ensure that even when a single communication link fails, the transmission of critical information can be maintained.

[0007] The prediction ability of the prediction system depends on the quantity and quality of the empirical knowledge base. Lacking self - learning, adaptive ability, and openness, it can only perform fault detection within the existing knowledge range. When beyond this range, such as when new faults occur, this method is no longer applicable. The fault diagnosis method based on parameter estimation requires designing multiple parameter deviation thresholds, and the design of these thresholds is also relatively difficult.

[0008] Existing intelligent fault diagnosis models have achieved remarkable results in industrial applications. However, when facing new unmanned vehicle conditions and low - resource, heterogeneous data, it is often difficult to maintain high performance. This is mainly due to the differences in data distribution and the finiteness and heterogeneity of training data. To solve these problems, a fault diagnosis method that can maintain robustness under different working conditions is required. Summary of the Invention

[0009] The object of the present invention is to propose a new type of fault detection system for driverless vehicles, which utilizes sensors and edge computing capabilities to improve the accuracy and real-time performance of fault detection, while reducing the dependence on diagnostic hardware, lowering the overall operating cost, and meeting the requirements for large-scale deployment of future driverless vehicles, so as to solve the problems existing in the above-mentioned prior art.

[0010] To achieve the above object, the present invention provides the following solution:

[0011] A new type of fault detection method for driverless vehicles, comprising:

[0012] Collecting physical quantity information of the driverless vehicle; wherein, the physical quantity information includes: noise data, image data, and surface shape feature data of each component of the vehicle;

[0013] Inputting the physical quantity information into a fault diagnosis model for fault diagnosis; wherein, the fault diagnosis model is obtained by training a meta-learning framework based on the historical physical quantity information of the driverless vehicle;

[0014] Generating a fault report for the fault diagnosis result; wherein, the fault report includes: fault type, fault time, fault location, and influence range.

[0015] Optionally, training the meta-learning framework based on the historical physical quantity information of the driverless vehicle includes:

[0016] Classifying signals from different sources in the historical physical quantity information;

[0017] Normalizing the classified signals by adopting a gradient alignment method;

[0018] For the normalized signals, calculating an approximate solution of L IDGM by adopting a gradient approximation method;

[0019] For the approximate solution of L IDGM performing semantic matching algorithm processing to complete the training of the meta-learning framework.

[0020] Optionally, normalizing the classified signals by adopting a gradient alignment method includes:

[0021] Training multiple fault data domains by adopting a two-layer meta-learning method, and regularizing the optimization direction of the fault diagnosis model through gradient alignment;

[0022] Wherein, feature learning of a specific fault domain is performed based on an optimization index of empirical risk minimization to obtain a model with an invariant gradient direction for different domains; considering a training data set K tr , which consists of S domains K s ={K 1 , ···, K S} consists of, where each domain's s has a characteristic of a data set x and y are random variables. The optimization metric is:

[0023]

[0024] In the formula, L ERM represents the empirical risk minimization metric, S represents the number of domains in the training data set, is the loss model u according to domain s;

[0025] When a fault diagnosis model based on ERM (empirical risk minimization) encounters unseen domain data, it may fail due to memorizing the unique characteristics of a specific domain. To avoid this situation, the model should have a domain-independent prediction strategy and impose constraints to prevent it from relying too much on specific domain characteristics and ensure generalization ability; among them, the constraint is: define the gradient of domain i as g i , the inner product between source domain pairs is denoted as gi·gj. When the optimization directions of the source domains conflict with each other, the inner product is negative, that is, gi·gj < 0. To align these gradient pairs with the explicit constraints, the inner product of gradients (GIP) regularization term L GIP As an optimization objective, take the inner product of gradients regularization term L GIP As the optimization objective:

[0026]

[0027] Among them, S represents the number of source domains, and αL GIP is the combination of the sum of all gradient pairs. The IDGM object is achieved by introducing the L GIP term as an additional optimization objective and assigning it a balancing weight α. L ERM represents the empirical risk minimization metric, is the model with loss weight u according to domain K s obtained; represents the s-th data set, and i, j represent the traversed source domain pairs. Minimizing the gradient alignment loss L IDGM aims to simultaneously reduce the ERM loss and improve the maximization effect of the GIP loss. However, when calculating this function, it can be seen that calculating the gradient of L GIP involves high-order derivative calculation problems, and its computational complexity is high. Therefore, an approximation method is proposed to solve it.

[0028] Optionally, an approximate solution for calculating L IDGM using the gradient approximation method includes:

[0029]

[0030] Among them, Denotes the gradient at the S-th iteration step in the inner loop, Denotes the loss of the updated model, U s Denotes the model parameters updated after S iteration steps in the inner loop, Denotes the expectation of the random variables x, y, k z Denotes the features of the dataset, Denotes the partial derivative operator.

[0031] Optionally, the semantic matching algorithm processing includes:

[0032] Converting the original fault labels from different domains into a common encoding format and inputting them into the fault label encoder;

[0033] The fault label encoder maps the converted encoding into the feature space;

[0034] Using the fault label decoder to reconstruct the original fault label from the fault label embedding in the feature space;

[0035] Using the learnable fault label embedding generated by the fault label encoder to guide the model training; where the total loss of the training process is:

[0036]

[0037] Where, Denotes the total loss of the training process, Denotes the empirical risk minimization metric, Denotes the alignment constraint expressed as a distance loss, Denotes the weight to avoid label information in the label encoding process, Denotes the gradient inner product regularization term;

[0038] Through the minimization metric of the total loss, obtain the fault label under the corresponding metric, and then obtain the fault diagnosis result.

[0039] A new type of unmanned vehicle fault detection system, the system includes: a data acquisition module, a preprocessing module, a storage module, an analysis module, a fault detection module, a warning module, a fault report module and a feedback module;

[0040] The data acquisition module is used to collect the physical quantity information of the unmanned vehicle; where the physical quantity information includes: noise data, image data and surface shape feature data of each component of the vehicle;

[0041] The preprocessing module is used to preprocess the physical quantity information;

[0042] The storage module is used to store the preprocessed physical quantity information;

[0043] The analysis module is used to construct a fault diagnosis model;

[0044] The fault detection module is used to input the preprocessed physical quantity information into the fault diagnosis model for fault diagnosis;

[0045] The early warning module is used to give an early warning of the fault diagnosis result;

[0046] The fault report module is used to generate a report on the fault diagnosis result; wherein, the fault report includes: fault type, fault time, fault location, and influence range;

[0047] The feedback module is used to give feedback on the fault diagnosis result.

[0048] Optionally, the preprocessing module preprocesses the physical quantity information including:

[0049] Removing interference noise and outliers; and converting all data into a unified format.

[0050] Optionally, the analysis module constructs a fault diagnosis model including:

[0051] Training the meta-learning framework based on the historical physical quantity information of the unmanned vehicle to obtain the fault diagnosis model.

[0052] Optionally, training the meta-learning framework based on the historical physical quantity information of the unmanned vehicle includes:

[0053] Classifying signals from different sources in the historical physical quantity information;

[0054] Using the gradient alignment method to normalize the classified signals;

[0055] For the normalized signals, using the gradient approximation method to calculate the approximate solution of L IDGM ;

[0056] For the approximate solution of L IDGM , performing semantic matching algorithm processing to complete the training of the meta-learning framework.

[0057] The beneficial effects of the present invention are:

[0058] The present invention first collects the physical quantity information of the driverless vehicle; then inputs the module physical quantity information into a fault diagnosis model for fault diagnosis; wherein, the module fault diagnosis model is obtained by training a meta-learning framework based on the historical physical quantity information of the driverless vehicle; finally, a fault report is generated for the fault diagnosis result. The present invention utilizes sensors and edge computing capabilities to improve the accuracy and real-time performance of fault detection, while reducing the dependence on diagnostic hardware, lowering the overall operating cost, and meeting the requirements for large-scale deployment of future driverless vehicles, so as to solve the problems existing in the above-mentioned prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0060] Figure 1 Schematic diagram of the DG framework based on meta-learning according to an embodiment of the present invention;

[0061] Figure 2 Schematic diagram of the fault detection process according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0063] In order to make the above-mentioned objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0064] This embodiment proposes a new method for fault detection of a driverless vehicle, including:

[0065] Collect the physical quantity information of the driverless vehicle; wherein, the module physical quantity information includes: noise data, image data, and surface shape feature data of each component of the vehicle;

[0066] Input the module physical quantity information into a fault diagnosis model for fault diagnosis; wherein, the module fault diagnosis model is obtained by training a meta-learning framework based on the historical physical quantity information of the driverless vehicle; wherein, the meta-learning framework is as Figure 1 shown;

[0067] Generate a fault report based on the fault diagnosis result; among them, the module fault report includes: fault type, fault time, fault location, and influence range.

[0068] Furthermore, training the meta-learning framework based on the historical physical quantity information of the unmanned vehicle includes:

[0069] Classify signals from different sources in the module historical physical quantity information;

[0070] Adopt the gradient alignment method to normalize the classified signals;

[0071] For the normalized signals, use the gradient approximation method to calculate the approximate solution of L IDGM ;

[0072] For the approximate solution of L IDGM Perform semantic matching algorithm processing on the approximate solution to complete the training of the meta-learning framework.

[0073] Furthermore, adopting the gradient alignment method to normalize the classified signals includes:

[0074] Use the two-layer meta-learning method to train multiple fault data domains, and regularize the optimization direction of the fault diagnosis model through gradient alignment;

[0075] Among them, perform feature learning on specific fault domains based on the optimization index of empirical risk minimization;

[0076] Adopt a domain-invariant strategy to impose constraints on the algorithm; among them, the module constraint is: represent the gradient of domain i as gi, and the inner product between source domains is represented as gi·gj. When the optimization directions of the source domains conflict, the inner product is negative, that is, gi·gj < 0;

[0077] Take the gradient inner product regularization term L GIP as the optimization objective.

[0078] Specifically, in this embodiment, the method of fault gradient alignment includes:

[0079] In order to alleviate the domain shift problem between different related fault source data domains, the present invention uses the two-layer meta-learning method to train multiple fault data domains, and regularizes the optimization direction of the fault diagnosis model through gradient alignment. Train the model with multi-source domain data to alleviate the problem of unbalanced fault training data. The present invention performs feature learning on specific fault domains based on the optimization index of empirical risk minimization (ERM), thereby significantly improving the algorithm performance on the source domain. By aggregating the data of all source domains and minimizing the loss equally, the parameters of the model are optimized. The index is as follows:

[0080]

[0081] However, when the ERM-based model is input with data from an unknown domain, since the model memorizes this part of the unknown specific domain features, the model is prone to crashing, resulting in system instability. Therefore, the fault diagnosis model of the present invention adopts a domain-invariant strategy to achieve stable fault diagnosis results, and prevents the fault diagnosis model from falling into the features of a specific unknown domain by imposing constraints on the algorithm.

[0082] Intra-domain gradient matching is a method for learning domain-invariant features between multiple data domains. Its core idea is to align the optimization directions of different source domains, which is achieved by maximizing the inner product between the gradients of each pair of source domains. Generally speaking, each unique source domain has its own optimal direction for the fastest parameter search, which can be regarded as a shortcut in the single-source domain training scenario. If these optimal directions deviate from each other during the training of multiple source domains, the model may not converge properly, and thus the generalization ability in the unseen domain may be poor. Therefore, a constraint is needed during the training process to regulate these optimal directions. Formally, the gradient of domain i is denoted as gi. The inner product between source domain pairs can be expressed as gi·gj. When the optimization directions of these source domains conflict with each other, the inner product is negative, that is, gi·gj < 0. In order to align these gradient pairs through explicit constraints, the gradient inner product regularization term L GIP appears as an optimization objective in Equation (2):

[0083]

[0084] where the first term is L ERM , the second term is αL GIP , d is the number of source domains, and the latter term is the sum of gradient pairs. The inner-loop gradient descent metric L IDGM The objective is formulated by treating the L GIP loss term as an additional optimization objective with a balanced weight α. Minimizing the gradient alignment loss L IDGM is equivalent to minimizing the empirical risk minimization loss and maximizing the GIP loss. It can be noted from the equation that calculating the gradient of the L GIP loss term involves high-order derivatives, which will greatly increase the computational complexity.

[0085] Furthermore, a gradient approximation method is adopted to calculate the approximate solution of L IDGM , including:

[0086] Updating the inner loop and the outer loop in a cycle.

[0087] Specifically, in this embodiment, the gradient approximation method includes:

[0088] Considering the above L IDGM metric in the fault model training, an effective method is proposed to avoid this problem, that is, estimating L in the double-layer meta-learning optimization frameworkGIP The gradient of, as Figure 1 shown on the upper side. The double-layer meta-learning framework involves two different levels of update loops: 1) the inner loop and 2) the outer loop.

[0089] For these two update loops, the update process and the corresponding gradient calculation formulas are as follows:

[0090]

[0091] where L((x, y); Us) is the loss of the updated model. In (3), represents the gradient at the S-th iteration step in the inner loop. The parameter U 0 represents the initial parameters of the model (i.e., the meta-learner), and Us represents the model parameters updated after S iteration steps in the inner loop. The learning rate for the inner loop update is denoted by ρ, and the learning rate for the outer loop update is denoted by η.

[0092] U S = U S-1 - ρv S-1 (4)

[0093] In (4), at each iteration, the parameter is moved a small step in the direction of the negative gradient, hoping to gradually approach the minimum of the function. The learning rate ε controls this step size. If the learning rate is too large, it may skip the minimum; if it is too small, more iterations may be required to find the minimum.

[0094]

[0095] After a total of S steps of iteration, the change in the parameter U s relative to the initial value U 0 is equal to the sum of the update amounts at each iteration. The update amount at each iteration is composed of the learning rate ε and the corresponding gradient estimate . In this way, the parameters can be gradually adjusted to move in the direction of reducing the loss function, hoping to find the minimum of the function.

[0096]

[0097] In the outer loop, the outer loop gradient is calculated according to (6). Here, v i represents the gradient of L((x, y); U s ) with respect to the initial model parameters U 0 . The parameter U represents the new model parameters after one outer loop update, and the matrix Hs is the second-order gradient of the loss metric.

[0098] U S = U 0 - ηv s(7)

[0099]

[0100] In the meta - learning framework, when the model receives data from different fault source domains for training, the model is actually optimized in the L IDGM way. The core of the inner - loop gradient - descent meta - learning method lies in using polynomial expansion to approximate the function. After applying the second - order expansion in a specific step of the inner - loop gradient descent, the first - order Taylor series is used to approximately calculate the gradient v s corresponding to Us.

[0101]

[0102] In (10), the second term can be replaced by the inner - loop update in (4), and L(U 0 ) and L(U 0 ) are predefined as V s and H s respectively. The first - order expansion is V s =U s +O(α). Using the first - order expansion to replace V k , the relationship between U s and V s is revealed in formula (11).

[0103] Then, according to the above approximate calculation method, the complex problem of L IDGM calculation is solved, and the outer - loop update calculation method is as follows:

[0104]

[0105]

[0106] Thus, an approximate value of L IDGM is calculated.

[0107] Furthermore, the semantic matching algorithm processing includes:

[0108] Converting the original fault labels from different domains into a common encoding format and inputting them into the fault label encoder;

[0109] The module fault label encoder maps the converted encoding to the feature space;

[0110] Using the fault label decoder to reconstruct the original fault label from the fault label embedding in the feature space;

[0111] Using the learnable fault label embedding generated by the fault label encoder to guide the model training; where the total loss of the training process is:

[0112] Obtain the fault label corresponding to the index through the minimization index of the total loss, and then obtain the fault diagnosis result.

[0113] Specifically, in this embodiment, the method of fault semantic matching is as follows:

[0114] To make full use of data from different domains and solve the low-resource problem, a fault feature matching technique is developed to handle the feature differences between fault heterogeneous source domains. This technique aims to separate the latent attributes of the labels and unify different label spaces by aligning the fault feature embeddings, the fault label embeddings in the feature space, and reconstructing the label representations.

[0115] Specifically, the original fault labels from different domains are converted into a common encoding format and then fed into a fault label encoder, which maps these common encodings to the feature space. Next, a fault label decoder is designed to reconstruct the original fault labels from the fault label embeddings in the feature space, so as to ensure that the information of the fault labels is not lost during the encoding process. The reconstruction process can be described as:

[0116]

[0117] The mean squared error (MSE) is used as a measure of the distance between the original label l and the reconstructed label Ue and Ud represent the parameters of the label encoder and decoder, respectively.

[0118] In the traditional learning process, the original fault labels are fixed and used to guide the model training and correct any deviated predictions. This optimization process is one-way. In the fault feature matching method, the learnable fault label embeddings generated by the fault label encoder are used to guide the model training. Specifically, the feature representations extracted by the feature extractor are mapped to the fault feature embeddings in a feature space using a multi-layer perceptron. To ensure the consistency of these two embeddings, the feature embedding is aligned with the label embedding. This alignment constraint is formulated as a distance loss in the following form:

[0119]

[0120] Here f s and l s represent the feature embedding and label embedding from different domains, respectively. U e is the total number of training samples in domain K s , and U f are the parameters of the feature extractor and the multi-layer perceptron.

[0121] By minimizing the distance loss, a two-way optimization process can be formed to match the fault feature space. The total loss of the entire training process can be expressed in the following form:

[0122]

[0123] Through the above minimization index, the fault labels corresponding to the index can be obtained, and then the fault diagnosis results can be obtained.

[0124] Finally, in order to test the performance of the proposed method, it can be verified by a visualization method.

[0125] This embodiment also proposes a new type of unmanned vehicle fault detection system, including: a data acquisition module, a preprocessing module, a storage module, an analysis module, a fault detection module, a warning module, a fault report module, and a feedback module;

[0126] The data acquisition module is used to collect the physical quantity information of the unmanned vehicle; among them, the physical quantity information of the module includes: noise data, image data, and surface shape feature data of each component of the vehicle;

[0127] The preprocessing module is used to preprocess the physical quantity information of the module;

[0128] The storage module is used to store the preprocessed physical quantity information of the module;

[0129] The analysis module is used to build a fault diagnosis model;

[0130] The fault detection module is used to input the preprocessed physical quantity information of the module into the fault diagnosis model of the module for fault diagnosis;

[0131] The warning module is used to warn of the fault diagnosis results;

[0132] The fault report module is used to generate a report on the fault diagnosis results; among them, the fault report of the module includes: fault type, fault time, fault location, and influence range;

[0133] The feedback module is used to feedback the fault diagnosis results.

[0134] Optionally, the preprocessing module preprocessing the physical quantity information of the module includes:

[0135] Removing interference noise and outliers; and converting all data into a unified format.

[0136] Furthermore, the analysis module building the fault diagnosis model includes:

[0137] Based on the historical physical quantity information of the unmanned vehicle, the meta-learning framework is trained to obtain the fault diagnosis model of the module.

[0138] Further, training the meta - learning framework based on the historical physical quantity information of the driverless vehicle includes:

[0139] Classifying signals from different sources in the module historical physical quantity information;

[0140] Using the gradient alignment method to normalize the classified signals;

[0141] For the normalized signals, using the gradient approximation method to calculate the approximate solution of L IDGM ;

[0142] Performing semantic matching algorithm processing on the approximate solution of L IDGM to complete the training of the meta - learning framework.

[0143] Specifically, in this embodiment, as Figure 2 shown, the data acquisition module: consists of multiple different types of sensors, which are respectively responsible for collecting various physical quantity information. For example, a sound sensor is used to monitor the noise generated during device operation; an image sensor takes photos of the on - site environment; a point cloud sensor obtains the surface shape characteristics of objects, etc.

[0144] The pre - processing module: receives the original signals sent from each sensor and performs preliminary processing on them. First, remove the interference noise, then check for the existence of outliers and eliminate them. Then convert all data into a unified format for subsequent calculation use.

[0145] The storage module: saves the pre - processed data to a specified location. It can be a local hard disk or a remote server. At the same time, in order to facilitate later retrieval and analysis, these files need to be organized and stored according to certain rules.

[0146] The analysis module: finds a representative heterogeneous fault sample set from a large number of historical records, such as signals of vehicle components' sound, vibration, image, current, etc., classifies the signals from different sources, performs normalization processing using the gradient alignment method, calculates L IDGM , and uses the gradient approximation method to calculate its approximate solution, and then obtains a fault diagnosis model based on the semantic matching algorithm.

[0147] The fault detection module: The fault detection module is responsible for real - time monitoring of the device status. The various types of data collected can be used through the fault diagnosis model of the analysis module to achieve fault diagnosis under heterogeneous small - sample data conditions and ensure the stable operation of the system. This module identifies potential faults or performance problems by analyzing data from various sensors.

[0148] Alarm Trigger Module: Ensure that any abnormalities in the device or system can be quickly identified and timely actions are taken to avoid greater losses. The fault alarm system can ensure that faults are detected and handled in a timely manner, minimizing equipment downtime, reducing maintenance costs, and ensuring the continuity and safety of production.

[0149] Fault Report Module: Generate a detailed fault report based on the fault content. The report content includes relevant information such as fault type, fault time, fault location, and impact scope, facilitating subsequent viewing and diagnostic analysis.

[0150] Take corresponding actions according to the specific content in the report to diagnose problems more efficiently and accurately.

[0151] Feedback Module: When a fault is detected in a certain component, immediately notify relevant personnel to take corresponding measures. At the same time, feedback the fault handling results to the data collection and analysis process for further improvement of the model and monitoring mechanism.

[0152] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A novel unmanned vehicle fault detection method, characterized in that: include: Collecting physical quantity information of the unmanned vehicle; wherein the physical quantity information includes: noise data, image data and surface shape feature data of various components of the vehicle; Inputting the physical quantity information into a fault diagnosis model to perform fault diagnosis; wherein the fault diagnosis model is obtained by training a meta-learning framework based on historical physical quantity information of the unmanned vehicle; Generate a fault report based on the fault diagnosis result; wherein the fault report includes: fault type, fault time, fault location, and impact range.

2. The novel unmanned vehicle fault detection method according to claim 1 is characterized in that: Based on the historical physical quantity information of the unmanned vehicle, the training of the meta-learning framework includes: Classifying signals from different sources in the historical physical quantity information; The classified signals are normalized using gradient alignment; For the normalized signal, the gradient approximation method is used to calculate the inner loop gradient descent index L IDGM An approximate solution of For L IDGM The approximate solution of is processed by semantic matching algorithm to complete the training of meta-learning framework.

3. The novel unmanned vehicle fault detection method according to claim 1 is characterized in that: And the gradient alignment method is used to normalize the classified signals, including: A two-layer meta-learning approach is used to train multiple fault data domains, and gradient alignment is used to regularize the optimization direction of the fault diagnosis model. Among them, the feature learning of specific fault domains is performed based on the optimization index of empirical risk minimization to obtain a fault diagnosis model with unchanged gradient direction for different fields; consider a training data set K tr , consisting of S domains K s ={K1, ···, K S }, where each domain s has a characteristic of a data set x,y are random variables; The optimization index based on empirical risk minimization is: Where, L ERM represents the empirical risk minimization metric, which simply aggregates all source domain data and minimizes the loss on average. S represents the number of training dataset domains. According to the domain K s The resulting model has a loss weight of u; A constraint is imposed on the fault diagnosis model, which is: the gradient of domain i is defined as g i , the inner product between source domain pairs is denoted as g i ·g j , when the optimization directions of the source domain conflict with each other, the inner product is negative, that is, g i ·g j < 0, in order to align these gradient pairs with the explicit constraints, the inner product regularization term L of the gradient GIP As an optimization objective; where the gradient inner product regularization term L GIP As optimization objectives: Among them, S represents the number of source domains, and αL GIP is the combination of the sum of all gradient pairs. The IDGM object introduces L GIP The term is taken as an additional optimization objective and given a balance weight α to achieve L ERM represents the empirical risk minimization index, According to the domain K s The resulting model has a loss weight of u; represents the sth dataset, i and j represent the traversed source domain pairs.

4. The novel unmanned vehicle fault detection method according to claim 1 is characterized in that: The gradient approximation method is used to calculate L IDGM Approximate solutions include: Perform an update cycle on the inner and outer loops; the update process and gradient calculation formula are as follows: in, represents the gradient of the Sth iteration step in the inner loop, Indicates the loss of the updated model, U s represents the model parameters updated after S iterations in the inner loop, represents the expectation of random variables x, y, k z Represents the characteristics of the dataset, represents the partial derivative operator.

5. The novel unmanned vehicle fault detection method according to claim 1 is characterized in that: The semantic matching algorithm processing includes: Convert the original fault labels from different fields into a common coding format and input them into the fault label encoder; The fault label encoder maps the converted code to the feature space; A fault label decoder is used to reconstruct the original fault label from the fault label embedding in the feature space; The learnable fault label embedding generated by the fault label encoder is used to guide model training; the total loss of the training process is: in, represents the total loss of the training process, represents the empirical risk minimization index, Denote the alignment constraint as a distance loss, Indicates the weight of avoiding label information in the label encoding process, represents the gradient inner product regularization term; By minimizing the total loss indicator, the fault label under the corresponding indicator is obtained, and then the fault diagnosis result is obtained.

6. A new type of unmanned vehicle fault detection system, characterized in that: Used to implement the method according to any one of claims 1 to 5, the system comprises: a data acquisition module, a preprocessing module, a storage module, an analysis module, a fault detection module, an early warning module, a fault reporting module and a feedback module; The data acquisition module is used to collect physical quantity information of the unmanned vehicle; wherein the physical quantity information includes: noise data, image data and surface shape feature data of various components of the vehicle; The preprocessing module is used to preprocess the physical quantity information; The storage module is used to store the preprocessed physical quantity information; The analysis module is used to construct a fault diagnosis model; The fault detection module is used to input the preprocessed physical quantity information into the fault diagnosis model to perform fault diagnosis; The early warning module is used to issue an early warning for the fault diagnosis result; The fault reporting module is used to generate a report on the fault diagnosis result; wherein the fault report includes: fault type, fault time, fault location, and impact range; The feedback module is used to provide feedback on the fault diagnosis result.

7. The novel unmanned vehicle fault detection system according to claim 6 is characterized in that: The preprocessing module preprocesses the physical quantity information, including: Remove interfering noise and outliers; and transform all data into a unified format.

8. The novel unmanned vehicle fault detection system according to claim 6 is characterized in that: The analysis module constructs a fault diagnosis model including: Based on the historical physical quantity information of the unmanned vehicle, the meta-learning framework is trained to obtain the fault diagnosis model.

9. The novel unmanned vehicle fault detection system according to claim 8 is characterized in that: Based on the historical physical quantity information of the unmanned vehicle, the training of the meta-learning framework includes: Classify signals from different sources in historical physical quantity information; The classified signals are normalized using gradient alignment; For the normalized signal, the gradient approximation method is used to calculate L IDGM An approximate solution of For L IDGM The approximate solution of is processed by semantic matching algorithm to complete the training of meta-learning framework.

Images