A vehicle running state real-time monitoring method and system based on the Internet of Things

Abstract

The application relates to the field of vehicle detection, and provides a vehicle operation state real-time monitoring method and system based on Internet of Things, which comprises the following steps: performing multistage pretreatment on original signals collected by a plurality of source sensors of a vehicle to obtain standardized multimodal data flow; inputting an edge computing unit, performing feature extraction on the standardized multimodal data flow through a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and performing abnormality detection through a manifold alignment algorithm to obtain a feature data packet with a priority label; performing channel allocation on a multi-channel communication interface through a game theory model to obtain a target transmission scheme, uploading the feature data packet to the cloud according to the target transmission scheme to obtain a time sequence feature database; and inputting the time sequence feature database into a cloud digital twin model, performing fault diagnosis on the time sequence feature database through a residual fusion algorithm to obtain a structured fault diagnosis report. The application improves the driving safety of the vehicle.

Description

Technical Field

[0001] This invention relates to the field of vehicle inspection technology, and in particular to a method and system for real-time monitoring of vehicle operating status based on the Internet of Things. Background Technology

[0002] With the continuous advancement of automotive electronics and intelligence, real-time monitoring of vehicle operating status has become a crucial means of ensuring driving safety and reducing maintenance costs. Modern vehicles are equipped with numerous heterogeneous sensors that continuously collect multi-dimensional operating data such as engine speed, tire pressure, chassis posture, and brake temperature. This data, aggregated via the vehicle's onboard bus, possesses the potential value to reflect the vehicle's health status. Simultaneously, the rapid development of edge computing, vehicle-to-everything (V2X) communication, and cloud-based deep learning technologies provides the technological foundation for constructing an edge-cloud collaborative vehicle health monitoring architecture.

[0003] Existing vehicle health monitoring methods typically extract statistical features from sensor data at the on-board or cloud level and set fixed thresholds to trigger fault warnings, or use spectral analysis to identify anomalies in signals such as vibration and temperature. However, these methods all make judgments on a single or a few feature dimensions at the numerical level, making it difficult to perceive changes in the overall distribution structure of multidimensional sensor data in high-dimensional space.

[0004] Secondly, under normal vehicle conditions, multidimensional sensor data exhibits a stable low-dimensional manifold structure in high-dimensional space. Early faults often manifest as slight distortions or local topological deviations in this manifold structure. Such structural changes are almost imperceptible at the level of single numerical features or simple statistics, resulting in insufficient early detection capability of existing methods in the budding stage of faults.

[0005] In addition, existing vehicle-to-everything (V2X) transmission solutions typically assign transmission channels statically based on data priority. In complex network environments with multiple channels coexisting, when multiple high-priority data packets compete for limited channel resources at the same time, the static routing strategy cannot dynamically perceive the real-time load status and transmission benefits of each channel, which can easily cause transmission delays or even loss of critical early warning data. It is difficult to ensure the timeliness of delivery of critical health data in scenarios of network congestion or unstable signals. Summary of the Invention

[0006] This invention provides a method and system for real-time monitoring of vehicle operating status based on the Internet of Things, in order to overcome the shortcomings of the prior art.

[0007] The first aspect of this invention provides a method for real-time monitoring of vehicle operating status based on the Internet of Things, comprising:

[0008] S1. Perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream;

[0009] S2. Input the standardized multimodal data stream into the edge computing unit, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package with priority marking.

[0010] S3. Based on the priority identifier in the feature data packet, channel allocation is performed on the multi-channel communication interface using a game theory model to obtain the target transmission scheme. The feature data packet is then uploaded to the cloud according to the target transmission scheme to obtain a time-series feature database.

[0011] S4. Input the time series feature database into the cloud digital twin model, and perform fault diagnosis on the time series feature database through the residual fusion algorithm to obtain a structured fault diagnosis report.

[0012] According to the method for real-time monitoring of vehicle operating status based on the Internet of Things provided by the present invention, step S1 further includes:

[0013] S11. The original acquisition signals from multiple sensors are processed by a low-pass filter circuit to remove high-frequency electromagnetic interference components and obtain a primary filtered signal.

[0014] S12. For vibration signals in the primary filtered signal, denoising is performed using a wavelet threshold denoising algorithm; for slowly varying signals in the primary filtered signal, smoothing is performed using a moving average filter to obtain a denoised signal set.

[0015] S13. Based on the denoised signal set, all channel signals are uniformly aligned to a preset sampling granularity using a time interpolation alignment algorithm, and then encapsulated into a standardized multimodal data stream.

[0016] According to the Internet of Things-based real-time vehicle operation status monitoring method provided by the present invention, step S2 further includes:

[0017] S21. Extract time-domain statistical features, frequency-domain power spectral density features, and time-frequency domain semantic features from the sliding window data of the standardized multimodal data stream to obtain three feature subsets.

[0018] S22. The three feature subsets are concatenated and fused to obtain the comprehensive feature vector;

[0019] S23. Anomaly detection is performed on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package.

[0020] According to the present invention, a method for real-time monitoring of vehicle operating status based on the Internet of Things, step S21 specifically includes:

[0021] From the time series data within the sliding window, the mean, variance, peak factor, waveform factor, and kurtosis are calculated to obtain the time-domain statistical characteristics.

[0022] The time-series data is subjected to spectral analysis by fast Fourier transform, and multiple frequency bands are divided using the Mel scale and the energy proportion of each frequency band is calculated to obtain the frequency domain power spectral density characteristics.

[0023] A time-frequency graph is generated by short-time Fourier transform, and the time-frequency graph is input into a lightweight convolutional neural network to extract the semantic feature vector in the time-frequency domain.

[0024] According to the method for real-time monitoring of vehicle operating status based on the Internet of Things provided by the present invention, step S23 further includes:

[0025] S231. Construct a k-nearest neighbor graph from the historical window data of the comprehensive feature vector using the incremental local linear embedding algorithm, perform manifold modeling on the topological structure of the data points in the low-dimensional space, and obtain the current manifold structure model.

[0026] S232. For newly arrived integrated feature vector data points, calculate the embedding distance and local curvature change relative to the current manifold structure model to obtain the manifold deviation metric.

[0027] S233. The manifold deviation metric and the anomaly score output by the isolated forest model are weighted and fused, and the health status is divided according to a preset threshold to obtain the local preliminary judgment label.

[0028] S234. The local preliminary judgment label and the manifold deviation metric are encapsulated to generate a feature data packet with priority tags.

[0029] According to the Internet of Things-based real-time vehicle operation status monitoring method provided by the present invention, step S3 further includes:

[0030] S31. Using the reciprocal of delay, channel reliability score, and transmission cost as components, dynamically adjust the weights of each component according to the priority identifier in the feature data packet to construct a multi-objective utility function.

[0031] S32. Treating C-V2X, cellular networks, Wi-Fi and low-power wide area networks as game participants, continuously optimize multiple channel strategies through iterative best response algorithm, solve for Nash equilibrium point, and obtain preliminary transmission scheme;

[0032] S33. When all channel utility values ​​are lower than the preset threshold, a transmission degradation strategy is negotiated through a bargaining game model on the dimensions of timeliness and data integrity to modify the preliminary transmission scheme and obtain the target transmission scheme.

[0033] S34. Upload the data according to priority levels based on the target transmission scheme to obtain a time-series feature database.

[0034] According to the Internet of Things-based real-time vehicle operation status monitoring method provided by the present invention, step S4 further includes:

[0035] S41. Based on the vehicle dynamics equations and thermodynamic models, the measured parameters in the time-series feature database are used to drive the simulation calculation of the physical mechanism sub-model to obtain the theoretical expected state values ​​of multiple vehicle components.

[0036] S42. Input the historical comprehensive feature vector sequence in the time series feature database into the time series Transformer network to learn the time series dependencies of multiple feature dimensions under normal driving mode and obtain the anomaly probability vector.

[0037] S43. Calculate the residual vector between the measured comprehensive feature vector in the time series feature database and the theoretical expected state value, and perform residual fusion of the residual vector and the anomaly probability vector through a multilayer perceptron to obtain a fault type probability distribution vector covering multiple fault modes.

[0038] S44. Generate the structured fault diagnosis report based on the probability distribution vector of the fault type combined with manifold evolution trajectory analysis.

[0039] A second aspect of the present invention provides a real-time vehicle operating status monitoring system based on the Internet of Things, comprising:

[0040] Preprocessing module: Used to perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream;

[0041] The detection module, configured based on an edge computing unit, is used to receive the standardized multimodal data stream, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data packet with priority marking.

[0042] Transmission module: used to allocate channels to the multi-channel communication interface according to the priority identifier in the feature data packet through a game theory model, obtain a target transmission scheme, and upload the feature data packet to the cloud according to the target transmission scheme to obtain a time series feature database;

[0043] The diagnostic module, configured based on a cloud-based digital twin model, is used to receive the time-series feature database, perform fault diagnosis on the time-series feature database using a residual fusion algorithm, and obtain a structured fault diagnosis report.

[0044] A third aspect of the present invention provides an Internet of Things (IoT)-based real-time vehicle operation status monitoring device, comprising: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the IoT-based real-time vehicle operation status monitoring device to execute an IoT-based real-time vehicle operation status monitoring method as described in any of the preceding claims.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for real-time monitoring of vehicle operating status based on the Internet of Things as described in any of the preceding claims.

[0046] This invention first performs hardware filtering, wavelet thresholding denoising, and time interpolation alignment on the raw signals from multiple sensors sequentially. This ensures that the generated standardized multimodal data stream has high consistency in format and timing, providing a reliable data input foundation for subsequent processing stages. Secondly, this invention introduces an incremental local linear embedding algorithm at the edge to construct a dynamic manifold structure model. By calculating the embedding distance and local curvature change of new data points, the perturbation degree of the current manifold topology is quantified. This allows for the effective capture of subtle changes in the multidimensional sensor data distribution structure during the fault initiation stage. Compared to traditional methods relying solely on numerical thresholds, this invention can provide an early warning response before the fault causes obvious numerical anomalies, significantly advancing the anomaly detection time window. Thirdly, this invention weightedly fuses the manifold deviation metric and the anomaly score output by the isolated forest model to generate a priority-labeled feature data package. This further reduces the probability of misjudgment due to sensor noise fluctuations in a single model, making the judgment results more robust and reliable. In the data transmission stage, this invention employs a channel allocation mechanism based on a multi-objective utility function and an iterative optimal response algorithm to solve for the Nash equilibrium point. This overcomes the limitation of traditional static priority routing, which is prone to channel congestion in multi-packet contention scenarios. It enables dynamic optimal allocation of resources for each communication channel, effectively ensuring the timeliness and reliability of critical early warning data delivery in network congestion or signal instability environments. When network quality deteriorates, the bargaining game model autonomously negotiates a compromise between timeliness and data integrity, preventing complete interruption of early warning information due to network anomalies. Furthermore, the cloud-based digital twin model of this invention integrates the theoretical expected state values ​​of the physical mechanism sub-model, the anomaly probability vector output by the temporal Transformer network, and the manifold evolution trajectory analysis results. After residual fusion, the output fault type probability distribution vector covers multiple fault modes, and the diagnostic conclusions balance mechanistic interpretability with data-driven generalization capabilities. Combined with the transfer adaptation mechanism of the model-independent meta-learning algorithm, an effective personalized baseline model can be established for new vehicles with relatively low accumulated mileage, solving the problem of insufficient diagnostic accuracy for vehicles with scarce data. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 A schematic diagram of a method for real-time monitoring of vehicle operating status based on the Internet of Things provided by the present invention;

[0049] Figure 2 This invention provides a schematic diagram of a vehicle operation status real-time monitoring system based on the Internet of Things. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0051] The embodiments of the present invention are described below with reference to the figures.

[0052] like Figure 1 As shown, the present invention provides a method for real-time monitoring of vehicle operating status based on the Internet of Things, comprising:

[0053] S1. Perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream.

[0054] Step S1 further includes:

[0055] S11. The original acquisition signals from multiple sensors are processed by a low-pass filter circuit to remove high-frequency electromagnetic interference components and obtain a primary filtered signal.

[0056] In step S1, the present invention first receives the raw acquisition signals. That is, during the actual operation of the vehicle, the signals from the vibration acceleration sensor, temperature sensor, and fuel pressure sensor in the engine compartment, the wheel speed sensor and tire pressure monitoring sensor at the wheel hub, the three-axis gyroscope and accelerometer at the chassis, and the engine speed, throttle opening, fault codes, etc. read from the OBD-II interface are collected and converged to the vehicle intelligent gateway through a hybrid topology of CAN bus and local Ethernet.

[0057] In step S11, the raw acquired signal undergoes hardware-level low-pass filtering before entering the digital processing module of the vehicle intelligent gateway. This low-pass filter has a cutoff frequency of 50Hz; any signal component with a frequency higher than 50Hz is considered high-frequency electromagnetic interference and filtered out. The principle is that electrical components such as the motor and ignition system inject high-frequency noise into the bus during operation. The frequency of this noise is much higher than the effective frequency range of the vehicle's mechanical status signals. This invention uses a low-pass filter to cut off this noise in the analog signal stage, thus completing the first stage of purification before the signal enters the digital domain. After the above processing, the analog signals output by each sensor are transformed into primary filtered signals that retain valid physical information and remove high-frequency interference, serving as the input for S12.

[0058] S12. For vibration-type signals in the primary filtered signal, denoising is performed using a wavelet threshold denoising algorithm; for slowly varying signals in the primary filtered signal, smoothing is performed using a moving average filter to obtain a denoised signal set.

[0059] Furthermore, the primary filtered signals are divided into two categories based on their physical characteristics, and in step S12, the present invention employs different denoising strategies for processing.

[0060] For vibration signals (i.e., signals output by vibration accelerometers), this invention employs a wavelet threshold denoising algorithm. Specifically, this invention decomposes the time-domain vibration signal into multiple frequency scales using wavelet transform, obtaining wavelet coefficients for each level. The wavelet coefficients corresponding to the true vibration information have larger amplitudes, while the wavelet coefficients corresponding to noise have smaller and more dispersed amplitudes. Subsequently, this invention selects the Daubechies db4 wavelet basis and performs a 5-level decomposition on the signal. A threshold is set at each level, and wavelet coefficients with amplitudes below the threshold are set to zero or reduced, while coefficients with higher amplitudes are retained. Then, an inverse wavelet transform is performed on the processed coefficients to reconstruct the denoised vibration signal. In denoising, the waveform characteristics of the db4 wavelet basis highly match the local transient characteristics of the mechanical vibration signal. The 5-level decomposition effectively separates the effective vibration information from noise in each frequency band, ensuring that the denoised vibration signal retains impact characteristics while suppressing random noise.

[0061] For slowly varying signals such as temperature and pressure, this invention employs a moving average filter for smoothing. Specifically, this invention uses a 16-frame window to calculate the arithmetic mean of the signal samples at the current moment and the previous 15 frames. This mean is used as the output value at the current moment, and then the window slides forward one frame, repeating the above calculation. Furthermore, slowly varying signals themselves change slowly, and the noise superimposed on them is mostly random disturbance. This invention uses a 16-frame window to take the mean, which can cancel out the random disturbances and output a smooth physical quantity estimate. After the two types of signals are processed separately, they are finally merged to form a denoised signal set.

[0062] S13. Based on the denoised signal set, all channel signals are uniformly aligned to a preset sampling granularity using a time interpolation alignment algorithm, and then encapsulated into a standardized multimodal data stream.

[0063] Because the sampling frequencies of the various signals in the noise reduction signal set differ significantly, the sampling frequency of the vibration acceleration sensor is as high as 1kHz, while the sampling frequency of some digital signals read by OBD-II is only 1Hz. Directly splicing signals with different sampling rates will cause the time axis to be misaligned and cannot reflect the overall state of the vehicle at the same moment.

[0064] Therefore, in step S13, this invention uses a time interpolation alignment algorithm to uniformly align all channel signals to a preset sampling granularity of 100ms. Specifically, for signals with a sampling frequency higher than 10Hz (such as a 1kHz vibration signal), the mean or representative statistics of all sampling points within each 100ms window are calculated as the alignment value for that 100ms moment; for signals with a sampling frequency lower than 10Hz (such as a 1Hz OBD digital signal), the estimated value of the intermediate moment is calculated by linear interpolation between two adjacent known sampling points at a time interval of 100ms. After the above processing, all channel signals have a uniform 100ms sampling granularity on the time axis.

[0065] Subsequently, the vehicle-mounted intelligent gateway encapsulates the aligned multi-channel signals in a structured manner according to signal type, acquisition time, physical dimensions, and quality flag bits, generating a standardized multimodal data stream. Each frame of this data stream contains a 32-dimensional feature vector, covering 8-dimensional powertrain parameters, 6-dimensional chassis attitude parameters, 6-dimensional environmental perception parameters, 8-dimensional tire status parameters, and 4-dimensional electrical system parameters. Data frames with the same format are output for different vehicle models and different sensor configurations.

[0066] S2. Input the standardized multimodal data stream into the edge computing unit, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package with priority marking.

[0067] Step S2 further includes:

[0068] S21. Extract time-domain statistical features, frequency-domain power spectral density features, and time-frequency domain semantic features from the sliding window data of the standardized multimodal data stream to obtain three feature subsets.

[0069] Specifically, step S21 includes:

[0070] The mean, variance, peak factor, waveform factor, and kurtosis are calculated from the time-series data within the sliding window to obtain the time-domain statistical characteristics.

[0071] Furthermore, for the extraction of time-domain statistical features, this invention directly calculates the mean (reflecting the DC component and overall level of the signal), variance (reflecting the fluctuation amplitude of the signal), peak factor (the ratio of peak value to root mean square value, reflecting the degree of impact), waveform factor (the ratio of root mean square value to absolute mean value, reflecting the fullness of the waveform), and kurtosis (the ratio of fourth central moment to squared variance, highly sensitive to impact faults) of the discrete sampling sequence of each signal channel within the window, for a total of 5 types of statistics, to obtain time-domain statistical features.

[0072] The time-series data is subjected to spectral analysis using Fast Fourier Transform, and multiple frequency bands are divided using the Mel scale. The energy percentage of each frequency band is calculated to obtain the frequency domain power spectral density characteristics.

[0073] Furthermore, regarding the extraction of frequency domain power spectral density features, this invention performs a Fast Fourier Transform on the time-series data within the window, converting the time-domain signal into a frequency-domain representation to obtain the power spectral density at each frequency point within the range of 0 to 500 Hz. Subsequently, this frequency range is nonlinearly divided using the Mel scale, forming 12 frequency bands. The Mel scale is more densely divided in the low-frequency band and more sparsely divided in the high-frequency band, matching the energy distribution characteristics of the mechanical vibration signal. Next, the power spectral density values ​​within each frequency band are summed and divided by the total power to obtain the energy proportion of each frequency band. The energy proportions of the 12 frequency bands together constitute the frequency domain power spectral density features.

[0074] A time-frequency graph is generated by short-time Fourier transform, and the time-frequency graph is input into a lightweight convolutional neural network to extract the semantic feature vector in the time-frequency domain.

[0075] Furthermore, regarding the extraction of time-frequency domain semantic features, this invention performs a short-time Fourier transform on the time-series data within the window, dividing the signal into several short-time segments. Each segment is then subjected to a Fourier transform, resulting in a two-dimensional time-frequency map with time as the horizontal axis, frequency as the vertical axis, and amplitude as the color depth. The obtained time-frequency map visually represents the energy evolution of each frequency component of the signal at different times. Subsequently, this invention inputs the time-frequency map into a pre-trained lightweight convolutional neural network, MobileNetV3-Small (quantized with INT8). The network's convolutional and pooling layers abstract the local texture and global structure in the time-frequency map layer by layer, finally outputting a 16-dimensional time-frequency domain semantic feature vector by a fully connected layer.

[0076] S22. The three feature subsets are concatenated and fused to obtain the comprehensive feature vector.

[0077] In step S22, the present invention directly concatenates the above-mentioned time-domain statistical features, frequency-domain power spectral density features, and time-frequency semantic feature vectors in a fixed order along the feature dimension to form a 33-dimensional comprehensive feature vector. This vector carries the time-domain statistical characteristics, frequency energy distribution, and time-frequency joint semantic information of the signal, serving as a unified input for subsequent manifold modeling and anomaly detection.

[0078] S23. Anomaly detection is performed on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package.

[0079] The concept of a manifold will be explained first. Under normal vehicle operation, the 33-dimensional composite feature vector is not randomly distributed, but rather constrained by the vehicle's physical mechanisms. All data points under normal conditions are actually distributed on a low-dimensional surface (i.e., a manifold) far below 33 dimensions. For example, when engine speed increases, multiple feature dimensions such as vibration amplitude, temperature, and oil pressure will change collaboratively according to defined physical laws. This collaborative change relationship gives the distribution of data points in high-dimensional space an inherent low-dimensional structure. Therefore, in step S23, this invention models this low-dimensional structure using an incremental locally linear embedding algorithm.

[0080] Step S23 further includes:

[0081] S231. Construct a k-nearest neighbor graph from the historical window data of the comprehensive feature vector using an incremental local linear embedding algorithm, perform manifold modeling on the topological structure of the data points in low-dimensional space, and obtain the current manifold structure model.

[0082] In step S231, this invention employs a local linear embedding algorithm. First, for N historical comprehensive feature vector data points within a sliding window, it finds the k nearest neighbors (k=12 in this invention) for each data point in 33-dimensional space, constructing a k-nearest neighbor graph. Then, for each data point, within the local neighborhood formed by its k nearest neighbors, it solves for a set of reconstruction weights, ensuring that the data point can be approximated by a weighted linear combination of its k nearest neighbors, minimizing the reconstruction error. Finally, using this set of weights as constraints, it solves for the embedding coordinates of each data point in low-dimensional space, ensuring that the low-dimensional coordinates also satisfy the aforementioned linear reconstruction relationship. After processing, the data points in the 33-dimensional space are mapped to low-dimensional manifold coordinates, while preserving the local topological adjacency relationships between the points.

[0083] Furthermore, this invention introduces an incremental update strategy based on the aforementioned locally linear embedding algorithm. Specifically, when a new comprehensive feature vector data point arrives, the algorithm only recalculates the embedding coordinates of the new point and its k nearest neighbors, without re-performing a full calculation for all historical points within the window, thus controlling the update time per frame to within 5ms. Through this processing, the edge computing unit maintains and continuously updates a dynamic manifold structure model reflecting the current normal driving state.

[0084] S232. For newly arrived integrated feature vector data points, calculate the embedding distance and local curvature change relative to the current manifold structure model to obtain the manifold deviation metric.

[0085] In step S232, for the newly arrived integrated feature vector data points, the present invention calculates two types of metrics relative to the current manifold structure model.

[0086] Specifically, the embedding distance is calculated as follows: First, the new data point is projected onto the existing manifold, that is, the new point is reconstructed using a linear combination of k nearest neighbors in the existing manifold. The reconstruction error (the Euclidean distance between the new point and its reconstructed approximation) is calculated, and the reconstruction error is weighted and summed with the Euclidean distance from the new point to its nearest neighbor manifold point to obtain the embedding distance. The larger the embedding distance obtained, the higher the degree to which the new data point deviates from the current normal manifold.

[0087] Specifically, the calculation of local curvature change is as follows: before adding a new point to its k-nearest neighbor local neighborhood, perform local principal component analysis (LPA) on the existing data points within that neighborhood, and record the variance ratio explained by the first two principal components. After adding the new point, perform LPA again on the expanded neighborhood, and recalculate the variance ratio of the first two principal components. The difference between the two variance ratios is then calculated as the local curvature change, reflecting the degree of perturbation to the manifold curvature within its neighborhood after the addition of the new point. If the new point causes a significant shift in the principal direction of the neighborhood data, the curvature change will increase significantly, indicating that the point has caused a structural perturbation to the local manifold topology.

[0088] Ultimately, the embedding distance and the change in local curvature together constitute the manifold deviation metric.

[0089] The expression for the embedding distance mentioned above is:

[0090]

[0091] in, For the newly arrived comprehensive feature vector data points, For newly arrived data points Relative to the embedding distance of the current manifold structure model, , These are the first and second weighting coefficients for the reconstruction error and the nearest neighbor Euclidean distance, respectively. For newly arrived data points manifold reconstruction error, For manifold point index values, This is the set of existing manifold points in the current manifold structure model. For the set of manifold points The Middle The coordinates of the manifold points.

[0092] The expression for the aforementioned local curvature change is:

[0093]

[0094] in, For new data points The change in curvature of its k-nearest neighbor local neighborhood after addition. For new data points After adding, the variance explained by the first principal component is obtained by re-performing local principal component analysis in the expanded neighborhood. For new data points After adding, the variance explained by the second principal component after re-performing local principal component analysis in the expanded neighborhood is... For new data points Before addition, the variance explained by the first principal component after performing local principal component analysis on the existing data points in its k-nearest neighbor region is... For new data points Before addition, the variance explained by the second principal component after performing local principal component analysis on the existing data points in its k-nearest neighbor neighborhood.

[0095] S233. The manifold deviation metric and the anomaly score output by the isolated forest model are weighted and fused, and the health status is divided according to a preset threshold to obtain a local preliminary judgment label; S234. The local preliminary judgment label and the manifold deviation metric are encapsulated to generate a feature data package with priority label.

[0096] Furthermore, the isolated forest model is a machine learning model for anomaly detection. Its core idea is that anomalous data points are located in sparse and isolated regions in the feature space, and are easier to isolate than normal data points. Therefore, the anomaly can be determined by measuring the effort required to isolate a data point.

[0097] In this invention, the isolated forest model is trained offline with 500,000 kilometers of normal driving data before leaving the factory. When running, it only needs to load the trained tree structure parameters and perform anomaly scoring on the 33-dimensional comprehensive feature vector of each real-time arrival on the vehicle edge computing unit.

[0098] During operation, when a new 33-dimensional comprehensive feature vector data point arrives, this invention sends it into each of the pre-trained isolation trees. Starting from the root node, it determines whether the point should enter the left or right child node based on the segmentation features and segmentation values ​​of each internal node. It traverses down the tree until the point reaches the leaf node, records the path length (i.e., the number of segmentations) traversed by the point from the root node to the leaf node, and obtains the average path length of the data point by averaging the path lengths of all isolation trees.

[0099] In the isolation forest model, the relationship between average path length and outlier score is as follows: if a data point is located in a dense region of normal data in the 33-dimensional feature space, the probability of separating it from other points in each random segment is low, requiring more segments to completely isolate it, resulting in a longer average path length. Conversely, if a data point is located in a sparse, marginal region of the feature space (i.e., an abnormal state deviating from the normal driving pattern), it can be separated from other data points with only a few random segments, resulting in a shorter average path length. This invention uses the reciprocal of the normalized average path length as the outlier score; the shorter the path, the higher the outlier score, indicating that the data point is more likely to deviate from the normal driving state.

[0100] To illustrate with a specific scenario: When an engine exhibits an early anomaly at a certain moment, with a slight increase in vibration amplitude and a slight decrease in oil pressure, the corresponding 33-dimensional comprehensive feature vector deviates from the densely distributed region of normal data in the feature space, entering a relatively sparse edge zone. In this case, during random segmentation, the isolation tree can separate this point from the normal data point with fewer segmentations, shortening the average path length and outputting a higher anomaly score. Conversely, the comprehensive feature vector under normal driving conditions is concentrated in the core dense region of the feature space, requiring multiple random segmentations to isolate it, resulting in a longer average path length and a lower anomaly score.

[0101] Subsequently, this invention performs a weighted fusion of the manifold deviation metric and the anomaly score output by the isolated forest model. The fusion weights are dynamically adjusted according to the current scenario. Specifically, in early warning scenarios, the weights of the manifold alignment model are dynamically increased to enhance the sensitivity to weak structural anomalies; the weights of the isolated forest model are relatively reduced when there is a lot of noise to suppress misjudgments.

[0102] After fusion, a comprehensive anomaly score is obtained. Subsequently, this invention classifies the health status according to a preset three-level threshold (green, yellow, and red) to obtain a local preliminary judgment label. Green indicates that the data point is within the normal manifold region, yellow indicates that the manifold structure has slight perturbation, and red indicates that the manifold structure has significant distortion or the embedding distance exceeds the upper limit.

[0103] Finally, the present invention encapsulates the local initial judgment label, manifold deviation metric (embedding distance and curvature change), 33-dimensional comprehensive feature vector, priority identifier and compressed original data fragment in a unified format to generate a feature data packet with priority label, and enters the transmission processing stage of step S3.

[0104] S3. Based on the priority identifier in the feature data packet, channel allocation is performed on the multi-channel communication interface using a game theory model to obtain the target transmission scheme. The feature data packet is then uploaded to the cloud according to the target transmission scheme to obtain a time-series feature database.

[0105] Step S3 further includes:

[0106] S31. Using the reciprocal of delay, channel reliability score, and transmission cost as components, dynamically adjust the weights of each component according to the priority identifier in the feature data packet to construct a multi-objective utility function.

[0107] In step S31, the present invention first constructs a multi-objective utility function for each feature data packet to be transmitted. This function consists of three components: the reciprocal of the delay, the channel reliability score, and the transmission cost. The reciprocal of the delay is the reciprocal of the end-to-end transmission delay, and the lower the delay, the higher the value of this component. The channel reliability score is calculated based on the recent historical packet loss rate and retransmission rate of the channel. The lower the packet loss rate and the lower the retransmission rate, the higher the score. The transmission cost includes both traffic cost and communication energy consumption. The higher the cost, the negative the contribution to the utility.

[0108] The three components correspond to the weighting coefficients respectively. , , This invention dynamically adjusts the three weights based on the priority identifier carried in the feature data packet: the red priority data packet will... Set as the maximum value. Secondly, Set to the minimum value, meaning the primary goal is to minimize latency, regardless of transmission cost; yellow priority data packets and A balanced configuration is adopted, taking into account both latency and reliability; green priority data packets will Set to a larger value The values ​​are set to a smaller value to save transmission costs and tolerate higher latency. The multi-objective utility function values ​​under the three weight configurations reflect the differentiated demands of data packets of different priorities on channel resources, serving as the input basis for the game solution in S32.

[0109] The expression for the above multi-objective utility function is:

[0110]

[0111] in, This is the channel index value. For packet priority identification, , , The time delay, reliability, and cost items are prioritized respectively. The dynamic weighting coefficients under, For channel End-to-end transmission delay, The nonlinear penalty exponent for time delay, when When the time delay penalty on utility increases superlinearly, it is suitable for red priority scenarios; when At this time, the increase in penalties slows down, making it suitable for green priority scenarios. For channel The latency jitter sensitivity coefficient, For channel The measured latency jitter at the current moment is defined as the current latency. Compared with the average delay within the recent sliding window The absolute value of the difference For channel Basic reliability score, The normalized variance parameter for the reliability score. For channel The current channel contention pressure coefficient is defined as the number of simultaneous requests to use the channel at the current time. Number of data packets and channel The ratio of maximum concurrent capacity For channel Competition penalty amplification factor For channel The basic transmission cost includes the cost per unit of data traffic. Power consumption of basic communication Weighted summation: ,in and These are the weighting coefficients for cost and energy consumption, respectively. For channel exist The dynamic transmission cost increment at any given moment is determined by the channel. Current signal-to-noise ratio The reciprocal of the current retransmission count Product estimation, This is a time-sensitive discount factor for dynamic cost increments, used to control the impact of real-time network status on cost items. For channel The load nonlinearity penalty index, For channel The current load rate is defined as the bandwidth currently occupied by the channel. With total available bandwidth The ratio.

[0112] S32. Treating C-V2X, cellular networks, Wi-Fi, and low-power wide-area networks as game participants, continuously optimize multiple channel strategies through iterative optimal response algorithms to find the Nash equilibrium point and obtain a preliminary transmission scheme.

[0113] In step S32, this invention treats the four types of wireless communication interfaces integrated in the vehicle communication module—C-V2X, cellular networks (4G / 5G), Wi-Fi 6, and LoRaWAN low-power wide-area network—as independent players in a game theory model. Each channel acts as a player in the game, with a strategy space of accepting or rejecting the current transmission task. The payoff for each player is defined as the difference between its utility function value and the current channel load cost; that is, accepting the task results in a positive payoff if the channel load is low, a negative payoff if the channel load is full, and a zero payoff if the task is rejected.

[0114] Subsequently, this invention solves for the Nash equilibrium point using an iterative optimal response algorithm. First, the strategy of each channel is initialized based on the current real-time load state of each channel. Then, with the strategies of the other three channels remaining constant, the payoffs for accepting and rejecting the data are calculated for each channel, and the strategy with the higher payoff is selected as the current optimal response for that channel. After updating the optimal response for each of the four channels sequentially, one iteration is completed. This invention continues to execute the above iterative process, using the condition that the strategies of the four channels remain unchanged for two consecutive iterations as the convergence criterion, or forcibly terminating when the maximum number of iterations (100) is reached. The strategy combination of each channel after convergence is the Nash equilibrium point, which determines which channels accept the current data packet transmission task, thus obtaining a preliminary transmission scheme.

[0115] Based on the initial transmission scheme, this invention further implements differentiated transmission configuration according to priority. Among them, red priority data packets are selected from the two channels with the highest utility value based on the Nash equalization solution and transmitted concurrently, with an end-to-end latency target of less than 100ms, and local SD card cyclic backup is started at the same time; yellow priority data packets select a single optimal channel based on the Nash equalization solution and use the MQTT-SN protocol QoS Level 1 to ensure at least one delivery; green priority data packets are aggregated into batches of 120 frames each, and are uploaded at low frequency via Wi-Fi 6 or 4G using the LwM2M protocol, and the LZ4 compression algorithm is enabled, with a compression ratio of about 3:1.

[0116] S33. When all channel utility values ​​are lower than the preset threshold, a transmission degradation strategy is negotiated through a bargaining game model on dimensions including timeliness and data integrity to modify the preliminary transmission scheme and obtain the target transmission scheme.

[0117] Furthermore, when the real-time utility values ​​of all four communication channels are lower than the preset threshold, it indicates that the overall quality of the current network environment has deteriorated, and the preliminary transmission scheme cannot be executed normally under the existing network conditions. The present invention then initiates a bargaining game model to correct the preliminary transmission scheme.

[0118] The bargaining game model adopts the Rubinstein alternating bid model, setting the edge device (vehicle communication module) as the bidder and the cloud device as the bidder. The edge device maintains a data value evaluation function, which takes time sensitivity and integrity weights as input to calculate the comprehensive value of the feature data packet to be transmitted; the cloud device maintains a receiving cost function, which takes storage overhead and processing latency as input to calculate the comprehensive cost of receiving the current data.

[0119] During the negotiation process, the edge device first proposes a transmission degradation plan, such as reducing the feature vector dimension (compressing it from 33 dimensions to a lower dimension) or transmitting only the comprehensive feature vector without the original data fragments, and sends the corresponding data value assessment function value to the cloud. The cloud calculates the net benefit of accepting the plan based on the receiving cost function. If the net benefit is positive, it accepts the plan; if the net benefit is negative, it rejects it and proposes a counter-plan (such as requiring a further reduction in transmission frequency or receiving only the two key data points: the initial local label and the embedding distance). After receiving the counter-plan, the edge device again determines whether to accept it based on the data value assessment function. Both parties alternately offer prices based on timeliness and data integrity, completing the negotiation process within 50ms. If both parties reach an agreement, the initial transmission plan is revised according to the negotiated degradation plan to obtain the target transmission plan. If no agreement is reached within the maximum negotiation round, the default degradation strategy is executed: red priority data packets are forced to transmit only the comprehensive feature vector and the warning label, while yellow and green priority data packets are locally cached and transmitted only after the network recovers.

[0120] S34. Upload data according to priority levels based on the target transmission scheme to obtain a time-series feature database.

[0121] In step S34, after performing hierarchical uploading according to the target transmission scheme, each feature data packet is sent along with an attached SHA-256 message digest and sequence number. Upon receiving the data, the cloud gateway performs digest verification on each packet to confirm that the data has not been tampered with or damaged during transmission, and reassembles out-of-order data packets according to their sequence numbers. For red and yellow priority data packets that fail verification or time out, the cloud triggers an automatic retransmission request.

[0122] After all data packets have undergone integrity verification and out-of-order reassembly, they are stored in the cloud according to the structure of vehicle ID, driving trajectory, multi-dimensional comprehensive feature vector sequence, manifold embedding distance sequence and local preliminary label sequence, thus obtaining a time-series feature database.

[0123] S4. Input the time series feature database into the cloud digital twin model, and perform fault diagnosis on the time series feature database through the residual fusion algorithm to obtain a structured fault diagnosis report.

[0124] Step S4 further includes:

[0125] S41. Based on the vehicle dynamics equations and thermodynamic models, the measured parameters in the time-series feature database are used to drive the simulation calculation of the physical mechanism sub-model to obtain the theoretical expected state values ​​of multiple vehicle components.

[0126] In step S41, the physical mechanism sub-model is constructed based on the vehicle dynamics equation and the thermodynamic model. The vehicle dynamics equation describes the mechanical relationship between engine output torque, transmission ratio, wheel driving force and vehicle speed, while the thermodynamic model describes the energy conservation relationship between cylinder temperature, coolant temperature and heat transfer during engine combustion.

[0127] The expression for the above vehicle dynamics equations is as follows:

[0128]

[0129] in, for Net driving force at the wheel at any given moment for The engine output torque at any given time is determined by the measured engine speed from the time-series feature database. With throttle opening The value is obtained by interpolation using the engine's universal characteristic diagram. The gear ratio corresponding to the current gear in the transmission. Main reducer transmission ratio, For the mechanical efficiency of the transmission system, The radius of the wheel's rolling motion. for The resultant force of the driving resistance at any given moment is composed of rolling resistance. air resistance With ramp resistance It consists of three parts: ,in The rolling resistance coefficient, For vehicle curb weight It is the acceleration due to gravity. This is the drag coefficient. The vehicle's frontal area. air density, for Real-time vehicle speed measurement for Road surface slope angle at all times.

[0130] The above thermodynamic model is expressed as:

[0131]

[0132] in, for At any given time, the theoretical expected temperature inside the engine cylinder. The cylinder temperature at the previous sampling time. For the heat capacity of the engine block. for The heat released during continuous combustion is determined by the fuel injection quantity. With low calorific value of fuel The calculation yielded: ,in For combustion efficiency, Throttle opening With engine speed Obtained by interpolation of the fuel injection pulse spectrum. for The heat removed by the cooling system at all times is expressed as: ,in The convective heat transfer coefficient of the coolant. For cooling contact area, The measured coolant temperature is used in the time-series feature database. for The heat carried away by the exhaust gas at any given time is expressed as: ,in The mass flow rate of the exhaust gas. The specific heat capacity of exhaust gas at constant pressure. The ambient temperature.

[0133] After establishment, this invention uses the measured parameters (including engine speed, throttle opening, vehicle speed, coolant temperature, etc.) at the current moment in the time-series feature database as driving inputs, substitutes them into the above equations for simulation calculation, and calculates the theoretical expected state values ​​that key components such as the engine, transmission, and braking system should exhibit under the current input conditions, such as the theoretical vibration amplitude range, theoretical oil pressure range, and theoretical braking temperature range. The difference between the measured parameters and the theoretical expected state values ​​will be used as a residual vector in S43.

[0134] S42. Input the historical comprehensive feature vector sequence from the time-series feature database into the time-series Transformer network to learn the time-series dependencies of multiple feature dimensions under normal driving mode and obtain the anomaly probability vector.

[0135] In step S42, the present invention is based on a temporal Transformer network. For each time step in the input sequence, the correlation weight between that time step and all other time steps in the sequence is calculated. The higher the weight, the stronger the temporal dependency between the feature changes of the two time steps. Then, the features of all time steps are summed according to the weight to obtain the feature representation of the current time step after global context modeling.

[0136] Subsequently, this invention inputs the vehicle's historical 33-dimensional comprehensive feature vector sequence (covering historical driving data throughout the vehicle's entire lifecycle) from the temporal feature database into the network. The network learns the coordinated change patterns of each feature dimension over time under normal driving conditions through a multi-head self-attention mechanism, such as the normal linkage between vibration and temperature features under specific vehicle speed and load conditions. For the 33-dimensional comprehensive feature vector input at the current moment, the network determines whether the values ​​of each feature dimension conform to the historical normal pattern based on the learned normal temporal dependencies, and outputs an anomaly probability vector covering 18 types of fault modes, where each dimension corresponds to the current anomaly probability estimate of a fault mode.

[0137] S43. Calculate the residual vector between the measured comprehensive feature vector in the time series feature database and the theoretical expected state value, and perform residual fusion of the residual vector and the anomaly probability vector through a multilayer perceptron to obtain a fault type probability distribution vector covering multiple fault modes.

[0138] In step S43, the present invention subtracts the measured 33-dimensional comprehensive feature vector of the current moment in the time series feature database from the theoretical expected state value output by the physical mechanism sub-model in S41 dimension by dimension to obtain the residual vector. Each dimension of the residual vector reflects the magnitude and direction of the measured value deviating from the expected value of the physical mechanism in a certain feature dimension, such as the degree to which the measured vibration amplitude exceeds the theoretical expected range, the magnitude of the measured oil pressure being lower than the theoretical expected value, etc.

[0139] Subsequently, the residual vector is concatenated with the anomaly probability vector output by the temporal Transformer network in S42, and then input into a multilayer perceptron for residual fusion processing. The multilayer perceptron consists of three fully connected layers with GELU activation. It receives the concatenated vector as input and, after three layers of nonlinear transformation, outputs a fault type probability distribution vector covering 18 fault modes. Each dimension of the vector represents the probability that the current vehicle state belongs to the corresponding fault type. The 18 fault modes cover engine abnormal combustion faults, engine lubrication system faults, engine cooling system faults, engine intake system faults, transmission shift shock faults, transmission oil overheating faults, abnormal driveshaft vibration faults, brake system thermal fade faults, brake pad wear exceeding limits faults, tire pressure loss faults, abnormal tire dynamic balance faults, abnormal suspension damper attenuation faults, suspension bushing aging faults, steering system abnormal faults, vehicle electrical system voltage abnormal faults, vehicle bus communication faults, fuel supply system faults, and chassis attitude abnormal faults.

[0140] S44. Generate the structured fault diagnosis report based on the probability distribution vector of the fault type combined with manifold evolution trajectory analysis.

[0141] This invention, based on the fault type probability distribution vector output by S43, further integrates the manifold embedding distance sequence and the local curvature change sequence uploaded over time from the time-series feature database to construct a manifold evolution trajectory diagram. The manifold evolution trajectory diagram uses time as the horizontal axis and manifold embedding distance and curvature change as the vertical axes, visually presenting the drift trend of the manifold structure of vehicle sensor data over time.

[0142] Subsequently, this invention performs trend analysis on the trajectory map to identify progressive failure modes. For example, a monotonically increasing manifold embedding distance over time indicates a continuous deterioration trend in a certain component, and a periodic increase in local curvature change under specific operating conditions indicates a structural anomaly under those conditions. The failure type probability distribution vector and the manifold evolution trajectory analysis results are combined and processed, and the root cause analysis tree is used to locate the specific component. A failure urgency score (0 to 100 points) is calculated, and recommended remedial measures are generated. Finally, a structured failure diagnosis report is output.

[0143] like Figure 2 As shown, the present invention also provides a vehicle operating status real-time monitoring system based on the Internet of Things, comprising:

[0144] Preprocessing module 100: used to perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream;

[0145] The detection module 200 is configured based on an edge computing unit and is used to receive the standardized multimodal data stream, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data packet with priority marking.

[0146] Transmission module 300: is used to allocate channels to the multi-channel communication interface according to the priority identifier in the feature data packet through a game theory model, obtain a target transmission scheme, and upload the feature data packet to the cloud according to the target transmission scheme to obtain a time series feature database;

[0147] The diagnostic module 400 is configured based on a cloud-based digital twin model and is used to receive the time-series feature database, perform fault diagnosis on the time-series feature database through a residual fusion algorithm, and obtain a structured fault diagnosis report.

[0148] The present invention also provides an Internet of Things (IoT)-based real-time vehicle operation status monitoring device, comprising: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the IoT-based real-time vehicle operation status monitoring device to execute an IoT-based real-time vehicle operation status monitoring method as described in any of the above.

[0149] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement a method for real-time monitoring of vehicle operating status based on the Internet of Things as described in any of the preceding claims.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0151] This invention enables intelligent monitoring and management of vehicle health status. A manifold alignment anomaly detection mechanism is introduced at the edge computing layer to capture weak anomaly signals from the geometric structure of data distribution, significantly improving the sensitivity and noise robustness of early fault warnings. A game theory-driven heterogeneous network collaboration mechanism is introduced at the network transmission layer to dynamically optimize channel resource allocation strategies, ensuring reliable delivery and low-latency transmission of critical warning data in network congestion scenarios. A digital twin model is constructed in the cloud and integrated with cross-vehicle group transfer learning, achieving intelligent fault diagnosis exceeding rule thresholds and effectively solving the cold start problem for new vehicles. Overall, this invention enables the system to maintain high-performance operation throughout its entire lifecycle, effectively reducing vehicle failure risks and improving driving safety.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time monitoring of vehicle operating status based on the Internet of Things, characterized in that, include: S1. Perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream; S2. Input the standardized multimodal data stream into the edge computing unit, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package with priority marking. Step S2 further includes: S21. Extract time-domain statistical features, frequency-domain power spectral density features, and time-frequency domain semantic features from the sliding window data of the standardized multimodal data stream to obtain three feature subsets. S22. The three feature subsets are concatenated and fused to obtain the comprehensive feature vector; S23. Perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package; Step S23 further includes: S231. Construct a k-nearest neighbor graph from the historical window data of the comprehensive feature vector using the incremental local linear embedding algorithm, perform manifold modeling on the topological structure of the data points in the low-dimensional space, and obtain the current manifold structure model. S232. For newly arrived integrated feature vector data points, calculate the embedding distance and local curvature change relative to the current manifold structure model to obtain the manifold deviation metric. S233. The manifold deviation metric and the anomaly score output by the isolated forest model are weighted and fused, and the health status is divided according to a preset threshold to obtain the local preliminary judgment label. S234. The local preliminary judgment label and the manifold deviation metric are encapsulated to generate a feature data packet with priority tags; S3. Based on the priority identifier in the feature data packet, channel allocation is performed on the multi-channel communication interface using a game theory model to obtain the target transmission scheme. The feature data packet is then uploaded to the cloud according to the target transmission scheme to obtain a time-series feature database. S4. Input the time series feature database into the cloud digital twin model, and perform fault diagnosis on the time series feature database through the residual fusion algorithm to obtain a structured fault diagnosis report. Step S4 further includes: S41. Based on the vehicle dynamics equations and thermodynamic models, the measured parameters in the time-series feature database are used to drive the simulation calculation of the physical mechanism sub-model to obtain the theoretical expected state values ​​of multiple vehicle components. S42. Input the historical comprehensive feature vector sequence in the time series feature database into the time series Transformer network to learn the time series dependencies of multiple feature dimensions under normal driving mode and obtain the anomaly probability vector. S43. Calculate the residual vector between the measured comprehensive feature vector in the time series feature database and the theoretical expected state value, and perform residual fusion of the residual vector and the anomaly probability vector through a multilayer perceptron to obtain a fault type probability distribution vector covering multiple fault modes. S44. Generate the structured fault diagnosis report based on the probability distribution vector of the fault type combined with manifold evolution trajectory analysis.

2. The method for real-time monitoring of vehicle operating status based on the Internet of Things according to claim 1, characterized in that, Step S1 further includes: S11. The original acquisition signals from multiple sensors are processed by a low-pass filter circuit to remove high-frequency electromagnetic interference components and obtain a primary filtered signal. S12. For vibration signals in the primary filtered signal, denoising is performed using a wavelet threshold denoising algorithm; for slowly varying signals in the primary filtered signal, smoothing is performed using a moving average filter to obtain a denoised signal set. S13. Based on the denoised signal set, all channel signals are uniformly aligned to a preset sampling granularity using a time interpolation alignment algorithm, and then encapsulated into a standardized multimodal data stream.

3. The method for real-time monitoring of vehicle operating status based on the Internet of Things according to claim 1, characterized in that, Step S21 specifically includes: From the time series data within the sliding window, the mean, variance, peak factor, waveform factor, and kurtosis are calculated to obtain the time-domain statistical characteristics. The time-series data is subjected to spectral analysis by fast Fourier transform, and multiple frequency bands are divided using the Mel scale and the energy proportion of each frequency band is calculated to obtain the frequency domain power spectral density characteristics. A time-frequency graph is generated by short-time Fourier transform, and the time-frequency graph is input into a lightweight convolutional neural network to extract the semantic feature vector in the time-frequency domain.

4. The method for real-time monitoring of vehicle operating status based on the Internet of Things according to claim 1, characterized in that, Step S3 further includes: S31. Using the reciprocal of delay, channel reliability score, and transmission cost as components, dynamically adjust the weights of each component according to the priority identifier in the feature data packet to construct a multi-objective utility function. S32. Treating C-V2X, cellular networks, Wi-Fi and low-power wide area networks as game participants, continuously optimize multiple channel strategies through iterative best response algorithm, solve for Nash equilibrium point, and obtain preliminary transmission scheme; S33. When all channel utility values ​​are lower than the preset threshold, a transmission degradation strategy is negotiated through a bargaining game model on the dimensions of timeliness and data integrity to modify the preliminary transmission scheme and obtain the target transmission scheme. S34. Upload data according to priority levels based on the target transmission scheme to obtain a time-series feature database.

5. A real-time vehicle operation status monitoring system based on the Internet of Things, characterized in that, include: Preprocessing module: Used to perform multi-level preprocessing on the raw signals collected by the vehicle's multi-source sensors to obtain a standardized multimodal data stream; The detection module, configured based on an edge computing unit, is used to receive the standardized multimodal data stream, extract features from the standardized multimodal data stream using a multi-domain feature extraction algorithm to obtain a comprehensive feature vector, and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data packet with priority marking. The detection module is further configured to: extract time-domain statistical features, frequency-domain power spectral density features, and time-frequency-domain semantic features from the sliding window data of the standardized multimodal data stream to obtain three feature subsets; concatenate and fuse the three feature subsets to obtain the comprehensive feature vector; and perform anomaly detection on the comprehensive feature vector using a manifold alignment algorithm to obtain a feature data package. When the detection module performs anomaly detection on the comprehensive feature vector using a manifold alignment algorithm and obtains a feature data packet, it further performs the following: constructs a k-nearest neighbor graph on the historical window data of the comprehensive feature vector using an incremental local linear embedding algorithm, performs manifold modeling on the topological structure of the data points in low-dimensional space, and obtains the current manifold structure model; for newly arrived comprehensive feature vector data points, it calculates the embedding distance and local curvature change relative to the current manifold structure model to obtain a manifold deviation metric. The manifold deviation metric is weighted and fused with the anomaly score output by the isolated forest model, and the health status is divided according to a preset threshold to obtain a local preliminary judgment label; the local preliminary judgment label and the manifold deviation metric are encapsulated to generate a feature data package with priority tags; Transmission module: used to allocate channels to the multi-channel communication interface according to the priority identifier in the feature data packet through a game theory model, obtain a target transmission scheme, and upload the feature data packet to the cloud according to the target transmission scheme to obtain a time series feature database; The diagnostic module, configured based on a cloud-based digital twin model, is used to receive the time-series feature database, perform fault diagnosis on the time-series feature database through a residual fusion algorithm, and obtain a structured fault diagnosis report. The diagnostic module is further configured to: based on vehicle dynamics equations and thermodynamic models, drive physical mechanism sub-model simulation calculations using measured parameters from the time-series feature database to obtain theoretical expected state values ​​for multiple vehicle components; input the historical comprehensive feature vector sequence from the time-series feature database into a time-series Transformer network to learn the temporal dependencies of multiple feature dimensions under normal driving conditions and obtain an anomaly probability vector; calculate the residual vector between the measured comprehensive feature vector from the time-series feature database and the theoretical expected state value, and perform residual fusion of the residual vector and the anomaly probability vector through a multilayer perceptron to obtain a fault type probability distribution vector covering multiple fault modes; and generate the structured fault diagnosis report based on the fault type probability distribution vector combined with manifold evolution trajectory analysis.

6. A vehicle operation status real-time monitoring device based on the Internet of Things, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause an Internet of Things (IoT)-based vehicle operation status real-time monitoring device to execute the IoT-based vehicle operation status real-time monitoring method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement a method for real-time monitoring of vehicle operating status based on the Internet of Things as described in any one of claims 1 to 4.

Images