Motor position sensor detection method for a steering system

Abstract

This invention provides a method for detecting motor position sensors in steering systems, comprising: acquiring real-time rotor position signals, torque signals, and speed signals of the steering motor; performing wavelet decomposition on the rotor position signal under normal operating conditions, statistically analyzing frequency band characteristics according to torque and speed intervals, and generating a reference probability distribution and uncertainty tensor; performing wavelet decomposition on the real-time signal, employing an asymmetric time window containing a short-time analysis window and a preceding discontinuous historical micro-pane, and solving the current and background probability density functions respectively; modulating the deviation contribution of the current state relative to the reference probability density based on the reference uncertainty tensor, calculating the absolute and relative dominance measures band by band, and concatenating them into a composite random dominance vector; mapping the vector to the fault feature space to obtain projected coordinates, adjusting the judgment superbody boundary by adjusting the recent background fluctuation measure determined by the historical micro-pane, and determining sensor fault if the coordinates exceed the boundary.

Description

Technical Field

[0001] This application belongs to the field of steering systems, and in particular relates to a method for detecting the position sensor of a motor in a steering system. Background Technology

[0002] The electric power steering system is a crucial component of modern automotive chassis control, and its operational stability and reliability directly impact vehicle handling performance and driving safety. In this system, the motor position sensor is used to collect the motor rotor position signal in real time. This signal is essential for the control unit to execute motor torque control, phase compensation, and closed-loop regulation. If the position sensor experiences signal distortion, drift, or malfunction, it can cause fluctuations in the motor's output torque, resulting in abnormally reduced steering assistance, or even system failure, posing safety risks. With the continuous improvement of automotive intelligence, advanced driver assistance systems (ADAS) and autonomous driving systems place higher demands on the reliability of underlying actuators. Therefore, effectively monitoring the operating status of the motor position sensor in the automotive steering system and promptly detecting potential faults has become a vital aspect of ensuring overall vehicle safety and a matter of ongoing concern in this field.

[0003] However, during actual vehicle operation, the steering system is constantly under dynamically changing conditions. Affected by factors such as changes in vehicle speed, sudden changes in road surface adhesion conditions, electromagnetic interference, and changes in driving operation, the motor speed and torque are prone to frequent fluctuations, causing the signals collected by the position sensor to exhibit significant non-stationarity, randomness, and time-varying characteristics. Existing sensor detection methods still have shortcomings under complex operating conditions, failing to fully characterize the probability distribution characteristics and inherent uncertainties of position signals under different operating states. When the vehicle undergoes a transient change in operating conditions, normal system fluctuations are easily misjudged as sensor malfunctions, leading to a high false alarm rate. Simultaneously, existing methods typically lack the ability to perform longitudinal correlation analysis of recent historical background states and the ability to fuse multi-dimensional features, making it difficult to dynamically adjust the judgment boundary according to changes in the actual operating environment. This makes it difficult to promptly identify early weak sensor anomalies, resulting in the risk of missed detections. Therefore, how to identify the statistical variation patterns of position signals under changing operating conditions and establish a fault detection method that can fuse multi-dimensional state features, suppress background fluctuation interference, and dynamically adjust the judgment boundary to solve the problems of low sensor detection accuracy and poor stability under changing and non-stationary operating conditions is a pressing technical problem in this field. Summary of the Invention

[0004] According to a first aspect of the present invention, a method for detecting motor position sensors in steering systems is proposed to address the problems of low detection accuracy and poor stability of sensors under variable and non-stationary operating conditions, comprising the following steps:

[0005] The system acquires real-time rotor position, torque, and speed signals of the automotive steering system motor; performs wavelet decomposition on the rotor position signals under historical normal operating conditions; performs feature statistical analysis on each frequency band signal based on the corresponding torque and speed signal intervals; and generates a baseline probability distribution tensor and a baseline uncertainty tensor. Wavelet decomposition is performed on the real-time rotor position signal. For each frequency band of real-time signal components, an asymmetric time window is used to extract data. The asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window. The current state probability density function within the short-time analysis window and the recent background probability density function obtained by weighted fusion of signals within the historical micro-panes are calculated respectively. Based on the benchmark uncertainty tensor, the deviation contribution of the current state probability density function from the benchmark probability density function is modulated with confidence. The absolute dominance measure of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance measure relative to the recent background probability density function are calculated band by band. The measures of each frequency band are concatenated to form a composite random dominance vector. The composite random dominance vector is projected onto a preset fault feature vector space to obtain fault projection coordinates; the recent background fluctuation metric determined by the historical micropane is used to adaptively adjust the boundary of the judgment superbody that defines the normal working range; when the fault projection coordinates are outside the judgment superbody, the sensor is determined to be faulty.

[0006] Optionally, acquiring the real-time rotor position signal, torque signal, and speed signal of the vehicle steering system motor includes: A continuous two-phase quadrature analog signal sequence is acquired by a resolver installed at the motor shaft end, and then analyzed into a real-time rotor position signal sequence by a resolver-to-digital converter. Extract speed and torque messages with consistent timestamps from the communication messages of the vehicle control communication network, and reconstruct real-value speed and torque signals according to the communication protocol. Real-time rotor position signals, speed signals, and torque signals are aligned and fused based on the same timestamp to construct an integrated multi-dimensional time-series data sequence.

[0007] Optionally, the wavelet decomposition of the rotor position signal under historical normal operating conditions includes: The Daubechies-4 wavelet basis function is used as the core scale analytical model of the filtering network to perform four-level multi-scale orthogonal wavelet decomposition on the rotor position signal under historical normal working conditions. Filter out low-frequency smooth approximation components and extract high-frequency detail components representing signal abrupt changes; The extracted high-frequency detail components of each layer are recombined according to the scale level to form an independent state vector sequence for each frequency band.

[0008] Optionally, the confidence modulation of the deviation contribution of the current state probability density function relative to the benchmark probability density function based on the benchmark uncertainty tensor includes: Using the representative values ​​of the torque signal and the representative values ​​of the speed signal corresponding to the short-time analysis window, as well as the current frequency band number, as indices, the corresponding discrete values ​​are extracted from the reference uncertainty tensor as basic parameters to be adjusted. Divide the dispersion value by the preset reference dispersion to obtain the dispersion ratio coefficient. Using the natural constant e as the base, and the negative number of the dispersion ratio coefficient as the exponent, calculate the attenuation weight factor with a value range greater than 0 and less than or equal to 1. The baseline probability density function is normalized and verified, and the decay weight factor is used as the confidence weight of the deviation contribution of the current state probability density function relative to the baseline probability density function. The larger the dispersion value, the smaller the decay weight factor and the lower the deviation contribution.

[0009] Optionally, the wavelet decomposition of the real-time rotor position signal and the data extraction of each frequency band real-time signal component using an asymmetric time window, wherein the asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window, including: In the real-time signal components of each frequency band, a sequence of 1000 consecutive sampling points is extracted from the current moment and traced back to the past to form the short-time analysis window; Starting from the historical boundary of the short-time analysis window, continue backtracking to the past, skipping the blank interval area with a width of 500 sampling points; After the blank interval area, extract a sequence of 200 sampling points as the first historical micro-pane, and repeat the operation of skipping 500 sampling points and then extracting a sequence of 200 sampling points three times to form four non-continuously distributed historical micro-panes.

[0010] Optionally, the step of calculating the absolute dominance metric of the current state probability density function relative to the normalized and verified baseline probability density function and the relative dominance metric relative to the recent background probability density function per frequency band includes: On a unified dimensionless feature domain, the cumulative distribution function difference between the current state probability density function and the normalized and verified baseline probability density function is calculated band by band, and the absolute value integral of the difference is used as the original absolute dominance metric. The absolute dominance metric is obtained by weighting the original absolute dominance metric with a decay weighting factor determined by the baseline uncertainty tensor. Calculate the cumulative distribution function difference between the current state probability density function and the recent background probability density function, and use the area of ​​the integral of the absolute value of the difference over a unified dimensionless feature domain as the relative dominance metric.

[0011] Optionally, the step of projecting the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and using the recent background fluctuation metric determined by the historical micropane to adaptively adjust the boundary of the judgment superbody defining the normal working range, and determining that the sensor is faulty when the fault projection coordinates are outside the judgment superbody, includes: In the fault feature vector space, a judgment hypersphere is established with the origin of the coordinate system as the center, and the hypersphere of the judgment hypersphere is used as the boundary of the normal working range. Calculate the Euclidean distance from the fault projection coordinates to the origin; Determine whether the Euclidean distance is greater than the radius of the determination hypersphere. If so, determine that the fault projection coordinates are outside the determination hypersphere and determine that the sensor is faulty. The radius is composed of a preset base radius and an adjustment term that is positively correlated with the recent background fluctuation metric determined by the historical micropane.

[0012] According to a second aspect of the present invention, a motor position sensor detection system for a steering system is provided, comprising the following modules: The generation module is used to acquire the real-time rotor position signal, torque signal, and speed signal of the automotive steering system motor; it performs wavelet decomposition on the rotor position signal under historical normal operating conditions, performs feature statistical analysis on the signals of each frequency band according to the division interval of the corresponding torque signal and speed signal, and generates the benchmark probability distribution tensor and the benchmark uncertainty tensor. The calculation module performs wavelet decomposition on the real-time rotor position signal. For each frequency band of real-time signal components, it uses an asymmetric time window to extract data. This asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window. It calculates the current state probability density function within the short-time analysis window and the recent background probability density function obtained by weighted fusion of signals within the historical micro-panes. Based on the benchmark uncertainty tensor, it performs confidence modulation on the deviation contribution of the current state probability density function relative to the benchmark probability density function. It calculates the absolute dominance metric of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance metric relative to the recent background probability density function band by band, and concatenates the metrics of each frequency band to form a composite random dominance vector. The determination module is used to project the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and to adaptively adjust the determination superbody boundary that defines the normal working range using the recent background fluctuation metric determined by the historical micropane. When the fault projection coordinates are located outside the determination superbody, the sensor is determined to be faulty.

[0013] Preferably, acquiring the real-time rotor position signal, torque signal, and speed signal of the vehicle steering system motor includes: A continuous two-phase quadrature analog signal sequence is acquired by a resolver installed at the motor shaft end, and then analyzed into a real-time rotor position signal sequence by a resolver-to-digital converter. Extract speed and torque messages with consistent timestamps from the communication messages of the vehicle control communication network, and reconstruct real-value speed and torque signals according to the communication protocol. Real-time rotor position signals, speed signals, and torque signals are aligned and fused based on the same timestamp to construct an integrated multi-dimensional time-series data sequence.

[0014] Preferably, the wavelet decomposition of the rotor position signal under historical normal operating conditions includes: The Daubechies-4 wavelet basis function is used as the core scale analytical model of the filtering network to perform four-level multi-scale orthogonal wavelet decomposition on the rotor position signal under historical normal working conditions. Filter out low-frequency smooth approximation components and extract high-frequency detail components representing signal abrupt changes; The extracted high-frequency detail components of each layer are recombined according to the scale level to form an independent state vector sequence for each frequency band.

[0015] Preferably, the confidence modulation of the deviation contribution of the current state probability density function relative to the benchmark probability density function based on the benchmark uncertainty tensor includes: Using the representative values ​​of the torque signal and the representative values ​​of the speed signal corresponding to the short-time analysis window, as well as the current frequency band number, as indices, the corresponding discrete values ​​are extracted from the reference uncertainty tensor as basic parameters to be adjusted. Divide the dispersion value by the preset reference dispersion to obtain the dispersion ratio coefficient. Using the natural constant e as the base, and the negative number of the dispersion ratio coefficient as the exponent, calculate the attenuation weight factor with a value range greater than 0 and less than or equal to 1. The baseline probability density function is normalized and verified, and the decay weight factor is used as the confidence weight of the deviation contribution of the current state probability density function relative to the baseline probability density function. The larger the dispersion value, the smaller the decay weight factor and the lower the deviation contribution.

[0016] Preferably, the wavelet decomposition of the real-time rotor position signal and the data extraction of each frequency band real-time signal component using an asymmetric time window, wherein the asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window, including: In the real-time signal components of each frequency band, a sequence of 1000 consecutive sampling points is extracted from the current moment and traced back to the past to form the short-time analysis window; Starting from the historical boundary of the short-time analysis window, continue backtracking to the past, skipping the blank interval area with a width of 500 sampling points; After the blank interval area, extract a sequence of 200 sampling points as the first historical micro-pane, and repeat the operation of skipping 500 sampling points and then extracting a sequence of 200 sampling points three times to form four non-continuously distributed historical micro-panes.

[0017] Preferably, the step of calculating the absolute dominance metric of the current state probability density function relative to the normalized and verified baseline probability density function and the relative dominance metric relative to the recent background probability density function per frequency band includes: On a unified dimensionless feature domain, the cumulative distribution function difference between the current state probability density function and the normalized and verified baseline probability density function is calculated band by band, and the absolute value integral of the difference is used as the original absolute dominance metric. The absolute dominance metric is obtained by weighting the original absolute dominance metric with a decay weighting factor determined by the baseline uncertainty tensor. Calculate the cumulative distribution function difference between the current state probability density function and the recent background probability density function, and use the area of ​​the integral of the absolute value of the difference over a unified dimensionless feature domain as the relative dominance metric.

[0018] Preferably, the step of projecting the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and using the recent background fluctuation metric determined by the historical micropane to adaptively adjust the boundary of the judgment superbody defining the normal working range, and determining that the sensor is faulty when the fault projection coordinates are outside the judgment superbody, includes: In the fault feature vector space, a judgment hypersphere is established with the origin of the coordinate system as the center, and the hypersphere of the judgment hypersphere is used as the boundary of the normal working range. Calculate the Euclidean distance from the fault projection coordinates to the origin; Determine whether the Euclidean distance is greater than the radius of the determination hypersphere. If so, determine that the fault projection coordinates are outside the determination hypersphere and determine that the sensor is faulty. The radius is composed of a preset base radius and an adjustment term that is positively correlated with the recent background fluctuation metric determined by the historical micropane.

[0019] This invention extracts steady-state and transient features of rotor position signals under different operating conditions and frequency bands through multi-source signal fusion and wavelet decomposition, constructing a benchmark probability distribution tensor and a benchmark uncertainty tensor to characterize the normal operating state of the motor. By setting an asymmetric time window that includes a short-time analysis window and discontinuous historical micro-windows, both the current instantaneous state and recent background trends are taken into account, improving the completeness of feature extraction and anti-interference capability. Furthermore, the confidence level of the deviation contribution of the current state from the benchmark probability distribution is modulated using the benchmark uncertainty, and a composite random dominance vector is generated by combining absolute dominance and relative dominance measures, reducing system noise and false alarms caused by operating condition switching. Finally, the composite random dominance vector is projected onto the fault feature space, and the judgment superbody boundary is dynamically adjusted according to the recent background fluctuation measure, realizing accurate identification of weak anomalies and potential faults of the position sensor, and improving the operational safety of the automotive steering system. Attached Figure Description

[0020] Figure 1 A flowchart of the first embodiment; Figure 2 This is a timing waveform diagram showing the rotor position, speed, and torque synchronization of an automotive steering motor. Figure 3 This is a schematic diagram of the time-domain distribution of the asymmetric time windows of the short-term analysis window and the historical micro-window. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the first embodiment, the present invention proposes a method for detecting motor position sensors for steering systems, such as... Figure 1 It includes the following steps: S1: Acquire the real-time rotor position signal, torque signal, and speed signal of the automotive steering system motor; perform wavelet decomposition on the rotor position signal under historical normal operating conditions; perform feature statistical analysis on each frequency band signal according to the corresponding torque signal and speed signal intervals; and generate a reference probability distribution tensor and a reference uncertainty tensor.

[0023] The real-time rotor position signal output by the motor resolver sensor is acquired through a resolver acquisition circuit, and the torque and speed signals are read through the vehicle control communication network. Before performing wavelet decomposition, the periodic rotor angle sequence within the range of 0 to 360 degrees is first subjected to phase expansion processing to obtain a continuous angle sequence, which is then used as the wavelet decomposition object. The rotor position signal under historical normal operating conditions is subjected to 4-level one-dimensional discrete wavelet decomposition to obtain low-frequency approximation coefficients and 4 high-frequency detail coefficients as signals for each frequency band. The amplitude ranges of the historical torque and speed signals are each equally divided into multiple preset intervals to form a two-dimensional operating condition grid. Within each two-dimensional operating condition grid, the corresponding frequency band signal data is extracted and Gaussian kernel density estimation is performed to calculate the probability density function, forming a baseline probability distribution tensor containing the dimensions of torque interval, speed interval, and frequency band number.

[0024] A resampling algorithm is used to resample the signal data of each frequency band within each operating condition grid. The confidence interval width of the corresponding probability density function at a 95% confidence level is calculated, and this confidence interval width is used as the uncertainty index for the corresponding operating condition and frequency band, forming a baseline uncertainty tensor. The confidence interval width is calculated on a unified dimensionless feature domain; therefore, the discrete value in the baseline uncertainty tensor is a dimensionless value. The preset reference discreteness adopts the calibration value under the same dimensionless feature domain. The resampling algorithm is a bootstrap resampling with replacement. For each frequency band signal data within each operating condition grid, resampling is repeated a preset number of times, such as 200 times. The sample size for each resampling is the same as the original sample size, and the probability density function is estimated on a preset discrete grid within a unified dimensionless feature domain. The 2.5% quantile and 97.5% quantile are extracted at each discrete grid point to form a 95% confidence band. The average of the confidence band widths at each discrete grid point is used as the uncertainty index for that operating condition and frequency band. The time-stamped synchronized timing waveforms of the automotive steering motor rotor position, speed, and torque are as follows: Figure 2 As shown.

[0025] In an optional embodiment, acquiring the real-time rotor position signal, torque signal, and speed signal of the vehicle steering system motor includes: A continuous two-phase quadrature analog signal sequence is acquired by a resolver installed at the motor shaft end, and then analyzed into a real-time rotor position signal sequence by a resolver-to-digital converter. Extract speed and torque messages with consistent timestamps from the communication messages of the vehicle control communication network, and reconstruct real-value speed and torque signals according to the communication protocol. Real-time rotor position signals, speed signals, and torque signals are aligned and fused based on the same timestamp to construct an integrated multi-dimensional time-series data sequence.

[0026] For the acquisition of resolver signals, a high-frequency sinusoidal excitation with a frequency of 10kHz and an amplitude of 7Vpp is applied to the resolver coil mounted on the motor shaft via the excitation source of the hardware control board. Simultaneously, the two-phase quadrature analog signal sequences of SIN and COS fed back from the sensors are acquired. Subsequently, a 16-bit resolver-to-digital converter chip, such as the AD2S1210, performs high-frequency sampling and decoding, outputting a digital rotor angle sequence ranging from 0 to 360 degrees in real time with a fixed sampling period of 0.1ms. On the network communication side, the CAN bus receiving node of the vehicle controller continuously identifies data frames with a baud rate of 500kbps. Based on vehicle communication protocols such as J1939, the effective byte payload is separated from the message data field identified by specific IDs, such as 0x0C5 and 0x0C6.

[0027] After extracting the payload, the control program uses the preset scaling factor and offset in the communication protocol matrix to perform numerical reconstruction calculations. For example, the torque message data is multiplied by a precision of 0.1 Nm / bit and an offset of -500 Nm is added; the speed message data is multiplied by 1 rpm / bit and an offset of -10000 rpm is added, reconstructing real-value signals such as 45.5 Nm and 1500 rpm. To address the issue of inconsistent time synchronization between signals from different hardware sources, an independent high-precision real-time clock module inside the microcontroller is used to timestamp all interrupt sampling or message reception actions. In the alignment and fusion process, the original rotor position sequence with a high sampling rate of 10kHz is retained as the data source for subsequent wavelet decomposition and short time window truncation. In addition, with a synchronization window width of 10ms as the reference, the window representative value is extracted from the rotor position sequence with a high sampling rate of 10kHz. The window representative value is then interpolated and aligned with the low-frequency 100Hz speed and torque signal points at the same reference time to construct a multi-dimensional integrated time series data sequence for working condition interval indexing, and then transferred to a dual-port RAM matrix for subsequent calculation and extraction.

[0028] In an optional embodiment, the wavelet decomposition of the rotor position signal under historical normal operating conditions includes: The Daubechies-4 wavelet basis function is used as the core scale analytical model of the filtering network to perform four-level multi-scale orthogonal wavelet decomposition on the rotor position signal under historical normal working conditions. Filter out low-frequency smooth approximation components and extract high-frequency detail components representing signal abrupt changes; The extracted high-frequency detail components of each layer are recombined according to the scale level to form an independent state vector sequence for each frequency band.

[0029] In the digital filtering network, a Daubechies-4 wavelet with tight support and 8 filter tap coefficients is configured as the core basis model. For one-dimensional time-series data of historical rotor positions under normal operating conditions with a time length of 10240 points, a sampling frequency of 10kHz, and a recording duration of approximately 1.024 seconds, a 4-stage cascaded discrete orthogonal decomposition is performed. The raw data is then passed through a first-stage high-pass and low-pass conjugate filter bank, dividing it into first-layer high-frequency detail coefficients with a bandwidth of 2.5kHz to 5kHz. and low-frequency approximation coefficient from 0 to 2.5 kHz For low frequency band Recursively split the bandwidth, successively separating bandwidths from 1.25kHz to 2.5kHz. 625Hz to 1.25kHz and 312.5Hz to 625Hz At the end of the decomposition process, extremely low-frequency approximation components remain. To remove the basic slowly varying baseline caused by the rotation of the motor's mechanical body and the system's own constant DC offset, the processing program will extract the fourth-layer low-frequency coefficients. All scalar values ​​are set to zero. After the low-frequency filtering operation, to The effective frequency bands for the four carrier system glitches, harmonic interference, and sudden edge transitions were subjected to zero-value interpolation and one-dimensional discrete inverse wavelet transform using a single-band reconstruction method. After reconstruction, the coefficients of each downsampled layer were mapped back to the original 10240-point time domain length. Four high-frequency detail subsequences with consistent dimensions and mutual orthogonality were extracted and mapped to independent state vectors. , , and These are concatenated and encapsulated into a 4-row, 10240-column frequency band feature state matrix. Therefore, the historical baseline probability distribution tensor, the baseline uncertainty tensor, and the subsequent real-time frequency band components all adopt the same four high-frequency band dimensions.

[0030] S2, perform wavelet decomposition on the real-time rotor position signal, and extract data for each frequency band's real-time signal component using an asymmetric time window. The asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window. Calculate the current state probability density function within the short-time analysis window and the recent background probability density function obtained by weighted fusion of signals within the historical micro-panes. Based on the benchmark uncertainty tensor, perform confidence modulation on the deviation contribution of the current state probability density function from the benchmark probability density function. Calculate the absolute dominance metric of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance metric relative to the recent background probability density function band by band. Concatenate the metrics of each frequency band to form a composite random dominance vector.

[0031] The real-time rotor position signal was decomposed into real-time signal components for each frequency band using a 4-level wavelet decomposition using db4 wavelets. For each frequency band's real-time signal component, a short-time analysis window was created by slicing the data to divide the current sampling point into 1000 sampling points. A 500-point blank interval was skipped before the short-time analysis window, and 200 sampling points were extracted after this blank interval as the first data segment. This process was repeated three times, skipping 500 sampling points and then extracting 200 sampling points, for a total of four data segments extracted as discontinuous historical micro-panes. Gaussian kernel density estimation was performed on the data within the short-time analysis window to obtain the current state probability density function. The data within the four historical micro-panes were assigned weights that decay exponentially over time, and these weights were normalized so that the sum of the weights for each historical micro-pane was 1. Weighted fusion was then performed to estimate the recent background probability density function. Based on the real-time acquired torque and speed signals, the corresponding two-dimensional operating condition grid is queried. Combined with the current frequency band number, the corresponding reference probability density function and reference uncertainty are extracted. The reference uncertainty is used to calculate a dispersion ratio coefficient according to a preset reference dispersion, and an attenuation weighting factor is calculated. The reference probability density function is then normalized and verified. The attenuation weighting factor is used as the confidence weight for the deviation contribution of the current state probability density function relative to the reference probability density function, ensuring that the deviation contribution of subsequent absolute dominance measures decreases as the reference uncertainty increases. Based on first-order stochastic dominance theory, the cumulative distribution functions of the current state probability density function, the normalized and verified reference probability density function, and the recent background probability density function are calculated. The absolute value of the difference between the current state cumulative distribution function and the normalized and verified reference cumulative distribution function is numerically integrated to obtain the original absolute dominance measure. The attenuation weighting factor is then used to weight the original absolute dominance measure to obtain the absolute dominance measure. The absolute value of the difference between the current state cumulative distribution function and the recent background cumulative distribution function is numerically integrated to obtain the relative dominance measure. The absolute and relative dominance measures of all frequency bands are concatenated to form a one-dimensional composite stochastic dominance vector. The time-domain distribution structure of short-time analysis windows and historical micro-windows with asymmetric time windows is as follows: Figure 3 As shown.

[0032] In an optional embodiment, the confidence modulation of the deviation contribution of the current state probability density function relative to the benchmark probability density function based on the benchmark uncertainty tensor includes: Using the representative values ​​of the torque signal and the representative values ​​of the speed signal corresponding to the short-time analysis window, as well as the current frequency band number, as indices, the corresponding discrete values ​​are extracted from the reference uncertainty tensor as basic parameters to be adjusted. Divide the dispersion value by the preset reference dispersion to obtain the dispersion ratio coefficient. Using the natural constant e as the base, and the negative number of the dispersion ratio coefficient as the exponent, calculate the attenuation weight factor with a value range greater than 0 and less than or equal to 1. The baseline probability density function is normalized and verified, and the decay weight factor is used as the confidence weight of the deviation contribution of the current state probability density function relative to the baseline probability density function. The larger the dispersion value, the smaller the decay weight factor and the lower the deviation contribution.

[0033] In the global two-dimensional operating condition domain, the torque operating range, for example, set to -50Nm to 50Nm, is divided into 20 vertical axis discrete gears at 5Nm intervals. Similarly, the speed operating range, for example, 0 to 3000rpm, is divided into 20 horizontal axis discrete gears in 150rpm increments. When the control cycle starts, assuming the measured average torque is 12.5Nm, falling into gear 13, and the current speed is 1600rpm, falling into gear 11; when processing the current frequency band number... At this time, the main control chip uses the torque gear, speed gear, and current frequency band number together as the address index, for example, forming (13, 11, ... The index coordinates are used to address the prior reference uncertainty tensor in the memory, from which the dispersion values ​​for the corresponding operating conditions and frequency bands are extracted as basic parameters to be tuned. Assume the dispersion value retrieved this time is D = 0.45. To achieve global consistency in parameter scaling, a reference dispersion reference constant, measured and calibrated on a laboratory standard test bench under static steady-state conditions, is introduced. If the factory default value is 0.30, perform a ratio calculation to remove the dimension and obtain the relative deviation K = 0.45 / 0.30 = 1.5.

[0034] Using the exponential weighted adjustment law, with the natural constant e as the base, the expression is calculated. The output yields a decay weighting factor α≈0.223 with values ​​in the half-open interval (0,1). Then, the initial discrete probability values ​​of the baseline probability density function in each statistical histogram interval are retrieved, for example, the local probability peak array P=[0.10,0.40,0.50]. This array is first normalized to verify that the sum of all discrete probability values ​​is 1. Subsequently, the integral of the absolute value of the difference between the current state cumulative distribution function and the baseline cumulative distribution function is calculated to obtain the original absolute dominance metric. If this calculation yields... =0.60, then the absolute dominance measure is obtained by weighting using a decay weighting factor α≈0.223. =0.60×0.223≈0.134. Therefore, when the baseline uncertainty is high, the deviation contribution of the current state from the baseline distribution is reduced, avoiding misjudging normal fluctuations under high uncertainty conditions as faults.

[0035] In an optional embodiment, the wavelet decomposition of the real-time rotor position signal is performed, and data is extracted from each frequency band real-time signal component using an asymmetric time window. The asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window, including: In the real-time signal components of each frequency band, a sequence of 1000 consecutive sampling points is extracted from the current moment and traced back to the past to form the short-time analysis window; Starting from the historical boundary of the short-time analysis window, continue backtracking to the past, skipping the blank interval area with a width of 500 sampling points; After the blank interval area, extract a sequence of 200 sampling points as the first historical micro-pane, and repeat the operation of skipping 500 sampling points and then extracting a sequence of 200 sampling points three times to form four non-continuously distributed historical micro-panes.

[0036] The control system manages high-speed data streams using a ring-shaped buffer storage area and memory access technology. Assuming a fixed sampling frequency of 10kHz and a sampling interval of 0.1ms, once the algorithm triggers an observation update task, the current trigger point is denoted as... The DMA control logic performs a reverse batch copy, completely reading from the latest time. Tracing back to the origins of history The data block. This array of 1000 completely adjacent discrete data points, spanning a continuous absolute duration of 100 ms in the time domain, is assembled into a short-time analysis window for identifying leading-edge transient impacts or instantaneous temperature drift effects. In the process of constructing the reference baseline, to prevent the inertial tail effect caused by anomalous impacts within the main window from penetrating into the reference background, the cursor pointer is positioned relative to the historical boundary point of the main window. Using the origin as the starting point, skip 500 address offsets in the memory buffer, and... to Sampling points covering a 50ms interval are considered blank and discarded. After crossing the blind zone, the first round of intermittent truncation immediately begins: completely capturing the next segment. to The first historical micro-pane, containing 200 data points and lasting 20ms, is then encapsulated; the cursor then skips 500 points across the isolation zone to capture... to Create the second pane within the interval; continue using the same intermittent sampling method until the fourth historical micro-pane is generated, covering... to Through the aforementioned step sampling mechanism, four discontinuous micro-segments distributed 50ms apart on the time axis were extracted, improving the decoupling performance for extracting low-frequency slow drift evolution background features.

[0037] In an optional embodiment, the step of calculating the absolute dominance metric of the current state probability density function relative to the normalized and verified baseline probability density function and the relative dominance metric relative to the recent background probability density function per frequency band includes: On a unified dimensionless feature domain, the cumulative distribution function difference between the current state probability density function and the normalized and verified baseline probability density function is calculated band by band, and the absolute value integral of the difference is used as the original absolute dominance metric. The absolute dominance metric is obtained by weighting the original absolute dominance metric with a decay weighting factor determined by the baseline uncertainty tensor. Calculate the cumulative distribution function difference between the current state probability density function and the recent background probability density function, and use the area of ​​the integral of the absolute value of the difference over a unified dimensionless feature domain as the relative dominance metric.

[0038] For the different sub-band feature channels obtained after wavelet-level separation, firstly, based on the feature value boundaries of historical normal operating condition samples or preset engineering boundaries, the feature variables of each band are uniformly mapped to the dimensionless standard interval [0,1]. Then, the global variable feature domain is divided into a standard grid scale space with a width of, for example, M=100, and the width of each discrete interval is Δx=1 / M. When processing a specific frequency band, such as... In frequency band analysis, the core current probability density function generated based on the short-time analysis window of this channel. By successively accumulating the probability weights of each interval on the spatial coordinates, the probability at each discrete independent variable anchor point is calculated. Cumulative distribution probability function value at ... and For this pair of cumulative step sequences, subtract the corresponding ordinate values ​​for each pair and take the absolute value, then multiply by the infinitesimal element on the horizontal axis using the area integral criterion of Riemann summation. Perform a global summation from k=1 to k=100. The output of this integral area calculation is a scalar in the range of 0 to 1, which is set as the original absolute dominance measure describing the distortion of the real-time waveform or the offset of the base center position. Subsequently, the attenuation weighting factor α, determined by the baseline uncertainty tensor, was used to... By weighting, the final absolute dominance measure is obtained. =α× Since the integral horizontal axis has been uniformly normalized to the [0,1] interval, the absolute value of the cumulative distribution function difference does not exceed 1. Therefore, the area result has a normalized meaning within the range of 0 to 1. In the relative contrast analysis channel, the recent background distribution, which is linearly mixed by four micro-panes from historical extraction backtracking according to set weights, is used to calculate the recent evolution characteristic baseline, i.e., the recent background cumulative distribution curve. Using the same discrete integration method, the current step curve... Discrete set sequence and background sequence Point-to-point differencing is performed on the dimensionless standard interval to obtain a sequence of absolute deviation values. Then, the areas of all small deviations within the dimensionless standard interval are summed. This result represents the area resulting from the change in the current actual operating condition relative to the recent background mean state, and is assigned the value of the relative dominance metric. For example, when the motor's rotary transformer is functioning normally, The integral area metric typically remains below 0.05, but when the filter capacitors within the decoder board degrade, causing a slow noise spike, this metric increases and exceeds 0.6. The final absolute dominance and relative dominance metrics calculated across all decomposed frequency bands are concatenated to obtain a high-dimensional composite parameter vector set driving the fault detection model.

[0039] S3, Project the composite random dominance vector onto the preset fault feature vector space to obtain the fault projection coordinates; Use the recent background fluctuation metric determined by the historical micropane to adaptively adjust the judgment superbody boundary that defines the normal working range; When the fault projection coordinates are located outside the judgment superbody, the sensor is determined to be faulty.

[0040] The composite random dominance vector is first centered using the mean vector of historical normal samples, then multiplied by a feature transformation matrix pre-trained through principal component analysis of historical samples for dimensionality reduction mapping, outputting the fault projection coordinates in low-dimensional space. The initial determination superbody defining the normal working range is a hypersphere with a preset base radius. The mean of the distribution difference between each historical micro-pane or the dispersion of the recent background probability density function is calculated as the background fluctuation coefficient. The preset base radius is added to the product of the background fluctuation coefficient and a proportionality constant to adjust the boundary threshold of the determination superbody. Specifically, the probability density function of each historical micro-pane is estimated and converted into a cumulative distribution function. The area of ​​integration of the absolute value of the difference between the cumulative distribution functions of any two historical micro-panes is calculated on a unified dimensionless feature domain. The arithmetic mean of the area of ​​integration corresponding to each pairwise combination of all historical micro-panes is used as the original background fluctuation measure. The original background fluctuation measure is divided by a preset background fluctuation reference value and limited to a preset upper limit to obtain the recent background fluctuation measure.

[0041] The real-time Euclidean distance from the fault projection coordinates to the origin of the fault feature vector space is calculated. This real-time Euclidean distance is compared with the adjusted threshold of the judgment hypersphere boundary. If the real-time Euclidean distance is greater than the judgment hypersphere boundary threshold, the fault projection coordinates are determined to be outside the judgment hypersphere. At this time, the controller outputs an alarm signal to determine that the position sensor is faulty. In this embodiment, since the judgment hypersphere is a hypersphere centered at the origin, the distance metric is uniformly Euclidean distance. If Mahalanobis distance were used, the corresponding judgment boundary would be a hyperellipsoid determined by the covariance matrix; this embodiment does not use this method.

[0042] In an optional embodiment, the step of projecting the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and adaptively adjusting the boundary of the judgment superbody defining the normal working range using the recent background fluctuation metric determined by the historical micropane, and determining that the sensor is faulty when the fault projection coordinates are outside the judgment superbody, includes: In the fault feature vector space, a judgment hypersphere is established with the origin of the coordinate system as the center, and the hypersphere of the judgment hypersphere is used as the boundary of the normal working range. Calculate the Euclidean distance from the fault projection coordinates to the origin; Determine whether the Euclidean distance is greater than the radius of the determination hypersphere. If so, determine that the fault projection coordinates are outside the determination hypersphere and determine that the sensor is faulty. The radius is composed of a preset base radius and an adjustment term that is positively correlated with the recent background fluctuation metric determined by the historical micropane.

[0043] The diagnostic system is based on linear algebraic transformation. Within a high-order Euclidean feature projection space with dimension d=4, a closed multidimensional hypersphere region is defined with the origin (0,0,0,0) as its centroid, and its hypersphere serves as the boundary of the normal operating range. The system receives the real-time coordinate point data matrix obtained from the output of the preceding model and after dimensionality reduction transformation, and extracts the coordinates of the projection vector points corresponding to the latest observation time from this matrix. Taking a certain calculation as an example, the measured components of the four orthogonal axes of the coordinate system are [0.85, -0.40, 1.30, 0.65]. For this vector array, according to the vector magnitude distance measurement formula in multidimensional space, the components are squared, the sum of squares of each term is calculated, and the square root is taken. The calculated Euclidean distance from the current vector position to the center origin is L≈1.73. This Euclidean distance represents the deviation of the steering motor position detection module's working state from the ideal balance reference point. To provide boundary scaling and fault tolerance self-adjustment functions, the MCU system pre-writes static initial boundary constants representing the sensor's basic limit tolerance into the Flash memory. The default base radius value is 1.85. In the adjustment calculation loop, the background fluctuation metric value calculated from the historical micro-pane is extracted, such as when it is captured... =0.80, and according to the compensation algorithm, the slope proportional gain k=0.5 is input, and the linear operator is applied to map out the adjustment term, that is... =0.5 × 0.80 = 0.40. (This is the base radius.) Add positively correlated adjustment terms Calculate the boundary radius =2.25. Execute the core threshold comparison logic: determine if the current Euclidean distance L (1.73) is less than the dynamic boundary radius. The safety diagnostic architecture determines that the fault projection coordinates are still within the detection super-body at this time, and therefore does not trigger a fault alarm. Only when the subsequently calculated Euclidean distance is greater than the dynamic boundary radius will the fault alarm flag be set to True, and a structural or electrical failure inside the steering motor position sensor will be reported via the high-speed communication bus, triggering the main controller to issue a steering torque auxiliary degradation command to shield the fault's impact.

[0044] In a second embodiment, the present invention also proposes a motor position sensor detection system for steering systems, comprising the following modules: The generation module is used to acquire the real-time rotor position signal, torque signal, and speed signal of the automotive steering system motor; it performs wavelet decomposition on the rotor position signal under historical normal operating conditions, performs feature statistical analysis on the signals of each frequency band according to the division interval of the corresponding torque signal and speed signal, and generates the benchmark probability distribution tensor and the benchmark uncertainty tensor. The calculation module performs wavelet decomposition on the real-time rotor position signal. For each frequency band of real-time signal components, it uses an asymmetric time window to extract data. This asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window. It calculates the current state probability density function within the short-time analysis window and the recent background probability density function obtained by weighted fusion of signals within the historical micro-panes. Based on the benchmark uncertainty tensor, it performs confidence modulation on the deviation contribution of the current state probability density function relative to the benchmark probability density function. It calculates the absolute dominance metric of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance metric relative to the recent background probability density function band by band, and concatenates the metrics of each frequency band to form a composite random dominance vector. The determination module is used to project the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and to adaptively adjust the determination superbody boundary that defines the normal working range using the recent background fluctuation metric determined by the historical micropane. When the fault projection coordinates are located outside the determination superbody, the sensor is determined to be faulty.

[0045] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting the position of a motor in a steering system, characterized in that, include: Acquire real-time rotor position, torque, and speed signals of the automotive steering system motor; Wavelet decomposition is performed on the rotor position signal under historical normal operating conditions. Based on the division intervals of the corresponding torque and speed signals, feature statistical analysis is performed on the signals of each frequency band to generate the reference probability distribution tensor and the reference uncertainty tensor. Wavelet decomposition is performed on the real-time rotor position signal. For each frequency band of real-time signal component, an asymmetric time window is used to extract the data. The asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located in front of the short-time analysis window. The current state probability density function in the short-time analysis window and the recent background probability density function of the weighted fusion of signals in the historical micro-panes are calculated respectively. Confidence modulation is performed on the deviation contribution of the current state probability density function from the benchmark probability density function based on the benchmark uncertainty tensor. The absolute dominance measure of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance measure relative to the recent background probability density function are calculated band by band. The measures of each band are concatenated to form a composite random dominance vector. The composite random dominance vector is projected onto a preset fault feature vector space to obtain fault projection coordinates; the recent background fluctuation metric determined by the historical micropane is used to adaptively adjust the boundary of the judgment superbody that defines the normal working range; when the fault projection coordinates are outside the judgment superbody, the sensor is determined to be faulty.

2. The method according to claim 1, characterized in that, The acquisition of real-time rotor position signal, torque signal, and speed signal of the vehicle steering system motor includes: A continuous two-phase quadrature analog signal sequence is acquired by a resolver installed at the motor shaft end, and then analyzed into a real-time rotor position signal sequence by a resolver-to-digital converter. Extract speed and torque messages with consistent timestamps from the communication messages of the vehicle control communication network, and reconstruct real-value speed and torque signals according to the communication protocol. Real-time rotor position signals, speed signals, and torque signals are aligned and fused based on the same timestamp to construct an integrated multi-dimensional time-series data sequence.

3. The method according to claim 1, characterized in that, The wavelet decomposition of the rotor position signal under historical normal operating conditions includes: The Daubechies-4 wavelet basis function is used as the core scale analytical model of the filtering network to perform four-level multi-scale orthogonal wavelet decomposition on the rotor position signal under historical normal working conditions. Filter out low-frequency smooth approximation components and extract high-frequency detail components representing signal abrupt changes; The extracted high-frequency detail components of each layer are recombined according to the scale level to form an independent state vector sequence for each frequency band.

4. The method according to claim 2, characterized in that, The confidence modulation of the deviation contribution of the current state probability density function from the benchmark probability density function based on the benchmark uncertainty tensor includes: Using the representative values ​​of the torque signal and the representative values ​​of the speed signal corresponding to the short-time analysis window, as well as the current frequency band number, as indices, the corresponding discrete values ​​are extracted from the reference uncertainty tensor as basic parameters to be adjusted. Divide the dispersion value by the preset reference dispersion to obtain the dispersion ratio coefficient. Using the natural constant e as the base, and the negative number of the dispersion ratio coefficient as the exponent, calculate the attenuation weight factor with a value range greater than 0 and less than or equal to 1. The baseline probability density function is normalized and verified, and the decay weight factor is used as the confidence weight of the deviation contribution of the current state probability density function relative to the baseline probability density function. The larger the dispersion value, the smaller the decay weight factor and the lower the deviation contribution.

5. The method according to claim 1, characterized in that, The real-time rotor position signal is decomposed using wavelet decomposition. For each frequency band of the real-time signal component, an asymmetric time window is used to extract data. This asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located before the short-time analysis window, including: In the real-time signal components of each frequency band, a sequence of 1000 consecutive sampling points is extracted from the current moment and traced back to the past to form the short-time analysis window; Starting from the historical boundary of the short-time analysis window, continue backtracking to the past, skipping the blank interval area with a width of 500 sampling points; After the blank interval area, extract a sequence of 200 sampling points as the first historical micro-pane, and repeat the operation of skipping 500 sampling points and then extracting a sequence of 200 sampling points three times to form four non-continuously distributed historical micro-panes.

6. The method according to claim 1 or 5, characterized in that, The calculation of the absolute dominance measure of the current state probability density function relative to the normalized and verified baseline probability density function and the relative dominance measure relative to the recent background probability density function, band by band, includes: On a unified dimensionless feature domain, the cumulative distribution function difference between the current state probability density function and the normalized and verified baseline probability density function is calculated band by band, and the absolute value integral of the difference is used as the original absolute dominance metric. The absolute dominance metric is obtained by weighting the original absolute dominance metric with a decay weighting factor determined by the baseline uncertainty tensor. Calculate the cumulative distribution function difference between the current state probability density function and the recent background probability density function, and use the area of ​​the integral of the absolute value of the difference over a unified dimensionless feature domain as the relative dominance metric.

7. The method according to claim 1, characterized in that, The process involves projecting the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; adaptively adjusting the boundary of the judgment superbody defining the normal operating range using the recent background fluctuation metric determined by the historical micropane; and determining that the sensor is faulty when the fault projection coordinates are outside the judgment superbody. In the fault feature vector space, a judgment hypersphere is established with the origin of the coordinate system as the center, and the hypersphere of the judgment hypersphere is used as the boundary of the normal working range. Calculate the Euclidean distance from the fault projection coordinates to the origin; Determine whether the Euclidean distance is greater than the radius of the determination hypersphere. If so, determine that the fault projection coordinates are outside the determination hypersphere and determine that the sensor is faulty. The radius is composed of a preset base radius and an adjustment term that is positively correlated with the recent background fluctuation metric determined by the historical micropane.

8. A motor position sensor detection system for a steering system, characterized in that, include: The generation module is used to acquire the real-time rotor position signal, torque signal, and speed signal of the automotive steering system motor; Wavelet decomposition is performed on the rotor position signal under historical normal operating conditions. Based on the division intervals of the corresponding torque and speed signals, feature statistical analysis is performed on the signals of each frequency band to generate the reference probability distribution tensor and the reference uncertainty tensor. The calculation module is used to perform wavelet decomposition on the real-time rotor position signal, and to extract data for each frequency band real-time signal component using an asymmetric time window. The asymmetric time window includes a short-time analysis window and multiple discontinuous historical micro-panes located in front of the short-time analysis window. The current state probability density function within the short-time analysis window and the recent background probability density function of the weighted fusion of signals within the historical micro-panes are calculated respectively. Confidence modulation is performed on the deviation contribution of the current state probability density function from the benchmark probability density function based on the benchmark uncertainty tensor. The absolute dominance measure of the current state probability density function relative to the normalized and verified benchmark probability density function and the relative dominance measure relative to the recent background probability density function are calculated band by band. The measures of each band are concatenated to form a composite random dominance vector. The determination module is used to project the composite random dominance vector onto a preset fault feature vector space to obtain fault projection coordinates; and to adaptively adjust the determination superbody boundary that defines the normal working range using the recent background fluctuation metric determined by the historical micropane. When the fault projection coordinates are located outside the determination superbody, the sensor is determined to be faulty.

9. The system according to claim 8, characterized in that, The acquisition of real-time rotor position signal, torque signal, and speed signal of the vehicle steering system motor includes: A continuous two-phase quadrature analog signal sequence is acquired by a resolver installed at the motor shaft end, and then analyzed into a real-time rotor position signal sequence by a resolver-to-digital converter. Extract speed and torque messages with consistent timestamps from the communication messages of the vehicle control communication network, and reconstruct real-value speed and torque signals according to the communication protocol. Real-time rotor position signals, speed signals, and torque signals are aligned and fused based on the same timestamp to construct an integrated multi-dimensional time-series data sequence.

10. The system according to claim 8, characterized in that, The wavelet decomposition of the rotor position signal under historical normal operating conditions includes: The Daubechies-4 wavelet basis function is used as the core scale analytical model of the filtering network to perform four-level multi-scale orthogonal wavelet decomposition on the rotor position signal under historical normal working conditions. Filter out low-frequency smooth approximation components and extract high-frequency detail components representing signal abrupt changes; The extracted high-frequency detail components of each layer are recombined according to the scale level to form an independent state vector sequence for each frequency band.

Images