A deep learning-based tire failure signal detection system and method

Abstract

The application discloses a kind of based on deep learning's tire failure signal detection system and method, it is related to tire vibration detection and fault diagnosis technical field, for solving the problem of low accuracy in tire test failure detection, early warning is not in time, through in rotary drum test bench deployment multichannel vibration sensor, acquisition time series signal and synchronous generation dynamic response sequence;Filtering and denoising are carried out on the edge side, and multi-scale energy curve is extracted, vibration energy mutation is analyzed to identify failure induction mode and probability distribution, combined with working condition label, the failure type and abnormal score are output by cloud deep learning model, and disposal strategy is generated according to risk assessment model;Sample is decoupled and recombined by characteristic space transformation, combined with virtual scene simulation synthesis virtual sample, adaptive augmentation of sample library is realized, accurate early identification and intelligent decision of tire failure are realized, and the accuracy and adaptive optimization ability of detection system are improved.

Description

Technical Field

[0001] This invention relates to the field of tire vibration detection and fault diagnosis technology, and more specifically, to a tire failure signal detection system and method based on deep learning. Background Technology

[0002] In high-speed tire durability testing and drum testing scenarios, tire vibration signals exhibit non-stationary and multi-scale characteristics due to multi-condition coupling. Traditional failure detection methods are mostly based on single-band energy thresholds or static template matching, simplifying fault judgment to vibration energy exceeding limits alarms with fixed thresholds. Furthermore, empirical formulas are used to describe the correspondence between vibration signals and failure types, without quantifying the dynamic correlation mechanism between multi-scale energy mutations and failure physical mechanisms.

[0003] In existing technologies, the detection model does not integrate the collaborative analysis capabilities of real-time feature extraction on the edge side and deep learning models in the cloud. It lacks a comprehensive judgment on the frequency, duration and multi-source working condition parameters of vibration energy mutations. This leads to a lag in the identification of early failure induction patterns and a large deviation in abnormal scoring. Consequently, the risk warning and handling strategies in the tire testing process deviate from the actual dynamic working conditions, exacerbating the risk of excessive tire wear or sudden failure. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a tire failure signal detection system and method based on deep learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A deep learning-based method for detecting tire failure signals includes the following steps:

[0007] In the high-speed operation of the drum and the test field environment, the vibration time sequence signals of multi-channel vibration sensors are collected, time-synchronized, and a vibration dynamic response sequence is generated and the working condition label is recorded.

[0008] Mechanical filtering and digital noise reduction are performed on the vibration dynamic response sequence at the edge to extract short-time energy, spectral band energy and envelope features, and multi-scale energy curves are calculated.

[0009] Based on multi-scale energy curve analysis, the frequency and duration of vibration energy mutations are analyzed to identify failure induction modes. Combined with working condition labels, the corresponding failure probability distribution is calculated. Through a cloud-based deep learning model, the corresponding failure type and anomaly score are output.

[0010] Based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated through a pre-set risk assessment model, and a response strategy is output.

[0011] Based on the handling strategy, the test bench was shut down or its load was reduced, and the multi-source monitoring parameters were archived;

[0012] Based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive, a sample augmentation and synthesis strategy is executed to supplement the sample size.

[0013] In a preferred embodiment, under high-speed drum operation and test environment, vibration time-series signals from multi-channel vibration sensors are acquired, synchronized in time, and a vibration dynamic response sequence is generated and a working condition label is recorded. Specific steps include:

[0014] Time synchronization calibration is performed using a preset synchronization signal to align the vibration timing signals of the multi-channel vibration sensor;

[0015] Under time synchronization conditions, the vibration time sequence signal is extracted according to the preset time window to generate a vibration dynamic response sequence, and the working condition label is recorded in each vibration dynamic response sequence.

[0016] Operating condition labels are markings that identify and describe the specific operating status and environmental conditions during the test, including speed, load, and temperature parameters.

[0017] In a preferred embodiment, mechanical filtering and digital noise reduction are performed on the vibration dynamic response sequence at the edge side to extract short-time energy, spectral band energy, and envelope features, and to calculate multi-scale energy curves. Specific steps include:

[0018] On the edge side, mechanical filtering and digital noise reduction techniques are used to denoise the vibration time-series signal of the vibration dynamic response sequence.

[0019] Short-time energy features representing the vibration state are extracted from the vibration time series signal. The spectral information of the vibration time series signal is analyzed, the spectral band energy features are extracted, and the envelope of the vibration time series signal is calculated and the envelope features are extracted.

[0020] Based on short-time energy characteristics, spectral band energy characteristics, and envelope characteristics, multi-scale energy curves are calculated and generated.

[0021] In a preferred embodiment, the specific steps for calculating and generating the multi-scale energy curve include:

[0022] An initial feature sequence is constructed based on short-time energy characteristics, spectral band energy characteristics, and envelope characteristics;

[0023] The initial feature sequence is decomposed into feature components using a signal decomposition method;

[0024] The energy distribution of each feature component at different time scales is calculated using an adaptive time window and presented as an energy sequence.

[0025] The energy sequences are combined and aligned in order of frequency from low to high to generate multi-scale energy curves.

[0026] In a preferred embodiment, the frequency and duration of vibration energy abrupt changes are analyzed based on multi-scale energy curves to identify failure induction modes. Combined with operating condition labels, the corresponding failure probability distribution is calculated. Then, a cloud-based deep learning model is used to output the corresponding failure type and anomaly score. Specific steps include:

[0027] Identify vibrational energy abrupt changes in multi-scale energy curves and identify vibrational energy abrupt change events.

[0028] The frequency of vibration energy mutation events was statistically analyzed, and the duration of each vibration energy mutation event was measured.

[0029] Based on the frequency and duration of vibration energy mutation events, failure induction modes are identified, including localized tire damage, tire cracks, tire deformation, tire peeling, and uneven tire wear.

[0030] The frequency and duration of different failure induction modes were statistically analyzed. Combined with the operating condition labels, the probability of each failure induction mode in all vibration energy mutation events was calculated to obtain the failure probability distribution.

[0031] The feature data extracted from the multi-scale energy curves, combined with the identified different failure induction modes and the calculated failure probability distribution, are used as input parameters and fed into the cloud-based deep learning model.

[0032] The cloud-based deep learning model outputs the corresponding failure category, determines the failure type, and calculates the anomaly score based on the input feature data, failure induction mode, and failure probability distribution.

[0033] In a preferred embodiment, identifying vibrational energy abrupt changes in a multi-scale energy curve and identifying vibrational energy abrupt change events includes the following steps:

[0034] Extract vibrational energy mutation sequences corresponding to different time scales from multi-scale energy curves;

[0035] Peak identification is used to detect peak values ​​in each vibration energy mutation sequence, and to identify the mutation point and mutation segment of vibration energy mutation.

[0036] The time period corresponding to each mutation point or mutation segment is defined, and the time period is regarded as the vibration energy mutation event.

[0037] In a preferred embodiment, based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated using a pre-set risk assessment model, and remedial recommendations are output. Specific steps include:

[0038] Different sensors collect multi-source monitoring parameters, including temperature, pressure, vibration characteristics, and ambient temperature;

[0039] The multi-source monitoring parameters are processed by feature extraction to form a feature vector of the multi-source monitoring parameters;

[0040] The feature vectors of failure type, anomaly score and multi-source monitoring parameters are input into the pre-established risk assessment model, and the output risk level assessment, potential failure probability and future risk change trend are generated.

[0041] Based on the risk level assessment, corresponding response strategies are automatically generated.

[0042] In a preferred embodiment, based on the handling strategy, the test bench is controlled to stop or be deloaded, and the multi-source monitoring parameters are archived. Specific steps include:

[0043] Based on the control instructions generated by the disposal strategy, receive control instructions in real time, including requirements to immediately stop the test bench operation or reduce the load.

[0044] The control commands are transmitted to the control unit of the test bench through the configured control interface, and the corresponding test bench shutdown or load adjustment program is started.

[0045] It automatically collects and archives multi-source monitoring parameters during the execution of control commands, while continuously monitoring the operating status.

[0046] In a preferred embodiment, based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution, and archives, a sample augmentation and synthesis strategy is executed to supplement the sample size. Specific steps include:

[0047] Based on failure type and anomaly score, the uniformity of sample distribution of various failure induction modes in the sample library is analyzed, and failure induction modes whose sample number proportion is lower than the preset sample number proportion threshold are identified as target failure induction modes that need to be sample augmented and synthesized.

[0048] Based on the target failure induction mode and failure probability distribution, extract the corresponding real samples from the archive as a reference base.

[0049] A sample augmentation strategy based on feature space transformation is adopted to decouple and reorganize the features of the reference basis and generate new samples with expanded feature dimensions.

[0050] The feature decoupling process separates the mixed features in the reference substrate into independent feature components through a reversible connection mechanism. The feature components include vibrational energy and spectral morphology.

[0051] The feature recombination process re-integrates and rearranges mutually independent feature components according to the physical mechanism of the target failure induction mode, and generates new samples through linear superposition, nonlinear fusion or cross-component weighted synthesis.

[0052] A sample synthesis strategy based on virtual scene simulation is adopted to simulate the inducing conditions and evolution path of target failure induction mode and synthesize virtual samples.

[0053] New samples and virtual samples are added to the sample library to supplement the number of samples for the target failure induction mode.

[0054] A deep learning-based tire failure signal detection system, used to implement the aforementioned deep learning-based tire failure signal detection method, includes:

[0055] The dynamic signal acquisition and synchronization module acquires vibration time-series signals from multi-channel vibration sensors in the high-speed operation of the drum and the test field environment, performs time synchronization, generates a vibration dynamic response sequence, and records working condition labels.

[0056] The signal preprocessing and feature extraction module performs mechanical filtering and digital noise reduction on the vibration dynamic response sequence at the edge, extracts short-time energy, spectral band energy and envelope features, and calculates multi-scale energy curves.

[0057] The mutation detection and failure mode identification module analyzes the frequency and duration of vibration energy mutations based on multi-scale energy curves, identifies failure induction modes, and calculates the corresponding failure probability distribution by combining working condition labels. Through a cloud-based deep learning model, it outputs the corresponding failure type and anomaly score.

[0058] The risk assessment and strategy formulation module, based on failure type, anomaly score and multi-source monitoring parameters, uses a preset risk assessment model to generate a risk level assessment and output a response strategy.

[0059] The data archiving and execution monitoring module controls the test bench to stop or reduce its load according to the handling strategy, and archives the multi-source monitoring parameters.

[0060] The sample augmentation and model optimization module, based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive, executes sample augmentation and synthesis strategies to supplement the number of samples.

[0061] The technical effects and advantages of this invention are as follows:

[0062] This invention achieves deep fusion of multi-source, multi-scale features in the field of tire vibration detection and fault early warning. By employing multi-scale energy curve extraction and intrinsic mode decomposition techniques, it fully captures the complex dynamic changes in vibration signals. Combined with a deep learning model, it performs multi-level failure mode identification and risk assessment, significantly improving the accuracy of fault identification and real-time response capability. Based on the combined effect of multi-source parameters, energy mutations, and failure probability distribution, it achieves early detection of fault precursors, reducing the false positive rate. Simultaneously, through automatic risk level classification and intelligent scheduling strategy output, it effectively optimizes tire maintenance resource allocation and provides early warning, thereby significantly improving the reliability and overall operational efficiency of the tire safety monitoring system. The innovative multi-scale feature decoupling and fusion mechanism enhances the sensitivity of fault identification and the control of misdiagnosis, providing a solid technical foundation for the refined management of tire health status, and has broad application prospects and practical promotion value. Attached Figure Description

[0063] Figure 1 This is a flowchart of a tire failure signal detection method based on deep learning according to the present invention.

[0064] Figure 2 This is a schematic diagram of the structure of a tire failure signal detection system based on deep learning according to the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1: As Figure 1 As shown, a tire failure signal detection method based on deep learning according to the present invention is presented, which includes the following steps:

[0067] In the high-speed operation of the drum and the test field environment, the vibration time sequence signals of multi-channel vibration sensors are collected, time-synchronized, and a vibration dynamic response sequence is generated and the working condition label is recorded.

[0068] Mechanical filtering and digital noise reduction are performed on the vibration dynamic response sequence at the edge to extract short-time energy, spectral band energy and envelope features, and multi-scale energy curves are calculated.

[0069] Based on multi-scale energy curve analysis, the frequency and duration of vibration energy mutations are analyzed to identify failure induction modes. Combined with working condition labels, the corresponding failure probability distribution is calculated. Through a cloud-based deep learning model, the corresponding failure type and anomaly score are output.

[0070] Based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated through a pre-set risk assessment model, and a response strategy is output.

[0071] Based on the handling strategy, the test bench was shut down or its load was reduced, and the multi-source monitoring parameters were archived;

[0072] Based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive, a sample augmentation and synthesis strategy is executed to supplement the sample size.

[0073] Step 1: Under the high-speed operation of the drum and in the test environment, collect vibration time-series signals from multi-channel vibration sensors, synchronize the signals, generate a dynamic vibration response sequence, and record the operating condition labels. The specific implementation is as follows:

[0074] In the high-speed operation of the drum and the test environment, during the acquisition and synchronization of tire vibration timing signals, a set of preset synchronization pulse signals is first generated using a specially designed synchronization signal generator. The synchronization pulse signals are characterized by a fixed frequency and period, with a relatively obvious and stable synchronization pulse in each period, which serves as a reference for time calibration. When the vibration timing signal acquisition begins, the multi-channel vibration sensor continuously monitors and records the vibration timing signal of each vibration sensor, while simultaneously detecting the synchronization pulse in the synchronization pulse signal. The position of the synchronization pulse in each channel is automatically identified, i.e., the marker point of each synchronization pulse signal, and the time deviation of each channel's marker point relative to the starting point of the synchronization pulse signal is calculated. By aligning the absolute time point of each channel's synchronization pulse with the known marker position in the synchronization pulse signal, the vibration timing signal of each channel is translated on the time axis using a time deviation correction algorithm, thereby achieving time alignment of the multi-channel vibration timing signals.

[0075] Next, the vibration timing signal is extracted according to the preset time window parameters. Specifically, the user sets the length and start position of the time window in advance, such as two seconds for each time window. Starting from the start point of the synchronized vibration timing signal, vibration response data is extracted once at fixed time intervals. In each calibrated channel, a continuous vibration timing signal is extracted at preset time intervals. The start time point of the extracted segment is determined according to the synchronization point of the synchronization pulse signal and the time window parameters. These extracted vibration response sequences cover different working conditions in the entire vibration acquisition process. In the vibration response sequence of each time period, the working condition label information at that time is also automatically recorded.

[0076] Operating condition labels are used to identify the operating status and environmental parameters during the test, including speed, load, temperature, etc. The operating condition label information comes from the real-time measurement data of the test equipment, which is integrated and stored in a one-to-one correspondence with the vibration response sequence.

[0077] Step 2: Perform mechanical filtering and digital noise reduction on the vibration dynamic response sequence at the edge side to extract short-time energy, spectral band energy, and envelope features, and calculate multi-scale energy curves. The specific implementation is as follows:

[0078] In the preprocessing of vibration time-series signals of the edge-side vibration dynamic response sequence, mechanical filtering is first used to filter the original vibration time-series signals of the vibration dynamic response sequence. Mechanical filtering is a filtering technique designed based on the characteristics of mechanical structures. It mainly introduces mechanical vibration filters or vibration isolation components of mechanical structures into the signal path to isolate and attenuate high-frequency noise and non-vibration-related interference signals, thereby obtaining a truly effective vibration time-series signal. Digital noise reduction technology uses digital signal processing methods to suppress high-frequency noise components and environmental noise appearing in the frequency domain through filter design, ensuring the clarity and signal-to-noise ratio improvement of the vibration time-series signal. Specific digital noise reduction algorithms include low-pass filtering, high-pass filtering, and band-pass filtering. The cutoff frequency and filtering bandwidth of the filter are designed and strictly set according to the characteristics of the collected vibration spectrum.

[0079] Based on filtering and noise reduction, short-time energy features representing the vibration state are extracted from the cleaned vibration time-series signal. Short-time energy features refer to the sum of energy of the vibration time-series signal within a continuous short time window. The calculation method is to square the vibration value of each sample and accumulate it within a preset fixed time window to reflect the energy amplitude of that time period. In order to analyze frequency domain information, the time domain signal is converted into spectrum data using fast Fourier transform, thereby obtaining the spectrum information of the vibration time-series signal. The spectrum band energy feature is calculated by segmenting the spectrum and calculating the sum of energy in each frequency band. The spectrum band energy feature reflects the energy distribution change of the vibration time-series signal in different frequency bands and is often used to distinguish different working conditions and identify abnormal vibration characteristics. In addition, envelope analysis technology is used to detect the envelope of the vibration time-series signal. The method is to calculate the absolute value of the vibration time-series signal, and then perform smoothing or low-pass filtering within a certain time window to extract the envelope line and obtain the envelope feature of the vibration time-series signal.

[0080] Step 3, the process of calculating and generating multi-scale energy curves, is specifically implemented as follows:

[0081] In the process of vibration time series signal processing, the short-time energy features, spectral band energy features, and envelope features extracted from the vibration time series signal are first constructed into an initial feature sequence. These three feature sequences capture the key performance of the vibration time series signal in terms of time, frequency, and amplitude changes: the short-time energy features can reflect the degree of energy concentration of the vibration time series signal within a short time window, the spectral band energy features show the energy distribution of the vibration time series signal in each frequency band, and the envelope features provide the overall amplitude profile of the vibration time series signal changes.

[0082] Next, signal decomposition methods are used to decompose the initial feature sequence into multiple feature components. Each feature component corresponds to different frequency components and energy changes of the signal. Commonly used decomposition methods include wavelet decomposition or empirical mode decomposition. These methods can perform multi-level and multi-band analysis on nonlinear and non-stationary signals in real time and extract representative feature components.

[0083] For each feature component that has been separated, an adaptive time window technique is used to calculate the energy distribution of each feature component at different time scales in order to obtain an accurate energy sequence. The adaptive time window dynamically adjusts the window length to adapt to the abrupt and slow variation characteristics of each feature component in the time dimension. The energy distribution within each adaptive time window is calculated and the energy distribution is statistically analyzed to form a detailed energy sequence of the feature component at multiple time scales.

[0084] Finally, the calculated energy sequence is combined and aligned in order from low frequency to high frequency according to the frequency corresponding to the characteristic component to generate a complete multi-scale energy curve. In the process of combination and alignment, the energy values ​​are matched step by step according to the frequency to ensure that the energy performance of each frequency band is connected in an orderly manner.

[0085] Step 4: Based on multi-scale energy curve analysis, analyze the frequency and duration of vibration energy abrupt changes, identify failure induction modes, and calculate the corresponding failure probability distribution by combining operating condition labels. Then, using a cloud-based deep learning model, output the corresponding failure type and anomaly score. The specific implementation is as follows:

[0086] By analyzing the occurrence of each vibration energy mutation event, the total number of vibration energy mutation events occurring within the overall monitoring time range is calculated, which is the frequency of vibration energy mutation events. At the same time, when each vibration energy mutation event occurs, the duration of the vibration energy mutation event is recorded, which is the time period from the beginning of the vibration energy mutation to the return to normal.

[0087] Next, based on the frequency and duration of vibration energy mutation events obtained from statistics, and combined with predefined failure induction modes, which include features such as local tire damage, cracks, deformation, peeling, and uneven wear, failure induction mode identification is performed for each vibration energy mutation event. The specific processing method is as follows: the feature parameters extracted from dynamic time-series signal analysis are matched with the failure induction mode feature template. Combined with the duration of the mutation and the frequency of the vibration energy mutation event, it is determined which failure induction mode the vibration energy mutation event belongs to. The frequency of occurrence of each failure induction mode is the number of times the mode appears in all mutation events, and the duration represents the time proportion of the failure induction mode in all events.

[0088] After statistically analyzing the total frequency, duration, and probability distribution of each type of failure-induced mode in the overall mutation events, and combining the operating condition label information, the characteristics of failure-induced modes under different environmental and operating conditions are analyzed for cause. The probability distribution of each failure-induced mode is obtained by dividing the frequency of occurrence of the failure-induced mode in all vibration energy mutation events by the total number of vibration energy mutation events. Statistical analysis is also performed on the duration and probability behavior of different failure-induced modes to obtain the failure probability distribution.

[0089] Using feature data extracted from multi-scale energy curves, failure induction modes, and failure probability distribution data as input parameters, the cloud-based deep learning model compares this data with the calibration data used in training to identify deviations and failure risks. In this process, the cloud-based deep learning model first calculates the overall deviation of the feature vector from the normal operating state. By comparing each feature data with the standard data, the cloud-based deep learning model accumulates the deviation values ​​of each feature based on the pre-trained weights and biases. Subsequently, the cloud-based deep learning model calculates the most likely failure type based on the matching degree of the failure induction mode, assigning an evaluation factor to each specific failure type, representing the relative probability and severity of each failure type. The evaluation factor is trained from historical failure data and, combined with the characteristics of the failure probability distribution, presents the risk of each failure type numerically.

[0090] Then, each feature deviation value is multiplied by its corresponding weight coefficient to form a weighted feature contribution value. The evaluation factors for each failure type are also weighted in the same way. Next, the products of all weighted feature deviation values ​​and evaluation factors are added one by one to obtain a total weighted sum, which is the anomaly score. To ensure the numerical stability and uniformity of the anomaly score, the anomaly score is linearly mapped or normalized to ensure that the anomaly score falls within a preset range, which is usually between 0 and 100.

[0091] In practical applications, assuming a high-speed train is in operation and collects continuous vibration data, after empirical mode decomposition, the system extracts energy characteristics in three frequency bands: 120 in the 20 Hz to 50 Hz band; 380 in the 500 Hz to 1000 Hz band; and 80 in the high-frequency band above 2000 Hz. Simultaneously, three potential failure-inducing modes are identified: tire cracking, localized damage, and uneven tread, which occurred 8, 10, and 12 times respectively in all abrupt events, with evaluation factors of 60%, 45%, and 50%. The energy values ​​and corresponding evaluation factors of each frequency band are weighted according to preset weighting coefficients. For example, the energy weight for the low-frequency band is 1, the mid-frequency band is 1.5, and the high-frequency band is 12. After summing all weighted values ​​and performing linear normalization, the resulting anomaly score is 75.

[0092] Step 5: Identify vibrational energy abrupt changes in the multi-scale energy curve, and identify vibrational energy abrupt change events. The specific implementation is as follows:

[0093] The energy change sequences corresponding to different frequency ranges extracted from the multi-scale energy curves are analyzed. First, the vibration energy mutation sequence of each frequency band is obtained by step-by-step calculation of the energy changes of the vibration time series signal at different time scales.

[0094] To detect abrupt changes, peak identification technology is used to detect local peaks in each vibration energy abrupt change sequence point by point, identify the peak points where the vibration energy changes drastically, and these peak points are the abrupt change points. The energy change trend between peak points is also observed, and the abrupt change segment is identified by combining the degree of amplitude jump.

[0095] In the identification of mutation regions, the occurrence time of each detected mutation point is first used as a benchmark, and a pre-set time window is extended forward and backward. For example, the time length of each mutation region is one second, and half a second is extended forward and half a second backward based on the time position of the mutation point. This ensures that the drastic changes in vibration energy within the mutation region are fully captured. All detected mutation regions will be used as the core data storage for vibration energy mutation events, providing a spatial basis for subsequent feature extraction and fault analysis. This process ensures that each mutation region, after the time is defined, represents a vibration energy mutation event.

[0096] For example, when monitoring the vibration of a high-speed train's tires, amplitude data is collected once per second. During one detection process, a sudden change in vibration energy is identified. This sudden change occurs at the 500th second on the time axis. To accurately capture the entire sudden change, this sudden change is used as a reference point, extending forward by 0.5 seconds to 499.5 seconds, and then extending backward by 0.5 seconds to 500.5 seconds. In this way, a sudden change region from 499.5 seconds to 500.5 seconds is defined, with a total duration of one second. The vibration signal within this sudden change region shows a change in energy that rapidly jumps from a normal state and then quickly falls back. The drastic change in vibration energy during this period is completely captured, forming a typical sudden change event in vibration energy.

[0097] Step 6: Based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated using a pre-set risk assessment model, and detailed handling recommendations are output. The specific implementation is as follows:

[0098] Multi-sensor synchronous acquisition of multi-source monitoring parameters is adopted. Specifically, these sensors are installed at key locations in tire testing, including pressure sensors, temperature sensors, vibration sensors inside the tire, and ambient temperature sensors outside the tire. These sensors collect temperature, pressure, vibration characteristics, and ambient temperature parameters in real time, forming multi-dimensional and dynamically changing real-time data information.

[0099] Next, feature extraction is performed on the collected multi-source monitoring parameters to form a feature vector representing the current state. The process of creating the feature vector uses data analysis algorithms to identify key parameters, such as vibration signal peak value, average temperature, maximum pressure, and ambient temperature change rate, and integrates them into a set of feature vectors. Each feature vector is normalized. The resulting multi-source monitoring parameter feature vector can reflect the dynamic state changes of the tire and its test environment.

[0100] The currently identified failure type, anomaly score, and feature vectors of multi-source monitoring parameters are input into a pre-established risk assessment model. The risk assessment model calculates the risk level assessment of the current tire condition through multi-dimensional analysis of the input data and combined with machine learning algorithms, predicts the possible failure probability and the risk change trend in the future. During the construction process, the risk assessment model is continuously adjusted by training data. The risk level, failure probability, and change trend are all derived by learning the relationship between historical data and current input data.

[0101] Finally, based on the risk level assessment results, corresponding handling strategies are automatically generated. In cases of low risk, it is recommended to continue monitoring and observe further. When the risk assessment level reaches the preset critical point, an emergency handling strategy is immediately generated and real-time commands are executed, such as immediately stopping the operation of the test bench to avoid further expansion of the risk, or controlling the test bench to adjust the load to restore it to a safe level. These strategies are implemented through the automated control of the device.

[0102] For example, in the tire management system of a large logistics company, the system, after monitoring and analysis, identified uneven wear as the tire failure type of a commercial truck, with an anomaly score of 80, indicating a high degree of abnormality. Simultaneously, feature vectors collected by multi-source monitoring sensors showed tire pressure of 600 kPa, average temperature of 45 degrees Celsius, peak vibration of 20 m / s², and ambient temperature of 28 degrees Celsius. These data, used as input parameters, were fed into a pre-established risk assessment model. Through comprehensive analysis of this multi-dimensional data and using machine learning algorithms, the risk assessment model calculated the current tire condition as high-risk and predicted a failure probability of 60%. Based on the analyzed trends, the risk assessment model predicted an increasing risk trend within the next week and recommended preventative measures. The training process of the risk assessment model utilized thousands of historical data points, continuously adjusting learning parameters and rules to ensure accurate responses to each condition.

[0103] Based on the output of the risk assessment model, the system generates corresponding handling strategies for truck tires. If the results indicate a high risk, the system recommends that the vehicle be taken out of service for tire inspection and plans to adjust the load distribution to reduce localized pressure. If abnormal vibrations or a rapid increase in temperature are detected during wheel rotation, more urgent actions will be automatically triggered, such as contacting a repair team for on-site repair or tire replacement. Through this intelligent decision-making, logistics companies can proactively address potential tire problems, reducing transportation delays and safety accidents caused by tire failures.

[0104] Step 7: According to the handling strategy, control the test bench to stop or reduce the load, and archive the multi-source monitoring parameters. The specific implementation is as follows:

[0105] During implementation, when a potential risk is detected in the tires or test bench, a control command for handling strategy is first generated. Upon receiving this control command, the command generated by the handling strategy is immediately transmitted to the control unit of the test bench through the configured control interface. The content of the control command may include immediately stopping the operation of the test bench or reducing the load operation requirements to deal with potential safety risks. These control commands are implemented through the equipped dedicated control interface and communication protocol, thereby activating the corresponding control program and completing the shutdown or load adjustment procedure of the test bench.

[0106] Throughout the operation, the test bench will automatically collect multi-source monitoring parameters during the execution of control commands, including key indicators such as temperature, pressure, vibration, and operating condition labels. These data are stored in real time to form a complete operation archive. The multi-source monitoring parameters in the archived data include the time point of this shutdown or load adjustment, the restored operating parameters, alarm information, and the execution status of control commands. All collected data are divided, organized, and stored in the system.

[0107] During operation, a real-time monitoring mechanism will be used to continuously monitor the operating status, including the degree of load reduction, changes in operating speed, temperature changes, and vibration and pressure indicators fed back by various sensors, to ensure that the control operation achieves the expected results. If any abnormal multi-source monitoring parameters exceed the preset safety range, an alarm will be issued in a timely manner, and backup emergency measures will be activated to ensure safety.

[0108] Step 8: Based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution, and archived data, execute a sample augmentation and synthesis strategy to increase the sample size. Specifically, the implementation is as follows:

[0109] In practice, the first step is to statistically analyze the sample distribution of various failure induction modes in the sample library based on the identified failure types and anomaly scores. Specifically, the sample size of each failure induction mode is calculated along with its proportion in the entire sample library to assess the uniformity of its distribution. The ratio of the sample size of each failure induction mode to the total sample size is used to determine the sample proportion value for each failure induction mode. The system sets a preset sample proportion threshold, such as 20%, as a detection standard. If the sample proportion of a certain failure induction mode falls below the preset threshold, it is considered that the sample size for that category is insufficient, and augmentation and synthesis of the target failure induction mode are performed to improve sample balance.

[0110] Based on the target failure induction mode and its failure probability distribution, corresponding real samples are extracted from archived data as reference bases. Archived data includes multi-source monitoring parameters recorded in historical tests, such as vibration time-series signals, temperature, pressure, and operating condition labels. During extraction, based on the target failure induction mode, all vibration time-series signals and associated features with the same failure induction mode are screened from the archive to ensure that the reference base is consistent with the target failure induction mode in terms of physical mechanism. For example, for the tire peeling mode, the short-time energy characteristics and spectral band energy characteristics of its corresponding vibration dynamic response sequence are extracted as reference bases.

[0111] The sample augmentation strategy for feature space transformation is implemented through a cloud-based deep learning model, mapping the original feature vectors of the reference basis to a high-dimensional latent space. The feature decoupling process uses a reversible connection mechanism, that is, the mixed features are separated into mutually independent feature components through an encoder network, including vibration energy components and spectral morphology components. The vibration energy component represents the intensity of the change in amplitude of the vibration time series signal over time, and the spectral morphology component represents the energy distribution profile of the vibration time series signal amplitude in the frequency domain. The feature recombination process is based on the physical mechanism of the target failure induction mode, and the mutually independent feature components after decoupling are reintegrated. For example, the vibration energy component and the spectral morphology component are combined according to weights by linear superposition, or the activation function is introduced by nonlinear fusion to achieve cross-component interaction, and finally new samples with expanded feature dimensions are generated.

[0112] The feature decoupling process specifically separates the mixed features in the reference substrate into independent feature components through reversible connections. The reversible connection mechanism adopts a symmetric neural network structure to ensure that the input mixed features can be reconstructed losslessly by the decoder after decoupling. The extraction of vibration energy components is based on short-time energy feature calculation, that is, the average value of the sum of squares of the signal in each frame is calculated after the vibration time-series signal is divided into frames. The extraction of spectral morphology components is based on spectral band energy features, that is, the energy integral of a specific frequency band is calculated after the vibration time-series signal is Fourier transformed. After decoupling, each feature component represents a different physical property of the signal.

[0113] For example, in practical applications, suppose a high-speed train is in operation. The vibration time-series signals collected by sensors are sent to a feature decoupling system. First, these vibration time-series signals are processed by frame segmentation, for example, extracting the signal once per second and calculating the sum of squares and average value of each frame to obtain the energy value of each frame. For example, within a certain time period, the sum of squares and average value of the vibration time-series signal in a certain frame is 350. This represents the basic characteristics of the vibration energy in that time frame. Next, a Fourier transform is performed on the complete vibration time-series signal to calculate the energy value of the signal in different frequency bands. For example, the energy is 200 in the 50 Hz to 100 Hz frequency band, 350 in the 100 Hz to 200 Hz frequency band, and 120 in the high-frequency band above 200 Hz. Using these data, the mixed feature signals are decoupled into two or more independent feature components under a reversible connection mechanism. Specifically, low-frequency vibration energy reflects the overall driving condition and larger amplitude vibration characteristics of the vehicle, while high-frequency energy reflects the condition of tire microcracks or minor local damage.

[0114] The feature recombination process reassembles and rearranges the independent feature components obtained after feature decoupling, based on the physical mechanism of the target failure induction mode, to generate a new sample. For the tire peeling induction mode, the physical mechanism is characterized by a sudden increase in high-frequency vibration energy and energy attenuation in a specific frequency band. During recombination, a cross-component weighted synthesis method is first adopted: a weighting coefficient of 0.7 is applied to the vibration energy component, and a weighting coefficient of 0.3 is applied to the spectral morphology component. The two weighted components are linearly superimposed to form a preliminary fused feature vector. On this basis, a nonlinear fusion mechanism is introduced. The linear superposition result is nonlinearly transformed by the sigmoid function. Utilizing the characteristic of this function to smoothly map any real number input to the interval between 0 and 1, the interaction intensity between the vibration energy and spectral morphology components is dynamically adjusted to simulate the nonlinear effects in the actual failure evolution process. Finally, the generated new sample is the feature vector after the above linear and nonlinear recombination.

[0115] A sample synthesis strategy based on virtual scene simulation is adopted. A digital twin model is constructed to simulate the induction conditions and evolution path of the target failure induction mode, thereby synthesizing virtual samples. The virtual scene simulation relies on a finite element analysis model, which is established based on the precise geometry and material parameters of the tire. This model can simulate the stress distribution, strain field, and thermodynamic behavior of the tire under specific operating conditions, such as a speed of 1200 RPM and a load of 3000 RPM. During the simulation, induction conditions are set according to the physical characteristics of the target failure induction mode, including pulse loads applied to the tire tread or temperature gradients acting on the internal structure of the tire. The simulation field is defined, and the evolution path of failure is defined. For example, the entire process of crack initiation from the tire shoulder area and extension towards the tire crown is simulated. The damage propagation state at each time step is calculated by the finite element model. When synthesizing virtual samples, the simulation parameters are dynamically adjusted according to the failure probability distribution. For example, the load amplitude or temperature gradient intensity is increased for high-probability failure induction modes, thereby generating the corresponding virtual vibration signal sequence. The final synthesized virtual sample is completely consistent with the real sample in terms of feature dimensions, but can contain more boundary cases that are difficult to obtain in real tests, such as failure waveforms under extreme high and low temperature environments or vibration characteristics under overload conditions.

[0116] The new samples generated by feature recombination and the virtual samples synthesized by virtual scene simulation are added to the sample library to supplement the number of samples of the target failure induction mode. Before adding, the failure induction mode of the new samples and virtual samples is verified to ensure that they correspond to the target failure induction mode. After the sample library is updated, the uniformity of sample distribution of various modes is recalculated and the augmentation and synthesis strategies are iteratively executed.

[0117] Example 2: A tire failure signal detection system based on deep learning, such as Figure 2 As shown, it specifically includes:

[0118] The vibration signal acquisition module is used to acquire vibration time-series signals from multi-channel vibration sensors in high-speed drum operation and test field environments, perform time synchronization, generate vibration dynamic response sequences, and record operating condition tags.

[0119] The signal preprocessing and extraction module is used to perform mechanical filtering and digital noise reduction on the vibration dynamic response sequence at the edge, extract short-time energy, spectral band energy and envelope features, and calculate multi-scale energy curves.

[0120] The failure mode identification module is used to analyze the frequency and duration of vibration energy mutations based on multi-scale energy curves, identify failure induction modes, and calculate the corresponding failure probability distribution by combining working condition labels. Through a cloud-based deep learning model, it outputs the corresponding failure type and anomaly score.

[0121] The strategy formulation module is used to generate a risk level assessment and output a response strategy based on the failure type, anomaly score and multi-source monitoring parameters through a preset risk assessment model.

[0122] The data archiving and monitoring module is used to control the test bench to stop or reduce the load according to the handling strategy, and to archive the multi-source monitoring parameters.

[0123] The sample supplementation module is used to supplement the number of samples by performing sample augmentation and synthesis strategies based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive.

[0124] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0125] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0126] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0128] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tire failure signal detection method based on deep learning, characterized in that, Includes the following steps: In the high-speed operation of the drum and the test field environment, the vibration time sequence signals of multi-channel vibration sensors are collected, time-synchronized, and a vibration dynamic response sequence is generated and the working condition label is recorded. Mechanical filtering and digital noise reduction are performed on the vibration dynamic response sequence at the edge to extract short-time energy, spectral band energy and envelope features, and multi-scale energy curves are calculated. Based on multi-scale energy curve analysis, the frequency and duration of vibration energy mutations are analyzed to identify failure induction modes. Combined with working condition labels, the corresponding failure probability distribution is calculated. Through a cloud-based deep learning model, the corresponding failure type and anomaly score are output. Based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated through a pre-set risk assessment model, and a response strategy is output. Based on the handling strategy, the test bench was shut down or its load was reduced, and the multi-source monitoring parameters were archived; Based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive, a sample augmentation and synthesis strategy is executed to supplement the sample size. Based on multi-scale energy curve analysis, the frequency and duration of vibration energy abrupt changes are analyzed to identify failure induction modes. Combined with operating condition labels, the corresponding failure probability distribution is calculated. Through a cloud-based deep learning model, the corresponding failure type and anomaly score are output. Specific steps include: Identify vibrational energy abrupt changes in multi-scale energy curves and identify vibrational energy abrupt change events. The frequency of vibration energy mutation events was statistically analyzed, and the duration of each vibration energy mutation event was measured. Feature parameters are extracted from vibration time-series signal analysis, and failure induction modes are identified by combining the frequency and duration of vibration energy mutation events. Failure induction modes include local tire damage, tire cracks, tire deformation, tire peeling, and uneven tire wear. The frequency and duration of different failure induction modes were statistically analyzed, and combined with the working condition labels, the probability of each failure induction mode in all vibration energy mutation events was calculated to obtain the failure probability distribution. The feature data extracted from the multi-scale energy curves, combined with the identified different failure induction modes and the calculated failure probability distribution, are used as input parameters and fed into the cloud-based deep learning model. The cloud-based deep learning model outputs the corresponding failure category, determines the failure type, and calculates the anomaly score based on the input feature data, failure induction mode, and failure probability distribution. The anomaly scoring calculation process includes: The feature data, failure induction mode and failure probability distribution data extracted from the multi-scale energy curve are used as input parameters and compared with the calibration data to identify deviations and failure risks. Based on the weights and biases of the cloud-based deep learning model after training, the bias values ​​of each feature are accumulated. All feature bias values ​​are weighted to obtain the feature deviation value. Each failure type is combined with the features of the failure probability distribution and weighted to obtain the evaluation factor. The products of all weighted feature deviation values ​​and evaluation factors are added one by one, and the resulting total weighted sum is used as the anomaly score.

2. The tire failure signal detection method based on deep learning according to claim 1, characterized in that: In a high-speed drum operation and test environment, vibration time-series signals from multi-channel vibration sensors are acquired, synchronized, and a dynamic vibration response sequence is generated. Operating condition labels are recorded. Specific steps include: Time synchronization calibration is performed using a preset synchronization signal to align the vibration timing signals of the multi-channel vibration sensor; Under time synchronization conditions, the vibration time sequence signal is extracted according to the preset time window to generate a vibration dynamic response sequence, and the working condition label is recorded in each vibration dynamic response sequence. Operating condition labels are markings that identify and describe the specific operating status and environmental conditions during the test, including speed, load, and temperature parameters.

3. The tire failure signal detection method based on deep learning according to claim 2, characterized in that: Mechanical filtering and digital noise reduction are performed on the vibration dynamic response sequence at the edge side to extract short-time energy, spectral band energy, and envelope features, and multi-scale energy curves are calculated. The specific steps include: On the edge side, mechanical filtering and digital noise reduction techniques are used to denoise the vibration time-series signal of the vibration dynamic response sequence. Short-time energy features representing the vibration state are extracted from the vibration time series signal. The spectral information of the vibration time series signal is analyzed, the spectral band energy features are extracted, and the envelope of the vibration time series signal is calculated and the envelope features are extracted. Based on short-time energy characteristics, spectral band energy characteristics, and envelope characteristics, multi-scale energy curves are calculated and generated.

4. The tire failure signal detection method based on deep learning according to claim 3, characterized in that: The specific steps for calculating and generating multi-scale energy curves include: An initial feature sequence is constructed based on short-time energy characteristics, spectral band energy characteristics, and envelope characteristics; The initial feature sequence is decomposed into feature components using a signal decomposition method; The energy distribution of each feature component at different time scales is calculated using an adaptive time window and presented as an energy sequence. The energy sequences are combined and aligned in order of frequency from low to high to generate multi-scale energy curves.

5. The tire failure signal detection method based on deep learning according to claim 4, characterized in that: Identifying vibrational energy abrupt changes in multi-scale energy curves and identifying vibrational energy abrupt change events involves the following steps: Extract vibrational energy mutation sequences corresponding to different time scales from multi-scale energy curves; Peak identification is used to detect peak values ​​in each vibration energy mutation sequence, and to identify the mutation point and mutation segment of vibration energy mutation. The time period corresponding to each mutation point or mutation segment is defined, and the time period is regarded as the vibration energy mutation event.

6. The tire failure signal detection method based on deep learning according to claim 5, characterized in that: Based on the failure type, anomaly score, and multi-source monitoring parameters, a risk level assessment is generated using a pre-set risk assessment model, and response recommendations are output. Specific steps include: Different sensors collect multi-source monitoring parameters, including temperature, pressure, vibration characteristics, and ambient temperature; The multi-source monitoring parameters are processed by feature extraction to form a feature vector of the multi-source monitoring parameters; The feature vectors of failure type, anomaly score and multi-source monitoring parameters are input into the pre-established risk assessment model, and the output risk level assessment, potential failure probability and future risk change trend are generated. Based on the risk level assessment, corresponding response strategies are automatically generated.

7. The tire failure signal detection method based on deep learning according to claim 6, characterized in that: According to the handling strategy, the test bench will be shut down or its load reduced, and the multi-source monitoring parameters will be archived. Specific steps include: Based on the control instructions generated by the disposal strategy, receive control instructions in real time, including requirements to immediately stop the test bench operation or reduce the load. The control commands are transmitted to the control unit of the test bench through the configured control interface, and the corresponding test bench shutdown or load adjustment program is started. It automatically collects and archives multi-source monitoring parameters during the execution of control commands, while continuously monitoring the operating status.

8. The tire failure signal detection method based on deep learning according to claim 7, characterized in that: Based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution, and archives, a sample augmentation and synthesis strategy is executed to increase the sample size. Specific steps include: Based on failure type and anomaly score, the uniformity of sample distribution of various failure induction modes in the sample library is analyzed, and failure induction modes with a sample number ratio lower than the preset sample number ratio threshold are identified as target failure induction modes that need to be sample augmented and synthesized. Based on the target failure induction mode and failure probability distribution, extract the corresponding real samples from the archive as a reference base. A sample augmentation strategy based on feature space transformation is adopted to decouple and reorganize the features of the reference basis and generate new samples with expanded feature dimensions. The feature decoupling process separates the mixed features in the reference substrate into independent feature components through a reversible connection mechanism. The feature components include vibrational energy and spectral morphology. The feature recombination process re-integrates and rearranges mutually independent feature components according to the physical mechanism of the target failure induction mode, and generates new samples through linear superposition, nonlinear fusion or cross-component weighted synthesis. A sample synthesis strategy based on virtual scene simulation is adopted to simulate the inducing conditions and evolution path of target failure induction mode and synthesize virtual samples. New samples and virtual samples are added to the sample library to supplement the number of samples for the target failure induction mode.

9. A deep learning-based tire failure signal detection system, used to implement the deep learning-based tire failure signal detection method according to any one of claims 1-8, characterized in that, include: The vibration signal acquisition module is used to acquire vibration time-series signals from multi-channel vibration sensors in high-speed drum operation and test field environments, perform time synchronization, generate vibration dynamic response sequences, and record operating condition tags. The signal feature extraction module is used to perform mechanical filtering and digital noise reduction on the vibration dynamic response sequence at the edge, extract short-time energy, spectral band energy and envelope features, and calculate multi-scale energy curves; The failure mode identification module is used to analyze the frequency and duration of vibration energy mutations based on multi-scale energy curves, identify failure induction modes, and calculate the corresponding failure probability distribution by combining working condition labels. Through a cloud-based deep learning model, it outputs the corresponding failure type and anomaly score. The strategy formulation module is used to generate a risk level assessment and output a response strategy based on the failure type, anomaly score and multi-source monitoring parameters through a preset risk assessment model. The data archiving and monitoring module is used to control the test bench to stop or reduce the load according to the handling strategy, and to archive the multi-source monitoring parameters. The sample supplementation module is used to supplement the number of samples by performing sample augmentation and synthesis strategies based on the failure type and anomaly score, combined with the identified failure induction mode, failure probability distribution and archive.

Images