Vehicle anomaly detection system and method based on multi-modal data analysis

Abstract

The invention relates to the technical field of vehicle fault diagnosis, and discloses a vehicle anomaly detection system and method based on multi-modal data analysis, and the system builds a time zero point alignment relation between control domain data and a high-frequency vibration signal through a hardware synchronous triggering unit by using a level pulse; performing angular domain resampling on the non-stationary vibration signal by adopting a coarse-fine coupling phase locking technology and combining rotating speed polynomial fitting and instantaneous frequency ridge line extraction; the system constructs a virtual reference model by using a deep neural network, and deduces a predicted ideal angular domain characteristic spectrum of a current working condition according to a dynamic state vector; and on the basis, calculating the residual error of the actual spectrum and the predicted spectrum, removing noise through order correlation mask filtering, and judging abnormity based on the mahalanobis distance of a residual error vector and a statistical threshold value. The method effectively solves the problems that the signal is not stable and weak fault features are difficult to extract under the variable speed working condition, and improves the detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of vehicle fault diagnosis technology, specifically to a vehicle anomaly detection system and method based on multimodal data analysis. Background Technology

[0002] With the rapid development of automotive intelligence and connectivity, the dimensions of vehicle operation data collection are becoming increasingly rich. Data-driven vehicle anomaly detection has become a key technology for ensuring driving safety and reducing maintenance costs. Modern vehicle diagnostic systems typically need to process operating parameter data from control buses (such as CAN buses) and high-frequency vibration data from additional sensors (such as accelerometers), attempting to identify potential mechanical faults or performance degradation through the fusion analysis of multi-source information.

[0003] However, in existing vehicle anomaly detection technologies, the timing alignment of multimodal data relies on timestamp synchronization at the host computer software level. Because control domain data and response domain data are transmitted through different physical links and are affected by uncertainties in operating system task scheduling jitter and bus transmission delays, this soft synchronization method is prone to millisecond-level or even higher time deviations under highly dynamic conditions. For high-speed rotating components, even a tiny timing misalignment translates into a significant phase error, causing subsequent data fusion analysis to be based on an incorrect spatiotemporal reference, thus weakening the reliability of correlation analysis.

[0004] Furthermore, vehicles operate in a non-stationary state with varying speeds and loads during actual driving. Traditional Fourier transform-based spectrum analysis methods are only applicable to stationary signals, and suffer from severe frequency ambiguity when processing signals with varying speeds. While order tracking technology can address this issue, existing high-precision order tracking schemes rely on additional photoelectric encoders or high-precision tachometers to obtain angular domain references, which is limited by installation space and hardware costs in mass-produced vehicles or non-invasive testing scenarios. Simply interpolating and resampling the low-sampling-rate CAN bus speed signal results in insufficient accuracy, leading to signal distortion and energy leakage in the reconstructed signal, making it difficult to extract clear fault characteristics from complex background noise.

[0005] Finally, in the anomaly decision-making stage, existing technologies employ fixed threshold judgment methods based on expert experience or simple statistical boundary methods. However, vehicle dynamics systems exhibit highly nonlinear and time-varying characteristics, and their normal vibration response fluctuates drastically with changes in speed, torque, and load. Under complex operating conditions, due to the lack of a benchmark model that can dynamically follow changes in vehicle operating status, fixed thresholds struggle to balance detection sensitivity and specificity, easily leading to false alarms under severe driving conditions or missed detections in the early stages of weak faults due to the signal being drowned out by background noise. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a vehicle anomaly detection system and method based on multimodal data analysis. It solves the problems in existing technologies, such as phase errors and timing misalignment caused by software synchronization mechanisms in multi-source heterogeneous data, difficulty in eliminating frequency ambiguity and energy leakage in the absence of high-precision encoders under non-stationary conditions of variable speed, and the inability of traditional fixed threshold discrimination methods to adapt to the complex dynamic characteristics of vehicles, resulting in weak fault features being submerged by noise or false alarms and missed alarms.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The first aspect of this invention provides a vehicle anomaly detection system based on multimodal data analysis. The system includes a data acquisition terminal and an edge computing processing host. The edge computing processing host logically integrates a data acquisition and preprocessing module, a phase-locked resampling module, a virtual reference model module, and a residual analysis and decision module.

[0009] To address the time alignment issue of multi-source heterogeneous data, the vehicle anomaly detection system employs a hardware synchronization triggering mechanism. By monitoring the control domain data stream, the first valid data frame is used as the system startup event, and a physical level pulse signal is generated to directly trigger the analog-to-digital conversion of the response domain data. This establishes a zero-point synchronization relationship between the control domain data and the high-frequency vibration signal in the time dimension at the hardware level, eliminating phase errors caused by transmission delays.

[0010] To address the non-stationary characteristics of signals under varying speed conditions, the vehicle anomaly detection system employs a coarse-fine coupled phase-locked loop strategy. First, a cubic polynomial fit is performed on the speed signal to construct a globally continuous coarse reference frequency, which serves as a priori constraint limiting the frequency search interval in the short-time Fourier transform energy spectrum. Within this interval, the instantaneous frequency ridge of the energy maximum is extracted, and the non-uniform resampling time is obtained through numerical integration and inverse function solving. Cubic spline interpolation is then used to reconstruct the time-domain vibration signal into an angular-domain sequence, achieving order tracking and stabilization of the non-stationary signal.

[0011] To address the dynamic changes in vehicle operating status, the vehicle anomaly detection system constructs a virtual reference model based on a deep neural network. By extracting signals such as engine speed, output torque, and load, a dynamic state vector is constructed. This vector is then used to excite the deep neural network model to infer the predicted ideal angular domain feature spectrum under the current operating conditions, providing a dynamically changing health benchmark for anomaly detection.

[0012] To identify subtle fault characteristics, the vehicle anomaly detection system combines order-related correlation masking filtering with statistical manifold distance analysis. By generating a weighted window covering the order of key targets and their harmonic components, an order mask vector is constructed and used to perform Hadamard product filtering on the residuals between the actual and predicted spectra, eliminating irrelevant noise interference. Then, the Mahalanobis distance of the filtered residual vector is calculated using the covariance matrix and compared with an anomaly decision threshold determined based on statistical distribution, achieving high-confidence anomaly decision-making.

[0013] A second aspect of this invention provides a vehicle anomaly detection method based on multimodal data analysis, applied to the aforementioned system. This method mainly includes: synchronously acquiring control domain and response domain data using hardware trigger signals; calculating the instantaneous phase by combining polynomial fitting of the reference frequency with time-frequency ridge extraction, and performing angular domain resampling on the vibration signal; inferring the ideal feature spectrum based on the vehicle's dynamic state using a neural network; calculating the difference between the actual and ideal features, and highlighting key frequency band information using an order mask vector; and finally calculating the statistical distance of the residual vector and combining it with a confidence threshold to determine the vehicle's abnormal state.

[0014] This invention provides a vehicle anomaly detection system and method based on multimodal data analysis. It has the following beneficial effects:

[0015] 1. This invention utilizes a hardware synchronization triggering mechanism in the data acquisition and preprocessing module to directly trigger the analog-to-digital conversion of response domain data using physical level pulse signals. This establishes a time zero-point alignment relationship between control domain data and high-frequency vibration signals at the hardware level, thereby eliminating phase errors caused by bus transmission delays and system scheduling jitter in traditional soft synchronization methods. This ensures accurate timing synchronization of multi-source heterogeneous data under high-speed dynamic conditions, providing a reliable data foundation for subsequent correlation analysis.

[0016] 2. This invention employs coarse-fine coupled phase-locked loop and angular domain resampling techniques, combined with cubic polynomial fitting to construct a reference frequency and a time-frequency ridge extraction method. This achieves accurate angular domain reconstruction of non-stationary signals with varying speeds without the need for additional high-precision photoelectric encoders. This technique solves the problems of frequency ambiguity and energy leakage in spectrum analysis under varying operating conditions, converting non-stationary time-domain signals into stationary angular domain sequences, thereby improving the focus of fault characteristic orders and the signal-to-noise ratio.

[0017] 3. This invention utilizes a deep neural network to construct a virtual reference model and combines it with order correlation masking filtering technology. It can dynamically infer the ideal feature spectrum based on the current dynamic state of the vehicle and remove background noise interference unrelated to the fault through Hadamard product filtering. This residual analysis mechanism based on dynamic benchmark and statistical manifold distance overcomes the shortcomings of the traditional fixed threshold method in adapting to complex working conditions and improves the system's sensitivity and reliability in identifying early weak fault features. Attached Figure Description

[0018] Figure 1 A schematic diagram of the overall architecture of a vehicle anomaly detection system based on multimodal data analysis provided in an embodiment of the present invention;

[0019] Figure 2 A flowchart illustrating the coarse-fine coupling phase locking and angular domain resampling processing provided in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the virtual reference model construction and deep neural network structure provided in the embodiments of the present invention;

[0021] Figure 4 This is a schematic diagram illustrating the principle of order correlation masking filtering and residual analysis provided in an embodiment of the present invention.

[0022] Figure 5 The overall flowchart of the vehicle anomaly detection method based on multimodal data analysis provided in the embodiments of the present invention is shown.

[0023] The system comprises: 10. Vehicle under test subsystem; 20. Sensor network layer; 21. Control domain sensing unit; 22. Response domain sensing unit; 30. Data acquisition terminal; 40. Edge computing processing host; 100. Data acquisition preprocessing module; 110. Control domain data acquisition unit; 120. Response domain data acquisition unit; 130. Hardware synchronization triggering unit; 200. Phase-locked resampling module; 210. Coarse reference frequency calculation unit; 220. Instantaneous frequency ridge extraction unit; 230. High-precision phase reconstruction unit; 240. Calculation order tracking execution unit; 300. Virtual reference model module; 310. Dynamic state vector construction unit; 320. Virtual reference model inference unit; 400. Residual analysis decision module; 410. Cross-modal residual calculation unit; 420. Order correlation mask filtering unit; 430. Manifold distance calculation and alarm unit. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See attached document Figure 1 The present invention provides a vehicle anomaly detection system based on multimodal data analysis. The vehicle anomaly detection system includes a vehicle subsystem 10 under test, a sensor network layer 20, a data acquisition terminal 30, and an edge computing processing host 40.

[0026] The vehicle subsystem 10 under test includes the vehicle powertrain, chassis structure, and Controller Area Network (CAN) bus interface. The CAN bus interface is configured to output vehicle operating status data.

[0027] The sensor network layer 20 includes a control domain sensing unit 21 and a response domain sensing unit 22. The control domain sensing unit 21 is physically connected to the CAN bus interface or on-board diagnostic (OBD) interface of the vehicle subsystem under test 10 to acquire low-frequency control domain signals. The response domain sensing unit 22 includes at least one accelerometer and at least one acoustic sensor. The accelerometer is an integrated piezoelectric (IEPE) sensor, rigidly connected to the engine mount, transmission housing, or subframe location of the vehicle subsystem under test 10 via bolts or magnetic bases. The acoustic sensor is positioned in a predetermined location inside the driver's cab or engine compartment of the vehicle subsystem under test 10.

[0028] The data acquisition terminal 30 is connected to the control domain sensing unit 21 and the response domain sensing unit 22 via signal transmission cables. The data acquisition terminal 30 includes an analog-to-digital converter (ADC) and a digital input / output (DIO) interface. The control domain sensing unit 21 is configured to send a frame header pulse signal to the digital trigger port of the data acquisition terminal 30, and the response domain sensing unit 22 is configured to send a continuous analog voltage signal to the analog input port of the data acquisition terminal 30. The data acquisition terminal 30 triggers the acquisition of analog signals based on the frame header pulse signal, ensuring that the control domain data and the response domain data are aligned at zero points on the time axis.

[0029] The edge computing processing host 40 is connected to the data acquisition terminal 30 via a high-speed data bus to receive digitized multimodal data streams. The edge computing processing host 40 includes a processor and a memory. The memory stores computer-executable instructions, which, when executed by the processor, enable the detection of abnormal vehicle conditions. Logically, the edge computing processing host 40 includes a data acquisition preprocessing module 100, a phase-locked resampling module 200, a virtual reference model module 300, and a residual analysis decision module 400.

[0030] The data acquisition preprocessing module 100 is configured to receive the raw data stream from the data acquisition terminal 30 and parse CAN protocol messages to obtain the control domain data sequence. The response domain data is then subjected to anti-aliasing filtering to obtain the response domain data sequence. Control domain data sequence Includes engine speed signal Response domain data sequence Includes high-frequency vibration signals .

[0031] Phase-locked resampling module 200 is connected to data acquisition and preprocessing module 100 to receive control domain data sequences. and response domain data sequence The phase-locked resampling module 200 is configured to first base its signal on the engine speed signal. Calculate the reference frequency Then, in high-frequency vibration signals Extracting instantaneous frequency ridges from the time-frequency distribution .

[0032] Phase-locked resampling module with 200 pairs of instantaneous frequency ridges Integral operation is performed to obtain the instantaneous cumulative phase. And according to the equal angle increment Calculate the corresponding non-uniform resampling time. The phase-locked resampling module 200 utilizes non-uniform resampling times. For high-frequency vibration signals Perform interpolation resampling to output the angular domain sequence. .

[0033] The virtual reference model module 300 is connected to both the data acquisition and preprocessing module 100 and the phase-locked resampling module 200. The virtual reference model module 300 is configured to acquire data from the control domain data sequence. Extract dynamic parameters to construct dynamic state vectors The virtual reference model module 300 contains a pre-trained deep neural network model, which uses dynamic state vectors... As input, output the predicted ideal angular domain feature spectrum under the current operating conditions. .

[0034] The residual analysis decision module 400 is connected to both the phase-locked resampling module 200 and the virtual reference model module 300. The residual analysis decision module 400 is configured to receive angular domain sequences. And convert it into the actual angular domain feature spectrum. The residual analysis decision module 400 calculates the actual angular domain characteristic spectrum. With the prediction of the ideal angular domain feature spectrum The residual vector between them, and apply the order mask vector. The residual vector is filtered. The residual analysis decision module 400 calculates the Mahalanobis distance of the filtered residual vector in the statistical manifold space. When the Mahalanobis distance exceeds a preset threshold, an abnormal alarm signal is generated.

[0035] The data acquisition and preprocessing module 100 of this invention includes a control domain data acquisition unit 110, which reads the differential voltage signal on the controller area network bus of the vehicle subsystem under test 10 via a physical layer connection. The control domain data acquisition unit 110 first loads a preset communication matrix description file (DBC file), which defines the CAN message identifier (CAN-ID), data segment length, signal start bit, bit length, and physical quantity conversion factor. The control domain data acquisition unit 110 is equipped with a hardware acceptance filter, which, according to the definition in the communication matrix description file, only allows messages containing key parameters of the powertrain system to pass through, filtering out irrelevant messages such as those related to body control and the entertainment system, thereby reducing the data processing load.

[0036] The control domain data acquisition unit 110 performs real-time parsing of the filtered CAN messages. For each selected CAN frame, the control domain data acquisition unit 110 extracts the raw hexadecimal value from the data domain and converts it into an engineering quantity with physical meaning according to a linear transformation formula. For any control signal... At any moment physical values The calculation is as follows:

[0037] ;

[0038] in, This is the original decimal integer value corresponding to the message data field; This is the scaling factor for the signal; This is the offset.

[0039] The control domain data acquisition unit 110 specifically analyzes and outputs the control domain data sequence. At least including: engine speed signal Its unit is revolutions per minute (rpm); engine load signal It represents the percentage ratio of the current intake air volume to the maximum intake air volume; throttle opening signal It represents the driver's pedal demand or the actual opening angle of the electronic throttle; and the engine output torque signal. The control domain data acquisition unit 110 assigns a receiving timestamp to each of the parsed physical data points. This timestamp is generated by the high-precision local clock of the data acquisition terminal 30 and is denoted as […]. .

[0040] Because CAN bus data transmission has event-triggered or non-uniform periodic characteristics, the control domain data acquisition unit 110 further performs zero-order hold processing or linear interpolation processing on the original discrete data sequence to generate a sampling frequency of... (e.g., 50Hz) equally spaced time series. During this process, the control domain data acquisition unit 110 executes physical range verification logic: setting the effective threshold range for each signal. The effective threshold range is directly extracted from the signal range definition domain in the preset communication matrix description file (DBC file). If the parsed... If the value exceeds this range, it is determined to be a communication error or bus interference. The control domain data acquisition unit 110 will use the valid value of the previous moment or use the moving average value to replace the abnormal point to ensure the continuity and validity of the data input to subsequent modules.

[0041] The data acquisition and preprocessing module 100 of the present invention further includes a response domain data acquisition unit 120, which is configured to process the high-frequency analog signal from the response domain sensing unit 22 and convert it into a digital sequence.

[0042] The response domain data acquisition unit 120 receives the analog voltage outputs from the accelerometer and acoustic sensor via a shielded coaxial cable. For sensors employing piezoelectric integrated circuit technology, the response domain data acquisition unit 120 integrates a constant current source excitation circuit, configured to provide a constant excitation current (e.g., 4mA) to the sensor to drive its internal impedance transformation circuit. Simultaneously, the response domain data acquisition unit 120 uses an AC coupling mode connection, filtering out the DC bias component in the signal through a series capacitor, retaining only the dynamically changing AC component, thereby eliminating the influence of sensor zero-point drift or gravity components on dynamic measurements.

[0043] Before the analog signal enters the analog-to-digital converter, the response domain data acquisition unit 120 passes the signal through a hardware anti-aliasing filter. This filter is an analog low-pass filter, such as a high-order Butterworth filter or an elliptic filter. The response domain data acquisition unit 120 sets the cutoff frequency of this filter. Less than a fixed sampling frequency Half of the frequency (i.e., the Nyquist frequency) is used to physically attenuate high-frequency noise signals above the Nyquist frequency, preventing them from folding into the baseband frequency range during digitization and causing spectral aliasing.

[0044] The analog signal, after anti-aliasing filtering, is sent to the analog-to-digital converter (ADC). The response domain data acquisition unit 120 is configured to use the ADC at a fixed sampling frequency. Continuous sampling (e.g., 25.6kHz) is performed to discretize the continuous voltage signal into a digital voltage sequence. The ADC has at least 24 bits of resolution to ensure that it can simultaneously capture weak early fault signals and large impulse signals over a wide dynamic range.

[0045] The response domain data acquisition unit 120 further processes the digital voltage sequence based on the sensor's sensitivity parameters. This is converted into a vibration acceleration sequence or sound pressure sequence with physical dimensions. For any sampling point... Its physical value The calculation formula is as follows:

[0046] ;

[0047] in, The digital voltage sequence acquired by the ADC (unit: mV); This is the sensitivity coefficient of the corresponding sensor (unit: mV / g or mV / Pa). The hardware gain factor set for the signal conditioning circuit.

[0048] Finally, the response domain data acquisition unit 120 outputs the response domain data sequence after physical quantity conversion. This sequence is a time-discrete one-dimensional array containing high-frequency dynamic response information of the vehicle's powertrain and chassis structure in the time domain. Before outputting this sequence, the response domain data acquisition unit 120 can also perform digital high-pass filtering (e.g., setting the cutoff frequency to 0.5Hz) on it to further remove integral drift and ensure that the signal mean is zero.

[0049] The data acquisition and preprocessing module 100 of the present invention further includes a hardware synchronization triggering unit 130. The hardware synchronization triggering unit 130 is configured to establish a zero-point alignment relationship between the control domain data sequence and the response domain data sequence in the time dimension, thereby eliminating the startup delay differences between heterogeneous data sources.

[0050] The hardware synchronization trigger unit 130 is connected to the digital trigger bus of the data acquisition terminal 30 via internal logic circuitry. The hardware synchronization trigger unit 130 monitors the data reception status of the control domain data acquisition unit 110 and defines the first valid data frame on the controller local area network bus as the system startup event. When the control domain data acquisition unit 110 detects a message containing a valid engine speed signal in the first frame, the hardware synchronization trigger unit 130 immediately generates a transistor and transistor logic (TTL) level pulse signal with a rising edge characteristic.

[0051] The hardware synchronization trigger unit 130 physically transmits the TTL level pulse signal to the analog-to-digital conversion trigger port of the response domain data acquisition unit 120. The response domain data acquisition unit 120 is in hardware wait mode, only starting the sampling clock when the rising edge of the TTL level pulse signal is detected. Through this mechanism, the first sampling point of the response domain data sequence... The first valid message in the control domain data sequence is physically locked at the same moment, denoted as system absolute zero. .

[0052] Based on the system's absolute zero time The hardware synchronization trigger unit 130 performs unified calibration of all subsequently acquired data points relative to the time axis. For the first data point in the response domain data sequence... Each sampling point, and its corresponding relative time Determined by the following formula:

[0053] ;

[0054] in, For sampling points ( , ), This represents the total number of sampling points in the current single analysis data block; This is the fixed sampling frequency for the response domain.

[0055] For the first in the control domain data sequence For each data frame, the hardware synchronization trigger unit 130 records its absolute system timestamp upon arrival in the buffer of the data acquisition terminal 30. And calculate its corresponding relative time. :

[0056] ;

[0057] in, This is the absolute system timestamp when the TTL level pulse signal occurs.

[0058] Through the above calculations, the hardware synchronization trigger unit 130 ensures... and Both sequences are Starting from a common time reference and sharing a unified time reference throughout the entire acquisition cycle, it provides a precise time synchronization basis for subsequent phase fusion calculations.

[0059] See attached document Figure 2 The phase-locked resampling module 200 of the present invention includes a coarse reference frequency calculation unit 210, which is configured to establish a time-continuous rotational speed reference model. Due to the control domain data sequence... It is based on discretely sampled data transmission, and its values ​​exhibit a stepped distribution on the time axis. The coarse reference frequency calculation unit 210 reads discrete time and speed pairs. ,in For the first The discrete time points corresponding to each data frame. For the first The engine speed recorded in each data frame.

[0060] To obtain a smooth speed estimate at any given time, the coarse reference frequency calculation unit 210 uses a cubic spline interpolation algorithm to construct a globally continuous speed function. The coarse reference frequency calculation unit 210 calculates the frequency at any two adjacent discrete time points. and The time interval formed between Define a cubic polynomial function. The polynomial function satisfies the following boundary constraints: the function in The value at is equal to ,exist The value at is equal to Furthermore, at all internal nodes, the first derivative (rate of change of velocity) and the second derivative (rate of change of acceleration) of two adjacent polynomials remain continuous and equal.

[0061] The coarse reference frequency calculation unit 210 calculates the frequency used to describe the time interval. The cubic polynomial function expression of the internal rotational speed variation is:

[0062] ;

[0063] in, The polynomial coefficients corresponding to this time interval are obtained by solving the continuity equations of the global nodes. Through this algorithm, the coarse reference frequency calculation unit 210 generates a smooth rotational speed curve with second-order continuous differentiability, thereby ensuring that the angular acceleration calculated subsequently will not have non-physical abrupt changes.

[0064] Based on continuous speed function (i.e., composed of each segment) The global function (composed of the coarse reference frequency calculation unit 210) further calculates the reference frequency associated with the engine's mechanical structure. The coarse reference frequency calculation unit 210 pre-stores the main order parameters of the engine of the vehicle under test. For a four-stroke internal combustion engine, this parameter is defined as half the number of cylinders, representing the number of ignition strokes that occur per revolution of the engine.

[0065] The rough reference frequency calculation unit 210 calculates the time according to the following formula. Reference frequency:

[0066] ;

[0067] in, This is a continuous speed function, with units of revolutions per minute (rpm). The reference frequency is measured in Hertz (Hz). These are the engine's main order parameters. The curve characterizes the central trend of the engine's main excitation energy changing over time, providing a central reference for the search path in subsequent precise location of instantaneous frequency in broadband high-frequency signals.

[0068] The phase-locked resampling module 200 of the present invention further includes an instantaneous frequency ridge extraction unit 220, which is configured to use the physical characteristics of high-frequency vibration signals to correct the low-frequency deviation of the control domain signal and obtain a high-precision physical instantaneous frequency.

[0069] The instantaneous frequency ridge extraction unit 220 receives high-frequency vibration signals from the response domain data sequence. The instantaneous frequency ridge extraction unit 220 applies a sliding window function in the time domain and performs short-time Fourier transform (STFT) processing on it. The long-time signal is segmented into a sequence of short-time frames with a predetermined overlap rate (e.g., 75% overlap). The instantaneous frequency ridge extraction unit 220 performs a fast Fourier transform on each frame signal and calculates the square of the magnitude of the transform result, thereby generating an energy spectral density matrix describing the change of signal energy with time and frequency. .

[0070] The instantaneous frequency ridge extraction unit 220 introduces the reference frequency output by the coarse reference frequency calculation unit 210. As a priori constraint, in the energy spectral density matrix Construct a constrained dynamic search path. For each time step... The instantaneous frequency ridge extraction unit 220 defines the frequency search interval. :

[0071] ;

[0072] in, The preset unilateral search bandwidth is set within a range smaller than the minimum frequency interval between adjacent engine orders to prevent the ridge extraction process from jumping to adjacent interference orders.

[0073] Instantaneous frequency ridge extraction unit 220 in frequency search range The maximum energy point is searched internally to preliminarily determine the discrete peak frequencies of the principal order ridges. To eliminate the picket fence effect caused by the discrete Fourier transform and improve frequency resolution, the instantaneous frequency ridge extraction unit 220 uses discrete peak frequencies. Centered on the peak point, select a neighborhood set containing the peak point and its left and right adjacent frequency points. Calculate continuous instantaneous frequency ridges using energy centroid correction. The calculation process is shown in the following formula:

[0074] ;

[0075] in, For the corresponding frequency point At any moment The energy spectral density matrix; Indicates the selected neighborhood set The frequency-weighted energy moment within the energy spectrum, i.e., the first-order origin moment of the energy spectrum; Indicates the selected neighborhood set The sum of the total energy amplitudes within the energy spectrum is the zeroth-order origin moment of the energy spectrum.

[0076] Instantaneous frequency ridge extraction unit 220 calculates the ridge line for the entire time period. The sequence is subjected to Savitsky-Gore smoothing filter to remove random micro-jitter during the calculation process, and finally outputs a time-continuous instantaneous frequency curve as the physical reference for subsequent phase reconstruction.

[0077] The phase-locked resampling module 200 of the present invention further includes a high-precision phase reconstruction unit 230, which is configured to convert time-varying frequency information into monotonically increasing phase information and establish a mapping relationship between non-uniform time sampling points and fixed angle increments.

[0078] The high-precision phase reconstruction unit 230 receives the time-continuous instantaneous frequency ridge output by the instantaneous frequency ridge extraction unit 220. Since this frequency corresponds to the engine's primary order (e.g., ignition order), in order to obtain the absolute phase corresponding to the physical rotation of the crankshaft, the high-precision phase reconstruction unit 230 first utilizes pre-stored engine primary order parameters. The instantaneous frequency is normalized, and then numerical integration is performed over time to obtain the instantaneous cumulative phase. .

[0079] The high-precision phase reconstruction unit 230 calculates the time according to the following formula. Instantaneous cumulative phase:

[0080] ;

[0081] in, Indicates at time The instantaneous cumulative phase of the vehicle's power rotating components relative to their initial position, expressed in radians (rad). This is the circumferential constant, used to convert the number of rotations into radians; This represents the lower limit of integration, corresponding to the absolute zero moment of the system; This indicates the upper limit of integration, i.e., the current moment when the phase is being calculated; Denotes the integral variable, representing the process from... arrive Any point in time within the interval; Indicates the integral variable Differential elements; This represents the engine's primary order parameter, used to restore the monitored order frequency to the fundamental frequency of the physical shaft. For a four-stroke engine, this value is typically half the number of cylinders.

[0082] In obtaining the instantaneous cumulative phase Subsequently, the high-precision phase reconstruction unit 230 determines the equal angle increment according to the preset angular domain resolution. This increment is determined by the number of sampling points per revolution. (For example, 1024 points / revolution) determines the calculation method. The high-precision phase reconstruction unit 230 defines the phase points of the equally spaced angular domain sequence as... ,in For corner sampling ( ),in This represents the total number of points in the corner domain sequence after resampling within the current analysis time window.

[0083] In order to map the data acquired in the time domain to the angular domain, the high-precision phase reconstruction unit 230 needs to determine the phase points at equal intervals. The corresponding physical time. The high-precision phase reconstruction unit 230 solves the inverse function equation. Calculate the corresponding number Non-uniform resampling time of each corner domain sampling .because It is a monotonically increasing function, and the equation has a unique solution. The high-precision phase reconstruction unit 230 samples all angular domains. Performing the above calculations generates a series of non-uniform resampling times. The non-uniform resampling time sequence is composed of.

[0084] The high-precision phase reconstruction unit 230 utilizes the calculated non-uniform resampling time The high-frequency vibration signal output by the response domain data acquisition unit 120 Perform interpolation and resampling. Because the calculated... Typically, the amplitude does not fall on the original sampling point. The high-precision phase reconstruction unit 230 uses an interpolation algorithm (such as cubic spline interpolation) to calculate the vibration amplitude at that moment, and finally outputs the angular domain sequence. The mathematical expression for this resampling process is as follows:

[0085] ;

[0086] in, This represents a numerical interpolation operation function; It is a high-frequency vibration signal; For the first The non-uniform resampling time corresponding to each corner domain sampling; This results in an output angularly spaced sequence. Through this step, the invention eliminates the influence of rotational speed fluctuations on the signal period, resulting in a more consistent output. It exhibits stable periodic characteristics over the angular domain.

[0087] The phase-locked resampling module 200 of the present invention further includes a calculation order tracking execution unit 240, which is configured to transform the steady-state time-series signal in the angular domain to the order domain in order to separate the vibration characteristics synchronized with the rotational speed.

[0088] The order-of-calculation tracking execution unit 240 receives the angular domain sequence output by the high-precision phase reconstruction unit 230. Because this sequence is based on equal angular increments. The periodicity of the sampled signal directly corresponds to the geometric periodicity of the rotating component, rather than the temporal periodicity. The order-tracking execution unit 240 first applies a digital window function to the angular domain sequence of a selected length. (e.g., Hanning window or Heming window) to suppress spectral leakage caused by the cutoff effect.

[0089] The order-tracking execution unit 240 performs an angular-domain discrete Fourier transform on the windowed angular-domain sequence, mapping the signal from the angular domain to the order domain. The order-tracking execution unit 240 calculates the complex sequence of the order spectrum according to the following formula. :

[0090] ;

[0091] in, Spectral line index representing the order spectrum ( ); Indicates the first The order spectrum complex sequence at each spectral line contains amplitude and phase information; Indicates the total number of points in the corner domain data block participating in the transformation; Indicates the angular domain sampling index; Indicates the first Each corner domain sample corresponds to a corner domain sequence; Indicates the sampling index in the corner domain The digital window function at the location; is the base of the natural logarithm; It is the imaginary unit.

[0092] The order tracing execution unit 240 calculates the order spectrum complex number sequence. Calculate the order amplitude spectrum To give the spectral lines a clear physical meaning, the order tracking execution unit 240 calculates the number of sampling points per revolution using a preset number of sampling points. Establish spectral index With physical order A linear mapping relationship between them.

[0093] The order tracing execution unit 240 calculates the order according to the following formula. Physical order corresponding to the spectral lines :

[0094] ;

[0095] in, Spectral line index representing the order spectrum ( ); Indicates the total number of points in the corner domain data block participating in the transformation; This represents the number of sampling points per revolution.

[0096] Based on this mapping relationship, the order tracing execution unit 240 determines the maximum analysis order of the system as follows: The final output of the computational order tracing execution unit 240 is in physical order. The order spectrum is plotted with the horizontal axis representing the vibration amplitude and the vertical axis representing the vibration energy distribution locked at engine speed.

[0097] See attached document Figure 3 The virtual reference model module 300 of this invention includes a dynamic state vector construction unit 310, which is configured to filter and reconstruct feature vectors that characterize the current operating conditions of the power system from a large amount of heterogeneous data generated by the vehicle bus, providing a standardized input benchmark for subsequent deep neural network inference. The dynamic state vector construction unit 310 is connected to the data acquisition and preprocessing module 100, receiving the cleaned and synchronized control domain data sequence therein. .

[0098] The dynamic state vector construction unit 310 is based on the vibration generation mechanism of the power transmission system, and extracts data from the control domain data sequence. Key operating parameters that have a decisive influence on vibration characteristics are selected. The parameters selected by the dynamic state vector construction unit 310 include at least engine speed, engine output torque, and engine load percentage. To eliminate the influence of the order-of-magnitude differences between different physical dimensions on subsequent neural network weight updates, the dynamic state vector construction unit 310 normalizes each parameter and constructs the time-varying parameters. dynamic state vector .

[0099] Dynamic state vector construction unit 310 constructs dynamic state vectors according to the following formula. :

[0100] ;

[0101] in, Indicates at time The constructed dynamic state vector is used as input layer data for the deep neural network; Representing time respectively The original engine speed signal (rpm), engine output torque signal (N), and engine load signal (%) were collected. These represent the preset maximum and minimum engine speed constants, respectively; These represent the preset maximum and minimum engine torque constants, respectively; These represent the preset maximum and minimum load constants, respectively. All of the above maximum and minimum constants are pre-programmed into the system memory based on the factory nameplate parameters of the engine of the vehicle under test or the limit physical values ​​of the bench calibration. These represent the normalized speed component, torque component, and load component, respectively, and their values ​​are all mapped to the [0,1] interval. This represents the transpose operation of a vector or matrix.

[0102] The dynamic state vector construction unit 310 is also configured to perform a time alignment check on the input data to ensure the construction of the dynamic state vector. The timestamp is strictly synchronized with the timestamp of the actual vibration signal to be compared later. The dynamic state vector construction unit 310 finally outputs a dynamic state vector containing current working condition information. The vector is then transmitted to a pre-trained deep neural network model to stimulate the network to output the corresponding fault-free ideal angular domain feature spectrum under the current operating condition.

[0103] The virtual reference model module 300 of the present invention also includes a virtual reference model inference unit 320, which is configured to use artificial intelligence algorithms to establish a nonlinear mapping relationship between vehicle operating conditions and vibration response, and infer the theoretical benchmark characteristics of the system under fault-free conditions based on the current operating conditions.

[0104] The virtual reference model inference unit 320 is pre-installed with a deep neural network (DNN) model trained by deep learning. This model uses the dynamic state vector output by the dynamic state vector construction unit 310. As the input layer excitation, since the deep neural network model has learned a large number of vibration response modes of fault-free vehicles under different speed, torque and load combinations during the training phase, the virtual reference model inference unit 320 can reconstruct the ideal vibration characteristics corresponding to the current working condition through forward propagation calculation.

[0105] The virtual reference model inference unit 320 performs model inference according to the following formula and outputs the predicted ideal angular domain feature spectrum. :

[0106] ;

[0107] in, This represents the predicted ideal angular domain feature spectrum, characterizing the feature spectrum at time [time value missing]. Under operating conditions, the order vibration amplitude distribution that the vehicle powertrain system should exhibit when it is in a healthy state; This represents the input dynamic state vector, which includes normalized speed, torque, and load information; This represents a nonlinear mapping function for a deep neural network, which is composed of a multi-layer neuron structure and activation functions (such as ReLU or Tanh). This represents the pre-trained optimal parameter set in a deep neural network, including the connection weight matrices of each layer. and bias vector .

[0108] The optimal parameter set used by the virtual reference model inference unit 320 It is obtained through offline training. During training, a massive amount of historical data from normal vehicles is used as samples, and backpropagation updates are performed with the objective function of minimizing the mean square error (MSE) between the predicted output and the actual healthy vibration spectrum. Therefore, the output of the virtual reference model inference unit 320... Abnormal vibration components caused by faults such as mechanical wear, loosening, or breakage were eliminated, and only the reference vibration energy determined by the engine combustion work and the inherent characteristics of the mechanical structure was retained.

[0109] The final output of the virtual reference model inference unit 320 is the predicted ideal angular domain feature spectrum. It has the same dimensions and physical meaning as the measured order amplitude spectrum output by the order tracking execution unit 240, thus providing a virtual reference with strict physical comparability for subsequent residual analysis and fault isolation.

[0110] See attached document Figure 4The residual analysis decision module 400 of the present invention includes a cross-modal residual calculation unit 410, which is configured to quantitatively evaluate the difference between the current actual operating state of the vehicle powertrain system and the ideal health benchmark in the statistical manifold space. The cross-modal residual calculation unit 410 is connected to the phase-locked resampling module 200 and the virtual reference model module 300, respectively.

[0111] The cross-modal residual calculation unit 410 receives the angular domain sequence output by the phase-locked resampling module 200. To achieve in-domain comparison with the prediction model, the cross-modal residual calculation unit 410 first performs an angular domain Fourier transform on the angular domain sequence, converting it into the actual angular domain feature spectrum. This characteristic spectrum contains the amplitude distribution information of vibrations of each order under the current operating condition. Simultaneously, the cross-modal residual calculation unit 410 receives the predicted ideal angular domain characteristic spectrum output by the virtual reference model module 300. .

[0112] To eliminate the interference of non-critical order noise on residual analysis, the cross-modal residual calculation unit 410 introduces an order mask vector. Spatial filtering is applied to the original residual. Order mask vector. It is a preset weight vector that has non-zero values ​​at the index positions corresponding to the engine's main order and its harmonic components, while its values ​​approach zero at other unrelated frequency band positions. The cross-modal residual calculation unit 410 calculates the residual vector after mask filtering. .

[0113] The cross-modal residual calculation unit 410 calculates the filtered residual vector according to the following formula. :

[0114] ;

[0115] in, This represents the residual vector after order mask filtering, which retains only the magnitude deviation at the order of interest; This represents the actual angular domain feature spectrum obtained by conversion based on measured data; This represents the predicted ideal angular domain feature spectrum obtained based on inference from a deep neural network; This represents a preset order mask vector used to filter key feature frequency bands; This represents the Hadamard product, which is the element-wise multiplication of vectors.

[0116] To comprehensively assess the anomaly degree of multidimensional residual characteristics, the cross-modal residual calculation unit 410 calculates the residual vector... Mapping to a statistical manifold space and calculating its Mahalanobis distance. Compared to the traditional Euclidean distance, the Mahalanobis distance can eliminate the interference of correlation between amplitudes of different orders and the difference in dimensional scale.

[0117] The cross-modal residual calculation unit 410 calculates the Mahalanobis distance according to the following formula. :

[0118] ;

[0119] in, The Mahalanobis distance of the system at the current moment is used as a statistical indicator to quantify the degree of anomaly. This represents the filtered residual vector obtained from the current calculation. This represents the mean vector of the residual vectors under historical normal operating conditions (obtained during the system calibration phase and usually set as the zero vector). The covariance matrix representing the residual vector under historical normal operating conditions. The inverse matrix of describes the statistical correlation between residuals of different orders; This represents the transpose operation of a vector or matrix.

[0120] The cross-modal residual calculation unit 410 finally outputs the calculated Mahalanobis distance. When the Mahalanobis distance When the vibration exceeds the preset statistical threshold, it indicates that the current vibration characteristics have significantly deviated from the healthy distribution cluster in the statistical manifold space, and the residual analysis decision module 400 generates an abnormal alarm signal.

[0121] The residual analysis decision module 400 of the present invention further includes an order correlation masking filter unit 420, which is configured to construct a spatial filter based on the inherent harmonic characteristics of the power transmission system in order to extract key feature components synchronized with the rotational speed from the full-band residual.

[0122] The order correlation masking filter unit 420 predefines a set of key target order based on the number of combustion strokes and cylinder arrangement of the engine. This set typically contains the engine's primary order parameters. And its integer multiples of harmonic components. To construct the order mask vector. The order correlation masking filter unit 420 generates a weighted window with a preset bandwidth at the center position of each target order, and superimposes all windows in the frequency domain.

[0123] The order correlation masking filter unit 420 calculates the order mask vector according to the following formula. The Middle Weight values ​​corresponding to spectral lines :

[0124] ;

[0125] in, Represents the order mask vector In the The weight value at each spectral line ranges from (0,1]. This indicates the total number of harmonic orders of primary interest (e.g., the top 10 harmonics). Indicating harmonic multiple index ( ); Indicates the engine's main order parameter; Indicates the first The physical order corresponding to each spectral line is given by the formula Sure; This represents the bandwidth parameter of the Gaussian window, which determines the passband width of the filter and is used to accommodate energy dissipation caused by numerical calculation leakage or small speed fluctuations. This represents an exponential function.

[0126] The order correlation masking filter unit 420 performs the above Gaussian superposition operation to generate an order mask vector. At all critical stages The vibration residuals exhibit high weighting in the vicinity and low weighting in the non-harmonic region. This construction method can effectively suppress unrelated vibration residuals caused by road surface excitation, random noise, or non-dynamic components.

[0127] The order correlation masking filter unit 420 is also configured to dynamically adjust the bandwidth parameters according to the current engine speed. Under high-speed operating conditions, the order correlation masking filter unit 420 should be appropriately increased. The value is adjusted to accommodate the spectral broadening effect caused by the relative decrease in sampling rate. Specifically, the order correlation masking filter unit 420 is based on the linear mapping formula. Calculate the bandwidth parameters at the current moment, where The basic bandwidth constant obtained from the bench calibration, The preset speed broadening factor, This represents the current engine speed. In this embodiment, the base bandwidth constant is... The dimension of is order, and its value range is set to [0.05, 0.25] to cover the leakage of basic numerical calculations; the rotational speed broadening factor The unit of measurement is order / revolution per minute, and its value range is set to [1×10]. −5 5×10 −4 This ensures that, at the engine's maximum speed, the width of the mask window is controlled within the range of adjacent half-orders, preventing energy interference between adjacent different orders. The order-related mask filtering unit 420 ultimately converts the constructed order mask vector... The data is transmitted to the cross-modal residual calculation unit 410 for point-to-point Hadamard product filtering.

[0128] The residual analysis decision module 400 of the present invention also includes a manifold distance calculation and alarm unit 430, which is configured to make a judgment on the quantified Mahalanobis distance index based on the statistical hypothesis testing principle, so as to identify the abnormal state of the power transmission system and generate corresponding control signals.

[0129] The manifold distance calculation and alarm unit 430 receives the Mahalanobis distance output by the cross-modal residual calculation unit 410. To achieve statistically significant fault detection, the manifold distance calculation and alarm unit 430 establishes an anomaly judgment threshold using the distribution characteristics of the squared Mahalanobis distance. Under the null hypothesis that the system is in a healthy operating state, the squared Mahalanobis distance... Approximately follows the degree of freedom of The chi-square distribution, where The degrees of freedom of the residual vector involved in the calculation.

[0130] The manifold distance calculation and alarm unit 430 calculates and alarms based on a preset confidence level (e.g., or Determine the anomaly judgment threshold The manifold distance calculation and alarm unit 430 calculates the threshold by querying the inverse cumulative distribution function of the chi-square distribution. The mathematical relationship is described as follows:

[0131] ;

[0132] in, The anomaly detection threshold represents the critical statistical boundary for determining whether a system failure has occurred. The inverse cumulative distribution function of the chi-square distribution; This represents the preset significance level or false alarm tolerance, usually taken as a very small value (e.g., 0.001). Indicates the confidence level; The degree of freedom of the residual vector is represented by the order of the mask vector. The total number of spectral lines with a weight value greater than a preset effective threshold (e.g., 0.1).

[0133] After determining the threshold, the manifold distance calculation and alarm unit 430 will receive the Mahalanobis distance in real time. With the abnormal judgment threshold The system compares the results and performs abnormal state decisions. To prevent false alarms caused by transient noise, the manifold distance calculation and alarm unit 430 uses a sliding window mechanism to statistically analyze the comparison results over multiple consecutive sampling periods, and finally outputs the system status flag. .

[0134] The manifold distance calculation and alarm unit 430 determines the system status flag bit according to the following formula. :

[0135] ;

[0136] in, Indicates the system status flags in the binary representation; This indicates a fault state, signifying that the current vibration characteristics deviate significantly from the healthy baseline cluster in the statistical manifold space, and the degree of deviation exceeds the statistically permissible range. This indicates a normal state, meaning the current vibration characteristics are within the expected statistical fluctuation range; This represents a conditional logical operator, used to indicate that when the following inequality condition is met, the corresponding numerical value will be assigned to it. .

[0137] when The output is At that time, the manifold distance calculation and alarm unit 430 generates an abnormal alarm signal and sends the signal to the vehicle's on-board diagnostic system (OBD) or human-machine interface (HMI). At the same time, it triggers the freezing frame storage of relevant vibration data for subsequent fault tracing analysis.

[0138] See attached document Figure 5 This invention provides a vehicle anomaly detection method based on multimodal data analysis, comprising the following steps:

[0139] In step S1, the vehicle anomaly detection method first performs synchronous acquisition and preprocessing of multimodal data. This step utilizes a high-frequency accelerometer to acquire vibration acceleration signals reflecting the mechanical response of the vehicle's powertrain system, and simultaneously acquires control domain data reflecting the system's operating conditions via the vehicle controller area network (CAN bus). The control domain data includes at least engine speed, engine torque, and engine load. To eliminate clock drift between heterogeneous data sources, this step timestamps all data streams according to the high-precision local clock or Network Time Protocol (NTP) of the data acquisition terminal 30, and removes invalid data frames caused by communication packet loss, constructing a time-synchronized raw dataset.

[0140] In step S2, the vehicle anomaly detection method performs angular domain resampling based on coarse-fine coupled phase-locked loop. This step does not rely on a constant sampling frequency, but rather performs isoangular resampling based on the rotation angle of the engine crankshaft. Specifically, this step first uses the engine speed pulse signal for coarse phase estimation, and then combines instantaneous frequency estimation techniques to perform fine phase correction, generating an accurate isoangular sampling time sequence. This step maps the time-domain non-stationary vibration signal into an angular-domain stationary sequence. This transforms the spectral characteristics that drift with rotational speed into characteristic spectral lines with fixed positions on the order spectrum, thus eliminating the spectral ambiguity caused by rotational speed fluctuations.

[0141] In step S3, the vehicle anomaly detection method constructs a virtual reference model and infers ideal features. This step is based on a data-driven deep learning strategy, extracting normalized dynamic state vectors from the control domain data. This step will transform the dynamic state vector The input is fed into a pre-trained deep neural network model, and through the network's forward propagation operation, the predicted ideal angular domain feature spectrum that should be presented when the system is in a fault-free and healthy state under the current operating conditions is inferred. The predicted spectrum dynamically reflects the theoretical baseline vibration distribution of the vehicle at the current specific speed and load.

[0142] In step S4, the vehicle anomaly detection method performs cross-modal residual calculation and order mask filtering. This step first converts the resampled actual angular domain sequence into an actual angular domain feature spectrum. And calculate its correlation with the predicted ideal angular domain feature spectrum. The difference between them. To focus on fault characteristics directly related to the mechanical connections of the powertrain, this step applies a preset order mask vector. Weighted filtering is applied to the original residual. Order mask vector. The engine's main order and its harmonics have high gain, thus filtering out irrelevant background noise such as road excitation and generating a filtered residual vector. .

[0143] In step S5, the vehicle anomaly detection method performs anomaly decision-making and alarm based on statistical manifold. This step involves filtering the residual vector... Projected onto the statistical manifold space, the Mahalanobis distance is calculated using the pre-stored health covariance matrix. The Mahalanobis distance As a unified measure of the current state's deviation from a healthy baseline, this step eliminates the influence of dimensions of different orders and correlations. This step further applies the calculated Mahalanobis distance... Anomaly detection threshold determined based on chi-square distribution Comparison, when detected When the threshold is exceeded, the system is determined to be abnormal and an alarm signal containing information on the time and severity of the fault is generated.

Claims

1. A vehicle anomaly detection system based on multimodal data analysis, characterized in that, It includes a data acquisition terminal (30) and an edge computing processing host (40), wherein the edge computing processing host (40) includes: The data acquisition preprocessing module (100) is configured to receive data from the data acquisition terminal (30) and output a control domain data sequence containing engine speed and a response domain data sequence containing high-frequency vibration signals. A phase-locked resampling module (200) is configured to calculate a reference frequency based on the engine speed, extract an instantaneous frequency ridge from the high-frequency vibration signal, and use the instantaneous frequency ridge to interpolate and resample the high-frequency vibration signal to output an angular domain sequence. The virtual reference model module (300) is configured to construct a dynamic state vector based on the control domain data sequence and output the corresponding predicted ideal angular domain feature spectrum using a deep neural network model. The residual analysis decision module (400) is configured to calculate the residual vector between the actual angular domain feature spectrum of the angular domain sequence transformation and the predicted ideal angular domain feature spectrum, and generate an abnormal alarm signal based on the Mahalanobis distance of the residual vector.

2. The vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The data acquisition and preprocessing module (100) includes a control domain data acquisition unit (110), a response domain data acquisition unit (120), and a hardware synchronization triggering unit (130). The data acquisition terminal (30) is connected to the control domain sensing unit (21) and the response domain sensing unit (22) in the sensing network layer (20) of the vehicle subsystem under test (10). The hardware synchronization triggering unit (130) is connected to the digital triggering bus of the data acquisition terminal (30), and is configured to monitor the data receiving status of the control domain data acquisition unit (110), and define the first valid data frame on the controller local area network bus as a system startup event; When the hardware synchronization triggering unit (130) recognizes a message containing a valid engine speed signal in the first frame, it generates a transistor and transistor logic level pulse signal with rising edge characteristics and transmits the transistor and transistor logic level pulse signal to the analog-to-digital conversion trigger port of the response domain data acquisition unit (120). The response domain data acquisition unit (120) starts the sampling clock when it detects the rising edge of the transistor and the transistor logic level pulse signal, and establishes the zero-point alignment relationship between the control domain data sequence and the response domain data sequence in the time dimension.

3. The vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The phase-locked resampling module (200) includes a coarse reference frequency calculation unit (210). The coarse reference frequency calculation unit (210) reads discrete data pairs consisting of discrete time points and engine speed, defines a cubic polynomial function between any two adjacent discrete time points, and solves for the cubic polynomial coefficients to construct a globally continuous speed function. The coarse reference frequency calculation unit (210) calculates the time-continuous reference frequency based on the globally continuous speed function and the pre-stored engine major order parameters.

4. The vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The phase-locked resampling module (200) includes an instantaneous frequency ridge extraction unit (220). The instantaneous frequency ridge extraction unit (220) receives the high-frequency vibration signal in the response domain data sequence and performs a short-time Fourier transform on the high-frequency vibration signal to generate an energy spectral density matrix; The instantaneous frequency ridge extraction unit (220) uses the reference frequency as a priori constraint, defines a restricted frequency search interval on the energy spectral density matrix, searches for the energy maximum point within the frequency search interval to determine the discrete peak frequency of the principal order ridge, selects a neighborhood set around the discrete peak frequency, and calculates the continuous instantaneous frequency ridge using energy centroid correction.

5. A vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The phase-locked resampling module (200) includes a high-precision phase reconstruction unit (230); The high-precision phase reconstruction unit (230) receives the instantaneous frequency ridge line with continuous time, performs normalization processing on the instantaneous frequency ridge line using the engine's main order constant, and performs numerical integration operation on time to obtain the instantaneous cumulative phase. The high-precision phase reconstruction unit (230) determines the equal angle increment based on the number of sampling points per revolution, calculates the non-uniform resampling time of the corresponding angular domain sampling by solving the inverse function equation, and uses the non-uniform resampling time to perform cubic spline interpolation resampling on the high-frequency vibration signal to generate the angular domain sequence.

6. The vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The phase-locked resampling module (200) includes a calculation order tracking execution unit (240). The order tracking execution unit (240) applies a digital window function to the angular domain sequence, performs an angular domain discrete Fourier transform on the windowed angular domain sequence, and calculates the order spectrum complex sequence and the order amplitude spectrum. The order calculation tracking execution unit (240) establishes a linear mapping relationship between the spectral index and the physical order using a preset number of sampling points per revolution, and outputs an order spectrum. The horizontal axis of the order spectrum is set to the physical order, and the vertical axis is set to the vibration amplitude.

7. The vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The virtual reference model module (300) includes a dynamic state vector construction unit (310) and a virtual reference model inference unit (320). The dynamic state vector construction unit (310) selects engine speed signal, engine output torque signal and engine load signal from the control domain data sequence, performs normalization processing on the engine speed signal, the engine output torque signal and the engine load signal respectively, and constructs the dynamic state vector; The virtual reference model inference unit (320) receives the dynamic state vector, transmits the dynamic state vector as an input layer excitation to the deep neural network model, and infers and outputs the predicted ideal angular domain feature spectrum.

8. A vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The residual analysis decision module (400) includes an order correlation mask filtering unit (420). The order correlation masking filter unit (420) defines a set of key target orders that includes the engine principal order constant and the harmonic components of the engine principal order constant, and generates a weighted window with a preset bandwidth at the center of each target order in the set of key target orders. The order correlation mask filtering unit (420) performs superposition synthesis on the weighted window in the frequency domain to construct an order mask vector, and dynamically adjusts the bandwidth parameter of the weighted window according to the current engine speed. The order mask vector is used to perform Hadamard product filtering on the residual vector.

9. A vehicle anomaly detection system based on multimodal data analysis according to claim 1, characterized in that, The residual analysis decision module (400) includes a cross-modal residual calculation unit (410) and a manifold distance calculation and alarm unit (430). The cross-modal residual calculation unit (410) uses the inverse matrix of the covariance matrix of the residual vector under historical normal operating conditions to calculate the Mahalanobis distance of the filtered residual vector; The manifold distance calculation and alarm unit (430) determines the anomaly decision threshold based on the preset confidence level and the inverse cumulative distribution function of the chi-square distribution, compares the Mahalanobis distance with the anomaly decision threshold, and outputs the anomaly alarm signal when the Mahalanobis distance is greater than the anomaly decision threshold.

10. A vehicle anomaly detection method based on multimodal data analysis, characterized in that, The vehicle anomaly detection system based on multimodal data analysis, as described in any one of claims 1-9, comprises the following steps: The data acquisition and preprocessing module (100) acquires the control domain data sequence from the control domain sensing unit (21) and the response domain data sequence from the response domain sensing unit (22) through the data acquisition terminal (30), and establishes a time zero-point alignment relationship using the hardware synchronization trigger unit (130). The phase-locked resampling module (200) receives the control domain data sequence and the response domain data sequence, calculates the reference frequency using the coarse reference frequency calculation unit (210), extracts the instantaneous frequency ridge using the instantaneous frequency ridge extraction unit (220), calculates the instantaneous cumulative phase using the high-precision phase reconstruction unit (230), and performs interpolation resampling on the response domain data sequence to generate the angular domain sequence. The virtual reference model module (300) extracts the dynamic state vector from the control domain data sequence using the dynamic state vector construction unit (310), and infers the predicted ideal angular domain feature spectrum corresponding to the current working condition using the virtual reference model inference unit (320). The residual analysis decision module (400) uses the cross-modal residual calculation unit (410) to calculate the difference between the actual angular domain feature spectrum obtained by the angular domain sequence conversion and the predicted ideal angular domain feature spectrum, and uses the order mask vector generated by the order correlation mask filtering unit (420) to filter the difference; The manifold distance calculation and alarm unit (430) calculates the Mahalanobis distance of the filtered residual vector in the statistical manifold space, compares the Mahalanobis distance with an anomaly decision threshold determined based on the statistical distribution, and generates an anomaly alarm signal when the Mahalanobis distance exceeds the anomaly decision threshold.

Images