Method and system for simulating working conditions of abnormal noise of automobile intermediate shaft and storage medium

Abstract

This application relates to the field of automotive engineering technology and discloses a method, system, and storage medium for simulating the operating conditions of abnormal noise from an automotive intermediate shaft. The method includes: acquiring multi-source data to obtain an initial heterogeneous dataset; applying a unified clock reference to determine alignment points; if the deviation of the alignment points exceeds a preset deviation threshold, obtaining a synchronized time-series dataset; if the correspondence between dynamic operating condition characteristics and the peak values ​​of the acquired abnormal noise signals shows a synchronization peak, it is marked as a related event; fusing vibration data spectra to determine the occurrence mode of the abnormal noise characteristics; performing cluster analysis on the determined occurrence modes to obtain the correlation clusters between transient operating conditions and abnormal noise; constructing a tracing matrix to obtain a complete abnormal noise mapping model; and adjusting the control parameters of a multi-axis robotic arm to simulate and reproduce the transient operating conditions that generate abnormal noise, thus realizing operating condition simulation and abnormal noise analysis. This application improves the simulation accuracy of transient operating conditions of abnormal noise from automotive intermediate shafts.

Description

Technical Field

[0001] This application relates to the field of automotive engineering technology, and in particular to a method, system and storage medium for simulating abnormal noise from an automotive intermediate shaft. Background Technology

[0002] In modern automotive transmission systems, the intermediate shaft, as a key component connecting the transmission and drive axle, directly affects the overall vehicle performance due to its dynamic characteristics. With the development of automotive electrification and lightweighting, the torque fluctuations and posture changes experienced by the intermediate bearing are becoming increasingly complex. Abnormal noise problems induced by factors such as torsional deformation, universal joint clearance, and bearing wear are becoming increasingly prominent, becoming a key quality indicator affecting driving comfort and product competitiveness.

[0003] However, the detection and diagnosis of intermediate axle noise faces significant technical challenges. On the one hand, intermediate axle noise is transient, sporadic, and involves multiple factors. Traditional road testing is limited by uncontrollable road conditions, difficulty in reproducing operating conditions, and incomplete data collection, making it difficult to accurately capture the boundary conditions for the occurrence of noise. On the other hand, existing laboratory testing methods mostly rely on single sensor data analysis, ignoring the comprehensive correlation between attitude adjustment records, acoustic signals, and vibration data. Furthermore, differences in the acquisition frequencies of different sensors lead to timing alignment deviations, making the correspondence between dynamic operating condition characteristics and the peak value of the noise signal unclear, and tracing the abnormal occurrence mode extremely difficult.

[0004] Furthermore, existing technologies lack effective means of reproducing operating conditions. The disconnect between the abnormal noise mapping model and the mechanical control system prevents the analysis conclusions from being verified experimentally, resulting in a broken closed loop in the "detection-analysis-optimization" process. Therefore, how to achieve accurate simulation of intermediate shaft abnormal noise conditions, synchronous fusion of multi-source data, and physically interpretable construction of the abnormal noise mapping model in a laboratory environment has become a key issue in improving the accuracy of fault diagnosis and R&D efficiency of automotive transmission systems. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, system, and storage medium for simulating the operating conditions of abnormal noise from an automotive intermediate shaft, thereby improving the simulation accuracy of transient operating conditions of abnormal noise from an automotive intermediate shaft.

[0006] Firstly, this application provides a method for simulating abnormal noise from an automotive intermediate shaft under operating conditions, the method comprising: The initial heterogeneous dataset is obtained by collecting multi-source data through a multi-axis robotic arm controller and sensor interface. A timestamp synchronization algorithm is used to apply a unified clock reference to the initial heterogeneous dataset to determine the alignment point of each data stream at a preset precision. If the deviation of the alignment point exceeds a preset deviation threshold, the missing time sequence is filled by interpolation to obtain the synchronized time sequence dataset. Based on the synchronized time-series dataset, dynamic operating condition features representing the intermediate shaft are extracted. The correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals is determined. If the correspondence shows a synchronized peak value, it is marked as an associated event. The marked associated events are obtained and the vibration data spectrum is fused to determine the occurrence mode of abnormal noise features under specific mechanical parameters. The determined occurrence patterns are clustered using a feature association algorithm to obtain association clusters between transient conditions and abnormal noises; a tracing matrix is ​​constructed for the association clusters to obtain a complete abnormal noise mapping model. Based on the abnormal noise mapping model, the control parameters of the multi-axis robotic arm are adjusted to simulate and reproduce the transient working conditions that generate abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

[0007] Secondly, this application provides a working condition simulation system for abnormal noise from the intermediate shaft of an automobile, the system comprising: The data acquisition unit is used to acquire multi-source data through the multi-axis robotic arm controller and sensor interface to obtain an initial heterogeneous dataset. The data processing unit is used to apply a unified clock reference to the initial heterogeneous dataset using a timestamp synchronization algorithm to determine the alignment point of each data stream at a preset precision; if the deviation value of the alignment point exceeds a preset deviation threshold, the missing timing sequence is filled by interpolation method to obtain the synchronized timing dataset. The feature analysis unit is used to extract dynamic operating condition features representing the intermediate shaft based on the synchronized time-series dataset, determine the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals, and mark the corresponding peak values ​​as associated events if the correspondence shows the synchronized peak values; obtain the marked associated events and fuse the vibration data spectrum to determine the occurrence mode of abnormal noise features under specific mechanical parameters. The model building unit is used to perform cluster analysis on the determined occurrence pattern through a feature association algorithm to obtain the association clusters between transient conditions and abnormal noises; and to construct a tracing matrix for the association clusters to obtain a complete abnormal noise mapping model. The simulation analysis unit adjusts the control parameters of the multi-axis robotic arm according to the abnormal noise mapping model to simulate and reproduce the transient working conditions that produce abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

[0008] A third aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned method for simulating abnormal noise from an automotive intermediate shaft.

[0009] Compared with the prior art, the beneficial effects of the present invention are at least as follows: First, a timestamp synchronization algorithm based on a network time protocol is used to calculate the time offset, applying a unified clock reference to attitude, acoustic, and vibration data. Missing time sequences are filled using linear interpolation and cubic spline interpolation, and vibration data spectral constraints are applied to verify the accuracy of the interpolation, controlling the synchronization precision to the millisecond level to ensure a precise correspondence between dynamic operating condition characteristics and abnormal signal peaks. Time-varying parameters such as attitude angular velocity and angular acceleration are calculated using first-order and second-order differential methods, and a time observation window is established by fusing acoustic peak data to mark synchronization peak-related events. Pearson correlation coefficients are used to screen strong correspondences, generating a correspondence spectrum. Frequent subgraph mining is used to identify recurring patterns, and vibration data is fused to verify reliability, effectively eliminating environmental noise interference and establishing a causal relationship between operating conditions and abnormal noises. Then, sub-patterns are extracted through temporal expansion and sliding windowing, and waveform similarity is calculated for clustering to obtain refined temporal features. K-means clustering combined with transient condition constraints and Mahalanobis distance boundary optimization ensures high similarity of conditions within clusters. False transition relationships are filtered out through cluster network stability assessment and graph entropy value determination, achieving accurate classification and reliable traceability of abnormal noise occurrence patterns. A traceability matrix is ​​constructed and latent mappings are extracted through non-negative matrix factorization, and singular value decomposition enhances model stability. Physical constraints such as monotonically non-decreasing abnormal noise loudness with torque and energy jumps at extreme attitude angles at specific frequencies are applied. Extreme condition data is used to verify the model's robustness, ensuring that the output results conform to the torsional mechanics of the intermediate shaft, thus improving the model's engineering credibility and interpretability. Finally, based on the specific torque-speed-angle parameter combination output by the abnormal noise mapping model, the motion trajectory command of the robotic arm is generated and driven to reproduce the noise. The reproduced signal is collected and compared with the original features to calculate the similarity. The model parameters are iteratively optimized to form a closed loop of "model prediction - mechanical reproduction - accuracy verification". This provides a reliable experimental verification method for the rapid location, mechanism analysis and design improvement of abnormal noise problems in chassis components, and significantly improves R&D efficiency and product quality. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a method for simulating abnormal noise from an automotive intermediate shaft, as described in this application. Figure 2 This is a schematic diagram of the correspondence between embodiments of this application; Figure 3 This is a schematic diagram comparing the spectra of the original abnormal noise and the reproduced abnormal noise in an embodiment of this application; Figure 4This is a schematic diagram of the working condition simulation system for abnormal noise from the intermediate shaft of an automobile, according to an embodiment of this application. Detailed Implementation

[0012] This application provides a method, system, and storage medium for simulating abnormal noise from an automotive intermediate shaft. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0013] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the working condition simulation method for abnormal noise from the intermediate shaft of an automobile, as described in this application, includes: Step S1: Collect multi-source data through the multi-axis robotic arm controller and sensor interface to obtain an initial heterogeneous dataset.

[0014] Furthermore, the initial heterogeneous dataset is obtained, including: The attitude adjustment records of the intermediate axis are collected by a multi-axis robotic arm controller, and acoustic signal sequences and vibration data streams are obtained from the sensor interface. The attitude adjustment records and acoustic signal sequences are fused together and combined with the vibration data stream to form multi-source data. The multi-source data is initially time-series labeled to determine the acquisition frequency differences of each data source. The data sampling rate is adjusted according to the acquisition frequency differences to generate an initial heterogeneous dataset with a unified format. Source identification labels are applied to the initial heterogeneous dataset, and the rate of change of spatial angle between adjacent time points in the attitude adjustment records is calculated. Time points with a rate of change of spatial angle exceeding a preset angle change threshold are selected as key event points. If the key event points correspond to the acoustic signal sequence, they are marked as potential abnormal noise intervals. The potential abnormal noise intervals are verified by the vibration data stream, and the peak features within the intervals are extracted. The peak features are integrated into the initial heterogeneous dataset to obtain an initial heterogeneous dataset containing multimodal information.

[0015] The intermediate shaft is a key rotating component in a car's transmission system, typically referring to the drive shaft connecting the transmission output and the drive axle input, responsible for transmitting engine torque to the wheels. Abnormal noise refers to unusual noise generated by the intermediate shaft under specific operating conditions, usually induced by factors such as torsional deformation, universal joint clearance, bearing wear, and poor spline fit. In a laboratory environment, a multi-axis robotic arm controller is used to drive the intermediate shaft to reproduce its dynamic posture and mechanical state under actual road conditions, thus simulating the operating conditions.

[0016] Specifically, a communication connection is established with the multi-axis robotic arm controller, and attitude adjustment records containing spatial coordinates and rotation angles are acquired according to a preset reading cycle. The multi-axis robotic arm controller is a dedicated control device used to control the movement of a multi-degree-of-freedom robotic arm. It uses kinematic algorithms to convert the motion requirements of the task space into drive commands in the joint space, achieving precise motion control of the robotic arm. The attitude adjustment records specifically include the three-dimensional spatial arrangement angles of the intermediate shaft relative to the vehicle body: pitch angle, roll angle, and yaw angle, as well as the real-time speed and torque values ​​of the drive system. The pitch angle is the rotation angle of the intermediate shaft around the vehicle's lateral axis, reflecting the degree of forward and backward tilt; the roll angle is the rotation angle of the intermediate shaft around the vehicle's longitudinal axis, reflecting the degree of left and right tilt; and the yaw angle is the rotation angle of the intermediate shaft around the vehicle's vertical axis, reflecting the degree of horizontal deflection. The speed is the rotational speed of the drive system output shaft, measured in revolutions per minute, reflecting the speed of rotation of the intermediate shaft; the torque is the rotational torque of the drive system output shaft, measured in Newton-meters, reflecting the load on the intermediate bearing.

[0017] The acoustic signal sequence is a time series of sound signals acquired by microphone arrays arranged on both sides of the intermediate shaft universal joint. The microphone arrays convert sound pressure signals into electrical signals using the piezoelectric effect or capacitance changes, and the array layout enables sound source localization and noise suppression. The vibration data stream is a time series of vibration acceleration acquired by a triaxial accelerometer mounted on the intermediate shaft bearing housing. The triaxial accelerometer is an inertial sensor capable of simultaneously measuring acceleration in three orthogonal directions. Based on the piezoelectric effect or microelectromechanical systems (MEMS) technology, it converts mechanical acceleration into measurable electrical signals.

[0018] Following the order of data reception, attitude adjustment records, acoustic signal sequences, and vibration data streams are stored in a temporary buffer to construct multi-source data containing multiple dimensions of physical quantities. The temporary buffer is a storage area used to temporarily store real-time acquired data. Multi-source data refers to a collection of data from different sensors that describe different dimensions of the same physical object; its characteristics include diverse data types, varying sampling rates, and independent time bases.

[0019] The arrival timestamp of each data point from the multi-source data is extracted as a preliminary time-series label. A timestamp is a time marker recording the moment data is generated or arrives, typically with millisecond or microsecond precision; its purpose is to establish a time index for the data. The first sampling rate is obtained by statistically analyzing the number of samples recorded by attitude adjustment per unit time, and the second sampling rate is obtained by statistically analyzing the number of samples of the acoustic signal sequence per unit time. The ratio of the first sampling rate to the second sampling rate is calculated to determine the acquisition frequency differences among the data sources. Determining these acquisition frequency differences provides a fundamental quantitative indicator for subsequent solutions to achieve accurate synchronization and alignment of multi-source heterogeneous data streams in the time dimension.

[0020] The data source with the highest sampling rate among the multi-source datasets is selected as the baseline frequency. Linear interpolation is then used to supplement data points from data sources with frequencies lower than the baseline frequency, ensuring that the sampling rates of all data sources are consistent with the baseline frequency. The adjusted sampling rates from each data source are then concatenated at a uniform sampling interval to generate an initial heterogeneous dataset with a unified format. The baseline frequency is the sampling rate used as a reference standard during data resampling; typically, the highest sampling rate among the multi-source datasets is selected to avoid the loss of high-frequency information. Linear interpolation is a mathematical method that constructs a linear equation using two known adjacent data points and calculates the value at any intermediate time. It utilizes the linear relationship between two points for proportional extrapolation, is computationally simple and computationally inexpensive, and is suitable for data with relatively flat trends. The initial heterogeneous dataset refers to multi-source data that has undergone preliminary integration but is not yet fully synchronized. Its characteristics include a unified data format and preliminary timeline alignment, but residual time offsets and missing data may exist between the data sources.

[0021] Independent data type identifiers are added to the acoustic, vibration, and attitude data in the initial heterogeneous dataset to complete the application of source identification labels. Source identification labels are used to distinguish different data sources, preventing type confusion during data fusion and ensuring the correct correspondence between modal data. The rate of change of spatial angle between adjacent time points in the attitude adjustment record is calculated. The rate of change of spatial angle is the difference between the torsion angle at the current moment and the torsion angle at the previous moment divided by the time interval, i.e.: ,in, This represents the difference in attitude angle between adjacent time points. The sampling time interval is defined as the spatial angle change rate, which measures the severity of the intermediate shaft's attitude change. A higher change rate indicates that the intermediate shaft has undergone greater attitude adjustments in a shorter period of time, often corresponding to extreme torsional conditions. When this change rate exceeds a set threshold, it indicates that the chassis components are in an extreme torsional state, which is often the root cause of abnormal noises. Therefore, these time points are identified as critical event points to provide time anchors for subsequent causal analysis.

[0022] Centered on the key event location, extract acoustic signal sequences of fixed time lengths both forward and backward. Calculate the effective sound pressure level (SPL) of the extracted acoustic signal sequences. It is the root mean square sound pressure level of the acoustic signal within the analysis window, and the calculation formula is: ,in The instantaneous sound pressure level (SPL) is represented by , and T is the analysis duration. Physically, it is equivalent to the DC sound pressure level of the same energy, objectively reflecting the intensity of the sound. If the effective sound pressure value exceeds a preset sound pressure threshold, this fixed-time interval is marked as a potential abnormal noise interval. The preset sound pressure threshold is a criterion determined based on the background noise level of the laboratory environment and the statistical characteristics of abnormal noise on the intermediate shaft. A value exceeding this threshold indicates an abnormal energy concentration in the acoustic signal, potentially corresponding to an abnormal noise event.

[0023] Within the potential abnormal noise range, vibration data streams synchronized with the acoustic signal sequence are extracted. The maximum amplitude value of the vibration data stream within this range is identified, and this maximum amplitude value and its corresponding time point are extracted as peak features. The maximum amplitude value is the absolute maximum value of the vibration signal within the analysis range, reflecting the intensity of mechanical vibration at that moment. Extreme points in the data sequence are identified through traversal comparison. Due to the possibility of background noise from other mechanical operations, relying solely on acoustic signals can easily lead to misjudgments. By traversing all vibration data points within the range and comparing them to obtain the maximum amplitude value, extracting this peak feature effectively confirms that physical vibration caused by torsional deformation has indeed occurred within this range, thus verifying the authenticity of the potential abnormal noise range and eliminating external acoustic interference. The extracted maximum amplitude value and its corresponding time point are added as new data columns to the initial heterogeneous dataset. Correlation pointers are established between the peak features and the attitude adjustment records and acoustic signal sequences at the corresponding time points, resulting in a complete initial heterogeneous dataset containing multimodal information. Correlation pointers are index links used to connect data at the same time point in different data tables. Their function is to enable rapid retrieval and joint analysis of cross-modal data and avoid redundant data storage. In this embodiment, multimodal information specifically refers to the fusion of three modal data: attitude angle, acoustic signal, and vibration signal. This fusion can comprehensively describe the operating status and abnormal noise characteristics of the intermediate shaft, providing a rich data foundation for subsequent synchronous analysis and pattern recognition.

[0024] Step S2: Apply a unified clock reference to the initial heterogeneous dataset using a timestamp synchronization algorithm to determine the alignment point of each data stream at a preset precision; if the deviation of the alignment point exceeds the preset deviation threshold, fill the missing timing sequence using an interpolation method to obtain the synchronized timing dataset.

[0025] Furthermore, the synchronized time-series dataset is obtained, including: The original device timestamps of each data stream are extracted from the initial heterogeneous dataset. A timestamp synchronization algorithm is used to calculate the time offset. Based on the time offset, the attitude adjustment recording and acoustic signal sequences are aligned. The timestamps of the vibration data stream are fused to generate a time axis under a globally unified clock reference. A preset precision calibration is applied to the time axis to determine the alignment point. If there is a time difference between the alignment points, the intermediate value is estimated by linear interpolation to obtain the calibrated alignment sequence. The matching degree between the calibrated alignment sequence and the original data stream is judged. The synchronization parameters are optimized based on the matching degree to determine the final alignment point. If the deviation value of the determined alignment point exceeds the preset deviation threshold, the missing time interval is identified, and the filling value is calculated by cubic spline interpolation to generate the interpolated time sequence chain. Vibration data spectrum constraints are applied to the time sequence chain. Vibration signals synchronized with the filling time interval are extracted and subjected to fast Fourier transform to obtain the characteristic frequency amplitude. It is determined whether the characteristic frequency amplitude falls within the theoretical resonance frequency range derived from the intermediate axis torsional mechanics model based on the interpolated attitude data to determine the accuracy of the filling and obtain the synchronized time sequence dataset.

[0026] Specifically, the raw timestamp is a time stamp recorded by each data acquisition device at the moment the data is generated. It is usually generated by the device's internal clock, and its accuracy depends on the device's crystal oscillator frequency and clock synchronization status. Since the clock sources of different devices operate independently, there is clock drift and accumulated error, which leads to systematic deviations between the raw timestamps of each data stream. Therefore, it is necessary to calibrate them through a timestamp synchronization algorithm.

[0027] The time offset is calculated using a timestamp synchronization algorithm based on the Network Time Protocol (NTP). NTP is an internet standard protocol for clock synchronization in distributed systems. Its working principle involves estimating network transmission delay and clock skew through bidirectional timestamp message exchange between the client and server. Specifically, this includes the client sending a request packet to the server and recording the sending time T1; the server receiving the request and recording the receiving time T2; the server sending a response packet and recording the sending time T3; and the client receiving the response and recording the receiving time T4. The time offset is calculated based on network latency. and clock offset The calculation gradually adjusts the local clock to keep it consistent with the reference clock. The network delay δ is the round-trip time of a data packet in the network, and the clock offset θ is the system deviation between the client clock and the server clock. The two are decoupled and calculated through four-point recording of bidirectional timestamps, avoiding confusion between delay and offset in unidirectional transmission. In this embodiment, the robotic arm controller is configured as an NTP server, and the acoustic acquisition card and vibration sensor are configured as NTP clients. Through the above bidirectional timestamp exchange, the offset θ and delay δ of each client relative to the server clock are calculated. After multiple iterations of filtering and clock drift compensation, the offset estimation error is controlled to the sub-millisecond level, thereby providing a reliable time reference for subsequent data alignment. For testing under high-speed conditions, the inherent time delay of the data acquisition system also needs to be considered. This delay is measured through pre-calibration experiments and added to the NTP offset as a fixed compensation amount to ensure the accuracy of time synchronization.

[0028] The calculated offset is used to adjust the timestamps of the attitude adjustment records, aligning them temporally with the acoustic signal sequence. For example, subtracting the offset ΔT from the timestamps of the attitude records achieves initial alignment. The timestamps of the vibration data stream are then integrated into the aligned attitude and acoustic sequences, creating a unified timeline through a merging operation. A timeline under a unified clock reference maps multi-source heterogeneous data onto the same reference axis, measured in milliseconds, ensuring that data streams with different sampling rates can be compared and fused on a unified time dimension. The generation process of the timeline essentially involves converting the local timestamps of each data stream into a global standard time. Through timestamp shifting and format standardization, time reference differences between data sources are eliminated, establishing a globally consistent time coordinate system.

[0029] A calibration function is applied to the generated unified timeline to adjust it to a preset accuracy. This preset accuracy is set according to the analysis requirements; in this embodiment, it is millisecond-level accuracy, meaning the smallest resolution unit of the timeline is 1 millisecond. Alignment points are determined, that is, the corresponding sampling points of each data stream on the unified timeline are identified. If there is a time difference between alignment points (i.e., the time interval between adjacent alignment points is greater than the target sampling interval), intermediate values ​​are estimated using linear interpolation to obtain the calibrated alignment sequence. Linear interpolation is a mathematical method that constructs a linear equation using two known adjacent data points and calculates the value at any intermediate time. Its working principle assumes that the data changes linearly between two known points, and the intermediate value is calculated through proportional relationships. The calculation is simple and ensures the continuity of the interpolation results. This ensures that the sampling density of the attitude adjustment records is consistent with the acoustic signal, forming a continuous alignment sequence.

[0030] The matching degree between the calibrated aligned sequence and the original data stream is calculated. The matching degree is a quantitative indicator measuring the consistency between the synchronized sequence and the original sequence; in this embodiment, the Pearson correlation coefficient is used. The Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two variables. The input to this correlation coefficient is two time series of equal length, and the output is a value between -1 and 1. The closer the absolute value is to 1, the stronger the linear correlation. Its working principle is to eliminate the influence of dimensions through covariance standardization, quantifying the degree of synchronization between the two sequences in terms of their changing trends. If the calculated matching degree is lower than a preset threshold, it indicates that the current synchronization parameters have not sufficiently aligned the data streams. The synchronization parameters, such as the weight of the offset calculation and the size of the filtering window, need to be adjusted, and the synchronization algorithm needs to be run iteratively until the matching degree meets the requirements, thereby determining the final alignment point. The final alignment point is the time marker that precisely corresponds to each data stream on a unified time axis after iterative optimization; its accuracy directly affects the reliability of subsequent feature extraction and pattern analysis.

[0031] The deviation value is extracted from the determined alignment points. The deviation value is the difference in timestamps of corresponding sampling points in different data streams at the same physical moment, reflecting the residual time error after synchronization. It is then determined whether the deviation value exceeds a preset deviation threshold, which is set according to the required analysis accuracy. If the deviation value exceeds the preset threshold, missing time intervals are identified. These missing time intervals are due to the low sampling rate of the robotic arm controller data compared to the extremely high sampling rate of the acoustic sensor. When aligned according to the high sampling rate time axis, gaps appear in the continuous data points of the robotic arm data sequence. Cubic spline interpolation is used to calculate the filling value. Cubic spline interpolation is a piecewise polynomial interpolation method that constructs a cubic polynomial on each sub-interval, requiring the function value, first derivative, and second derivative to be continuous at the nodes. This ensures that the interpolation curve not only passes through all known points but also has smooth transition characteristics, smoothly approximating the trajectory of angle changes during the continuous movement of the robotic arm. In this embodiment, cubic spline interpolation is used for the sequence describing the joint angles of the robotic arm. Based on the nearest real attitude data points before and after the missing interval, an interpolation function is constructed to calculate the attitude data filling value corresponding to each high-sampling moment within the missing time period, generating the interpolated time series chain. The time series chain is a continuous sequence arranged in chronological order, containing original data and interpolated data, characterized by uniform time intervals and complete data.

[0032] The vibration signal synchronized with the filling time period is extracted, and a Fast Fourier Transform (FFT) is performed on this vibration signal to obtain its spectral characteristics, especially the characteristic frequency amplitudes related to the torsional resonance of the component. The Fast Fourier Transform (FFT) is an efficient algorithm for converting time-domain signals into frequency-domain representations. Its input is a sequence of time-domain vibration signals, and its output is each frequency component along with its corresponding amplitude and phase. Its working principle utilizes the periodicity and symmetry of the signal to quickly extract its spectral characteristics. Spectral characteristics represent the energy distribution of the signal in the frequency domain, with frequency on the horizontal axis and amplitude or power on the vertical axis, revealing hidden periodic components and resonance phenomena in the time-domain signal.

[0033] Simultaneously, based on the interpolated attitude data, the theoretical dominant frequency range of the component under this attitude is derived through the intermediate shaft torsional mechanics model. The intermediate shaft torsional mechanics model is based on the material properties, geometric dimensions, and boundary constraints of the intermediate shaft, and its general form is: ,in, For the equivalent moment of inertia, The damping coefficient is... To increase torsional stiffness, For external torque, The model takes the interpolated attitude data (torsion angle) as input and outputs the theoretical resonant frequency range. Its working principle is to simplify the intermediate shaft into a torsional spring-mass system, establish equations of motion based on Hooke's law and Newton's second law, solve the eigenvalue problem to obtain the system's natural frequencies, and determine the theoretical resonant frequency range considering frequency fluctuations. In this embodiment, the theoretical resonant frequency of the intermediate shaft at the interpolated torsion angle is calculated using the model.

[0034] The synchronized time series dataset is a complete data set after time synchronization, interpolation padding, and physical verification. Its characteristics are that all data streams share a unified time axis, have consistent sampling rates, and have no missing time series. Furthermore, the interpolation results have passed the physical consistency test of spectral constraints, and can be reliably used for subsequent dynamic operating condition feature extraction and abnormal noise pattern analysis.

[0035] Step S3: Extract dynamic operating condition features representing the intermediate shaft based on the synchronized time series dataset, determine the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals, and mark the corresponding peak values ​​as associated events if the correspondence shows the synchronized peak values; obtain the marked associated events and fuse the vibration data spectrum to determine the occurrence mode of abnormal noise features under specific mechanical parameters.

[0036] Furthermore, determining the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals includes: The attitude angle sequence, rotational speed sequence, and torque sequence are synchronously extracted from the synchronized time-series dataset. The attitude angle sequence is calculated using first-order difference to obtain the attitude angular velocity, and second-order difference to obtain the attitude angular acceleration. The attitude angular velocity and attitude angular acceleration are used as time-varying parameters and fused with the attitude angle, rotational speed, and torque to form dynamic operating condition characteristics. For these dynamic operating condition characteristics, acoustic signal peak data is fused to determine the peak occurrence time. A time observation window of a preset duration is established centered on the peak occurrence time. The correspondence between the local maxima of the time-varying parameters and the acoustic signal peak within the time observation window is determined. If the correspondence shows a synchronized peak, it is marked as a related event. The operating condition feature vector is analyzed based on the related events. The operating condition feature vector is composed of rotational speed values, torque values, three-dimensional attitude angle values, and attitude angular acceleration values ​​concatenated in a fixed-dimensional order. The Pearson correlation coefficient between the vectors is calculated, and strong correspondences are filtered through a correlation threshold to generate a correspondence map. Pattern recognition is performed on the correspondence map to determine recurring correspondence patterns. Vibration data is fused to verify the reliability of the correspondence patterns. If the reliability is higher than a preset reliability threshold, the determined correspondence is output.

[0037] Furthermore, the occurrence mode of abnormal noise characteristics under specific mechanical parameters is determined, including: Extract the associated event sequence from the determined correspondence, obtain the vibration data spectrum as the fusion input, calculate the spectral features through Fourier transform, align the spectral features with the correspondence, and determine the frequency domain representation of the abnormal noise features; determine the distribution of the frequency domain representation under a specific torque-speed-angle parameter combination, if the distribution shows peak clustering, it is identified as the occurrence mode; analyze the parameter dependence based on the occurrence mode, generate the mode parameter mapping, fuse multi-source data to verify the accuracy of the mapping, and if the accuracy meets the conditions, output the determined occurrence mode.

[0038] Specifically, the correspondence refers to the correlation and mapping relationship between dynamic operating condition characteristics and abnormal noise signal peaks. Multi-source test data aligned to a unified time axis is read, and the speed and torque change sequences recorded by the robotic arm control system are separated. The attitude angle sequences of the tested chassis components in three-dimensional space are then extracted from the multi-source test data. Synchronous extraction refers to the operation of simultaneously acquiring multiple data sequences from a time-series dataset in a unified format under the same time reference. A unified time index ensures that the sampled values ​​of each data sequence at the same physical moment are retrieved simultaneously, avoiding analysis errors caused by time misalignment.

[0039] The formula for calculating the first-order difference is: ,in, The current attitude and angle. The posture and angle at the previous moment. The sampling time interval is used. The first-order difference is used to calculate the rate of change of attitude angle between adjacent time points. The method takes a discrete attitude angle sequence as input and outputs an attitude angular velocity sequence. Its working principle is to approximate the instantaneous angular velocity by dividing the angle difference between the current and previous moments by the time interval, reflecting the rate of change of the intermediate axis attitude. The second-order difference calculation in the time dimension is performed on the attitude angle sequence, as shown in the formula: The instantaneous angular acceleration is approximately represented by dividing the difference in angular velocity between the current moment and the previous moment by the time interval. This reflects the degree of drastic change in the attitude of the intermediate axis, i.e., the rate of change of angular velocity.

[0040] Time-varying parameters are dynamic characteristic quantities that change over time, distinct from static state quantities such as attitude angle, rotational speed, and torque. They can characterize the dynamic changes of the intermediate shaft during motion. Dynamic operating condition characteristics are a set of multi-dimensional parameters describing the operating state of the intermediate shaft at a specific moment. Through the comprehensive characterization of multi-dimensional parameters, the intensity of dynamic changes of chassis components during torsion is quantified, providing a comprehensive operating condition description for abnormal noise correlation analysis.

[0041] The acoustic signal sequence was traversed synchronously, and envelope analysis was used to extract local maxima where the sound pressure level exceeded the ambient background noise baseline. These local maxima were defined as acoustic signal peak data, and their corresponding absolute timestamps were recorded as the peak occurrence times. Envelope analysis is a method for extracting the signal amplitude envelope. Its input is the acoustic signal sequence, and its output is the signal amplitude envelope. The working principle is to obtain the contour line of the signal amplitude changing over time through Hilbert transform or sliding window maximum extraction, thereby identifying local peaks. Hilbert transform is a mathematical transformation that converts a real signal into an analytic signal. The instantaneous amplitude, i.e., the envelope, of the signal can be obtained through the modulus of the analytic signal. Acoustic signal peak data are local maxima in the acoustic signal that are significantly higher than the surrounding environment, representing possible abnormal noise events. The threshold setting for exceeding the background noise is to ensure that the peak has a sufficient signal-to-noise ratio and avoids interference from environmental noise.

[0042] Centered on the peak occurrence time, a time observation window with a preset duration (extended by 50 milliseconds each) is established, forming a total analysis period of 100 milliseconds. Time-varying parameters within this time window are extracted. Data within a specific interval is extracted using a time index, enabling refined analysis of local time periods and avoiding computational redundancy and spurious correlations caused by searching the entire time period. If the values ​​of attitude angular velocity or attitude angular acceleration also show local maxima within the time observation window, indicating a synchronous peak, the operating condition at that time point and the acoustic peak are treated as a whole and marked as a correlated event. A synchronous peak refers to the temporal overlap between the acoustic signal peak and the dynamic operating condition characteristic peak. Its physical meaning is that the occurrence of abnormal noise and the drastic change in the intermediate shaft attitude are synchronized in time, suggesting a possible causal relationship between the two, rather than accidental independent events. Correlated events are data structures that bind the operating condition and acoustic peak as a whole. Through the constraint of the time window, operating condition-acoustic feature pairs with temporal correlation are selected to provide samples for subsequent statistical analysis.

[0043] The operating condition feature vector is a column vector formed by arranging multiple operating condition parameters in a fixed order. It has six dimensions (speed, torque, pitch angle, roll angle, yaw angle, and attitude angular acceleration). Through structured encoding, complex transient operating conditions are compressed into computable and comparable mathematical objects. The Pearson correlation coefficient is calculated between the feature vectors of all related events. The Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two variables. The input to this correlation coefficient is two feature vectors, and the output is a value between -1 and 1. The closer the absolute value is to 1, the stronger the linear correlation. The covariance of the two vectors is calculated and divided by the product of their standard deviations to quantify the linear dependence of the vectors in terms of their changing trends after eliminating the influence of dimensions. A strong correspondence means that two related events have a high degree of linear similarity in operating condition features, suggesting that they may be triggered by similar operating conditions, providing a stable sample basis for pattern recognition. Using each associated event as a node in the graph, and selecting strong correspondences as edges connecting the nodes, the weights of the edges are assigned the corresponding Pearson correlation coefficient values, thus generating a correspondence graph reflecting the intrinsic relationship between different dynamic operating conditions and abnormal noise occurrences. The correspondence graph is a graph-structured data type where nodes represent associated events, edges represent strong correspondences between events, and edge weights are correlation coefficients. It visualizes and structures complex operating condition-abnormal noise relationships, revealing clustering trends and transmission patterns among events through the graph's topological properties.

[0044] A frequent subgraph mining algorithm is used to traverse the generated correspondence graph, identifying recurring node connection topologies and defining them as recurring correspondence patterns. The frequent subgraph mining algorithm is a computational method used to find subgraph structures in large amounts of graph data whose frequency exceeds a specific support threshold. Its inputs are the correspondence graph and the support threshold, and its output is the frequently occurring subgraph pattern. The working principle is to discover stable, recurring structural patterns in the data through graph isomorphism detection and frequency statistics. The support threshold is the minimum frequency standard for determining the frequency of a subgraph. This represents a stable combination of abnormal noise conditions in multiple cyclic tests, excluding accidental environmental interference. For each recurring correspondence pattern, the timestamps of all associated events are extracted. Vibration acceleration sensor data corresponding to these timestamps are retrieved from the synchronized time-series dataset, and the root mean square (RMS) value of the vibration signal within the corresponding time window is calculated. If the RMS value exceeds the set vibration benchmark threshold, the abnormal noise is determined to be accompanied by significant mechanical vibration. The proportion of events accompanied by significant vibration under this mode is statistically analyzed, and this proportion is used as the mode reliability score. If the mode reliability score is higher than a preset reliability threshold, it is confirmed that the corresponding mode reflects real mechanical structural interference or friction, and the corresponding relationship after judgment is output. The mode reliability score is a probabilistic index that measures the physical authenticity of the corresponding mode. Through cross-validation of multiple physical quantities, pure acoustic interference is filtered out to ensure that the output corresponding relationship focuses on the real stress and abnormal noise of the chassis components.

[0045] For example, Figure 2 This diagram illustrates the correlation profile, showing six related events (E1~E6) and their Pearson correlation coefficients. Nodes represent different related events, and the width and grayscale of the edges indicate the magnitude of the correlation coefficient (wider edges and darker colors indicate stronger correlations). For example, the correlation coefficient between E1 and E2 is 0.92, indicating a strong correlation and suggesting highly similar dynamic operating conditions. In contrast, the correlation coefficient between E3 and E4 is only 0.55, indicating a weaker correlation. This diagram allows for the intuitive identification of clusters of strongly correlated events.

[0046] Specifically, the sequence of associated events of the chassis components during the stress deformation process is extracted from the determined correspondence. Vibration signals of the chassis components are collected, and the vibration data spectrum is obtained as the fusion input for subsequent processing. The vibration data spectrum is a representation of the energy distribution of the vibration signal in the frequency domain. Spectral analysis can reveal the frequency composition and resonance characteristics of abnormal noises. Fourier transform is applied to the acquired vibration data to calculate the spectral characteristics. Fourier transform is a mathematical process that converts a continuous vibration signal in the time domain into different frequency components and their corresponding amplitudes in the frequency domain. Through integration, the signal is decomposed into a superposition of sine waves of different frequencies, thereby extracting specific frequency components hidden in the complex vibration waveform. The calculated spectral characteristics are aligned with the extracted sequence of associated events on the time axis. Using the timestamp of the attitude data as an anchor point, the vibration spectral characteristics within each preset time period before and after the anchor point are bound to the spatial arrangement angle of the robotic arm at that instant. This ensures that the frequency of abnormal noise at a certain moment can accurately correspond to the specific torsional posture, thus determining the frequency domain representation of the abnormal noise characteristics. Frequency domain representation is a mathematical description of abnormal sound characteristics in the frequency dimension. It redistributes time-domain energy to the frequency axis through orthogonal transformation, revealing the periodic structure and resonance characteristics of the signal.

[0047] The determined frequency domain representation is determined by the numerical distribution of the values ​​under a specific torque-speed-angle parameter combination experienced by the chassis components. This specific torque-speed-angle parameter combination is a parameter division method that discretizes continuous mechanical parameters into interval segments. If this distribution shows a peak clustering phenomenon, it is identified as an occurrence mode. Peak clustering refers to a state in which the amplitude of a specific frequency is significantly higher than the surrounding frequencies within a specific torque or speed range, and this high-amplitude phenomenon occurs densely when the mechanical parameters are repeated multiple times. Its physical meaning is that the occurrence of abnormal noise has a stable triggering relationship with specific mechanical conditions, rather than random noise. Identifying this peak clustering phenomenon that occurs stably under specific mechanical parameters as an occurrence mode successfully establishes a precise mapping relationship between dynamic operating conditions and abnormal noise characteristics, allowing the accurate localization of transient abnormal noise phenomena that were previously difficult to separate.

[0048] Based on the identified occurrence patterns, their dependence on different mechanical parameters and attitude angles is calculated. Parameter dependence measures the sensitivity of the probability of abnormal noise occurrence to parameter changes, and is usually quantified by statistically analyzing the frequency of abnormal noise occurrence within different parameter ranges. A pattern parameter mapping is generated between the occurrence patterns and key parameters based on the magnitude of the dependence. The pattern parameter mapping is a mapping table or function describing the dependence between occurrence patterns and key mechanical parameters. Statistical analysis methods are used to quantify the influence weight of different parameters on abnormal noise occurrence, forming a queryable and predictable mapping relationship. Acoustic signals, vibration signals, and robotic arm attitude data are fused, and the generated pattern parameter mapping is cross-validated. Cross-validation is a method that uses independent datasets to test the model's generalization ability. Its inputs are the pattern parameter mapping and the validation dataset, and the output is the validation accuracy. The working principle is to divide the data into training and validation sets, using the training set to construct the mapping and the validation set to test the predictive ability of the mapping, avoiding overfitting. If the cross-validation accuracy meets the preset conditions, the final determined occurrence pattern is output.

[0049] Step S4: Perform cluster analysis on the determined occurrence patterns using the feature association algorithm to obtain the association clusters between transient conditions and abnormal noises; construct a traceability matrix for the association clusters to obtain a complete abnormal noise mapping model.

[0050] Furthermore, the correlation clusters between transient operating conditions and abnormal noises are obtained, including: The determined occurrence patterns are extended in the time domain, and multiple sub-patterns are extracted using a sliding window. The waveform similarity between each sub-pattern and the overall occurrence pattern is calculated. Based on the waveform similarity, the multiple sub-patterns are clustered and divided. The sub-pattern grouping results are used as the time domain refinement features of the occurrence patterns to form an enhanced occurrence pattern set.

[0051] Clustering feature vectors are extracted from the set of enhanced generation modes. The clustering feature vectors include acoustic frequency amplitude, vibration acceleration peak value and time domain refinement features. The distance between vectors is calculated using a feature association algorithm and clustering is performed to generate initial cluster groups.

[0052] Transient operating condition constraints are applied to the initial cluster group. The spatial arrangement angle of the intermediate axis and the system speed and torque data aligned with the feature vectors in the initial cluster group on the same time axis are obtained. The variance of the transient operating condition parameters corresponding to each feature vector in the initial cluster group is calculated. If the variance is lower than the preset operating condition fluctuation threshold, the adjacent initial cluster groups with similar operating conditions are merged. The Mel frequency cepstral coefficients of the acoustic and vibration signals in the merged cluster group are extracted as anomalous features. The cluster group boundary is redefined according to the Mahalanobis distance of the covariance matrix. The anomalous features are fused to optimize the cluster group boundary and determine the final associated cluster group.

[0053] Furthermore, a complete abnormal noise mapping model is obtained, including: Cluster elements are obtained from the associated clusters to construct a traceability matrix. Each row of the traceability matrix represents a transient operating condition, characterized by normalized intermediate shaft pitch, roll, yaw angles, and speed and torque parameters. Each column represents an abnormal noise characteristic mode, characterized by normalized acoustic loudness, sharpness, wave intensity psychoacoustic parameters, specific resonance band energy proportion, root mean square value of vibration acceleration in each axis, peak factor, and envelope spectrum characteristic frequency amplitude. The mutual information between each transient operating condition and each abnormal noise characteristic mode is calculated based on the row and column positions and used as the correlation value to fill the matrix, thus obtaining the traceability matrix.

[0054] The traceability matrix is ​​decomposed into a potential operating condition pattern matrix and a potential abnormal noise feature pattern matrix using a non-negative matrix factorization method. The reconstruction error is minimized by an iterative optimization algorithm to extract the potential mapping. The consistency between the potential mapping and the original cluster group is judged. If the consistency is higher than the preset consistency threshold, an initial mapping model is generated. The model parameters are calibrated by integrating multi-source data to determine the calibrated complete mapping model.

[0055] The transient operating conditions are simulated by a complete mapping model to obtain the matching degree between the simulated output and the actual abnormal noise. The matrix elements are iteratively optimized based on the matching degree. Singular value decomposition is performed on the optimized matrix to extract the dominant feature vector. The dominant feature vector is then integrated into the mapping model to generate an enhanced model.

[0056] Physically based constraints are imposed on the enhanced model. These constraints include the model's predicted abnormal noise loudness monotonically increasing with torque and the vibration energy at a specific frequency increasing after the spatial attitude angle exceeds a preset threshold. After applying the constraints, the enhanced model is validated using test data covering various extreme and boundary conditions to determine its robustness under the constraints. If the model's robustness meets the requirements, a complete abnormal noise mapping model is output.

[0057] Specifically, the occurrence pattern is extended in the time domain by extending it forward and backward on the time axis. The time domain extension refers to adding a fixed-length time interval in both the past and future directions, with the core time point corresponding to the occurrence pattern as the center, so as to include the complete physical process before and after the abnormal sound is generated, and avoid missing the key information of energy accumulation and decay by only analyzing the peak moment.

[0058] A sliding window refers to setting a fixed time interval for data extraction and moving it sequentially from left to right across the expanded time-domain data according to a set step size. Each move extracts data within the window as a sub-pattern. This method can divide a complex dynamic anomaly process into multiple continuous, fine-grained transient slices, effectively overcoming the analysis challenges posed by the continuous movement of robotic arms.

[0059] Waveform similarity is a measure of the similarity in shape between two signal waveforms. It quantifies the degree of matching in waveform shape by calculating the point-to-point difference or correlation between the two signals. The value is between -1 and 1. It is calculated by dividing the dot product of two vectors by the product of their magnitudes, which reflects the cosine of the angle between the vectors. The closer the value is to 1, the more consistent the direction and the more similar the waveforms.

[0060] Clustering is a method of grouping data objects based on similarity. By using a similarity threshold or clustering algorithm, highly similar sub-patterns are aggregated into the same group, ensuring that sub-patterns within the same group have similar waveform characteristics, representing similar stages in the evolution of abnormal sounds. The sub-pattern grouping result provides a refined description of the temporal structure within a single occurrence pattern, revealing the dynamic evolution of abnormal sounds from their generation to their dissipation. The sub-pattern grouping results are then appended to the corresponding occurrence patterns to form an enhanced occurrence pattern set. This enhanced set enriches the descriptive dimensions of the occurrence patterns by adding refined features, enabling subsequent clustering analysis to simultaneously consider the frequency domain characteristics and temporal evolution of abnormal sounds, thus improving classification accuracy and physical interpretability.

[0061] Clustering feature vectors are extracted from multiple occurrence patterns in the enhanced occurrence pattern set. These feature vectors are data point vectors used for cluster analysis, and their dimensions include acoustic frequency amplitude, peak vibration acceleration, and the aforementioned temporal refinement features. The multidimensional features of each occurrence pattern are compressed into a computable mathematical object, giving the distance metric between different patterns a clear geometric meaning. In this embodiment, Euclidean distance is used as the feature association algorithm. Its input is two clustering feature vectors, and its output is a distance value, quantifying the degree of difference between feature vectors through spatial distance.

[0062] K-means clustering is an unsupervised clustering algorithm based on distance partitioning. It uses the rate of error descent to determine the optimal number of clusters. Before the inflection point, the error decreases rapidly, and after the inflection point, the rate of error descent slows down, meaning that increasing the number of clusters yields diminishing returns. K cluster centers are randomly initialized, and each feature vector is assigned to the nearest cluster center to form an initial cluster. The cluster center positions are iteratively updated until the change in cluster center positions is less than a preset convergence threshold or the maximum number of iterations is reached, thus generating the initial clusters. Its characteristic is that data points within the same cluster are close in the feature space, representing similar abnormal sound patterns. However, there is a possibility that similar sounds under different operating conditions may be misclassified into the same cluster.

[0063] The robot arm spatial arrangement angle and system rotation speed and torque data, aligned with the feature vectors within the initial cluster on a unified time axis, are obtained as transient operating condition constraints. These constraints impose operating condition limitations on the clustering results. By statistically analyzing the dispersion of operating condition parameters, it ensures that data within the same cluster originates from similar dynamic operating conditions, avoiding confusion between similar sounds under different operating conditions due to reliance solely on acoustic features. The variance of the transient operating condition parameters corresponding to each feature vector within the initial cluster is calculated, serving as a statistic measuring the degree of data dispersion. The input to this statistic is the operating condition parameters of each feature vector within the cluster; a smaller variance indicates more concentrated data and more similar operating conditions. If the variance is below a preset operating condition fluctuation threshold, the operating conditions within that cluster are considered highly similar, and adjacent initial clusters with similar operating conditions are merged. The preset operating condition fluctuation threshold is the upper limit of variance for determining operating condition similarity; a variance below this threshold indicates a high degree of consistency in operating conditions.

[0064] Mel-frequency cepstral coefficients (MFCCs) are an acoustic feature extraction method that simulates the auditory characteristics of the human ear. By using a Mel filter bank to simulate the human ear's sensitivity to different frequencies, the signal undergoes a series of processing steps including pre-emphasis, framing, windowing, fast Fourier transform, Mel filtering, logarithmic compression, and discrete cosine transform to extract perceptually meaningful spectral features, effectively characterizing the auditory quality of abnormal sounds. The mean and covariance matrix of all Mel-frequency cepstral coefficients within the merged cluster are calculated, and the joint variability between each dimension is determined to characterize the overall distribution shape of the data. Using the mean as the center and the Mahalanobis distance of the covariance matrix as the boundary, noisy data points outside the boundary are removed, thus accurately locking the set of abnormal sound features under specific distortion postures and determining the final associated clusters. Mahalanobis distance is the standardized distance between a data point and the distribution mean. The associated clusters are the final clustering results after working condition constraints and abnormal feature optimization. Their characteristics are that the data in the same cluster not only have similar acoustic features, but also come from similar transient working conditions. At the same time, free noise is excluded, and they can reliably represent the specific working condition-abnormal noise association pattern.

[0065] Specifically, a cluster element is a set of data points in an associated cluster formed after cluster analysis and operating condition constraint optimization. Each data point contains complete transient operating condition parameters and corresponding abnormal noise characteristic parameters.

[0066] Key parameters characterizing the transient torsional state of chassis components are extracted from the associated clusters. Three Euler angles—pitch, roll, and yaw—relative to the vehicle body of the intermediate shaft are selected as spatial attitude descriptions, while the real-time speed and torque values ​​of the drive system are selected as dynamic state descriptions. The pitch angle is the rotation angle of the intermediate shaft about the vehicle's lateral axis, reflecting the degree of forward and backward tilt; the roll angle is the rotation angle of the intermediate shaft about the vehicle's longitudinal axis, reflecting the degree of left and right tilt; and the yaw angle is the rotation angle of the intermediate shaft about the vehicle's vertical axis, reflecting the degree of horizontal deflection. These three Euler angles, together with speed and torque, constitute a six-dimensional operating condition parameter space, comprehensively describing the operating state of the intermediate shaft at a specific moment. After normalization, these parameters are grouped according to their physical meaning to form row representation vectors, with each row corresponding to a comprehensive transient operating condition state. The normalization process is a method of converting data of different dimensions and magnitudes to a unified numerical range.

[0067] Key indicators characterizing the auditory properties and vibration patterns of abnormal sounds are extracted from associated clusters. Psychoacoustic parameters such as loudness, sharpness, and wave intensity, as well as the energy proportion of specific resonant frequency bands, are extracted from acoustic signals. Loudness, a psychoacoustic parameter reflecting the perceived volume of a sound, is calculated by simulating the differences in human ear sensitivity to different frequencies and the nonlinear response of sound pressure levels, resulting in a volume index consistent with subjective hearing. Sharpness, a psychoacoustic parameter reflecting the harshness of high-frequency sounds, is quantified by weighting the energy proportion of high-frequency components to determine the subjective perception of harshness or sharpness. Wave intensity, a psychoacoustic parameter reflecting the degree of amplitude fluctuation in sound, is quantified by analyzing modulation frequency and modulation depth to determine the subjective perception of sound fluctuations. The root mean square (RMS) values ​​of vibration acceleration along each axis, peak factor, and characteristic frequency amplitudes in the envelope spectrum are extracted from the vibration signal. The RMS value, a statistical measure of the energy magnitude of the vibration signal, is calculated by taking the square root of the mean of the squared values ​​of the signal to reflect the overall energy level of the vibration. The peak factor is an indicator measuring the impact characteristics of a vibration signal, defined as the ratio of the peak value to the root mean square value. It identifies the impact component in the signal by comparing extreme peak values ​​with the average energy level. The envelope spectrum characteristic frequency amplitude is obtained by extracting the vibration signal envelope using Hilbert transform and then performing spectral analysis on the envelope. Demodulation analysis identifies the periodic impact component in the vibration signal, revealing the fault characteristic frequencies of components such as bearings and gears. These indicators are also normalized to form a list of characteristic vectors, with each column corresponding to a comprehensive abnormal noise or vibration characteristic mode.

[0068] Based on the established row-list feature structure, the correlation strength between each transient operating condition and each abnormal noise feature pattern is calculated. The correlation value is filled by calculating the mutual information of the corresponding operating condition data sequence and the feature data sequence after time alignment. Mutual information is an information-theoretic measure of the degree of interdependence between two random variables. It quantifies the amount of information shared by the two variables by calculating the relative entropy between the product of the joint probability distribution and the marginal probability distribution. A larger mutual information value indicates a stronger statistical dependence and a higher degree of correlation. If a certain operating condition and a certain abnormal noise feature show a strong statistical correlation in multiple test periods, the correlation value at the corresponding position in the matrix is ​​high. The correlation value is filled by calculating mutual information to obtain a traceability matrix. This traceability matrix is ​​a complete data table recording the correlation between operating conditions and abnormal noises. The scattered operating condition-abnormal noise correspondences are integrated into a structured matrix form through statistical correlation measures.

[0069] Let the tracing matrix be denoted as V, and it can be decomposed into the product of two non-negative matrices W and H, namely: The column vectors of matrix W can be interpreted as basic latent operating condition patterns, and the row vectors of matrix H can be interpreted as basic latent abnormal noise characteristic patterns. By constraining the non-negativity of the decomposition results, the latent patterns acquire interpretable physical meaning, avoiding the negative component problem that may occur in traditional principal component analysis, as negative operating condition patterns or abnormal noise characteristics are physically difficult to interpret. The iterative optimization algorithm minimizes V and... The reconstruction error between them works by ensuring non-negativity through a multiplication update rule while adjusting matrix elements along the descent direction of the reconstruction error, so that the product... The original matrix V is gradually approximated. After decomposition, the original, high-dimensional, and potentially sparse traceability matrix V is represented as a linear combination of a few latent patterns (W and H). This helps to discover deep-seated relationships between operating conditions and features that are not directly observed but collectively cause abnormal noises. The latent mapping is a low-dimensional essential structure extracted from high-dimensional observational data. Through non-negative constrained matrix decomposition, the complex operating condition-abnormal noise correlation is decoupled into interpretable latent factors, revealing the physical mechanisms hidden behind the data.

[0070] Calculate the latent mapping relationship obtained from the decomposition, i.e. The consistency between the reconstructed matrix and the original traceability matrix V is measured, for example, by calculating the cosine similarity of the two across all elements. If the calculated similarity is higher than a preset consistency threshold, the extracted latent mapping is considered to have effectively preserved the key associations between the original data clusters. Matrix W and H are then used as parameters of the initial mapping model to generate the initial mapping model. The initial mapping model is a mathematical model that defines a linear mapping relationship from the latent operating condition space to the latent abnormal noise feature space. Through a linear combination of latent pattern matrices, quantitative prediction from operating condition parameters to abnormal noise features is achieved.

[0071] A calibration set was created using a portion of test data acquired synchronously but not used to construct the initial traceability matrix. This calibration set contained complete time-series data, including robotic arm posture angle sequences, speed and torque sequences, high-sampling-rate acoustic signal sequences, and vibration signal sequences. The operating condition data in the calibration set were converted into operating condition vectors according to the row index definition, and the corresponding abnormal noise feature vectors were predicted using the initial mapping model. Calibration is the process of optimizing model parameters using independent datasets. By comparing the differences between model predictions and actual observations, the model parameters were adjusted to better reflect the actual physical characteristics. The predicted abnormal noise feature vectors were compared with the actual abnormal noise feature vectors extracted from the calibration set, the prediction error was calculated, and the least squares method was used to fine-tune the elements in matrices W and H based on the prediction error. After calibration, the model parameters better reflect the characteristics of the current test bench and the tested component, thus determining the calibrated complete mapping model. The complete mapping model is a data-validated mathematical model that establishes a quantitative correlation between dynamic operating condition parameters and abnormal noise feature parameters. Its characteristics include high prediction accuracy, strong physical interpretability, and reliable applicability for operating condition simulation and abnormal noise prediction.

[0072] In historical test data, the actual test segment closest to the input working condition is found, and the actual abnormal noise feature vector of this segment is extracted. The matching degree between the predicted feature vector and the actual feature vector is calculated, which can be measured using weighted Euclidean distance or cosine similarity. The higher the matching degree, the more accurate the model's prediction for this type of working condition. If the matching degree is lower than expected, it indicates that the model's predictive ability under this type of working condition is insufficient. Based on the matching degree error, gradient descent is used to make small adjustments to the local elements in the tracing matrix V corresponding to this type of working condition and associated abnormal noise features. Gradient descent is an optimization algorithm that updates parameters in the opposite direction of the gradient of the error function. By calculating the partial derivative of the error with respect to the parameters, the direction and step size of parameter updates are determined, gradually approaching the minimum error. Then, starting from the adjusted matrix V, the process of matrix decomposition and parameter calibration is repeated, but the iteration range can be limited to the affected potential modes to achieve local optimization of the model and iterative optimization of matrix elements.

[0073] Singular Value Decomposition (SVD) is a matrix factorization method that decomposes a matrix into a product of three matrices. The decomposition form is as follows: ,in, It is a left singular vector matrix. It is a singular value diagonal matrix. The SVD decomposition is the transpose of the right singular vector matrix. The input is the tracing matrix V, and the output is three decomposition matrices. Through orthogonal transformations, the matrix is ​​decomposed into a rotation-scaling-rotation form, extracting the dominant change directions in the data. The magnitude of the singular values ​​reflects the importance of each direction. The dominant eigenvectors are the basis vectors explaining the main directions of variation in the data. Their input is the SVD decomposition result, and the output is the dimensionality-reduced eigenbase. By selecting the directions corresponding to large singular values, the main trends in the data are preserved, and the influence of noise and minor fluctuations is suppressed.

[0074] The first k left singular vectors (the first k columns of U) are extracted and used as the new working condition basis vectors, and the first k right singular vectors ( The first k rows of the singular vectors are used as new anomalous feature basis vectors, forming a more stable, dimension-reduced mapping space. The enhanced model projects the original operating condition data onto this subspace spanned by the left singular vectors, and then uses the singular value matrix... The scaling is mapped to a singular feature subspace spanned by the right singular vectors. This processing enhances the model's ability to capture the main trends in the data and suppresses the influence of noise, generating an enhanced model. The enhanced model is an optimized mapping model that incorporates the dominant feature vectors, improving the model's stability and generalization ability through dimensionality reduction projection and feature enhancement.

[0075] Constraints are engineering physics knowledge transformed into mathematical inequalities or boundary conditions to limit the reasonable range of model output. The physical basis for the monotonically increasing loudness of abnormal noise with increasing torque is that increased torque means increased load on the intermediate bearing, leading to intensified structural deformation and friction, thus increasing, rather than decreasing, the noise energy. The physical basis for the jump in vibration energy at a specific frequency after the spatial attitude angle exceeds a preset threshold is that under extreme torsional posture, the non-uniform transmission characteristics of the universal joint intensify, causing the excitation of vibration modes at specific frequencies. These constraints, in the form of inequalities or boundary conditions, are integrated into the model evaluation process to ensure that the model output conforms to known physical laws and avoids counterintuitive predictions.

[0076] After applying constraints, the enhanced model is validated using extensive test data covering various extreme and boundary conditions, including maximum torque, maximum speed, extreme attitude angles, and rapid attitude switching. The aim is to comprehensively verify the model's performance within its operating range. The model's predictions under all these conditions are checked to ensure they meet the aforementioned physical constraints, while also assessing whether its prediction accuracy remains within an acceptable range of fluctuation. Model robustness measures the model's ability to maintain prediction accuracy and physical consistency under different input conditions. Stress testing verifies the model's stability and reliability. If the model maintains high prediction accuracy and stability under constraints, meaning its robustness meets engineering analysis requirements, the model is considered reliable, and a complete anomaly mapping model is output for subsequent analysis. The complete abnormal noise mapping model is a comprehensive data model that integrates multi-source data association, latent pattern mining, parameter calibration, feature enhancement, and physical constraint verification. Its key feature is that it can accurately, reliably, and interpretably associate and map the spatial attitude and mechanical parameters of dynamically changing chassis components simulated by a multi-axis robotic arm in the laboratory with acoustic and vibration abnormal noise characteristics collected at a high sampling rate. This provides a core analytical tool for quickly locating the specific working conditions and underlying mechanisms of abnormal noise. At the same time, it supports adjusting the robotic arm control parameters based on the model output to achieve the reproduction and verification of abnormal noise conditions.

[0077] Step S5: Based on the abnormal noise mapping model, adjust the control parameters of the multi-axis robotic arm to simulate and reproduce the transient working conditions that generate abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

[0078] Furthermore, adjust the control parameters of the multi-axis robotic arm, including: Based on the specific torque-speed-angle parameter combination output by the abnormal noise mapping model, a motion trajectory command for the robotic arm is generated; the trajectory command is sent to the multi-axis robotic arm controller to drive the intermediate axis to move along the reproduced trajectory; acoustic signals and vibration data during the reproduction process are collected and compared with the original abnormal noise characteristics to verify the reproduction accuracy.

[0079] Specifically, by querying the complete abnormal noise mapping model and inputting the target abnormal noise feature or target operating condition, the model outputs the corresponding specific torque-speed-angle parameter combination. This specific torque-speed-angle parameter combination is a transient operating condition parameter predicted or inverted by the abnormal noise mapping model that can trigger a specific abnormal noise. It includes the torque setpoint and speed setpoint of the drive system, as well as the three-dimensional spatial arrangement angles of the intermediate shaft (pitch angle, roll angle, yaw angle). This parameter combination can be obtained in two ways: forward query and backward inversion. Forward query refers to inputting known operating condition parameters, and the model outputs the predicted abnormal noise feature; backward inversion refers to inputting the target abnormal noise feature and using an optimization algorithm to search for the operating condition parameter combination that can generate that feature.

[0080] The motion trajectory commands of a robotic arm are sequences of commands describing the position, velocity, and acceleration of each joint over time. The inputs are operational parameters (intermediate axis attitude, rotational speed, torque) in Cartesian or task space, and the outputs are angle, angular velocity, and angular acceleration commands in joint space. The working principle is to convert the end effector's pose requirements into drive commands for each joint through inverse kinematics calculations, and then generate smooth and continuous motion commands through trajectory planning algorithms. Inverse kinematics is a mathematical problem of solving for joint angles from the end effector's pose. Its input is the end effector pose (position + attitude), and the output is the joint angle solution. The working principle is to establish a mapping relationship between the end effector pose and joint angles through geometric relationships or numerical iterations, and solve for the joint configuration that satisfies the pose requirements. Trajectory planning algorithms are methods for generating smooth trajectories that satisfy kinematic and dynamic constraints. Commonly used algorithms include polynomial interpolation, spline curve planning, and trapezoidal velocity curve planning. The inputs are the pose, velocity, and acceleration constraints of the start and end points, and the output is a time-parameterized trajectory function. The working principle is to describe the changes in joint variables over time through mathematical functions, ensuring that the trajectory is continuous, smooth, and satisfies physical constraints.

[0081] The generated trajectory commands are sent to the multi-axis robotic arm controller via a communication interface. This interface is a physical or logical channel connecting the host computer and the lower-level controller, using a standardized communication protocol to ensure orderly data transmission and error checking. The multi-axis robotic arm controller parses the commands and drives each servo motor to move along a predetermined trajectory, reproducing the transient working condition that caused the abnormal noise on the intermediate shaft. The multi-axis robotic arm controller receives higher-level commands and converts them into lower-level motor control signals. Through motion control algorithms, it interpolates the trajectory commands into high-frequency motor position or torque commands, combining this with servo control loops (position loop, speed loop, current loop) to achieve precise motion tracking. The servo motors are the driving devices that execute the trajectory motion, converting electrical energy into mechanical energy through electromagnetic induction. Combined with encoder feedback, they achieve closed-loop control, ensuring that the actual motion matches the commands.

[0082] During the reproduction process, acoustic and accelerometer sensors are activated simultaneously, acquiring acoustic signal sequences and vibration data streams at the same sampling rate and precision as the original acquisition. Synchronous acquisition ensures time alignment of multi-sensor data; this is achieved through hardware triggering or software synchronization mechanisms, ensuring that each sensor begins sampling at the same moment, guaranteeing data temporal consistency. Abnormal noise characteristics from the acquired acoustic signals and vibration data are extracted. This extraction follows the same feature calculation process as the original analysis, including calculating spectral characteristics using Fast Fourier Transform, extracting specific frequency amplitudes, and calculating psychoacoustic parameters (loudness, sharpness, vibration intensity), etc. The same signal processing flow ensures the comparability of the reproduced characteristics with the original characteristics.

[0083] For example, Figure 3This is a schematic diagram comparing the spectra of the original abnormal noise and the reproduced abnormal noise. Figure 3 In this study, the acoustic signals collected during the reproduction process were subjected to spectral analysis and compared with the original abnormal noise spectrum. The black solid line represents the original abnormal noise spectrum, and the dark gray dashed line represents the reproduced abnormal noise spectrum. The amplitude distribution of the two in the main resonance frequency bands around 850Hz and 2100Hz is highly consistent, and the peak frequency deviation is less than 5Hz. This proves that the torque-speed-angle parameter combination output by the abnormal noise mapping model can drive the multi-axis robotic arm to accurately reproduce the original transient working condition.

[0084] The above describes a method for simulating abnormal noise from an automotive intermediate shaft according to an embodiment of this application. The following describes a system for simulating abnormal noise from an automotive intermediate shaft according to an embodiment of this application. Please refer to [link to relevant documentation]. Figure 4 One embodiment of a working condition simulation system for abnormal noise from an automotive intermediate shaft, as described in this application, includes: The acquisition unit is used to acquire multi-source data through the multi-axis robotic arm controller and sensor interface to obtain an initial heterogeneous dataset.

[0085] The data processing unit is used to apply a unified clock reference to the initial heterogeneous dataset using a timestamp synchronization algorithm to determine the alignment point of each data stream at a preset precision. If the deviation value of the alignment point exceeds the preset deviation threshold, the missing timing sequence is filled by interpolation method to obtain the synchronized timing dataset.

[0086] The feature analysis unit is used to extract dynamic operating condition features representing the intermediate shaft based on the synchronized time-series dataset, determine the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals, and mark the corresponding peak values ​​as associated events if the correspondence shows the synchronized peak values; obtain the marked associated events and fuse the vibration data spectrum to determine the occurrence mode of abnormal noise features under specific mechanical parameters.

[0087] The model building unit is used to perform cluster analysis on the determined occurrence patterns through feature association algorithms to obtain the association clusters between transient conditions and abnormal noises; a traceability matrix is ​​constructed for the association clusters to obtain a complete abnormal noise mapping model.

[0088] The simulation analysis unit adjusts the control parameters of the multi-axis robotic arm according to the abnormal noise mapping model to simulate and reproduce the transient working conditions that produce abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

[0089] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the method for simulating abnormal noise from an automotive intermediate shaft.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0091] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for simulating abnormal noise from an automotive intermediate shaft under operating conditions, characterized in that, The method includes: The initial heterogeneous dataset is obtained by collecting multi-source data through a multi-axis robotic arm controller and sensor interface. A timestamp synchronization algorithm is used to apply a unified clock reference to the initial heterogeneous dataset to determine the alignment point of each data stream at a preset precision. If the deviation of the alignment point exceeds a preset deviation threshold, the missing time sequence is filled by interpolation to obtain the synchronized time sequence dataset. Based on the synchronized time-series dataset, dynamic operating condition features representing the intermediate shaft are extracted. The correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals is determined. If the correspondence shows a synchronized peak value, it is marked as an associated event. The marked associated events are obtained and the vibration data spectrum is fused to determine the occurrence mode of abnormal noise features under specific mechanical parameters. The determined occurrence patterns are clustered using a feature association algorithm to obtain association clusters between transient conditions and abnormal noises; a tracing matrix is ​​constructed for the association clusters to obtain a complete abnormal noise mapping model. Based on the abnormal noise mapping model, the control parameters of the multi-axis robotic arm are adjusted to simulate and reproduce the transient working conditions that generate abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

2. The method according to claim 1, characterized in that, The initial heterogeneous dataset is obtained, including: The attitude adjustment records of the intermediate axis are collected by the multi-axis robotic arm controller, and the acoustic signal sequence and vibration data stream are obtained from the sensor interface; the attitude adjustment records and the acoustic signal sequence are fused together and combined with the vibration data stream to form multi-source data; Preliminary time-series labeling is performed on multi-source data to determine the differences in collection frequency among various data sources. The data sampling rate is adjusted based on the differences in collection frequency to generate an initial heterogeneous dataset in a unified format. Source identification labels are applied to the initial heterogeneous dataset. The rate of change of spatial angle between adjacent time points in the attitude adjustment record is calculated. Time points whose rate of change of spatial angle exceeds a preset angle change threshold are selected as key event points. If the key event points correspond to the acoustic signal sequence, they are marked as potential abnormal noise intervals. The potential abnormal noise range is verified by vibration data stream, the peak features within the range are extracted, and the peak features are integrated into the initial heterogeneous dataset to obtain the initial heterogeneous dataset containing multimodal information.

3. The method according to claim 1, characterized in that, The synchronized time-series dataset includes: The original device timestamps of each data stream are extracted from the initial heterogeneous dataset. The time offset is calculated using a timestamp synchronization algorithm based on the network time protocol. The attitude is aligned and the recording and acoustic signal sequences are adjusted according to the time offset. The timestamps of the vibration data stream are fused to generate a time axis under a globally unified clock reference. A preset precision calibration is applied to the time axis to determine the alignment point. If there is a time difference between the alignment points, the intermediate value is estimated by linear interpolation to obtain the calibrated alignment sequence. The matching degree between the calibrated alignment sequence and the original data stream is determined, and the synchronization parameters are optimized based on the matching degree to determine the final alignment point. If the deviation of the determined alignment point exceeds the preset deviation threshold, the missing time series interval is identified, and the filling value is calculated using the cubic spline interpolation method to generate the interpolated time series chain. Vibration data spectrum constraints are applied to the time series chain, vibration signals synchronized with the filling time period are extracted and subjected to fast Fourier transform to obtain characteristic frequency amplitudes, and it is determined whether the characteristic frequency amplitudes fall within the theoretical resonance frequency range to determine the accuracy of the filling and obtain the synchronized time series dataset.

4. The method according to claim 1, characterized in that, Determine the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals, including: Simultaneously extract attitude angle sequence, speed sequence, and torque sequence from the synchronized time-series dataset; The attitude angle sequence is calculated by first-order difference to obtain attitude angular velocity, and the attitude angular acceleration is calculated by second-order difference. The attitude angular velocity and the attitude angular acceleration are used as time-varying parameters and fused with attitude angle, rotational speed and torque to form dynamic operating condition characteristics. Based on the characteristics of dynamic operating conditions, acoustic signal peak data are integrated to determine the peak occurrence time point; a time observation window of preset duration is established with the peak occurrence time point as the center, and the correspondence between the local maximum value of the time-varying parameter and the acoustic signal peak value within the time observation window is determined. If the correspondence shows a synchronous peak value, it is marked as an associated event. Based on the characteristic vectors of the associated events, the Pearson correlation coefficient between the vectors is calculated, and strong correspondences are filtered through the correlation threshold to generate a correspondence map. The correspondence map is then subjected to pattern recognition to determine the recurring corresponding patterns. Vibration data is fused to verify the reliability of the corresponding patterns. If the reliability is higher than the preset reliability threshold, the determined correspondence is output.

5. The method according to claim 1, characterized in that, Determine the occurrence mode of abnormal noise characteristics under specific mechanical parameters, including: The associated event sequence is extracted from the correspondence after judgment, the vibration data spectrum is obtained as the fusion input, the spectral features are calculated by Fourier transform, the spectral features are aligned with the correspondence, and the frequency domain representation of the abnormal noise features is determined. The frequency domain representation is determined under a specific combination of torque-speed-angle parameters. If the distribution shows peak clustering, it is identified as an occurrence mode. Based on the occurrence pattern analysis parameter dependency, a pattern parameter mapping is generated. Multi-source data is fused to verify the accuracy of the mapping. If the accuracy meets the conditions, the determined occurrence pattern is output.

6. The method according to claim 1, characterized in that, The correlation clusters between transient operating conditions and abnormal noises are obtained, including: The determined occurrence pattern is extended in the time domain, and multiple sub-patterns are extracted using a sliding window. The waveform similarity between each sub-pattern and the overall occurrence pattern is calculated. The multiple sub-patterns are clustered based on the waveform similarity, and the sub-pattern grouping results are used as the time domain refinement features of the occurrence pattern to form an enhanced occurrence pattern set. Clustering feature vectors are extracted from the enhanced generation mode set. The clustering feature vectors include acoustic frequency amplitude, vibration acceleration peak value and time domain refinement features. The distance between vectors is calculated using a feature association algorithm, and initial clusters are generated by clustering. Apply transient operating condition constraints to the initial cluster group, obtain the intermediate axis spatial arrangement angle and system speed and torque data that are aligned with the feature vectors in the initial cluster group on the same time axis, calculate the variance of the transient operating condition parameters corresponding to each feature vector in the initial cluster group, and if the variance is lower than the preset operating condition fluctuation threshold, merge the adjacent initial cluster groups with similar operating conditions. Mel frequency cepstral coefficients of acoustic and vibration signals within the merged clusters are extracted as anomalous features. The cluster boundaries are redefined based on the Mahalanobis distance of the covariance matrix. The anomalous features are fused to optimize the cluster boundaries, and the final associated clusters are determined.

7. The method according to claim 1, characterized in that, Obtain the complete abnormal noise mapping model, including: Cluster elements are obtained from the associated clusters to construct a traceability matrix. The traceability matrix is ​​decomposed into a potential operating condition pattern matrix and a potential abnormal noise feature pattern matrix using a non-negative matrix factorization method. The reconstruction error is minimized through an iterative optimization algorithm to extract the potential mapping. Determine the consistency between the potential mapping and the original cluster group. If the consistency is higher than a preset consistency threshold, generate an initial mapping model. Integrate multi-source data to calibrate the model parameters and determine the calibrated complete mapping model. The complete mapping model is used to simulate transient operating conditions, obtain the matching degree between the simulated output and the actual abnormal noise, and iteratively optimize the matrix elements based on the matching degree; singular value decomposition is performed on the optimized matrix to extract the dominant feature vector, and the dominant feature vector is integrated into the mapping model to generate an enhanced model; Physically based constraints are applied to the enhanced model. After applying the constraints, the enhanced model is validated using test data covering various extreme and boundary conditions to determine the robustness of the model under the constraints. If the robustness of the model meets the requirements, the complete abnormal sound mapping model is output.

8. The method according to claim 1, characterized in that, Adjusting the control parameters of the multi-axis robotic arm includes: Based on the specific torque-speed-angle parameter combination output by the abnormal noise mapping model, a motion trajectory command for the robotic arm is generated; the trajectory command is sent to the multi-axis robotic arm controller to drive the intermediate axis to move along the reproduced trajectory; acoustic signals and vibration data during the reproduction process are collected and compared with the original abnormal noise characteristics to verify the reproduction accuracy.

9. A system for simulating abnormal noise from an automotive intermediate shaft, used to implement a method for simulating abnormal noise from an automotive intermediate shaft as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition unit is used to acquire multi-source data through the multi-axis robotic arm controller and sensor interface to obtain an initial heterogeneous dataset. The data processing unit is used to apply a unified clock reference to the initial heterogeneous dataset using a timestamp synchronization algorithm to determine the alignment point of each data stream at a preset precision; if the deviation value of the alignment point exceeds a preset deviation threshold, the missing timing sequence is filled by interpolation method to obtain the synchronized timing dataset. The feature analysis unit is used to extract dynamic operating condition features representing the intermediate shaft based on the synchronized time-series dataset, determine the correspondence between the extracted dynamic operating condition features and the peak values ​​of the acquired abnormal noise signals, and mark the corresponding peak values ​​as associated events if the correspondence shows the synchronized peak values; obtain the marked associated events and fuse the vibration data spectrum to determine the occurrence mode of abnormal noise features under specific mechanical parameters. The model building unit is used to perform cluster analysis on the determined occurrence pattern through a feature association algorithm to obtain the association clusters between transient conditions and abnormal noises; and to construct a tracing matrix for the association clusters to obtain a complete abnormal noise mapping model. The simulation analysis unit adjusts the control parameters of the multi-axis robotic arm according to the abnormal noise mapping model to simulate and reproduce the transient working conditions that produce abnormal noise, thereby realizing working condition simulation and abnormal noise analysis.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements a working condition simulation method for abnormal noise of the intermediate shaft of an automobile as described in any one of claims 1-8.

Images