EMB system-oriented high-frequency electric signal feature classification method and system

Abstract

The invention relates to the technical field of automobile electronic braking, discloses a high-frequency electric signal feature classification method and system oriented to an EMB system, and aims to improve the operation safety and control accuracy of the EMB system. According to the method, high-frequency electric signals generated during operation of the EMB system are collected through a sensor, continuous electric signal flow is formed and then preprocessed in real time, noise is restrained through a digital filter, the signals are smoothed, and purified electric signals are obtained. And then quality evaluation is carried out on the purified electric signal, a signal-to-noise ratio and a stability coefficient are calculated, an evaluation result is generated, and when the result meets a preset standard, a multi-scale feature extraction process is started. In the process, time domain analysis, frequency domain transformation and statistical feature calculation are carried out on the electric signal to obtain a feature vector, the feature vector is input into a pre-trained neural network classifier for pattern recognition, and a classification label is output. And updating the weight of the neural network classifier on line according to the performance index of the classification label.

Description

Technical Field

[0001] This invention relates to the field of automotive electronic braking technology, specifically to a method and system for classifying high-frequency electrical signal characteristics for EMB systems. Background Technology

[0002] Traditional braking systems, as the foundation of automotive braking technology, have long played a crucial role in ensuring driving safety. Their working principle is based on hydraulic transmission. When the driver depresses the brake pedal, brake fluid in the master cylinder, under pressure, is transmitted through the brake lines to the wheel cylinders of each wheel, pushing the brake pads or calipers and causing them to rub against the rotating brake drums or discs. This converts the vehicle's kinetic energy into heat energy, achieving deceleration or stopping. This braking method relies on the coordinated work of mechanical and hydraulic components, resulting in a relatively complex structure, mainly including the brake pedal, master cylinder, wheel cylinders, brake lines, and brake drums or discs.

[0003] However, with the rapid development of the automotive industry and the increasing demands for driving safety, comfort, and intelligence, the limitations of traditional braking systems are becoming increasingly apparent. Regarding response speed, because braking signals are transmitted through hydraulic fluid, the flow rate of the fluid and the resistance in the lines cause a delay in braking response. When a vehicle encounters a sudden situation requiring emergency braking at high speed, this delay can significantly increase the braking distance, seriously threatening driving safety. According to relevant research, in emergency braking situations, the response time of traditional braking systems can reach hundreds of milliseconds. This brief time difference, at high speeds, can cause the vehicle to travel several meters or even further, greatly increasing the risk of an accident.

[0004] In terms of control precision, traditional braking systems primarily rely on the driver's pressure on the brake pedal to indirectly control braking force, making it difficult to precisely distribute braking force based on the vehicle's real-time driving conditions (such as vehicle speed, road conditions, and vehicle load). Under different road conditions, such as dry, wet, or icy surfaces, the required braking force varies. Traditional braking systems often fail to detect these changes promptly and accurately and make corresponding adjustments, easily leading to wheel lock-up or insufficient braking force, affecting vehicle handling stability and braking performance.

[0005] Traditional braking systems also have high maintenance costs. Their hydraulic systems contain numerous seals, lines, and moving parts, which are susceptible to wear, corrosion, and aging over long-term use, requiring regular inspection, replacement, and maintenance. Brake fluid needs to be changed periodically to ensure stable performance; brake lines may experience leaks or blockages, requiring timely repair or replacement; friction components such as brake pads and discs also have a limited lifespan and must be replaced after a certain degree of wear. These maintenance tasks not only consume time and manpower but also increase the operating costs for vehicle owners.

[0006] Traditional braking systems are proving inadequate in the face of the intelligent and electric development trends in modern automobiles. In electric vehicles, traditional braking systems cannot effectively coordinate with energy recovery systems, leading to energy waste and impacting vehicle range. Furthermore, with the rise of autonomous driving technology, traditional braking systems struggle to meet the high precision, rapid response, and intelligent control requirements of autonomous driving, becoming a major obstacle to its development. EMB (Electromechanical Braking) is a technology that breaks away from traditional hydraulic transmission structures, directly driving the brake actuator with an electric motor to achieve vehicle braking. Its core is the conversion of electrical energy into braking mechanical energy, eliminating the need for hydraulic components such as the master cylinder, brake lines, and brake fluid. This avoids the response delay issues associated with hydraulic transmission, improves braking force control precision, and is compatible with the energy recovery systems and intelligent control requirements of autonomous driving in electric vehicles. It represents a key direction for overcoming the limitations of traditional braking systems. This invention collects high-frequency electrical signals from an EMB system during operation using sensors. After real-time preprocessing, quality assessment, and multi-scale feature extraction, a neural network classifier is used to achieve electrical signal pattern recognition. The classifier weights are updated online based on the performance indicators of the classification labels. This allows for the accurate capture of electrical signal features corresponding to the operating status of the EMB system, providing reliable technical support for monitoring the operating status and identifying potential faults in the EMB system. Subsequent inventions based on this technology can further enrich the dimensions of EMB system electrical signal analysis, optimize intelligent classification and decision-making mechanisms, and help improve the operational safety, stability, and ease of maintenance of the EMB system, thus ensuring the efficient application of EMB technology in the automotive braking field. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for classifying high-frequency electrical signal characteristics for EMB systems, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a high-frequency electrical signal feature classification method for EMB systems, the method comprising: The high-frequency electrical signals generated during the operation of the EMB system are collected by sensors to form a continuous electrical signal stream; the electrical signal stream is preprocessed in real time, including noise suppression and signal smoothing using digital filters, to obtain a purified electrical signal; The purified electrical signal is subjected to quality assessment, the signal-to-noise ratio and stability coefficient of the signal are calculated, and a quality assessment result is generated; when the quality assessment result meets the preset standard, the multi-scale feature extraction process is initiated. The multi-scale feature extraction process includes: performing time-domain analysis, frequency-domain transformation, and statistical feature calculation on the purified electrical signal to obtain a set of feature vectors; inputting the feature vectors into a pre-trained neural network classifier for pattern recognition and outputting classification labels; The weights of the neural network classifier are updated online based on the performance metrics of the classification labels.

[0009] Preferably, the specific steps for quality assessment of the purified electrical signal include: extracting the signal amplitude sequence and frequency trajectory from the purified electrical signal; calculating the fluctuation variance of the signal amplitude sequence and the offset of the frequency trajectory; comparing the fluctuation variance with the amplitude tolerance value and comparing the offset with the frequency tolerance value; when the fluctuation variance exceeds the amplitude tolerance value or the offset exceeds the frequency tolerance value, the quality assessment result is marked as unqualified; otherwise, it is marked as qualified.

[0010] Preferably, the specific steps for performing time-domain analysis on the purified electrical signal include: dividing the purified electrical signal into multiple time windows, calculating the signal mean, variance, peak factor, and zero-crossing rate for each time window to obtain a time-domain feature set; and the specific steps for performing frequency-domain transformation on the purified electrical signal include: applying short-time Fourier transform to each time window to extract the main frequency band energy and spectral entropy of the power spectral density to obtain a frequency-domain feature set.

[0011] Preferably, the specific steps for calculating the statistical features of the purified electrical signal include: calculating the skewness, kurtosis, and autocorrelation function of the purified electrical signal within a sliding window to obtain a statistical feature set; and combining the time-domain feature set, frequency-domain feature set, and statistical feature set into the feature vector.

[0012] Preferably, the specific steps for inputting the feature vector into a pre-trained neural network classifier for pattern recognition include: using a multilayer perceptron network to perform forward propagation on the feature vector and calculating the output probability of each category; selecting the category with the highest output probability as the classification label; and simultaneously calculating the classification confidence score, triggering a reclassification process when the classification confidence score is lower than a threshold.

[0013] Preferably, the specific steps for online updating of the weights of the neural network classifier include: calculating the gradient of the loss function based on the error between the classification label and the true label; adjusting the hidden layer weights and bias terms of the neural network classifier using the stochastic gradient descent algorithm; and dynamically adapting the update step size based on historical errors.

[0014] Preferably, the specific steps for calculating the zero-crossing rate for each time window include: For a discrete signal sequence within a time window, calculate the number of sign changes between adjacent sampling points; Divide the total number of sign changes by the total number of sampling points within the time window, and then normalize to obtain the standardized zero-crossing rate; The zero-crossing rate, together with the signal mean, variance, and peak factor, constitutes the time-domain feature set.

[0015] Preferably, the specific steps for calculating the autocorrelation function include: Within the sliding window, calculate the correlation coefficient between the signal and itself at different time delays; The position of the first maximum point of the autocorrelation function is extracted as a periodicity feature of the signal, and the decay rate of the autocorrelation function near the zero delay point is calculated as a stability feature of the signal. The skewness, kurtosis, location of the first maximum point, and decay rate together constitute the statistical feature set.

[0016] Preferably, after obtaining the time-domain feature set and the frequency-domain feature set, nonlinear dynamic feature extraction is further performed on the purified electrical signal. Specific steps include: Reconstruct the phase space of the electrical signal, and calculate the time delay and embedding dimension; Based on the reconstructed phase space, the maximum Lyapunov exponent and correlation dimension are calculated to characterize the chaotic properties of the signal. The maximum Lyapunov exponent and the correlation dimension are added to the feature vector as nonlinear features.

[0017] Preferably, the present invention also includes a high-frequency electrical signal feature classification system for EMB systems, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described high-frequency electrical signal feature classification method for EMB systems.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires high-frequency electrical signals generated by the EMB system during operation using sensors, forming a continuous electrical signal stream. This method can accurately and comprehensively reflect the operating status of the EMB system, recording every subtle change and characteristic completely, providing the most original and reliable basis for subsequent signal processing and analysis. Compared with traditional signal acquisition methods, the continuous electrical signal stream avoids data interruptions and loss, making the monitoring of the system's operating status more accurate and timely. During vehicle braking, the electrical signal stream can capture key information such as the motor's operating status and changes in braking force in real time, providing a solid data foundation for subsequent signal processing and analysis.

[0019] The acquired electrical signal stream undergoes real-time preprocessing, employing digital filters for noise suppression and signal smoothing. Digital filters accurately identify and remove noise components from the signal, effectively suppressing both electromagnetic interference from the external environment and electronic noise generated within the system. Signal smoothing makes signal fluctuations more stable, reducing glitches and abrupt changes and improving overall signal quality. The purified electrical signal exhibits more distinct characteristics, facilitating subsequent analysis and processing, laying a solid foundation for accurate signal feature extraction and pattern recognition. In practical applications, the purified electrical signal more accurately reflects the true operating status of the EMB system, reducing misjudgments and erroneous analyses caused by noise interference.

[0020] The purified electrical signal undergoes a quality assessment by calculating its signal-to-noise ratio (SNR) and stability coefficient, generating a comprehensive and objective quality evaluation result. The SNR reflects the ratio of effective information to noise in the signal; a higher SNR indicates better signal quality and more prominent effective information. The stability coefficient measures the signal's stability over time; a higher stability coefficient indicates less signal fluctuation and greater stability and reliability. By calculating these two key indicators, it is possible to accurately determine whether the signal quality meets preset standards. Only when the signal quality assessment result meets the preset standards is the subsequent multi-scale feature extraction process initiated. This ensures that the signals input to the classifier are of high quality, significantly improving the accuracy and reliability of classification. In complex real-world application scenarios, this scientific quality assessment mechanism can effectively filter out high-quality signals, avoiding classification errors caused by low-quality signals and improving the overall system performance and stability.

[0021] The multi-scale feature extraction process performs time-domain analysis, frequency-domain transformation, and statistical feature calculation on the purified electrical signal. Time-domain analysis reveals the signal's variation patterns over time, such as amplitude, period, rising edge, and falling edge characteristics. Frequency-domain transformation converts the signal from the time domain to the frequency domain, analyzing its frequency composition and energy distribution to uncover its characteristics across different frequency bands. Statistical feature calculation calculates the signal's mean, variance, and peak power factor from a statistical perspective, further enriching the signal's feature description. Through this series of analyses and calculations, a set of feature vectors containing rich information is obtained. These feature vectors comprehensively reflect the various characteristics of the electrical signal, providing ample data support for subsequent pattern recognition. In practical applications, multi-scale feature extraction can deeply explore the value of signals from different perspectives, enabling classifiers to more accurately identify signal types and characteristics, and improving the system's ability to process complex signals. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of the high-frequency electrical signal feature classification method for EMB systems described in this invention. Figure 2 A flowchart for assessing the quality of the purified electrical signal; Figure 3 This is a flowchart for pattern recognition and classification label determination of feature vectors. Detailed Implementation

[0023] The technical solutions of 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.

[0024] Please see Figure 1This invention provides a method and system for high-frequency electrical signal feature classification in EMB systems. The method includes using various high-precision sensor devices to collect high-frequency electrical signals generated by the EMB system during actual operation. After collection, these high-frequency electrical signals form a continuous and uninterrupted electrical signal stream. Subsequently, this continuous electrical signal stream enters a real-time preprocessing stage. In this stage, digital filter technology is used to effectively suppress noise components in the signal and smooth signal fluctuations, thereby obtaining a high-quality electrical signal after purification. A comprehensive quality assessment is then performed on the purified electrical signal. During the quality assessment, the signal-to-noise ratio and stability coefficient of the electrical signal are accurately calculated, and a detailed quality assessment result report is generated. When the generated quality assessment result meets the pre-set standard requirements, a multi-scale feature extraction process is triggered and started. In the multi-scale feature extraction process, the purified electrical signal is first analyzed in depth from multiple dimensions. This process includes detailed time-domain analysis of the electrical signal; simultaneous transformation to the frequency domain using appropriate mathematical methods; and statistical feature calculations for various attributes of the signal. Through this series of operations, a set of feature vectors that fully characterize the original electrical signal is obtained. These feature vectors are then input into a neural network classifier pre-trained on a large amount of data for complex pattern recognition tasks. After processing, the neural network classifier outputs corresponding classification labels. Finally, based on the performance indicators of the output classification labels, a specific algorithm is used to update the weight parameters within the neural network classifier online, thereby continuously improving the classifier's performance and accuracy.

[0025] Example 1: See Figure 2In the quality assessment stage of a high-frequency electrical signal feature classification method for EMB systems, the signal amplitude sequence and frequency trajectory are extracted from the purified electrical signal. The variance of the signal amplitude sequence is obtained by calculating the dispersion of amplitude changes at consecutive sampling points in the sequence, and the offset of the frequency trajectory is calculated based on the cumulative deviation of the instantaneous frequency of the signal. The variance of the variance is compared with a preset amplitude tolerance value, and the offset of the frequency trajectory is compared with a frequency tolerance value. If the variance of the variance exceeds the amplitude tolerance value or the offset exceeds the frequency tolerance value, the quality assessment result is marked as unqualified; otherwise, it is marked as qualified. After the quality assessment is qualified, time-domain analysis and frequency-domain transformation are performed. The time-domain analysis divides the purified electrical signal into multiple time windows, and calculates the signal mean, variance, peak factor, and zero-crossing rate for each time window. The signal mean represents the arithmetic mean of the sampling points within the window, the variance reflects the dispersion of the signal energy, the peak factor characterizes the ratio of the signal peak value to the root mean square value, and the zero-crossing rate is obtained by statistically analyzing the frequency at which the signal crosses zero. Frequency domain transformation applies a short-time Fourier transform to each time window, converting the time-domain signal into a frequency-domain representation and extracting the dominant frequency band energy and spectral entropy of the power spectral density. The dominant frequency band energy focuses on the frequency range where the signal energy is concentrated, while the spectral entropy quantifies the degree of disorder in the frequency distribution. The time-domain feature set and the frequency-domain feature set together constitute the fundamental components of multi-scale feature extraction.

[0026] In time-domain analysis, a fixed-length sliding window is used for time window segmentation, with the window size adaptively adjusted according to the signal sampling rate. The signal mean is calculated by summing all samples within the window and dividing by the number of samples; variance is calculated based on the sum of squares of the differences between the mean and each sample. The peak factor is determined by the ratio of the maximum absolute value within the window to the root mean square value; zero-crossing rate calculation involves detecting the number of sign changes between adjacent samples and normalizing it to the window length. In frequency-domain transformation, the short-time Fourier transform parameters include the window function and overlap rate. Power spectral density is calculated using the squared modulus of the Fourier coefficients; the dominant frequency band energy is selected from the integral value of the band with the highest energy proportion; and the spectral entropy is derived from the information entropy formula of the power spectral probability distribution. Amplitude and frequency tolerance values ​​for quality assessment are calibrated based on historical EMB system operating data to ensure that the threshold settings conform to the actual application scenario. The signal amplitude sequence is extracted using an envelope detection algorithm, and the frequency trajectory is estimated using Hilbert transform to determine the instantaneous frequency. Fluctuation variance is calculated using the squared standard deviation method, and the offset is based on the mean square error between the frequency trajectory and the reference frequency. The data structures of the time-domain feature set and the frequency-domain feature set are multi-dimensional arrays, which facilitates the subsequent integration of feature vectors.

[0027] During time-domain analysis, the time window length for signal segmentation is set to milliseconds to accommodate the transient characteristics of high-frequency signals. Time-domain feature calculations for each window are performed in parallel to improve processing efficiency. The short-time Fourier transform in the frequency domain uses a Hanning window to reduce spectral leakage, and the overlap rate is set to 50% to ensure frequency domain continuity. Power spectral density analysis focuses on the typical fault frequency band of the EMB system, and the spectral entropy calculation uses a logarithmic function to handle probability values. Unqualified markers in the quality assessment results trigger a signal re-acquisition process to avoid low-quality data affecting classification accuracy. Qualified markers activate the feature extraction module, and the generated time-domain and frequency-domain feature sets are temporarily stored in a buffer. The dimension of the feature vector is proportional to the number of time windows to ensure comprehensive capture of signal dynamics. The entire process is implemented on an embedded processor to meet real-time requirements. Signal amplitude sequence extraction in the quality assessment stage uses an envelope detection algorithm. The envelope detection algorithm constructs an analytic signal through Hilbert transform, and the magnitude of the analytic signal constitutes the signal amplitude sequence. Frequency trajectory extraction is based on instantaneous phase differential calculation, and the instantaneous phase is obtained from the phase angle of the analytic signal. The variance fluctuation is calculated using the squared standard deviation within a sliding window, with the amplitude tolerance set based on the signal amplitude fluctuation range under normal EMB system operating conditions. The frequency trajectory offset is calculated using the cumulative absolute difference method, with the frequency tolerance referencing the allowable deviation range of the system's rated frequency. When a quality assessment result is marked as unqualified, the system automatically discards the current data segment and starts a new acquisition cycle. A qualified result allows data to enter the feature extraction pipeline, and the time-domain analysis module initializes the time window segmentation parameters.

[0028] The time window segmentation employs a fixed-duration strategy, with the window length matching the period of the highest frequency component of the signal. The signal mean within each time window is calculated using the arithmetic mean formula, while the variance is calculated using an unbiased estimator. Peak factor calculation requires first obtaining the absolute maximum value of the signal, then calculating the root mean square value of the sampling points within the window. Zero-crossing rate detection compares the sign difference between adjacent sampling points, counts the number of sign changes, and normalizes by dividing by the window length. The time-domain feature set forms a four-dimensional vector sequence, with each time window corresponding to a set of feature values. The frequency domain transformation module receives the same window data; the Hanning window is selected as the window function for the short-time Fourier transform, and the transform length is set to twice the window size. Power spectral density calculation involves taking the square of the Fourier coefficients modulo and normalizing them. Main frequency band energy extraction requires a preset frequency band range, determined based on the characteristic frequencies of the EMB system. Spectral entropy calculation first converts the power spectral density into a probability distribution, then sums it using the information entropy formula. The quality assessment module employs a dual-threshold comparison mechanism, with the fluctuation variance threshold and offset threshold set independently. The sampling interval of the signal amplitude sequence remains consistent with the original signal, and the sampling points of the frequency trajectory are aligned through resampling. Fluctuation variance calculation uses a moving window variance algorithm to adapt to the non-stationary characteristics of the signal. Offset calculation employs the integral absolute error method to reflect the overall deviation of the frequency trajectory. Peak factor calculation in the time domain avoids the influence of the signal's DC component by first performing detrending processing. Zero-crossing rate calculation uses a hysteresis threshold to prevent false zero-crossing counts for quantization noise. The short-time Fourier transform in the frequency domain is accelerated using the fast Fourier transform algorithm, and power spectral density smoothing uses a moving average filter. The integral interval for calculating the main frequency band energy is dynamically adjusted to follow changes in the signal's fundamental frequency. Spectral entropy calculation avoids zero-probability values ​​by adding a small regularization term.

[0029] The overlap rate setting for time window segmentation considers the requirement for feature continuity, with adjacent windows overlapping by 30%. Signal mean calculation uses a recursive formula to reduce computational load, while variance calculation uses the Welford online algorithm. Peak factor calculation combines peak detection and RMS calculation, and zero-crossing rate calculation adds an amplitude threshold to filter out minor fluctuations. The window function length of the short-time Fourier transform affects frequency resolution; the optimal parameters are determined experimentally. Power spectral density normalization ensures energy conservation, and the percentage of energy in the main frequency band is calculated relative to the total energy. Spectral entropy is normalized to between zero and one for easier comparison of different signals. The tolerance update mechanism of the quality assessment module supports online learning and dynamically adjusts according to the system's operating status. Envelope detection of the signal amplitude sequence uses a low-pass filtering alternative after full-wave rectification, and frequency trajectory extraction uses phase-locked loop technology. Fluctuation variance calculation adds outlier removal logic, and offset calculation introduces a weighting factor to emphasize recent data.

[0030] The temporal alignment of the time-domain and frequency-domain feature sets is achieved through timestamp synchronization, and feature vectors are assembled according to the window index order. The decision logic for quality assessment results adopts a hard-decision rule; only when both conditions are met is the result marked as qualified. The calculation of the fluctuation variance of the signal amplitude sequence ignores the DC component, and the calculation of the frequency trajectory offset uses the root mean square error standard. Mirror expansion is used for boundary processing of time window segmentation to avoid edge effects. Fixed-point arithmetic is used to optimize the speed of signal mean calculation, and piecewise parallel computation is used for variance calculation. Peak factor calculation performs amplitude limiting processing on saturated signals, and a jitter reduction algorithm is added to zero-crossing rate calculation. Logarithmic scaling is used to enhance the dynamic range of the short-time Fourier transform spectrum, and the periodogram method is used for power spectral density calculation. Main frequency band energy analysis supports simultaneous extraction of multiple frequency bands, and radix-2 logarithm calculation ensures numerical stability. A double-buffering mechanism is used between the quality assessment module and the feature extraction module to achieve pipelined parallel processing.

[0031] Example 2: A High-Frequency Electrical Signal Feature Classification Method for EMB Systems. In the statistical feature calculation stage, the skewness, kurtosis, and autocorrelation function of the purified electrical signal within a sliding window are calculated. Skewness characterizes the degree of asymmetry in the signal amplitude distribution, and its calculation relies on the normalization operation of the third central moment and the cube of the standard deviation of the signal sample values. Kurtosis measures the difference between the signal distribution shape and the normal distribution, and is calculated by subtracting three from the ratio of the fourth central moment to the square of the variance. The autocorrelation function reveals the self-similarity pattern of the signal under different time delays, and its calculation involves cross-correlation operations between the signal sequence and its time-shifted replicas. The location of the first maximum point of the autocorrelation function reflects the period length of the main periodic components of the signal, and the decay rate of the autocorrelation function near the zero-delay point characterizes the signal's ability to maintain correlation. The skewness calculation results, kurtosis calculation results, the data of the first maximum point location, and the decay rate data are combined to form a statistical feature set. This statistical feature set, along with the time-domain feature set obtained from time-domain analysis and the frequency-domain feature set obtained from frequency-domain transformation, undergoes a vector concatenation operation to form the complete feature vector fed into the neural network classifier.

[0032] The construction of the sliding window is completely consistent with the time window used in the previous time-domain analysis in terms of length and sliding step size. The specific calculation process of skewness begins by removing the mean from all signal sample values ​​within the sliding window, then calculating the third moment of these centered sample values, and finally dividing the third moment by the cube of the sample standard deviation to obtain the dimensionless skewness coefficient. The calculation path of kurtosis is similar to that of skewness. It is also based on the centered signal sample values, calculating their fourth central moment, then dividing the fourth central moment by the square of the variance, and subtracting the number three from the calculation result to facilitate comparison with the kurtosis value of the normal distribution. The calculation of the autocorrelation function adopts the direct sliding product method. For each given time delay, the corresponding points of the original signal segment and its delayed signal segment are multiplied and summed. The sum is then normalized by dividing the signal variance and the window length adjustment factor to obtain the autocorrelation function value sequence under different delays. Extracting the first maximum value from the calculated autocorrelation function value sequence requires a search process. The search range starts after the point with zero time delay, excluding the maximum value at zero delay. The search proceeds forward along the delay axis to identify the point where the autocorrelation function value first reaches a local maximum. The delay time corresponding to this local maximum is recorded as a periodic feature. The calculation of the decay rate focuses on the shape of the autocorrelation function in the region near the zero-delay point. A segment of the function curve from the zero-delay point to the first zero-crossing point is selected. After a logarithmic transformation of this segment, linear regression is applied, and the absolute value of the slope of the fitted line is used as the quantitative indicator of the decay rate. The data within the statistical feature set are arranged in a fixed order according to skewness, kurtosis, the location of the first maximum value, and decay rate. The concatenation with the external time-domain and frequency-domain feature sets follows a predefined dimensional order to ensure consistent feature vector structure in each generation.

[0033] The sliding window mechanism uses a fixed-step movement method. The number of sampling points the window moves forward each time is controlled by the overlap rate parameter, and the overlap rate setting needs to strike a balance between feature update frequency and computational load. In skewness calculation, the treatment of the third moment considers sample bias correction, and a Bessel correction factor is used to improve the unbiasedness of the estimation under small sample windows. In kurtosis calculation, subtracting the number three makes the kurtosis baseline of the normal distribution zero. A positive kurtosis value indicates that the distribution is more peaked and heavy-tailed than the normal distribution, while a negative kurtosis value indicates that the distribution is flatter. Although the direct calculation method of the autocorrelation function has higher computational complexity, it is superior to the indirect method based on Fourier transform in that it avoids estimation errors caused by spectral leakage, making it particularly suitable for short sequence analysis. The autocorrelation function value sequence needs to be smoothed and filtered before feature extraction. A moving average filter is used to suppress high-frequency fluctuations caused by noise, facilitating accurate location of the first maximum point. In decay rate calculation, the logarithmic transformation converts the exponential decay trend into a linear trend, and the linear regression uses the least squares method to solve for the slope. The goodness of fit of the regression line serves as an implicit indicator of the reliability of the decay feature. After the statistical feature set is generated, it needs to be standardized using the same scaling factor as the spatiotemporal features. Typically, the Z-score standardization method is used to transform each feature dimension into a distribution with a mean of zero and a variance of one. The final dimension of the feature vector is the sum of the number of time-domain features, the number of frequency-domain features, and the number of statistical features, multiplied by the number of time windows, forming a high-dimensional feature vector for pattern recognition.

[0034] The magnitude and sign of the skewness coefficient convey important information about the signal amplitude distribution. A significantly positive skewness greater than zero indicates the presence of large-amplitude pulses in the signal, with a long tail on the right side of the distribution; a significantly negative skewness less than zero suggests a leftward tilt in the distribution. The kurtosis coefficient is highly sensitive to the signal's impulse characteristics. A high kurtosis value indicates the presence of pulse components sharper than Gaussian noise, often associated with arc discharge or insulation damage in EMB system fault diagnosis. The first maximum point of the autocorrelation function directly corresponds to the period of the strongest periodic component in the signal, and its reciprocal is the characteristic frequency of the pulse component. The decay rate is inversely proportional to the signal bandwidth; faster decay indicates a wider frequency range and a faster loss of correlation. The implementation of the statistical characteristic calculation module needs to consider computational efficiency. Skewness and kurtosis calculations can reuse intermediate results from variance calculations. The calculation of the autocorrelation function can be accelerated using the Fast Fourier Transform (FFT). By padding the signal to zero to twice its length, calculating its power spectral density, and then performing an inverse Fourier transform to obtain the cyclic autocorrelation function, the effective portion is truncated to obtain a linear autocorrelation function estimate. The algorithm for detecting the first maximum point needs to set a minimum delay constraint to avoid misclassifying small fluctuations near the zero-delay point as periodic peaks. The choice of the linear fitting interval in the decay rate calculation affects the stability of the results; an interval that is too small will be highly susceptible to noise interference, while an interval that is too large will introduce nonlinear errors.

[0035] The fusion of statistical feature sets and spatiotemporal feature sets enables multi-perspective representation of signals at the feature level. Skewness and kurtosis supplement the temporal moment features from the perspective of amplitude distribution morphology, while autocorrelation function features enhance the frequency domain spectral features from the perspective of temporal correlation. The high-dimensional representation formed after concatenating feature vectors requires dimensionality reduction processing. Principal component analysis or linear discriminant analysis are optional solutions, but regularization methods are more commonly used in online learning frameworks to prevent overfitting. The entire statistical feature calculation process is embedded in the real-time signal processing chain, and the computational latency needs to be controlled within the system's allowable range. Real-time constraints are met through algorithm optimization and parallel computing. The computational stability of statistical features is ensured through outlier handling mechanisms, such as truncating extreme skewness or kurtosis values ​​to prevent individual outliers from having an excessive impact on the overall features. The output of the statistical feature calculation module is strictly synchronized in time with the outputs of the upstream time-domain analysis module and frequency-domain transformation module, with each time window corresponding to a complete set of feature values. The computational accuracy of statistical features is affected by the window length and signal stability. In non-stationary signal segments, the time-varying nature of statistical features can itself become an important state indicator. The order of the statistical feature set and the spatiotemporal feature set in the feature vector needs to be completely matched with the node connection order of the input layer of the neural network classifier. This structure is fixed during the inference phase after it is determined during the model training phase.

[0036] Example 3: See Figure 3This high-frequency electrical signal feature classification method for EMB systems uses a pre-trained neural network classifier for pattern recognition in its pattern recognition stage. The classifier employs a multilayer perceptron (MLP) network structure to perform nonlinear transformations on the feature vectors. The MLP network consists of an input layer, hidden layers, and an output layer. The number of nodes in the input layer strictly corresponds to the dimension of the feature vector. The hidden layer uses a combination of linear weighted summation and the ReLU activation function, while the output layer applies a linear transformation followed by the Softmax activation function. The forward propagation process begins at the input layer. The feature vector is multiplied by the weight matrix from the input layer to the hidden layer, and then the hidden layer bias vector is added. The result is then processed by the ReLU activation function to generate the hidden layer output value. The hidden layer output value is further multiplied by the weight matrix from the hidden layer to the output layer and then the output layer bias vector is added. The resulting linear value is then converted into a probability distribution for each category using the Softmax function. The final classification label is determined by the category with the highest output probability. The classification confidence is the highest probability value compared to a preset threshold. If the classification confidence falls below the threshold, a reclassification process is triggered, which may involve secondary processing of the same feature vector or require re-acquiring signal data. The mathematical expression of forward propagation involves a multi-level combination of weight matrix multiplication, bias vector addition, and activation function transformations. The ReLU activation function sets negative values ​​to zero while keeping positive values ​​unchanged. The Softmax function transforms any real-valued vector into a probability distribution through exponential operations and normalization. Output probability calculation follows the axiom of probability: the sum of all class probabilities is always one, and the maximum probability value reflects the model's certainty about the current classification result. Threshold setting is based on the classification performance curve on the validation set, balancing the risk of misclassification with the cost of reclassification.

[0037] During the training phase of a multilayer perceptron network, labeled historical data is used, and network parameters are optimized through backpropagation. During the inference phase, the network weights are fixed, and only forward propagation is performed. Feature vectors are standardized before input, using the same mean and standard deviation as the training data. The number of hidden layer nodes is determined based on feature complexity and the number of classification categories, typically a power of two to optimize computational efficiency. The weight matrix is ​​initialized using the He initialization method to accommodate the characteristics of the ReLU activation function. The bias vector is initialized to zero and gradually adjusted during network training.

[0038] Forward propagation computation can be represented as:

[0039]

[0040] in: This represents the input feature vector. This represents the weight matrix from the input layer to the hidden layer. This represents the hidden layer bias vector. This represents the hidden layer output vector. This represents the weight matrix from the hidden layer to the output layer. This represents the output layer bias vector. ReLU represents the output probability vector. The ReLU function is defined as follows: , To find the maximum value, the Softmax function has the following... The output is ,in It is the first linear transformation of the output layer. One value, It represents the total number of categories.

[0041] The dimensionality of the hidden layer output vector determines the network's representational power; too low a dimensionality may lead to underfitting, while too high a dimensionality may cause overfitting. Each element of the output probability vector corresponds to a class membership degree, and the maximum value index is the predicted class. The classification confidence threshold is typically set between 0.7 and 0.9; a threshold that is too low makes the reclassification mechanism overly sensitive, while a threshold that is too high reduces its protective effect. The reclassification process can be configured with various strategies, including using a backup classifier, increasing feature dimensions, or requesting manual intervention.

[0042] The inference speed of a multilayer perceptron network impacts system real-time performance. Optimization is achieved through parallel computation of matrix multiplication and vectorization of activation functions. Changes in feature vector dimensions require retraining the network or adjusting the input layer structure to maintain a consistent number of nodes in the input layer with the feature dimension. The number of hidden layers can be expanded to a multilayer deep network, increasing non-linear expressive power but also increasing computational complexity and the risk of overfitting. The number of output layer nodes is strictly equal to the number of classes, with each node corresponding to the probability output of one class. Exponential operations in probability calculations may cause numerical overflow; a technique of subtracting the maximum value is used to maintain numerical stability. Classification label selection is achieved by finding the index of the maximum probability vector; confidence calculation only requires recording the maximum value, without sorting the entire vector. Threshold comparison uses a simple comparison instruction; when reclassification is triggered, a system flag is set and a copy of the current feature vector is saved. The reclassification process may introduce additional latency, requiring a trade-off between real-time performance and accuracy. Persistent storage of network weights supports rapid system startup; the pre-trained model file contains the values ​​of the weight matrix and bias vector. Memory requirements during inference are determined by the network size and batch size; memory usage needs optimization for embedded deployments. The computational complexity of forward propagation is proportional to the number of network connections; using sparse matrix representation can reduce computation. The choice of activation function affects network performance; the ReLU function avoids gradient vanishing but may cause neurons to die, while the Softmax function ensures that the output conforms to a probability distribution.

[0043] Outliers in feature vectors can distort network outputs; input standardization can mitigate this issue. The distribution characteristics of hidden layer outputs reflect the effect of feature transformation, and visualizing hidden layer activation values ​​helps in understanding network behavior. The entropy value of the output probability measures prediction uncertainty; high entropy indicates model indecisiveness, while low entropy indicates model confidence in the prediction. The classification confidence threshold can be adjusted online and dynamically optimized based on system operating conditions. The trigger frequency of the reclassification process monitors system performance; frequent reclassification suggests the model is not adapted to current operating conditions. The implementation of multilayer perceptron networks relies on numerical computation libraries, such as BLAS, for efficient matrix operations. The results of each step of forward propagation are cached in memory for use during backpropagation training. The inference phase only requires forward computation, resulting in low memory consumption. The mapping from feature vectors to output labels constitutes a non-linear decision boundary, capable of recognizing complex signal patterns. Increasing network depth and width improves expressive power but also increases computational resource requirements, requiring careful balancing on embedded devices. Model deployment typically involves converting to fixed-precision computation to reduce storage space and improve computation speed.

[0044] Example 4: A High-Frequency Electrical Signal Feature Classification Method for EMB Systems. In the online update stage, the gradient of the loss function is calculated based on the error between the classification label and the true label. The loss function adopts the cross-entropy loss form. The gradient calculation is implemented through the backpropagation algorithm, which propagates the error signal backward from the output layer layer by layer. The stochastic gradient descent algorithm is responsible for adjusting the hidden layer weights and bias terms of the neural network classifier. The update step size is dynamically adapted based on the exponential moving average of historical errors. This adaptive mechanism can balance convergence speed and stability. The gradient calculation of the hidden layer weights relies on the chain rule, which propagates the output layer error backward along the network structure to each layer. The gradient of the bias term comes directly from the error signal of the previous layer. The dynamic adjustment of the update step size adopts the core idea of ​​the Adam optimizer, combining the first-order moment estimation and second-order moment estimation to adaptively adjust the learning rate of each parameter. The gradient calculation of the loss function starts from the difference between the predicted probability of the output layer and the true label. The gradient of the cross-entropy loss function with respect to the input of the output layer is represented as the difference between the predicted probability and the true label. This error signal propagates back to the hidden layer, is multiplied by the transpose of the weight matrix from the hidden layer to the output layer, and then multiplied by the derivative of the hidden layer activation function to obtain the hidden layer error signal. The gradient of the hidden layer weights is calculated by the outer product of the input feature vector and the hidden layer error signal, and the gradient of the bias term is the hidden layer error signal itself. The stochastic gradient descent algorithm uses mini-batch data to update parameters; the batch size affects the variance of the gradient estimation and the computational efficiency.

[0045] The dynamic adaptation strategy maintains two state variables, recording the exponential moving averages of the first and second moments of the gradient. The first moment estimate represents the direction of the gradient, while the second moment estimate reflects the trend of gradient magnitude changes. Both moment estimates undergo bias correction to eliminate the influence of initial zero values. The update step size is obtained by dividing the initial learning rate by the square root of the second moment estimate plus the numerical stability constant, and then multiplying by the first moment estimate. This adaptive method allows frequently occurring parameters to have a smaller effective learning rate, while sparse parameters have a larger effective learning rate. The online update process is executed asynchronously in a dedicated thread to avoid impacting the time performance of real-time classification tasks. Historical error records are stored using a fixed-length circular buffer, with new error values ​​overwriting the oldest, maintaining a constant memory window size. Gradient calculation involves numerous matrix operations, which are accelerated using GPU parallel computing capabilities. A gradient pruning strategy is implemented before weight updates to prevent gradient explosion and numerical instability. Bias term updates are performed synchronously with weight updates, using the same adaptive learning rate mechanism. The dimension of the hidden layer weight matrix determines the number of parameters; larger networks require more training data to avoid overfitting. The bias vector has the same dimension as the number of hidden layer nodes, and each hidden layer node corresponds to a bias parameter. The magnitude of the gradient value of the loss function reflects the difference between the current parameter settings and the optimal solution, and the gradient direction indicates the direction of parameter updates. The randomness of the stochastic gradient descent algorithm comes from the random sampling of mini-batch data. This noise helps to escape local minima but may increase convergence fluctuations.

[0046] The hyperparameter settings of the dynamic adaptation strategy have a significant impact on performance, including the initial learning rate, first-order moment decay factor, second-order moment decay factor, and numerical stability constant. These hyperparameters need to be adjusted according to the specific application scenario. See Table 1, which shows typical hyperparameter settings: Table 1: Adam Optimizer Hyperparameter Settings

[0047] During online updates, the weight changes of the neural network classifier need to be constrained to prevent over-adaptation to new samples and forgetting old knowledge. The elastic weight consolidation method protects important parameters by calculating their importance scores and adding a regularization term to the loss function. The importance scores are calculated based on the diagonal elements of the Fisher information matrix of the parameters over historical data, allowing for relatively small variations in important parameters.

[0048] The derivative of the hidden layer activation function needs to be stored along with the activation values ​​during forward propagation. The derivative of the ReLU function is a piecewise linear function, with a derivative of 1 in the positive input region and 0 in the negative input region. When gradients flow through the hidden layer, due to the derivative characteristics of the ReLU function, the gradient of neurons corresponding to negative inputs becomes zero, leading to neuron death. Leaking ReLU functions, by setting a small positive slope in the negative region, can alleviate this problem and maintain gradient flow. The momentum term in the stochastic gradient descent algorithm can accelerate convergence and reduce oscillations; the momentum coefficient controls the degree to which the historical gradient direction influences the current update. Nesterov momentum updates parameters temporarily based on the momentum term before calculating the gradient, and this forward-looking update has better theoretical properties. In practical implementations, the Adam optimizer already includes a momentum mechanism, eliminating the need for a separate momentum term. The frequency of online updates needs careful design; overly frequent updates can lead to excessive computational load, while excessively long update intervals reduce model adaptability. Update triggering conditions can be based on fixed time intervals, sample size, or changes in model performance. Model performance monitoring uses a reserved validation set to calculate an accuracy metric, and triggers a model update when the accuracy drops below a threshold.

[0049] Weight updates require maintaining numerical precision; single-precision floating-point numbers are sufficient in most cases, while mixed-precision training uses half-precision floating-point numbers to accelerate computation. Gradient accumulation techniques allow for batch sizes larger than physical memory limits, accumulating gradients through multiple forward and backward propagations before updating parameters. Distributed training distributes data across multiple computing nodes, with each node calculating local gradients and then aggregating them to update the global model. Online updates of neural network classifiers need to consider the catastrophic forgetting problem, as training on new data may overwrite knowledge learned from older data. Experience replay methods save small batches of historical data and mix them with new data for training, maintaining the model's memory of historical distributions. Elastic weight consolidation methods do not require storing original data and protect important knowledge through parameter importance constraints, making them more suitable for data privacy-sensitive scenarios. The initialization of state variables in dynamic adaptation strategies affects the initial update step size; first-order and second-order moment estimates are typically initialized to zero vectors. In the early stages of training, due to the small state variable bias correction factor, the effective learning rate may be large, requiring a warm-up phase to gradually increase the learning rate. Learning rate scheduling strategies such as cosine annealing or step descent can further improve convergence performance, complementing adaptive learning rate mechanisms.

[0050] The magnitude of hidden layer weight updates reflects the model's adaptation to new data; large updates indicate a mismatch between the current model and the data distribution. The loss function value change curve during the update process can monitor training progress; a stable loss value may indicate convergence or the need to adjust hyperparameters. Gradient histogram visualization provides an intuitive understanding of the gradient distribution, helping to diagnose vanishing or exploding gradient problems. The computational complexity of the online update process is proportional to the number of network parameters; updating large networks requires more computational resources. Model compression techniques such as weight quantization, pruning, or knowledge distillation can reduce model size and improve update efficiency. Online updates on edge computing devices require a trade-off between model performance and resource constraints, selectively updating important parameters or using lightweight update strategies. The online update mechanism for high-frequency electrical signal feature classification methods for EMB systems enables neural network classifiers to adapt to changes in system operating states, maintaining classification performance without degradation over time. Dynamic adaptation strategies ensure stable convergence of the update process, avoiding the hassle of manually adjusting the learning rate.

[0051] Example 5: A High-Frequency Electrical Signal Feature Classification Method for EMB Systems. In the zero-crossing rate calculation stage, the number of sign changes between adjacent sampling points of the discrete signal sequence within the time window is calculated. The total number of sign changes is divided by the total number of sampling points within the time window, and then normalized to obtain the standardized zero-crossing rate. The zero-crossing rate, along with the signal mean, variance, and peak factor, constitutes the time-domain feature set. After obtaining the time-domain and frequency-domain feature sets, nonlinear dynamic features are extracted from the purified electrical signal. Nonlinear dynamic feature extraction includes reconstructing the phase space of the electrical signal, calculating the time delay, and the embedding dimension. Based on the reconstructed phase space, the maximum Lyapunov exponent and correlation dimension are calculated to characterize the chaotic properties of the signal. The maximum Lyapunov exponent and correlation dimension are added to the feature vector as nonlinear features. Taking a specific high-frequency electrical signal generated by an EMB system under brake sticking fault conditions as an example, the signal sampling rate is 100kHz, and the time window length is set to 10ms containing 1000 sampling points. Zero-crossing rate calculation begins by detecting the sign change between each sampling point and the next sampling point. The sign change is defined as the negative product of two adjacent sampling points. For discrete signal sequences, the number of sign changes is counted by iterating through all adjacent sampling point pairs. The original zero-crossing rate is obtained by dividing the total number of changes by the total number of sampling points, and then normalized by mapping to zero and one through a linear transformation. Nonlinear dynamic feature extraction is performed on signal segments within the same 10ms time window. Phase space reconstruction uses the time delay method. The time delay is determined by the first minimum value of the average mutual information function, and the embedding dimension is calculated using the spurious nearest neighbor method. The reconstructed phase space points consist of a multidimensional vector composed of the original signal and its time-delayed copies. The maximum Lyapunov exponent is calculated by tracking the divergence rate of adjacent orbits in phase space, and the correlation dimension is calculated by analyzing the fractal characteristics of the phase space point distribution.

[0052] In practical implementation, zero-crossing rate calculation requires preprocessing to eliminate the influence of DC components. A first-order high-pass filter is applied to the original signal to remove baseline drift. Sign change detection uses a digital comparator circuit to compare the sign bits of adjacent sampling points. Normalization uses a lookup table method to map the original zero-crossing rate value to a standard range; the mapping table is pre-calibrated according to signal characteristics. The zero-crossing rate calculation result is combined with the signal mean, variance, and peak factor calculated simultaneously to form a four-dimensional time-domain feature vector. Nonlinear dynamic feature extraction is implemented on a digital signal processor. Phase space reconstruction requires determining two key parameters first. Time delay is calculated by scanning the autocorrelation function of the signal and selecting the delay amount corresponding to the first drop of the autocorrelation function to a specific value. The embedding dimension is determined by gradually increasing the dimension until the proportion of spurious nearest neighbors is below a set threshold. The reconstructed phase space point set is used to calculate dynamic features; the maximum Lyapunov exponent reflects the system's sensitivity to initial conditions, and a positive exponent value indicates chaotic characteristics. The correlation dimension quantifies the fractal dimension of phase space attractors; non-integer values ​​characterize the existence of strange attractors.

[0053] Taking a fault signal segment from an actual EMB system as an example, the signal contains significant nonlinear oscillatory components. Zero-crossing rate calculation shows that the zero-crossing rate value increases significantly under fault conditions, reflecting an increase in high-frequency components of the signal. The phase space reconstruction parameters are determined to be a specific time delay and embedding dimension. The reconstructed three-dimensional projection of the phase space shows that the attractor structure changes from a limit cycle to a strange attractor. The calculated value of the maximum Lyapunov exponent is positive, and the calculated value of the correlation dimension is non-integer, confirming the existence of chaotic dynamics in the signal. The zero-crossing rate calculation module is implemented in parallel on the FPGA, including differential circuits, a symbol detector, and a counter array. Nonlinear feature extraction is performed serially on the DSP, using a sliding window method to update the phase space point set. In the feature vector combination stage, the time-domain feature set, frequency-domain feature set, and nonlinear feature set are concatenated in a fixed order to form a multi-dimensional feature vector containing the zero-crossing rate, signal mean, variance, peak factor, main frequency band energy, spectral entropy, maximum Lyapunov exponent, and correlation dimension. Before the feature vector is input into the neural network classifier, it is standardized, using a standardization method to transform the values ​​of each dimension to zero mean and unit variance. In another case study of insulation aging in an EMB system, the zero-crossing rate variation exhibited a different pattern. Insulation aging led to partial discharge signals, a decrease in the zero-crossing rate, and a significant increase in signal variance. Nonlinear feature analysis revealed an increase in the maximum Lyapunov exponent and the correlation dimension, indicating more complex system dynamics. These feature variation patterns, learned by a neural network classifier, were able to accurately distinguish different types of fault states.

[0054] Noise interference in zero-crossing rate calculation is suppressed by setting an amplitude threshold; only signals exceeding the noise threshold are considered valid zero-crossing points. Nonlinear feature calculation is sensitive to data length; shorter signal segments may lead to estimation errors, necessitating the selection of an appropriate time window length based on signal characteristics. Feature vector dimensionality optimization is achieved through principal component analysis, preserving key features while reducing computational complexity. In real-time implementation, zero-crossing rate calculation is updated at a fixed period, while nonlinear feature calculation is updated at a longer period, balancing feature update rate and computational load. Zero-crossing rate features are sensitive to signal frequency changes, effectively detecting abnormal fluctuations in switching frequency within the EMB system. Nonlinear dynamic features capture the implicit dynamic system characteristics within the signal, exhibiting sensitivity to subtle nonlinear changes caused by early faults. These two types of features provide complementary information based on time-frequency features, enhancing the classifier's ability to identify complex fault modes. Feature calculation parameters need to be adjusted according to the specific EMB system model, learning optimal parameter settings through historical data. Computational resource allocation during implementation must consider real-time requirements; simpler zero-crossing rate calculations are allocated to low-power microprocessors, while more complex nonlinear feature calculations are allocated to high-performance DSPs. The data stream employs a pipeline structure, with zero-crossing rate calculation and nonlinear feature extraction performed in parallel. Feature vector transmission utilizes a double-buffering mechanism to ensure data integrity and real-time performance. During the system calibration phase, signal samples are collected under normal and various fault conditions to optimize feature extraction parameters and classifier thresholds.

[0055] The hardware implementation of zero-crossing rate calculation employs dedicated digital circuitry, including shift registers, XOR gates, and adder trees. The software implementation of nonlinear feature extraction utilizes optimized numerical algorithms, employing lookup tables to accelerate exponential and logarithmic operations. A feature vector combination module is responsible for synchronizing the timestamps of each feature value, handling the asynchronous issues arising from different feature update rates. The final feature vector is transmitted to the neural network accelerator via a direct memory access channel to complete the real-time pattern classification task. This extraction method, combining zero-crossing rate and nonlinear dynamic features, has been validated for effectiveness on multiple EMB system experimental platforms. Zero-crossing rate provides an intuitive measure of signal frequency content, while nonlinear features reveal the dynamic mechanism of signal generation; their combination provides richer feature representations for fault diagnosis. In practical deployment, feature weights need to be adjusted according to specific application scenarios to optimize classification performance. The entire feature extraction pipeline adopts a modular design, facilitating portability and optimization across different hardware platforms. Parameter settings involved in feature calculation need to be recorded in configuration files to support system debugging and performance optimization. Nonlinear dynamic feature calculation requires significant computational resources; in resource-constrained environments, simplified algorithms or reduced update frequencies can be employed. Zero-crossing rate calculation has low computational complexity, making it suitable for real-time operation on edge devices. The fusion of these two types of features improves the adaptability of the fault diagnosis system to complex operating conditions.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for classifying high-frequency electrical signal features for an EMB system, characterized in that, The method comprises: Collecting high-frequency electrical signals generated by the EMB system during operation through a sensor to form a continuous electrical signal stream; performing real-time preprocessing on the electrical signal stream, including noise suppression and signal smoothing processing using a digital filter to obtain purified electrical signals; Performing quality assessment on the purified electrical signals, calculating the signal-to-noise ratio and stability coefficient of the signals, and generating a quality assessment result; when the quality assessment result meets the preset standard, starting a multi-scale feature extraction process; The multi-scale feature extraction process comprises: performing time domain analysis, frequency domain transformation and statistical feature calculation on the purified electrical signals to obtain a set of feature vectors; inputting the feature vectors into a pre-trained neural network classifier for pattern recognition and outputting a classification label; According to the performance indicators of the classification label, the weights of the neural network classifier are updated online.

2. The EMB system-oriented high-frequency electrical signal feature classification method according to claim 1, characterized in that, The specific steps of quality assessment on the purified electrical signals include: extracting the signal amplitude sequence and frequency trajectory from the purified electrical signals, calculating the fluctuation variance of the signal amplitude sequence and the offset of the frequency trajectory; comparing the fluctuation variance with the amplitude tolerance value and comparing the offset with the frequency tolerance value; when the fluctuation variance exceeds the amplitude tolerance value or the offset exceeds the frequency tolerance value, marking the quality assessment result as unqualified, otherwise marking as qualified.

3. The method of claim 2, wherein the method further comprises: The specific steps of time domain analysis on the purified electrical signals include: dividing the purified electrical signals into multiple time windows, calculating the signal mean, variance, peak factor and zero-crossing rate for each time window to obtain a time domain feature set; the specific steps of frequency domain transformation on the purified electrical signals include: applying short-time Fourier transform to each time window, extracting the main frequency band energy and spectral entropy of the power spectral density to obtain a frequency domain feature set.

4. The method of claim 3, wherein the method further comprises: The specific steps of statistical feature calculation on the purified electrical signals include: calculating the skewness, kurtosis and autocorrelation function of the purified electrical signals in the sliding window to obtain a statistical feature set; the time domain feature set, frequency domain feature set and statistical feature set are combined into the feature vector.

5. The method of claim 4, wherein the EMB system-oriented high-frequency electrical signal feature classification method is characterized by, The specific steps of inputting the feature vector into the pre-trained neural network classifier for pattern recognition include: using a multilayer perceptron network to forward propagate the feature vector, calculating the output probability of each class; selecting the class with the highest output probability as the classification label; at the same time, calculating the classification confidence, when the classification confidence is lower than the threshold, triggering the reclassification process.

6. The method of claim 5, wherein the method further comprises: The specific steps of online updating the weights of the neural network classifier include: calculating the loss function gradient according to the error between the classification label and the true label; adjusting the hidden layer weights and bias terms of the neural network classifier using the stochastic gradient descent algorithm, and the update step is dynamically adapted according to the historical error.

7. The method of claim 3, wherein the method further comprises: The specific steps of calculating the zero-crossing rate for each time window include: For the discrete signal sequence in the time window, calculate the number of sign changes of adjacent sampling points; Divide the total number of sign changes by the total number of sampling points in the time window, and then normalize to obtain the normalized zero-crossing rate; The zero-crossing rate, together with the signal mean, variance and peak factor, constitutes the time-domain feature set.

8. The method of claim 4, wherein the EMB system-oriented high-frequency electrical signal feature classification method is characterized by, The specific steps for calculating the autocorrelation function include: Within the sliding window, the correlation coefficient of the signal with itself at different time delays is calculated; The position of the first maximum point of the autocorrelation function is extracted as the periodicity feature of the signal, and the decay rate of the autocorrelation function near the zero delay point is calculated as the stability feature of the signal; The skewness, kurtosis, first maximum point position and decay rate constitute the statistical feature set.

9. The method of claim 3, wherein the method is characterized by, After obtaining the time-domain feature set and the frequency-domain feature set, the purified electrical signal is also subjected to nonlinear dynamics feature extraction, and the specific steps include: The phase space of the electrical signal is reconstructed, and the time delay and embedding dimension are calculated; Based on the reconstructed phase space, the maximum Lyapunov exponent and the correlation dimension are calculated to characterize the chaotic characteristics of the signal; The maximum Lyapunov exponent and the correlation dimension are taken as nonlinear features and supplemented to the feature vector.

10. A high-frequency electrical signal feature classification system for an EMB system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the EMB system-oriented high-frequency electrical signal feature classification method according to any one of claims 1 to 9.

Images