New energy vehicle fault diagnosis system and method based on fractional order neural network

Abstract

The invention relates to the technical field of computers, and particularly discloses a new energy vehicle fault diagnosis system and method based on a fractional order neural network, a hidden layer adopts a Caputo type alpha-order fractional order differential equation to update a neuron state, all historical state information before a t moment is weighted and accumulated in an integral form, and a new energy vehicle fault diagnosis result is obtained. Both the short-term and long-term operation data participate in feature calculation, the parameter fluctuation trend in the continuous time period of the slowly-changing fault can be captured, missing detection of the early fault due to normal instantaneous signals is avoided, and the fault missing report rate is reduced; a ReLU fractional order activation function enhances nonlinear expression of historical features, complex working condition fluctuation and real fault signals can be distinguished in combination with full connection layer mapping and Softmax classification, due to no continuous historical trend support, the complex working condition fluctuation and the real fault signals can be recognized as interference by a system through a memory effect, and the fault false alarm rate is reduced; through the mode, the memory effect of the fractional order neural network is utilized, the limitation of only depending on the instantaneous change rate is broken through, and the diagnosis accuracy and the early warning capability are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a fault diagnosis system and method for new energy vehicles based on fractional neural networks. Background Technology

[0002] Fractional derivatives are an extension of integer derivatives, their core being the breaking of the limitations of integer-order differentiation / integration, allowing for arbitrary real-number orders. They describe the cumulative effect of the rate of change of a function, rather than merely reflecting the instantaneous rate of change. Since the core of first-order or higher derivatives is the instantaneous limit, while the core of fractional derivatives is integral accumulation (essentially definite integrals), and their mathematical expression is in integral form rather than integer-order difference form, this global integration logic naturally makes them the cumulative result of historical states, thus exhibiting a "memory effect." Because neural networks rely on integer-order derivatives for state updates and gradient calculations, they can only capture instantaneous or local information, making them ill-suited for complex data with long-term dependencies and dynamic evolution. Fractional neural networks, on the other hand, integrate fractional derivatives into the state update, weight optimization, or feature extraction processes. Their memory and nonlocality compensate for this deficiency while retaining the nonlinear fitting ability of neural networks.

[0003] The core power system of new energy vehicles consists of power batteries, drive motors, electronic control units, and charging modules. Its fault modes exhibit dynamic, coupled, and slowly changing characteristics, placing extremely high demands on the real-time performance and accuracy of fault diagnosis systems. Fault diagnosis technologies for new energy vehicles primarily rely on integer-order calculus to construct diagnostic models, such as rule-based reasoning, integer-order neural networks, and instantaneous signal feature extraction. Their common drawback is that they all use the instantaneous rate of change as the core judgment criterion, essentially focusing on local changes at a specific moment. The core of the instantaneous rate of change lies in the physical meaning of the integer-order derivative, whose mathematical definition is based on the limit difference of an infinitesimally small neighborhood. It can be calculated using only data from the current moment and the infinitely closest previous moment, completely ignoring earlier historical information. For example, the instantaneous rate of change of the battery voltage in a new energy vehicle only reflects the voltage difference between time t and time t-Δt (Δt approaches 0), without considering the voltage fluctuation trend before t-Δt or the cumulative effects of related parameters such as current and temperature during this period. The development of faults (such as battery cell degradation and motor bearing wear) is the result of the continuous evolution of historical states. The instantaneous rate of change at a certain moment can only capture a slice of the fault evolution and cannot connect the state relationships at different time points, so it is impossible to form a memory of the historical process.

[0004] Moreover, most faults in new energy vehicles are characterized by slow change and accumulation. A fault's development from its initial stage to its manifestation requires the continuous accumulation of historical conditions, such as the cumulative capacity decay caused by the cumulative number of battery charge-discharge cycles, and bearing wear caused by the accumulation of motor vibration energy. Instantaneous rate of change only focuses on the current rate of change and cannot quantify this cumulative effect. Even if the instantaneous rate of change is within the normal range at a certain moment, if the trend of change over a period of time has consistently deviated from the norm, it may be an early sign of a fault, but the instantaneous rate of change may ignore this risk because it meets the standard at a single moment. For example, the instantaneous value of motor vibration acceleration may always be within the normal threshold, but if the vibration rate of change shows a slow upward trend over a continuous hour, this is an early cumulative characteristic of bearing wear. A diagnostic system that relies solely on the instantaneous rate of change will miss this fault because the instantaneous value is normal at every moment. Furthermore, the development of faults has a clear temporal correlation; the current state is a continuation of the historical state, and the future state is an evolution of the current state. The instantaneous rate of change can only judge whether the current state is abnormal in isolation and cannot establish a logical connection between the historical and current states. For example, in a minor leakage fault of a power battery, the instantaneous voltage change rate is normal during the initial single charge-discharge cycle. However, after multiple consecutive charge-discharge cycles, the fluctuation range of the voltage change rate gradually increases. This logic of accumulating multiple historical states leading to a gradual change in the current state cannot be captured by the instantaneous change rate. The system cannot record voltage change rate data from past charge-discharge cycles, nor can it analyze the evolutionary relationship between this historical data and the current data. This forces the diagnostic system to react only to the immediate situation, unable to predict fault development from the perspective of historical accumulation. Essentially, it lacks the ability to memorize historical states.

[0005] In summary, the core limitation of instantaneous change rate lies in its focus on a single moment, its severing of historical connections, its neglect of cumulative effects, and its lack of ability to remember and cumulatively analyze historical states during fault development. The gradual, cumulative, and temporally correlated nature of fault development precisely necessitates the memory and cumulative analysis of historical states; this is an inherent flaw that diagnostic methods relying solely on instantaneous change rate cannot overcome. Therefore, there is an urgent need for a fault diagnosis system and method for new energy vehicles that can integrate historical fault state information, characterize the cumulative effects of fault evolution, overcome the limitations of relying solely on instantaneous change rate, and improve diagnostic accuracy and early warning capabilities. Summary of the Invention

[0006] This invention provides a fault diagnosis system and method for new energy vehicles based on fractional neural networks. It can integrate historical fault state information, characterize the cumulative effect of fault evolution, break through the limitation of relying only on instantaneous change rate, and improve diagnostic accuracy and early warning capability.

[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: A fault diagnosis system for new energy vehicles based on fractional neural networks includes a data acquisition module, a data preprocessing module, a fractional neural network diagnosis module, and a fault output module. The data acquisition module collects multi-dimensional raw data during vehicle operation; The data preprocessing module performs outlier removal and standardization on the multi-dimensional raw data to obtain standardized feature data. The fractional neural network diagnostic module takes standardized feature data as input. The hidden layer neuron state update of the fractional neural network diagnostic module employs a Caputo-type α-order fractional differential equation, which is: ,in, It is the α-order fractional derivative. Let be the state value of the i-th hidden layer neuron at time t. The neuron decay coefficient, Let be the connection weight between the j-th input neuron and the i-th hidden layer neuron. It is a ReLU fractional activation function. As a bias term, the hidden layer output feature data is calculated through the fractional differential equation; the fractional neural network diagnostic module performs fully connected layer mapping and Softmax classification on the hidden layer output feature data to obtain fault type probability distribution data. The fault output module compares the fault type probability distribution data with a preset fault threshold. When the probability value of a certain fault type is greater than the preset fault threshold, the corresponding fault diagnosis result is output.

[0008] The basic principle and beneficial effects of the scheme are as follows: The hidden layer uses Caputo-type α-order fractional differential equations to update the neuron state. Essentially, it is a weighted cumulative calculation of historical state data, rather than relying solely on the instantaneous changes at the current moment. Unlike integer derivatives, which focus on the instantaneous difference between the current moment and the infinitely close previous moment, fractional derivatives naturally encompass standardized feature data from all historical periods before the current moment. Through weighted allocation of near-large and far-small, both recent and long-term fault-related data participate in the neuron state update. For example, the current state of a slow-changing fault such as battery cell degradation or motor bearing wear is the result of multiple charge-discharge cycles and long-term operating load accumulation. Fractional differential equations can capture this continuous evolution process through memory effects, connecting and integrating parameter change trends at different time points to form a complete characterization of fault development. By weighted accumulation of all historical state information before time t in integral form, both recent and long-term operating data participate in feature calculation, thus achieving memory of the fault evolution process. This allows for the capture of parameter fluctuation trends within the continuous period of a slow-changing fault, avoiding the missed detection of early faults due to normal instantaneous signals and reducing the false alarm rate.

[0009] Meanwhile, the ReLU fractional activation function enhances the nonlinear expressive power of historical features. Combined with fully connected layer mapping and Softmax classification, it can accurately extract the correlation logic between historically accumulated fault features and the current state. For example, the instantaneous value of motor vibration acceleration may be within the normal range, but the fractional neural network can integrate the cumulative trend of vibration changes over the past few hours through the memory effect, identifying early fault signals that are instantaneously normal but gradually changing over a long period, avoiding missed detections due to ignoring historical cumulative effects. This ability to remember and accumulate historical states can overcome the limitations of relying solely on instantaneous change rates, comprehensively integrating multi-dimensional historical information in the fault evolution process, accurately characterizing the cumulative effects of faults such as battery degradation and insulation aging. This not only reduces the false alarm rate of early, slowly changing faults but also effectively distinguishes between instantaneous fluctuations and real faults under complex operating conditions, reducing the false alarm rate and improving diagnostic accuracy. Furthermore, after preprocessing, the raw data is input into the fractional network. The memory effect can integrate the historical correlation of parameters, and coupled faults can be accurately located through the cooperative change characteristics of historical parameters. Compared with integer-order neural network diagnosis, it can achieve early warning and accurate identification of faults, improving early warning capabilities.

[0010] Furthermore, the data acquisition module continuously collects multi-dimensional raw data at a preset sampling frequency during vehicle movement, and the preset sampling frequency is dynamically adjusted according to the vehicle's speed. The data preprocessing module performs real-time streaming processing on the continuously collected multi-dimensional raw data, sequentially completing outlier removal and standardization to generate a continuous standardized feature data stream. The fractional neural network diagnostic module receives the standardized feature data stream, updates the state of hidden layer neurons in real time using Caputo-type α-order fractional differential equations, accumulates historical feature data during vehicle movement using the memory effect of fractional derivatives, and performs joint calculations on the current and historical standardized feature data at preset diagnostic intervals to obtain the hidden layer output feature data for the corresponding period. This data is then processed by fully connected layer mapping and Softmax classification to generate real-time fault type probability distribution data. The fault output module dynamically compares the real-time fault type probability distribution data with a preset fault threshold to achieve fault self-detection during vehicle movement. When the detected fault type probability value is greater than the preset fault threshold, the fault diagnosis result is output.

[0011] The beneficial effects are as follows: First, the sampling frequency is dynamically adjusted according to the driving speed. High-frequency sampling at high speeds ensures that instantaneous fluctuation data is not lost, while low-frequency sampling at low speeds balances performance and energy consumption, ensuring the accuracy and efficiency of data collection during travel. Second, real-time streaming processing and real-time status updates are combined with the fractional derivative memory effect to jointly calculate current and historical features each cycle. This not only captures the dynamic evolution trend of faults during travel but also avoids the loss of historical information caused by real-time processing, shortening the response time to sudden faults during travel. Dynamic comparison enables immediate warnings during travel. Compared with static diagnosis, it can detect potential faults in advance, preventing the escalation of faults during vehicle travel and improving driving safety.

[0012] Furthermore, during vehicle operation, the fractional neural network diagnostic module, based on continuously received standardized feature data streams, constructs a fault evolution time series model using historical state data accumulated through Caputo-type α-order fractional differential equations. This model predicts the trend of standardized feature data within a preset future time period, yielding predicted feature data. The fractional neural network diagnostic module then substitutes the predicted feature data into the α-order fractional differential equations for calculation, obtaining predicted hidden layer output feature data. This data is then mapped through a fully connected layer and processed by Softmax classification to generate fault type prediction probability distribution data within the preset future time period. The fault output module compares the fault type prediction probability distribution data with a preset warning threshold. When the predicted probability value of a certain fault type exceeds the preset warning threshold, it outputs the fault prediction result.

[0013] The beneficial effects are as follows: by using historical data accumulated through fractional derivative memory effect to build an evolutionary model, the prediction accuracy is improved, and the slowly changing faults can be accurately captured; by calculating and classifying predictive feature data, fault prediction results can be output in advance, which can give drivers sufficient time to deal with the faults compared to diagnosis after they occur, and reduce the risk of sudden faults while driving; by setting pre-set warning thresholds and risk level classifications, the prediction results can be made more practical, helping drivers to take targeted measures and improve the driving safety and reliability of new energy vehicles.

[0014] Furthermore, before performing outlier removal and standardization on the multi-dimensional raw data, the data preprocessing module first performs correlation analysis on the multi-dimensional raw data: calculating the Pearson correlation coefficient between different types of raw data and selecting strongly correlated data pairs; constructing cross-features on the strongly correlated data pairs to generate correlated feature data; the data preprocessing module fuses the correlated feature data with the single-dimensional standardized data after outlier removal and standardization to obtain fused standardized feature data; the fractional neural network diagnosis module takes the fused standardized feature data as input, calculates the hidden layer output feature data through Caputo-type α-order fractional differential equations, and then performs fully connected layer mapping and Softmax classification to obtain fault type probability distribution data of fused correlated features; the fault output module completes fault diagnosis based on this fault type probability distribution data.

[0015] The beneficial effects are as follows: mining strong correlations in multi-dimensional data and constructing cross-features can compensate for the one-sidedness of single-dimensional data information and improve the ability to identify coupled faults; the input of fused correlation features enables fractional neural networks to capture the multi-parameter coordinated change patterns of faults, which improves the accuracy of coupled fault diagnosis compared with single-dimensional feature diagnosis and can avoid missed diagnoses caused by ignoring parameter correlations; the Pearson correlation coefficient screening ensures the effectiveness of correlation features, thereby improving diagnostic efficiency.

[0016] Furthermore, after the fault output module outputs the fault diagnosis result, the data acquisition module collects the actual fault information confirmed by subsequent vehicle repairs as feedback data. The fault output module compares the feedback data with the corresponding fault type probability distribution data and calculates the diagnostic deviation value, which is the absolute value of the difference between the actual fault type probability and the predicted probability. When the diagnostic deviation value > a preset deviation threshold, the fault output module transmits the feedback data and the diagnostic deviation value to the fractional neural network diagnostic module. The fractional neural network diagnostic module uses the feedback data as samples to analyze the neuron decay coefficients in the Caputo-type α-order fractional differential equation. and connection weights The process involves iterative correction, employing fractional gradient descent to minimize the diagnostic bias. Once the correction is complete, the fractional neural network diagnostic module processes the newly input standardized feature data based on the updated parameters.

[0017] The beneficial effects are as follows: by constructing a feedback loop through actual fault information from maintenance, the neural network parameters are iteratively corrected to prevent the model from losing diagnostic accuracy due to changes in operating conditions during long-term use; the corrected parameters make the model converge more smoothly, and after feedback optimization, the fault diagnosis accuracy gradually improves with the length of use; the threshold setting of the diagnostic deviation value ensures that the correction is targeted, avoids ineffective iteration, and ensures that high diagnostic reliability is maintained for a long time under complex operating conditions.

[0018] Furthermore, after outputting the feature data from the hidden layer, the fractional neural network diagnostic module performs a consistency check on the feature data, calculating the Euclidean distance between the feature data and the normal feature template under similar historical conditions. When the Euclidean distance is greater than a preset distance threshold, it is determined that the feature data has a deviation. The fractional neural network diagnostic module then calls a built-in correction algorithm to correct the deviation of the current hidden layer output feature data based on the historical normal feature template. The correction formula is as follows: ,in, For the corrected feature data, For deviation characteristic data, λ is the mean of the historical normal feature template and λ is the correction coefficient. After correction, the fractional neural network diagnosis module performs fully connected layer mapping and Softmax classification processing on the corrected feature data to obtain the corrected fault type probability distribution data. The fault output module outputs the final fault diagnosis result based on the corrected fault type probability distribution data.

[0019] The beneficial effects are as follows: by verifying the deviation of the hidden layer feature data and making targeted corrections in combination with historical normal templates, feature distortion caused by sensor drift, electromagnetic interference and other factors can be effectively eliminated, and false fault signal interference can be reduced; by taking reasonable values ​​for the correction coefficients, the authenticity of features and the effectiveness of correction can be balanced, the fluctuation range of the probability distribution data of fault types after correction can be reduced, and the stability of diagnostic results can be improved.

[0020] Furthermore, the fault output module presets three levels of fault thresholds: a warning threshold T1, an alarm threshold T2, and an emergency threshold T3, where T1 < T2 < T3. The fault output module compares the fault type probability distribution data output by the fractional neural network diagnostic module with the three levels of fault thresholds: when T1 < fault type probability value ≤ T2, a first-level diagnostic result is output, which includes the fault type and a recommendation for regular inspection; when T2 < fault type probability value ≤ T3, a second-level diagnostic result is output, which includes the fault type, affected components, and a prompt repair reminder; when the fault type probability value > T3, a third-level diagnostic result is output, which includes the fault type, urgency level, and an immediate stop and repair instruction. The fault output module displays the diagnostic results hierarchically on the in-vehicle display screen, and triggers different frequencies of audio and visual prompts corresponding to different levels of results.

[0021] The beneficial effects are as follows: the three-level threshold division makes the severity of faults visible, enabling drivers to quickly distinguish between minor hidden dangers and emergency faults, avoiding improper handling due to vague information; different levels of results correspond to differentiated prompts and instructions, with level one results guiding preventive maintenance, level two results prompting timely repairs, and level three results ensuring emergency safety, achieving precise control over the entire life cycle of faults; the combination of audible and visual prompts and hierarchical display improves the efficiency of result communication and reduces the safety risks caused by untimely fault handling in new energy vehicles.

[0022] Furthermore, the data acquisition module, while collecting multi-dimensional raw data, simultaneously collects specific time environment data of the vehicle, including real-time time information and environmental parameters. The data preprocessing module encodes the specific time environment data and fuses it with the outlier-removed and standardized multi-dimensional raw data to obtain spatiotemporal fusion standardized feature data. The fractional neural network diagnostic module has built-in fault feature weight matrices for different time environments. When the spatiotemporal fusion standardized feature data is used as input, it calls the weight matrix of the corresponding time environment, calculates the hidden layer output feature data by combining it with Caputo-type α-order fractional differential equations, and then performs fully connected layer mapping and Softmax classification processing to obtain fault type probability distribution data for the specific time environment. The fault output module outputs fault diagnosis results adapted to the specific time environment based on this probability distribution data.

[0023] The beneficial effects are as follows: the integration of temporal and environmental data with operational data prevents differences in fault characteristics under different scenarios; the built-in scenario-based weight matrix enables the fractional-order network to focus on key fault characteristics under specific environments, which can improve the accuracy of fault identification under specific time environments compared with general diagnosis and avoid misjudgment caused by scenario differences; the spatiotemporal fusion features enhance the adaptability of the diagnostic model to environmental changes and improve the reliability and practicality of fault diagnosis.

[0024] A fault diagnosis method for new energy vehicles based on fractional neural networks includes the following steps: Step S1: Collect multi-dimensional raw data during vehicle operation; Step S2: Perform outlier removal and standardization on the multi-dimensional raw data to obtain standardized feature data; Step S3: Fractional Neural Network Diagnostic Step. Using standardized feature data as input, the hidden layer neuron states of the fractional neural network are updated using a Caputo-type α-order fractional differential equation. The α-order fractional differential equation is: ,in, It is the α-order fractional derivative. Let be the state value of the i-th hidden layer neuron at time t. The neuron decay coefficient, Let be the connection weight between the j-th input neuron and the i-th hidden layer neuron. It is a ReLU fractional activation function. As a bias term, the hidden layer output feature data is calculated through the fractional differential equation; the hidden layer output feature data is then subjected to fully connected layer mapping and Softmax classification to obtain fault type probability distribution data. Step S4: Compare the fault type probability distribution data with the preset fault threshold. When the probability value of a certain fault type is greater than the preset fault threshold, output the corresponding fault diagnosis result.

[0025] The basic principles and beneficial effects of the solution are as follows: The application of Caputo-type α-order fractional differential equations enables the neuron state update to accumulate historical operational data. Compared with integer-order equations that rely solely on instantaneous change rates, this can capture the evolution trend of slowly changing faults and reduce the early fault false alarm rate. Standardization eliminates the differences in the dimensions of multi-dimensional data, providing high-quality input for subsequent diagnosis. Combined with ReLU fractional activation function and Softmax classification, it can effectively distinguish between complex operating condition fluctuations and real faults, reducing the false alarm rate in scenarios such as rapid acceleration and low-temperature start-up. By integrating historical information through fractional derivatives, the diagnostic accuracy of coupled faults is improved, enabling both early fault warning and providing accurate basis for maintenance, thereby reducing vehicle operation and maintenance costs and safety risks. Attached Figure Description

[0026] Figure 1 This is a system block diagram of Embodiment 1 of a fault diagnosis system for new energy vehicles based on fractional neural networks; Figure 2 This is a flowchart of Example 2 of the fault diagnosis method for new energy vehicles based on fractional neural networks. Detailed Implementation

[0027] The following detailed description illustrates the specific implementation method: Example 1

[0028] This invention discloses a fault diagnosis system for new energy vehicles based on fractional neural networks, as shown in the attached figure. Figure 1 As shown, the specific implementation process is as follows: The data acquisition module collects multi-dimensional raw data in real time during vehicle operation via the vehicle's CAN bus. This includes: individual cell voltage, total current, and cell temperature of the power battery; real-time speed, output torque, and casing vibration acceleration of the drive motor; and input voltage, output PWM signal duty cycle, and chip temperature of the electronic control unit. The data preprocessing module processes the collected multi-dimensional raw data: outliers are removed using the 3σ criterion, and then all valid data are mapped to the [0,1] interval using min-max normalization, ultimately obtaining standardized feature data. The fractional neural network diagnostic module takes the standardized feature data as input. It has three hidden layers, each containing 64 neurons. Neuron state updates use a Caputo-type 0.96th-order fractional derivative (α=0.96, within the reasonable range of 0-1). The ReLU fractional activation function performs a nonlinear transformation on the neuron output. For example, when the neuron state value is 0.3, the output after activation function processing is 0.28, simulating a nonlinear enhancement effect. By calculating the fractional differential equation and integrating historical and current state information, the hidden layer output feature data is obtained. After mapping through a fully connected layer (two layers with 32 and 16 nodes respectively), the data is input into a Softmax classifier to obtain fault type probability distribution data. For example, the probability of unbalanced voltage of individual power battery cells is 95%, the probability of abnormal motor vibration is 3%, and the probability of abnormal signal of the electronic control unit is 2%. The fault output module presets a fault threshold of 80%. The probability distribution data is compared with this threshold. Since the probability of unbalanced voltage of individual power battery cells is 95% greater than 80%, the corresponding fault diagnosis result is output, including the fault type (unbalanced voltage of individual power battery cells), fault confidence (95%), and fault development trend prediction (the voltage difference may increase by 0.1V in the next hour).

[0029] More specifically, the data acquisition module continuously collects data at a preset sampling frequency during vehicle movement. The sampling frequency is dynamically adjusted according to the driving speed. For example, when the vehicle speed is ≥60km / h, the sampling frequency is set to 100Hz, meaning 100 sets of data are collected per second, ensuring that data is not lost under high motor load and high battery current output scenarios at high speeds. When the vehicle speed is <60km / h, the sampling frequency is set to 50Hz, reducing the system's computational load while ensuring data validity. The data preprocessing module adopts a real-time streaming processing method, performing outlier removal and standardization on each set of continuously collected multi-dimensional raw data. For example, when the vehicle is traveling at low speeds, 50 sets of motor torque data are received per second. Each set of data is processed according to the 3σ criterion and normalization method to generate a continuous standardized feature data stream. After receiving this data stream, the fractional neural network diagnostic module performs joint calculations on the standardized feature data of the current and past 5 seconds every 5 seconds (preset diagnostic cycle), utilizing the memory effect of the fractional derivative to accumulate historical feature data within this period. For example, during vehicle acceleration, battery current and motor torque data are continuously collected. The dynamic correlation between these two data points is captured through joint calculations, yielding hidden layer output feature data for the corresponding period. This data is then processed by a fully connected layer and Softmax classification to generate real-time fault type probability distribution data. Compared to diagnostic methods that rely solely on instantaneous data from a single moment, this solution explicitly defines a time window encompassing the current moment and the past 5 seconds, performing joint calculations on standardized feature data. This breaks through the limitations of isolated instantaneous judgments from a data perspective. This time window setting aligns with the dynamic evolution of faults in new energy vehicles. Whether it's parameter fluctuations caused by changes in operating conditions such as vehicle acceleration and deceleration, or the initial gradual changes in early faults such as battery degradation and motor wear, their key characteristics require at least several seconds of time-series data to be fully presented. For example, during vehicle acceleration, the battery current rapidly increases from 100A to 180A, and the motor torque increases from 150N·m to 280N·m. The parameter changes in this process are not isolated, but rather there is a dynamic correlation between the current increase and the torque increase. The 5-second time window can perfectly cover the coordinated parameter changes under this type of operating condition, providing complete data support including temporal correlation for subsequent feature extraction, and avoiding the feature bias caused by extracting data from a single moment.

[0030] Because the core advantage of the fractional derivative lies in its inherent memory property, in the joint calculation within a 5-second time window, this memory effect can weighted accumulate historical feature data within this time window, rather than simply adding up the values. Specifically, the fractional derivative, through a weighting mechanism of near-large and far-small, ensures that recent data (such as the current moment or data from the previous second) within the time window maintains a higher weight, while older data (such as data from the previous 4-5 seconds) retains its effective contribution, ensuring that the evolutionary trend of historical features is accurately remembered. For example, during acceleration, if the battery current shows a steady increase in the first 3 seconds and exhibits abnormal fluctuations in the last 2 seconds, the memory effect of the fractional derivative will highlight the abnormal trend in the last 2 seconds in the cumulative calculation, while simultaneously relating it to the stable base of the first 3 seconds, clearly depicting the gradual process from normal to abnormal. In contrast, traditional instantaneous rate of change judgment can only capture fluctuations at a certain moment and cannot relate them to the evolutionary logic of the preceding and following data. This cumulative effect can keenly capture subtle trends in parameter changes within 5 seconds, and it has a strong ability to perceive the initial gradual characteristics of early faults, such as the slight increase in motor vibration acceleration during acceleration.

[0031] Based on joint calculation of 5-second data and fractional-order memory accumulation, the system can accurately capture the dynamic correlation between multiple parameters. Further processing via fully connected layer mapping and Softmax classification enhances the real-time performance and accuracy of fault identification. Taking vehicle acceleration as an example, the normal dynamic correlation between battery current and motor torque follows a specific pattern: for example, a 20A increase in current corresponds to a 30N·m increase in torque. However, if the correlation becomes unbalanced due to aging of the motor winding insulation (e.g., a 20A increase in current results in only a 15N·m increase in torque), the abnormal correlation can be quickly identified through joint calculation of the two types of data within 5 seconds, combined with historical correlation features accumulated via fractional-order memory. After enhancing feature discrimination through fully connected layer mapping, the Softmax classifier can accurately output the probability distribution data of fault types related to motor insulation aging, ensuring the reliability of real-time diagnostic results. This design avoids interference from instantaneous fluctuations under complex operating conditions, such as brief parameter jitters during acceleration which are considered normal due to the lack of a sustained correlation trend. It also accurately identifies real faults with temporal correlations, achieving the dual goals of real-time response and accurate identification, significantly improving the timeliness and accuracy of fault diagnosis during driving.

[0032] This collaborative mechanism, utilizing joint data calculation within a time window and the accumulation of fractional-order memory effects, enables precise capture of the dynamic evolution characteristics of faults. The fault output module dynamically compares real-time probability data with a preset fault threshold (80%), achieving self-detection of faults while the vehicle is in motion. For example, if the data shows a real-time probability of "abnormal motor vibration acceleration" of 88% while the vehicle is in motion, which is greater than the threshold, the fault diagnosis result is immediately output, ensuring that sudden faults during travel can be identified in a timely manner.

[0033] More specifically, during vehicle operation, the fractional neural network diagnostic module, based on continuously received standardized feature data streams, utilizes historical state data accumulated over the past 10 minutes using a Caputo-type 0.6th order fractional differential equation to construct a fault evolution time-series model. This model, based on trend features in historical data (such as battery voltage fluctuations and motor vibration energy changes), predicts trends in standardized feature data for the next 5 minutes, obtaining predicted feature data. For example, it predicts that the voltage difference of a certain battery cell will increase from 0.2V to 0.35V within the next 3 minutes. This predicted feature data is then substituted into the α-order fractional differential equation for calculation, yielding the predicted hidden layer output feature data. After mapping by a fully connected layer and softmax classification, it generates a fault type prediction probability distribution data for the next 5 minutes; for example, the predicted probability for "uneven voltage of individual power battery cells" is 86%. The fault output module has a preset warning threshold of 75%. It compares the predicted probability of 86% with this threshold. Since it is greater than the warning threshold, it outputs the fault prediction result, including the predicted fault type (uneven voltage of power battery cells), the predicted occurrence time (the probability of fault will exceed 90% within the next 3-5 minutes), and the risk level (medium risk).

[0034] More specifically, before performing outlier removal and standardization on the multi-dimensional raw data, the data preprocessing module first performs correlation analysis: calculating the Pearson correlation coefficient between different types of raw data. For example, the correlation coefficient between the output current of the power battery and the input torque of the motor is 0.78, which is greater than 0.6, and is therefore selected as a strongly correlated data pair; the correlation coefficient between motor speed and vibration acceleration is 0.65, also belonging to a strongly correlated data pair. Cross-feature construction is then performed on these strongly correlated data pairs: for the battery output current-motor input torque data pair, the product feature, difference feature, and ratio feature are calculated; for the motor speed-vibration acceleration data pair, the above three types of cross-features are also constructed to generate correlated feature data. The data preprocessing module then merges these correlated feature data with the single-dimensional standardized data (such as standardized data of battery temperature and electronic control unit voltage) after outlier removal and standardization, obtaining fused standardized feature data. This data contains both single-dimensional basic information and multi-dimensional correlated information. The fractional neural network diagnostic module takes fused standardized feature data as input, calculates the hidden layer output feature data through the above-mentioned Caputo type 0.96 fractional differential equation, and then performs fully connected layer mapping and Softmax classification to obtain fault type probability distribution data with fused associated features. For example, the probability of "abnormal coupling between motor torque and current" is 92%. The fault output module completes fault diagnosis based on this data.

[0035] In this embodiment, by calculating the Pearson correlation coefficient between multi-dimensional raw data, strongly correlated data pairs with an absolute correlation coefficient ≥ 0.6 are selected, achieving accurate extraction of fault correlation information. Since faults in new energy vehicles are often coupled, such as the highly coordinated changes in parameters like the power battery output current and motor input torque, and motor speed and vibration acceleration, abnormal fluctuations in battery current can directly lead to imbalances in motor torque output, while unstable motor speed can cause abnormal vibration acceleration. The selection of strongly correlated data pairs with correlation coefficients of 0.78 and 0.65 effectively eliminates interference from uncorrelated or weakly correlated data, such as weakly correlated data like environmental humidity and electronic control unit voltage. This focuses on core parameter combinations that play a crucial role in fault evolution, providing a high-value data foundation for subsequent feature construction and avoiding invalid data consuming computational resources and interfering with diagnostic logic.

[0036] For strongly correlated data pairs, three types of cross-features—product, difference, and ratio—are constructed to transform the collaborative fault information hidden among multiple parameters into quantifiable feature data. Single-dimensional raw data can only reflect its own changing state, while cross-features can accurately characterize the dynamic matching relationship between parameters. For example, the product feature of battery output current and motor input torque can quantify the energy conversion efficiency of current and torque. If the conversion efficiency decreases due to aging of the motor windings, this product feature will show an abnormally low trend. The difference feature reflects the parameter response delay, while the ratio feature reflects the matching proportion between the two. Similarly, the cross-feature of motor speed and vibration acceleration can capture the inherent correlation between speed and vibration. If bearing wear causes an abnormal rate of vibration increase with speed, the ratio feature will show a significant deviation. This construction of cross-features allows fault identification to move beyond the superficial identification of single-parameter anomalies to the essential judgment of multi-parameter collaborative anomalies, improving the recognizability of fault features.

[0037] By fusing correlated feature data with single-dimensional standardized data, fused standardized feature data is obtained, achieving full coverage of both collaborative correlated features and independent basic features. Single-dimensional data (such as battery temperature and electronic control unit voltage) can reflect the operating status of a single component, while correlated feature data can characterize the coupling relationship between multiple components. The feature system formed by the fusion of the two covers both independent fault signals of a single component (such as single-parameter anomalies of excessively high battery temperature) and coupled fault signals of multiple components (such as correlated anomalies of current-torque mismatch). This comprehensive feature input allows the fractional neural network diagnostic module to capture fault information from both independent state and collaborative relationship dimensions, avoiding missed detections of coupled faults due to reliance on only single-dimensional data or misjudgments of independent faults due to a lack of basic features. This provides more comprehensive and accurate feature support for subsequent diagnosis, laying a solid foundation for improving diagnostic accuracy from the data preprocessing stage.

[0038] More specifically, after the fault output module outputs the diagnostic result of "uneven voltage of individual power battery cells," the vehicle is inspected by a repair shop, and the actual fault is confirmed as "degradation of power battery cell No. 3." The data acquisition module collects the actual fault information confirmed by the repair as feedback data. The fault output module compares the actual fault information with the corresponding fault type probability distribution data and calculates the diagnostic deviation value: the actual fault "degradation of cell No. 3" is included in the "uneven voltage of individual power battery cells" category in the original prediction, with an original predicted probability of 95%. The probability corresponding to the actual fault is calculated as 100%, and the diagnostic deviation value is |100%-95%|=5%. If the preset deviation threshold is 8%, and 5% < 8%, then no correction is triggered; if in a certain diagnosis, the actual fault is "electronic control unit signal interference," the original predicted probability is 70%, and the deviation value is 30% > 8%, the fault output module transmits the feedback data and the 30% diagnostic deviation value to the fractional neural network diagnostic module. The fractional neural network diagnostic module uses this feedback data as a sample and employs the fractional gradient descent method to analyze the neuron decay coefficient α (originally 0.98) and connection weight w in the Caputo-type α-order fractional differential equation. ij (The original weight is 0.32) Iterative correction is performed. If the attenuation coefficient α is corrected to 0.95, the connection weight w ij The value was adjusted to 0.36, and the diagnostic bias was minimized through multiple iterations. After the adjustment, the module processed the newly input standardized feature data based on the updated parameters to improve the accuracy of subsequent diagnoses.

[0039] More specifically, after outputting the hidden layer's feature data, the fractional neural network diagnostic module performs a consistency check on the data: it extracts normal feature data from historical conditions of similar operation (e.g., same vehicle speed, same battery SOC, same ambient temperature), constructs a historical normal feature template, and calculates the Euclidean distance between the current hidden layer's output feature data and this template. For example, if the current operating conditions are a vehicle speed of 50 km / h, a battery SOC of 60%, and an ambient temperature of 25℃, the corresponding historical normal feature template average is 0.45, and the current hidden layer's output feature data is 0.68, the calculated Euclidean distance is 0.23. If the preset distance threshold is 0.2, and 0.23 > 0.2, then the feature data is considered to have a deviation. At this point, the built-in correction algorithm is invoked, with the correction coefficient λ set to 0.3. Based on the historical normal feature template average of 0.45, the current deviated feature data of 0.68 is corrected to obtain the corrected feature data.

[0040] After correction, the fractional neural network diagnosis module performs fully connected layer mapping and Softmax classification on the corrected feature data to obtain the corrected fault type probability distribution data. For example, the original probability of "abnormal motor vibration" is 82%, which is reduced to 76% after correction. The fault output module outputs the final fault diagnosis result based on the corrected data to avoid misjudgment caused by feature distortion.

[0041] More specifically, the fault output module presets three levels of fault thresholds: warning threshold T1=60%, alarm threshold T2=80%, and emergency threshold T3=90%, and T1<T2<T3. When the probability distribution data of fault types output by the fractional neural network diagnostic module shows that the probability of "power battery SOC estimation deviation" is 75%, which is greater than T1 (60%) and less than or equal to T2 (80%), a first-level diagnostic result is output, including the fault type (power battery SOC estimation deviation) and a recommendation for regular inspection (it is recommended to calibrate the SOC algorithm at a service station within 1 month); when the probability of "motor bearing wear" is 85%, which is greater than T2 (80%) and less than or equal to T3 (90%), a second-level diagnostic result is output, including the fault type (motor bearing wear), affected components (drive motor bearing), and a prompt repair reminder (it is recommended to replace the bearing within 7 days); when the probability of "electronic control unit short circuit" is 93%, which is greater than T3 (90%), a third-level diagnostic result is output, including the fault type (electronic control unit short circuit), the urgency level (extremely high), and an immediate stop and repair instruction (please pull over immediately, turn off the power, and contact roadside assistance). The fault output module displays diagnostic results in stages on the vehicle's central control display screen: Level 1 results are displayed in green text, Level 2 results are displayed in yellow text with flashing indicators, and Level 3 results are displayed in red text with continuous flashing indicators. At the same time, different frequencies of audio and visual prompts are triggered. Level 1 results have no audio prompt, Level 2 results emit a "beep" sound once every 3 seconds, and Level 3 results emit two rapid "beep beep" sounds per second to ensure that the driver can quickly identify the fault level.

[0042] More specifically, the data acquisition module collects specific time-related environmental data simultaneously while acquiring multi-dimensional raw data: real-time time information includes weekdays / weekends and peak / off-peak hours; environmental parameters include ambient temperature, humidity, and precipitation status. The data preprocessing module encodes the specific time-related environmental data: weekdays are coded as 1, weekends as 0; peak hours as 1, off-peak hours as 0; ambient temperature is coded by range (-40℃-0℃ is 0, 0℃-25℃ is 1, 25℃-85℃ is 2); humidity is coded by range (0%-30% is 0, 30%-70% is 1, 70%-100% is 2); precipitation status is coded as 0 (no rain), 1 (light rain), 2 (moderate rain), and 3 (heavy rain). The coded time-related environmental data is then fused with the outlier-removed and standardized multi-dimensional raw data to obtain spatiotemporally fused standardized feature data. The fractional neural network diagnostic module has built-in fault feature weight matrices for different time environments. For example, "high temperature environment (code 2) + peak period (code 1)" corresponds to a set of weight matrices, in which the feature weights of motor temperature and battery current are higher than those of other operating conditions. When spatiotemporal fusion standardized feature data is input, the module automatically identifies the current time environment (such as weekday peak period, ambient temperature 35℃, no rain), calls the corresponding weight matrix, and calculates the hidden layer output feature data by combining the Caputo type 0.6 fractional differential equation. Then, after mapping by a fully connected layer and softmax classification processing, the probability distribution data of fault types under this specific time environment is obtained. For example, the probability of "motor insulation aging under high temperature and high load" is 89%. The fault output module outputs fault diagnosis results adapted to the current environment based on this data. Example 2

[0043] The only difference from Example 1 is that, based on this, a fault diagnosis method for new energy vehicles based on fractional-order neural networks is disclosed, as shown in the appendix. Figure 2 As shown, the specific implementation process is as follows: Step S1: Data Acquisition. A communication connection is established with the vehicle's power battery management system, motor controller, vehicle controller, and various onboard sensors via the vehicle's CAN bus to collect multi-dimensional raw data during vehicle operation. This includes: the total voltage of the power battery, charging and discharging current, and temperature of each module; the drive motor's speed, output torque, stator winding temperature, and vibration acceleration; and the input power supply voltage, output control signal frequency, and chip operating temperature of the electronic control unit. The acquisition process is continuous, with a sampling interval of 10ms to ensure the real-time nature and continuity of the data.

[0044] Step S2: Data Preprocessing. The collected multi-dimensional raw data is preprocessed: First, outliers are removed using the 3σ criterion. For example, if the collected temperature of a certain group of motor stator windings is 150℃, far exceeding the 3σ boundary value of 130℃ within the normal operating range, this data is identified as an outlier and removed. For missing data (such as a missing set of battery current data due to a momentary interruption of the CAN bus), the average of the five adjacent sets of data is used to fill in the missing data. Then, using the min-max normalization method, all valid data are uniformly mapped to the [0,1] interval. For example, the total battery voltage of 200V corresponds to 0, 400V corresponds to 1, and a total voltage of 300V at a certain moment is normalized to 0.5; the motor vibration acceleration of 0m / s² corresponds to 0, 25m / s² corresponds to 1, and a vibration acceleration of 12.5m / s² at a certain moment is normalized to 0.5. Finally, standardized feature data is obtained, eliminating the influence of differences in data units.

[0045] Step S3: Fractional Neural Network Diagnostic Step. Using the obtained standardized feature data as input, a fractional neural network model is constructed. This model includes an input layer, hidden layers, and an output layer. The input layer has 18 nodes (corresponding to 18 categories of multi-dimensional raw data), the hidden layer has 4 layers with 128 nodes each, and the output layer has 12 nodes (corresponding to 12 common fault types). The hidden layer neuron state update uses a Caputo-type 0.5th order fractional derivative (α=0.95). This fractional differential equation integrates information from current and historical standardized feature data. For example, when calculating the state value of a hidden layer neuron at time t, not only the current input standardized feature data is considered, but also historical data from the past minute is included. If the standardized battery current data shows a gradually increasing trend over the past minute, this trend will be accumulated in the current neuron state calculation through the memory effect of the fractional derivative. The ReLU fractional activation function performs a nonlinear transformation on the output of each neuron. For example, if the output of a neuron calculated by the fractional differential equation is 0.4, the output after processing by the activation function is 0.35 (simulating the effect of nonlinear enhancement of feature expression), thus enhancing the model's ability to capture complex fault features. After obtaining the hidden layer output feature data through the above calculation, it is input into a fully connected layer (set to 2 layers, with 64 and 32 nodes respectively) for feature mapping, and finally input into a Softmax classifier to obtain fault type probability distribution data. For example, the probability of the classification result as "uneven temperature of the power battery module" is 91%, the probability of "abnormal vibration of the main drive motor" is 4%, the probability of "MCU signal transmission delay" is 3%, and the sum of the probabilities of other fault types is 2%.

[0046] Step S4: Fault Output Step. A preset fault threshold of 85% is used. The obtained fault type probability distribution data is compared with this threshold one by one. Since the probability of "uneven temperature of the power battery module" is 91%, which is greater than the preset fault threshold of 85%, the fault is determined to be valid, and the corresponding fault diagnosis result is output. The diagnosis result specifically includes: fault type (uneven temperature of the power battery module), fault confidence (91%), and fault development trend prediction value (within the next 30 minutes, the temperature difference between modules may increase from the current 5℃ to 8℃; it is recommended to check the cooling system promptly). If the probability values ​​of all fault types are less than 85%, the result "The current vehicle has no obvious fault; it is recommended to continue normal driving and perform regular maintenance and inspection" is output.

[0047] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A fault diagnosis system for new energy vehicles based on fractional neural networks, characterized in that, It includes a data acquisition module, a data preprocessing module, a fractional neural network diagnostic module, and a fault output module; The data acquisition module collects multi-dimensional raw data during vehicle operation; The data preprocessing module performs outlier removal and standardization on the multi-dimensional raw data to obtain standardized feature data. The fractional neural network diagnostic module takes standardized feature data as input. The hidden layer neuron state update of the fractional neural network diagnostic module employs a Caputo-type α-order fractional differential equation, which is: ,in, It is the α-order fractional derivative. Let be the state value of the i-th hidden layer neuron at time t. The neuron decay coefficient, Let be the connection weight between the j-th input neuron and the i-th hidden layer neuron. It is a ReLU fractional activation function. As a bias term, the hidden layer output feature data is calculated through the fractional differential equation; the fractional neural network diagnostic module performs fully connected layer mapping and Softmax classification on the hidden layer output feature data to obtain fault type probability distribution data. The fault output module compares the fault type probability distribution data with a preset fault threshold. When the probability value of a certain fault type is greater than the preset fault threshold, the corresponding fault diagnosis result is output.

2. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 1, characterized in that, The data acquisition module continuously collects multi-dimensional raw data at a preset sampling frequency during vehicle movement. The preset sampling frequency is dynamically adjusted according to the vehicle's speed. The data preprocessing module performs real-time streaming processing on the continuously collected multi-dimensional raw data, sequentially completing outlier removal and standardization to generate a continuous standardized feature data stream. The fractional neural network diagnostic module receives the standardized feature data stream, updates the state of hidden layer neurons in real time using Caputo-type α-order fractional differential equations, accumulates historical feature data during vehicle movement using the memory effect of fractional derivatives, and performs joint calculations on the current and historical standardized feature data at preset diagnostic intervals to obtain the hidden layer output feature data for the corresponding period. This data is then mapped by a fully connected layer and classified using Softmax to generate real-time fault type probability distribution data. The fault output module dynamically compares the real-time fault type probability distribution data with the preset fault threshold to realize fault self-detection during vehicle movement. When the detected fault type probability value is greater than the preset fault threshold, the fault diagnosis result is output.

3. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 2, characterized in that, During vehicle operation, the fractional neural network diagnostic module constructs a fault evolution time series model based on continuously received standardized feature data streams and historical state data accumulated using Caputo-type α-order fractional differential equations. It then predicts the trend of standardized feature data within a preset future time period to obtain predicted feature data. The module substitutes the predicted feature data into the α-order fractional differential equations for calculation, obtaining predicted hidden layer output feature data. This data is then mapped by a fully connected layer and processed by Softmax classification to generate a predicted probability distribution of fault types within the preset future time period. The fault output module compares the predicted probability distribution data of fault types with the preset warning threshold. When the predicted probability value of a certain fault type is greater than the preset warning threshold, the fault prediction result is output.

4. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 3, characterized in that, Before performing outlier removal and standardization on the multi-dimensional raw data, the data preprocessing module first performs correlation analysis on the multi-dimensional raw data: calculating the Pearson correlation coefficient between different types of raw data and selecting strongly correlated data pairs; constructing cross-features on the strongly correlated data pairs to generate correlated feature data; the data preprocessing module then merges the correlated feature data with the single-dimensional standardized data after outlier removal and standardization to obtain fused standardized feature data; the fractional neural network diagnosis module takes the fused standardized feature data as input, calculates the hidden layer output feature data through Caputo-type α-order fractional differential equations, and then performs fully connected layer mapping and Softmax classification to obtain fault type probability distribution data of fused correlated features; the fault output module completes fault diagnosis based on this fault type probability distribution data.

5. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 4, characterized in that, After the fault output module outputs the fault diagnosis result, the data acquisition module collects the actual fault information confirmed by subsequent vehicle maintenance as feedback data. The fault output module compares the feedback data with the corresponding fault type probability distribution data to calculate the diagnostic deviation value, which is the absolute value of the difference between the actual fault type probability and the predicted probability. When the diagnostic deviation value > a preset deviation threshold, the fault output module transmits the feedback data and the diagnostic deviation value to the fractional neural network diagnostic module. The fractional neural network diagnostic module uses the feedback data as a sample to analyze the neuron decay coefficients in the Caputo-type α-order fractional differential equation. and connection weights The process involves iterative correction, employing fractional gradient descent to minimize the diagnostic bias. Once the correction is complete, the fractional neural network diagnostic module processes the newly input standardized feature data based on the updated parameters.

6. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 5, characterized in that, After outputting feature data from the hidden layer, the fractional neural network diagnostic module performs a consistency check on the feature data, calculating the Euclidean distance between the feature data and the normal feature template under similar historical conditions. When the Euclidean distance exceeds a preset distance threshold, it is determined that the feature data has a deviation. The fractional neural network diagnostic module then calls a built-in correction algorithm to correct the deviation of the current hidden layer output feature data based on the historical normal feature template. The correction formula is as follows: ,in, For the corrected feature data, For deviation characteristic data, λ is the mean of the historical normal feature template and λ is the correction coefficient. After correction, the fractional neural network diagnosis module performs fully connected layer mapping and Softmax classification processing on the corrected feature data to obtain the corrected fault type probability distribution data. The fault output module outputs the final fault diagnosis result based on the corrected fault type probability distribution data.

7. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 6, characterized in that, The fault output module presets three levels of fault thresholds, namely the early warning threshold T1, the alarm threshold T2, and the emergency threshold T3, where T1 < T2 < T3. The fault output module compares the fault type probability distribution data output by the fractional neural network diagnostic module with the three-level fault thresholds: when T1 < fault type probability value ≤ T2, a first-level diagnostic result is output, which includes the fault type and recommendations for regular inspection. When T2 < fault type probability value ≤ T3, output secondary diagnostic results, which include fault type, affected components, and prompt repair reminders. When the probability value of the fault type is greater than T3, a three-level diagnostic result is output. The three-level diagnostic result includes the fault type, urgency level, and an immediate stop and repair instruction. The fault output module displays the diagnostic results in a hierarchical manner through the vehicle display screen, and triggers different frequencies of audio and visual prompts corresponding to different levels of results.

8. The new energy vehicle fault diagnosis system based on fractional neural network according to claim 7, characterized in that, The data acquisition module collects multi-dimensional raw data while simultaneously acquiring specific time-environment data of the vehicle, including real-time time information and environmental parameters. The data preprocessing module encodes the specific time-environment data and fuses it with the outlier-removed and standardized multi-dimensional raw data to obtain spatiotemporal fusion standardized feature data. The fractional neural network diagnostic module has built-in fault feature weight matrices for different time environments. When the spatiotemporal fusion standardized feature data is used as input, it calls the weight matrix of the corresponding time environment, calculates the hidden layer output feature data by combining it with Caputo-type α-order fractional differential equations, and then performs fully connected layer mapping and Softmax classification processing to obtain fault type probability distribution data for the specific time environment. The fault output module outputs fault diagnosis results adapted to the specific time environment based on this probability distribution data.

9. A fault diagnosis method for new energy vehicles based on fractional neural networks, characterized in that, Includes the following steps: Step S1: Collect multi-dimensional raw data during vehicle operation; Step S2: Perform outlier removal and standardization on the multi-dimensional raw data to obtain standardized feature data; Step S3: Fractional Neural Network Diagnostic Step. Using standardized feature data as input, the hidden layer neuron states of the fractional neural network are updated using a Caputo-type α-order fractional differential equation. The α-order fractional differential equation is: ,in, It is the α-order fractional derivative. Let be the state value of the i-th hidden layer neuron at time t. The neuron decay coefficient, Let be the connection weight between the j-th input neuron and the i-th hidden layer neuron. It is a ReLU fractional activation function. The bias term is used to calculate the hidden layer output feature data through the fractional differential equation. The hidden layer output feature data is mapped by a fully connected layer and classified using Softmax to obtain fault type probability distribution data; Step S4: Compare the fault type probability distribution data with the preset fault threshold. When the probability value of a certain fault type is greater than the preset fault threshold, output the corresponding fault diagnosis result.

10. The fault diagnosis method for new energy vehicles based on fractional neural networks according to claim 9, characterized in that, The state update of neurons in fractional neural networks uses the Caputo type 0.96th order fractional derivative.

Images