A system and method for detecting the operating status of AMT bearings under active-passive switching conditions.

Abstract

This invention relates to the field of automotive automatic transmission technology, and discloses an AMT bearing operating status detection system and method under active / passive switching conditions. The method collects real-time data using vibration acceleration, current, and temperature sensors, as well as a speed encoder. After preprocessing such as envelope demodulation and Kalman filtering, a time-frequency feature matrix is ​​generated using variational mode decomposition and Hilbert transform. The matrix is ​​input into a probabilistic neural network model employing a sliding time window mechanism, which outputs a bearing health status probability value. A multi-parameter state assessment model optimized using a quantum genetic algorithm is constructed to obtain the optimal feature combination. Combined with a hierarchical diagnostic control model (including a data acquisition layer, analysis layer, and execution layer), a confidence rule base inference engine outputs warning levels and maintenance suggestions. The system also includes a signal verification module to ensure data reliability. This invention achieves multi-source data fusion and dynamic adaptive diagnosis, improving the accuracy and real-time performance of AMT bearing status detection under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of automotive automatic transmission technology, specifically to an AMT bearing operating status detection system and method under active / passive switching conditions. Background Technology

[0002] In the process of the automotive industry moving towards intelligence and automation, automated manual transmissions (AMTs) have been widely used due to their advantages such as high transmission efficiency and low cost. However, monitoring the operating status of their bearings under complex working conditions has always been a challenge for the industry. When AMTs switch between active and passive operating conditions (such as starting, shifting, and braking), the load, speed, and impact load on the bearings change drastically, making it difficult for traditional detection methods to effectively capture early fault characteristics, which can lead to transmission failure or even safety accidents.

[0003] In existing technologies, single-sensor detection methods (such as those relying solely on vibration signals) are easily affected by operating conditions and cannot comprehensively reflect the bearing's operating status. Diagnostic strategies based on fixed thresholds lack adaptability to real-time operating conditions, often resulting in false alarms or missed alarms during active / passive switching. Traditional feature extraction algorithms (such as Fourier transforms) have limitations when processing non-stationary signals, making it difficult to accurately characterize time-frequency domain feature changes caused by bearing faults. Furthermore, the high feature redundancy and low computational efficiency during multi-parameter fusion make it difficult for existing systems to meet the real-time and reliability requirements of AMT transmissions.

[0004] With the development of intelligent algorithms, some studies have attempted to apply neural networks to bearing fault diagnosis. However, traditional neural network models suffer from low parameter update efficiency under dynamic operating conditions, and their fixed activation thresholds for hidden layer nodes make them unable to adapt to the real-time load changes required by AMT bearings. Genetic optimization algorithms suffer from premature convergence in feature selection, making it difficult to achieve a globally optimal search in a multi-parameter space. Furthermore, in the reasoning process of confidence rule bases, fixed rule weights result in a lack of dynamic adjustment capabilities in the diagnostic strategy, hindering the full utilization of historical diagnostic data to optimize the diagnostic process.

[0005] At the data preprocessing level, existing methods are insufficient in extracting modulation features from vibration signals, current signals are susceptible to electromagnetic interference, and noise suppression of temperature and speed data is inadequate, resulting in low feature quality of the input model. In the hierarchical diagnostic system, the linkage between multi-level diagnostic strategies is insufficient, and there is a lack of a predictive mechanism for the evolution trend of bearing health status, making it difficult to trigger effective maintenance measures at the incipient stage of failure.

[0006] Existing technologies for detecting the operating status of AMT bearings suffer from problems such as insufficient fusion of multi-source data, inadequate feature extraction accuracy, weak model adaptability, and poor real-time diagnostic performance. There is an urgent need for an intelligent detection system and method that can adapt to active-passive switching conditions, fuse multi-sensor data, and have dynamic optimization capabilities to improve the reliability and service life of AMT transmissions. Summary of the Invention

[0007] The purpose of this invention is to provide an AMT bearing operating status detection system and method under active-passive switching conditions, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting the operating status of an AMT bearing under active-passive switching conditions, the method comprising: acquiring real-time operating data of the gearbox bearing through a multi-channel sensing module, the multi-channel sensing module including a vibration acceleration sensor, a current sensor, a temperature sensor, and a speed encoder; extracting intrinsic mode components from the real-time operating data based on a variational mode decomposition algorithm, and generating a time-frequency feature matrix through Hilbert transform; inputting the time-frequency feature matrix into a pre-trained probabilistic neural network model, the probabilistic neural network model employing a sliding time window mechanism, adjusting the activation threshold of hidden layer nodes based on real-time operating data, and outputting a bearing health status probability value; and based on the bearing health status probability... A multi-parameter condition assessment model is constructed using rate values. This model aims to maximize fault sensitivity and minimize feature redundancy. A quantum genetic optimization algorithm is employed to perform a global search on a subset of features, incorporating adaptive rotation angles and dynamic crossover probabilities. Based on this model, the optimal feature combination data is output. A hierarchical diagnostic control model is then established using this optimal feature combination data. This model comprises an acquisition layer, an analysis layer, and an execution layer. The acquisition layer updates the operating condition mode based on the time-frequency feature matrix. The analysis layer generates multi-level diagnostic strategies based on the bearing health status probability values. The execution layer maps diagnostic commands to warning levels and maintenance recommendation signals using a confidence rule base inference engine.

[0009] Preferably, the probabilistic neural network model employing a sliding time window mechanism includes: constructing a probabilistic network structure comprising vibration feature nodes, current feature nodes, and health state nodes, wherein the vibration feature nodes and the current feature nodes are parent nodes, and the health state nodes are child nodes; updating the probability density function of the vibration feature nodes based on real-time acquired data; performing frequency band energy analysis on the current feature nodes using a wavelet packet decomposition algorithm to obtain the posterior probability of the feature frequency band energy distribution; introducing a dynamic threshold adjustment mechanism to dynamically adjust the inference step size of the probabilistic network according to changes in bearing load, wherein the time window width is negatively correlated with the bearing rotational speed; and using an expectation-maximization algorithm to update the parameters of the probabilistic neural network online, optimizing the hidden layer node weight parameters by maximizing the likelihood function.

[0010] Preferably, the quantum genetic optimization algorithm introducing adaptive rotation angle and dynamic crossover probability includes: constructing a genetic optimization objective function, which includes a feature sensitivity term and a feature independence term. The feature sensitivity term is calculated using the inter-class distance of faulty samples, and the feature independence term is calculated using the correlation coefficient between feature vectors; designing an adaptive rotation angle update rule, where the rotation angle changes exponentially with the number of generations, maintaining a large search range in the initial stage and converging to the neighborhood of the optimal solution in the later stage; and designing a dynamic crossover probability adjustment mechanism, where the crossover probability is dynamically adjusted according to the population fitness variance. When the population variance is higher than a threshold, the frequency of crossover operations is increased; otherwise, the frequency of mutation operations is increased.

[0011] Preferably, the confidence rule base inferencer maps diagnostic instructions to warning levels and maintenance recommendation signals, including: constructing an input activation module, wherein the inputs include bearing health status probability values, temperature change gradients, and bearing operating time, and assigning confidence levels to the input variables using a Gaussian membership function; constructing a rule base containing confidence rules; designing an evidence fusion module, which uses DS evidence theory to fuse multi-source rule outputs into a comprehensive confidence distribution, wherein the outputs include warning level codes and maintenance recommendation codes; and introducing a rule weight correction mechanism to dynamically update the rule confidence coefficients based on historical diagnostic accuracy and maintenance records.

[0012] Preferably, the rule construction of the confidence rule base inferencer includes: establishing a fault mode knowledge base, which contains vibration spectrum feature templates for three typical faults: bearing pitting, cracking, and wear.

[0013] Preferably, the data preprocessing of the multi-channel sensing module includes: extracting fault characteristic frequency components from vibration acceleration data using an envelope demodulation algorithm; generating a spectrum diagram from current sensor data using a fast Fourier transform and suppressing background noise using an adaptive threshold filter; eliminating measurement fluctuations from temperature sensor data using an exponential weighted moving average algorithm; and fusing angular velocity and angular acceleration data from speed encoder data using a Kalman filter algorithm to output the instantaneous bearing speed and speed fluctuation rate.

[0014] Preferably, the analysis layer of the hierarchical diagnostic control model generates a multi-level diagnostic strategy including: dividing the health level into three levels: normal, warning, and fault; the normal level triggers periodic monitoring; the warning level triggers feature tracking monitoring; and the fault level triggers a shutdown protection protocol; constructing a state transition matrix, which performs state transitions based on continuous changes in the health level; if the health level decreases for three consecutive sampling periods, an upgraded diagnostic process is triggered; and introducing a trend prediction mechanism, which initiates an active testing mode when the rate of change of the bearing health status probability value exceeds a preset gradient threshold.

[0015] Preferably, the constraints of the multi-parameter state assessment model include: feature dimension constraint, limiting the maximum dimension of the feature vector to no more than the upper limit threshold of computing resource allocation; operating condition coverage constraint, verifying the operating condition adaptability of the feature combination through typical operating condition simulation tests; and real-time constraint, limiting the maximum delay of a single diagnostic process to no more than 20% of the control cycle.

[0016] Preferably, the method further includes: constructing a signal verification module to detect signal anomalies through multi-sensor data correlation analysis; and initiating data reconstruction if the correlation coefficient between a certain channel's data and the associated channel is lower than a set threshold.

[0017] Preferably, the present invention also includes an AMT bearing operating status detection system under active / passive switching conditions, the system comprising: a multi-channel sensing module for acquiring real-time operating data of the gearbox bearing, including a vibration acceleration sensor, a current sensor, a temperature sensor, and a speed encoder; a feature extraction module for extracting intrinsic mode components from the real-time operating data acquired by the multi-channel sensing module based on a variational mode decomposition algorithm, and generating a time-frequency feature matrix through Hilbert transform; and a state evaluation module for inputting the time-frequency feature matrix into a pre-trained probabilistic neural network model, which employs a sliding time window mechanism to adjust the activation threshold of hidden layer nodes based on real-time operating data. The system outputs the bearing health status probability value; the optimization decision module constructs a multi-parameter status assessment model based on the bearing health status probability value, with the optimization objectives of maximizing fault sensitivity and minimizing feature redundancy. It employs a quantum genetic optimization algorithm incorporating adaptive rotation angles and dynamic crossover probabilities to perform a global search on the feature subset, outputting the optimal feature combination data; the diagnostic execution module is established based on the optimal feature combination data, including an acquisition layer, an analysis layer, and an execution layer. The acquisition layer updates the operating condition mode based on the time-frequency feature matrix, the analysis layer generates multi-level diagnostic strategies based on the bearing health status probability value, and the execution layer maps diagnostic instructions into warning levels and maintenance suggestion signals through a confidence rule base inference engine.

[0018] Compared with existing technologies, the advantages of this invention are as follows: In the data acquisition and preprocessing stage, the multi-channel sensing module integrates vibration acceleration, current, temperature, and speed encoders, achieving comprehensive perception of the bearing's operating status. For data characteristics from different types of sensors, algorithms such as envelope demodulation, fast Fourier transform, exponentially weighted moving average, and Kalman filtering are used for preprocessing, effectively suppressing noise and extracting key features. For example, the envelope demodulation algorithm for vibration signals can enhance the frequency components of fault characteristics, and the adaptive threshold filtering for current signals can effectively suppress electromagnetic interference, ensuring that the feature data of the input model has a high signal-to-noise ratio and physical representation capability.

[0019] In the feature extraction and state assessment stages, the combination of variational mode decomposition (VMD) and Hilbert transform enables accurate extraction of intrinsic mode components from non-stationary operating data, generating a time-frequency feature matrix and overcoming the shortcomings of traditional Fourier transform in processing nonlinear signals. The probabilistic neural network model introduces a sliding time window mechanism, dynamically adjusting the activation threshold and inference step size of hidden layer nodes, allowing the model to adaptively update parameters based on real-time operating conditions (such as bearing speed and load changes). Dynamic updates of the probability density functions of vibration and current feature nodes, along with wavelet packet energy analysis, enhance the model's sensitivity to feature changes under active / passive switching conditions, providing a quantitative basis for subsequent assessment of the output bearing health state probability value.

[0020] The multi-parameter state assessment model aims to maximize fault sensitivity and minimize feature redundancy, employing a quantum genetic optimization algorithm that incorporates adaptive rotation angle and dynamic crossover probability. The adaptive rotation angle decays exponentially with each generation, balancing global search capability in the initial stage with local convergence accuracy in later stages. The dynamic crossover probability is adjusted based on the population fitness variance, avoiding premature convergence issues common in traditional genetic algorithms and ensuring the optimal feature combination is found in the high-dimensional feature space. This model simultaneously satisfies constraints on feature dimension, operating condition coverage, and real-time performance, improving diagnostic accuracy while reducing computational complexity, thus meeting the stringent real-time requirements of AMT transmissions.

[0021] The hierarchical diagnostic control model constructs a multi-layered diagnostic system through the collaborative work of the acquisition, analysis, and execution layers. The analysis layer classifies health levels into three levels: normal, warning, and fault. Combined with a state transition matrix and trend prediction mechanism, it achieves dynamic strategy switching from periodic monitoring to shutdown protection. A continuous decline in health level triggers an upgraded diagnostic process, and a probability value change rate exceeding a threshold initiates an active testing mode. This proactive design provides timely warnings in the early stages of a fault. The execution layer's confidence rule base inference engine integrates multi-source inputs (health status probability values, temperature change gradients, and runtime), achieves evidence fusion through DS evidence theory, and dynamically adjusts rule weights based on historical data, significantly improving the reliability of diagnostic conclusions and the relevance of maintenance recommendations.

[0022] The introduction of a signal verification module is another innovation of this invention. It detects signal anomalies through multi-sensor data correlation analysis. When the correlation coefficient between a certain channel's data and the associated channel is lower than a threshold, data reconstruction is initiated to ensure the reliability of the data source input to the system and avoid misdiagnosis caused by a single sensor failure.

[0023] This invention constructs a comprehensive testing system adapted to the active / passive switching conditions of AMT bearings by organically combining multi-source data fusion, intelligent algorithm optimization, dynamic threshold adjustment, and hierarchical diagnostic strategies. This solution not only solves the adaptability problem of traditional methods under non-stationary conditions but also improves diagnostic accuracy through feature optimization and model adaptation mechanisms. It provides advanced technical means for ensuring the reliability of AMT transmissions, demonstrating significant engineering application value and economic benefits. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the working principle of the AMT bearing operating status detection method under active / passive switching conditions described in this invention.

[0025] Figure 2 This is a diagram illustrating the working principle of the sliding time window mechanism in a probabilistic neural network model.

[0026] Figure 3 This is a design diagram for a confidence rule base inference engine. Detailed Implementation

[0027] 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.

[0028] Please see Figures 1-3 This invention relates to a method for detecting the operating status of an AMT bearing under active / passive switching conditions, which achieves accurate detection and diagnosis of the bearing's operating status through multi-module collaboration. Specifically, it includes the following steps:

[0029] Real-time operating data of the gearbox bearing is acquired through a multi-channel sensing module, which includes a vibration acceleration sensor, a current sensor, a temperature sensor, and a speed encoder. The sensors are respectively positioned in the gearbox bearing housing, the motor input circuit, the bearing cavity, and the end of the drive shaft. The vibration acceleration sensor uses a three-way mounting method to obtain complete vibration information; the current sensor uses a Hall effect element to monitor motor current fluctuations in real time; the temperature sensor uses a PT100 resistance temperature detector (RTD) element embedded in the bearing lubrication cavity; and the speed encoder is rigidly connected to the drive shaft via an incremental code disk.

[0030] The intrinsic mode components (EMCs) of real-time operating data are extracted using the Variational Mode Decomposition (VMD) algorithm, and a time-frequency feature matrix is ​​generated through Hilbert transform. Specifically, the VMD algorithm parameters are set as follows: number of modes K=6, quadratic penalty factor α=2000, and adaptive decomposition of each mode component is achieved through frequency domain masking and quadratic programming. The time-frequency feature matrix is ​​input into a pre-trained probabilistic neural network model. This model employs a sliding time window mechanism, adjusting the activation threshold of hidden layer nodes based on real-time operating data, and outputting the bearing health status probability value. The initial value of the time window width is set to 1024 sampling points. The bearing speed is calculated in real-time by a speed encoder, and the window width is dynamically adjusted according to a negative correlation (window width = 1024 × reference speed / real-time speed).

[0031] A multi-parameter condition assessment model is constructed based on the bearing health status probability values. With the optimization objectives of maximizing fault sensitivity and minimizing feature redundancy, a quantum genetic optimization algorithm is employed to perform a global search on the feature subset. During optimization, the initial population size is set to 50, and the number of generations is 100. An adaptive rotation angle and dynamic crossover probability mechanism are introduced to improve search efficiency. The final output contains the optimal feature combination data, including parameters such as vibration energy entropy, current harmonic distortion rate, and temperature gradient.

[0032] A hierarchical diagnostic control model is established based on the optimal feature combination data. This model includes a data acquisition layer, an analysis layer, and an execution layer. The data acquisition layer identifies the operating mode (switching between primary and passive operating modes) based on the time-frequency feature matrix. The analysis layer generates a three-level diagnostic strategy (normal, warning, fault) based on the health status probability value. The execution layer maps diagnostic commands to warning levels (Level I, Level II, Level III) and maintenance suggestion signals (continue operation, periodic inspection, immediate shutdown) through a confidence rule base inference engine.

[0033] The present invention will be further described below with reference to Examples 1 to 5:

[0034] Example 1: In the implementation of the sliding time window mechanism in the probabilistic neural network model, a probabilistic network structure is first constructed, including vibration feature nodes, current feature nodes, and health state nodes. The vibration feature nodes and current feature nodes serve as parent nodes, and the health state nodes serve as child nodes. The vibration feature nodes receive signal data collected by the vibration acceleration sensor, specifically extracting statistical features such as the root mean square value and kurtosis value of the vibration signal as input parameters. These parameters reflect the energy intensity and impact characteristics of the vibration signal. The current feature nodes receive current signal data collected by the current sensor, extracting features such as the fundamental component and harmonic energy of the current signal as input parameters. The fundamental component reflects the basic operating state of the current, while the harmonic energy reflects electromagnetic anomalies during motor operation. Based on the parameters input from the parent node, the health state node outputs a health probability value between 0 and 1 through internal network calculation and inference. This value characterizes the probability that the bearing is currently in a healthy state; a higher value indicates a better bearing health.

[0035] Based on the real-time acquired data, the probability density function of the vibration feature nodes needs to be updated. Specifically, kernel density estimation is used to divide the vibration data into several equidistant intervals according to certain rules, for example, 50 equidistant intervals. Then, for the sample points within each interval, the weighted sum of their Gaussian kernel functions is calculated. This generates a continuous probability density curve, which more accurately describes the probability distribution of vibration features, enabling the probabilistic neural network model to adapt to changes in the vibration signal in real time.

[0036] For current characteristic nodes, wavelet packet decomposition algorithm is required for frequency band energy analysis. First, the number of wavelet packet decomposition levels is determined; for example, setting it to 3 levels yields 8 characteristic frequency bands. Next, the energy distribution of each characteristic frequency band is calculated, and the posterior probability corresponding to each frequency band's energy is further calculated using Bayes' theorem. This approach decomposes the current signal into different frequency bands for analysis, extracting more diagnostically valuable frequency band energy features and providing richer information for assessing the bearing's health status.

[0037] To enable the probabilistic neural network model to better adapt to changes in bearing load, a dynamic threshold adjustment mechanism is introduced. Changes in bearing load can be characterized by the root mean square value of the current signal. When the load change rate exceeds a set threshold (e.g., 5%), it indicates a significant change in the bearing's operating state. In this case, the inference step size of the probabilistic network needs to be adjusted, shortening it from a longer time (e.g., 10ms) to a shorter time (e.g., 5ms) to more timely analyze and judge the bearing state.

[0038] During model operation, the expectation-maximization algorithm is used to update the parameters of the probabilistic neural network online. This algorithm calculates the maximum likelihood estimate of the hidden layer node weight parameters iteratively. Specifically, it first calculates the expectation of the data based on the current parameter estimates, and then updates the parameters by maximizing this expectation. This process is repeated continuously, allowing the model parameters to be continuously optimized based on real-time data, thereby improving the model's accuracy and adaptability.

[0039] The implementation of the sliding time window mechanism also involves the dynamic adjustment of the time window width. The time window width is negatively correlated with the bearing speed; that is, the higher the bearing speed, the narrower the time window width, and vice versa. In practical applications, the bearing speed data is first acquired in real time using a speed encoder. Then, the current time window width is calculated according to a preset negative correlation formula (e.g., the time window width equals the initial width multiplied by the reference speed and divided by the real-time speed; the initial width can be set to include a certain number of sampling points, such as 1024 sampling points). This dynamic adjustment method ensures that the time window can reasonably capture data under different speed conditions, guaranteeing both data timeliness and the ability to fully capture the changing characteristics of the bearing's operating state.

[0040] When constructing the probabilistic network structure, it is necessary to clarify the connection relationships and calculation methods between the nodes. Parent nodes (vibration feature nodes and current feature nodes) and child nodes (health status nodes) are connected through weights, the magnitude of which reflects the degree of influence of the parent node's features on the bearing's health status. During model training, these weights are learned and optimized using a large amount of sample data, enabling the model to accurately output reasonable health status probability values ​​based on the parent node's feature inputs.

[0041] During data processing, the raw data collected by vibration accelerometers and current sensors need to be preprocessed to remove noise and interference, thereby improving data quality. For example, vibration data can be processed by removing the mean and filtering to eliminate DC components and high-frequency noise; current data can be normalized to make current signals of different amplitudes comparable. The preprocessed data is then input into a probabilistic network structure for feature extraction and state assessment.

[0042] Furthermore, to ensure the stability and reliability of the sliding time window mechanism, the step size of the time window needs to be set appropriately. The step size is typically smaller than the width of the time window to ensure some overlap between data points and prevent the loss of important information due to data truncation. For example, if the time window width is 1024 sampling points, the step size can be set to 512 sampling points. This way, each time the time window is moved, half of the data will be new and half will be old, ensuring both real-time performance and data continuity.

[0043] During the online operation of the model, it is necessary to monitor the bearing's operating status in real time and issue early warnings or take corresponding measures in a timely manner based on changes in the health status probability value. For example, when the health status probability value is lower than a certain set threshold, it indicates that the bearing may have an abnormality and further diagnosis and analysis are required; when the probability value continues to decline, it may indicate the occurrence of bearing failure, and maintenance personnel need to be notified in a timely manner for inspection and repair.

[0044] Example 2: An adaptive rotation angle and dynamic crossover probability mechanism are introduced into the quantum genetic optimization algorithm to improve the efficiency and accuracy of global search of feature subsets, so as to construct a multi-parameter state evaluation model that satisfies the objectives of maximizing fault sensitivity and minimizing feature redundancy. The specific implementation is as follows:

[0045] First, a genetic optimization objective function is constructed, comprising two parts: a feature sensitivity term and a feature independence term. The feature sensitivity term is calculated using the inter-class distance of fault samples. The inter-class distance measures the degree of separation between samples of different fault categories in the feature space; a larger distance indicates higher sensitivity of the feature for fault classification. The feature independence term is calculated using the correlation coefficient between feature vectors. The correlation coefficient reflects the degree of linear association between features; a smaller coefficient indicates lower redundancy between features. By combining these two indicators, an optimization objective for the feature subset is formed, ensuring that the selected features can effectively distinguish fault categories while minimizing information overlap between features.

[0046] The design of adaptive rotation angle update rules needs to consider the needs of different stages of the evolutionary process. In the early stages of evolution, the population is not yet close to the optimal solution and requires a larger search range to explore a wider feature space. Therefore, a larger rotation angle is set in the initial stage, enabling the algorithm to perform a global search over a larger range and avoid getting trapped in local optima too early. As the number of generations increases, the population gradually converges towards the neighborhood of the optimal solution. At this point, the rotation angle is reduced, making the search step size finer, thus performing a more refined local search near the optimal solution and improving optimization accuracy. The change in rotation angle follows a specific curve pattern, such as an exponential decay curve with increasing number of generations. In this way, the search range is adaptively adjusted from global to local, balancing the breadth and depth of the search.

[0047] The core of the dynamic crossover probability adjustment mechanism is to dynamically adjust the frequency of crossover and mutation operations based on the population fitness variance. Population fitness variance reflects the degree of difference between individuals in the population. When the variance is higher than a set threshold, it indicates high diversity among individuals in the population, with many different feature combinations. In this case, the frequency of crossover operations is increased to combine the superior features of different individuals, further exploring new feature spaces and enhancing global search capabilities. When the variance is lower than the threshold, it indicates that the individuals in the population tend to be similar, diversity decreases, and the population is prone to getting trapped in local optima. In this case, the frequency of mutation operations is increased to introduce new feature combinations, breaking the stagnation of the population and improving local search accuracy. Through this dynamic adjustment method, the algorithm can automatically switch search strategies according to the actual state of the population, improving optimization efficiency.

[0048] In the specific implementation process, the basic parameters of the genetic algorithm need to be determined first, such as the initial population size and the number of generations. The initial population size can be set to a certain number of individuals, each representing a subset of features. For example, the initial population size can be set to 50 individuals. Each individual consists of a binary code indicating whether a feature is selected (e.g., 1 indicates that the feature is selected, and 0 indicates that it is not selected). The number of generations should be set to a sufficiently large value to ensure that the algorithm has enough time for searching and optimization. For example, the number of generations can be set to 100.

[0049] In each generation of evolution, the fitness value of each individual is first calculated. This is achieved by calculating the combined score of fault sensitivity and feature independence of the corresponding feature subset based on the genetic optimization objective function. Individuals with higher fitness values ​​have feature subsets that are closer to the optimization objective. Then, selection operations are performed based on the fitness values. Methods such as roulette wheel selection and tournament selection are used to select individuals with higher fitness from the population as parents, providing a foundation for subsequent crossover and mutation operations.

[0050] In the crossover operation, a dynamic crossover probability adjustment mechanism determines whether to perform a crossover operation on selected parent individuals. The crossover operation generates new offspring individuals by exchanging partial codes of two parent individuals, thus producing new feature combinations. For example, for two parent individuals with binary codes, single-point crossover or multi-point crossover can be used, exchanging their code fragments at randomly selected crossover points to generate two new offspring individuals. Since the crossover probability is dynamically adjusted based on the population fitness variance, a higher variance results in a higher crossover probability, leading to more parent individuals undergoing crossover and generating more new feature combinations; conversely, a lower variance results in a lower crossover probability, reducing the frequency of crossover operations and preventing overexploration that could decrease search efficiency.

[0051] Mutation operations randomly alter an individual's encoding, such as changing a 0 to a 1 or a 1 to a 0 in binary encoding, thereby introducing new features or removing selected features. The probability of mutation operations is usually set to a low value to avoid disrupting desirable feature combinations. However, when the population's fitness variance is low, increasing the frequency of mutation operations can effectively improve population diversity. For example, when the variance is below a threshold, the mutation probability can be appropriately increased from an initial low value (e.g., 0.01) to (e.g., 0.05) to increase the probability of generating new feature combinations.

[0052] Throughout the evolutionary process, the adaptive rotation angle mechanism continuously adjusts the size of the rotation angle, altering the search direction and step size of the individual in the feature space. For example, in the early stages of evolution, a larger rotation angle results in a larger step size for the individual in the feature space, enabling rapid exploration of different feature regions. As the number of generations increases, the rotation angle gradually decreases, and the individual's step size becomes finer, gradually focusing on feature regions near the optimal solution for fine-tuning. This mechanism allows the algorithm to automatically adjust its search strategy at different stages of evolution, improving search efficiency and optimization quality.

[0053] In addition, attention should be paid to the impact of constraints such as feature dimension, operating condition coverage, and real-time performance on the optimization process. Feature dimension constraints limit the maximum dimension of the feature vector, for example, it must not exceed the upper limit threshold of computing resource allocation. During optimization, it is necessary to ensure that the number of selected features does not exceed this threshold to avoid computational complexity exceeding the system's capacity due to excessively high feature dimensions. Operating condition coverage constraints require verification of the adaptability of feature combinations through simulation tests under typical operating conditions. Therefore, during optimization, the performance of features under different operating conditions must be considered to ensure that the selected features can effectively reflect the bearing's operating status under various typical operating conditions. Real-time performance constraints limit the maximum latency of a single diagnostic process. Therefore, when designing the algorithm, it is necessary to optimize the algorithm's computational efficiency to ensure that each step in the evolution process can be completed within a reasonable time, meeting the system's real-time requirements.

[0054] By introducing an adaptive rotation angle and dynamic crossover probability mechanism, the quantum genetic optimization algorithm can automatically adjust its search strategy based on the evolutionary stage and population state during the global search of feature subsets, effectively balancing global and local search capabilities and improving the efficiency and quality of feature optimization. This mechanism enables the multi-parameter state assessment model to quickly select the optimal feature combination data, providing reliable feature input for subsequent hierarchical diagnostic control, thereby improving the accuracy and reliability of the entire AMT bearing operating condition detection system. In practical applications, by reasonably setting algorithm parameters and constraints, this mechanism can effectively adapt to the complex requirements of bearing operating condition detection under active-passive switching conditions, providing strong support for bearing fault diagnosis and maintenance.

[0055] Example 3: In the process of mapping diagnostic instructions to warning levels and maintenance suggestion signals by the confidence rule base inference engine, an input activation module needs to be constructed first. The input parameters of this module include the bearing health status probability value, temperature change gradient, and bearing running time. The bearing health status probability value is output by the probabilistic neural network model, ranging from 0 to 1, and is used to characterize the probability that the bearing is in a healthy state. The temperature change gradient is calculated from temperature sensor data, with units of °C / min, reflecting the rate of change of bearing temperature over time. The bearing running time is obtained by accumulating the operating time after bearing startup, with units of h, and is used to assess the impact of the bearing's service life on its condition.

[0056] For the above input parameters, a Gaussian membership function is used to assign confidence scores, converting precise input data into fuzzy confidence values ​​to facilitate reasoning by the rule base. The general form of the Gaussian membership function is: in, For the specific numerical value of the input variable, The center value of the corresponding fuzzy subset represents the typical characteristic value of the fuzzy subset; The width parameter determines the "thickness" of the membership function curve, reflecting the size of the fuzzy subset. For example, for the probability value of bearing health status, three fuzzy subsets can be defined: "normal," "warning," and "fault." The center value of the "normal" subset... Set to 0.8, width Setting it to 0.1 means that when the probability value of a healthy state is close to 0.8, the confidence level is highest for the "normal" subset, and the confidence level gradually decreases as the value deviates from 0.8; the center value of the "faulty" subset. Set to 0.3, width Set to 0.1 to describe a poor health condition. For temperature change gradient and bearing running time, corresponding fuzzy subsets and center values ​​and width parameters should also be set based on engineering experience and historical data. For example, the temperature change gradient can be divided into three subsets: "low", "medium" and "high", and the bearing running time can be divided into three subsets: "short", "medium" and "long", each corresponding to different center values ​​and width parameters.

[0057] Next, a rule base is constructed, containing multiple confidence rules. Each rule describes the mapping relationship between input variables and output results in the form of "IF condition THEN conclusion". For example, a typical rule can be expressed as: "IF the bearing health status probability value is 'normal' AND the temperature change gradient is 'low' AND the bearing operating time is 'short' THEN the warning level is Level I, with a confidence level of 0.8". The rules in the rule base need to be constructed based on the fault mode knowledge base and domain expert experience. The fault mode knowledge base contains vibration spectrum feature templates for three typical bearing faults: pitting, cracking, and wear. For example, the characteristic frequency corresponding to pitting faults is 1.5 times the rotational frequency, cracking faults correspond to 2 times the rotational frequency, and wear faults correspond to 0.8 times the rotational frequency. These feature templates can serve as important bases for rule conditions. The rule base initially contains a certain number of rules, such as 96 rules, covering diagnostic scenarios under different input combinations. The confidence coefficient of each rule (such as 0.8 in the example above) is initially set according to expert experience, with a value range of 0.7 to 0.9, indicating the degree of confidence that the rule is valid.

[0058] When designing the evidence fusion module, Dempster's evidence theory is used to fuse the outputs of multiple rules into a comprehensive confidence distribution. Dempster's evidence theory is a reasoning method for handling uncertain information. It fuses the outputs of multiple rules by calculating the basic probability assignment function, the trust function, and the likelihood function. Specifically, the process is as follows: First, for each rule's output (such as warning level and confidence level), a corresponding basic probability assignment (BPA) is generated, representing the degree of support the rule provides for a given warning level. Then, the multiple basic probability assignments are synthesized using Dempster's synthesis rule to obtain a comprehensive basic probability assignment. Finally, the trust function and likelihood function are calculated based on the comprehensive basic probability assignment to generate a comprehensive confidence distribution, which reflects the overall support of all rules for each warning level. Through this evidence fusion module, information from multiple rules can be effectively integrated, improving the reliability and accuracy of diagnostic results.

[0059] To enable the confidence rule base inference engine to adapt to changes in actual operating conditions, a rule weight correction mechanism is introduced to dynamically update the rule confidence coefficient based on historical diagnostic accuracy and maintenance records. Specifically, the correction method is as follows: when a rule's diagnostic result is verified as correct through maintenance records, the rule's confidence coefficient is increased; if the diagnostic result is incorrect, its confidence coefficient is decreased. The mathematical expression for the correction is: in, This represents the updated rule confidence coefficient. The confidence coefficient before the update. The learning rate controls the magnitude of the correction; its value typically ranges from 0 to 1 (e.g., ...). ); The historical diagnostic accuracy rate is calculated by dividing the number of correct diagnoses made by the rule in historical cases by the total number of diagnoses, with a value ranging from 0 to 1. Through this dynamic correction mechanism, the rule base can continuously learn from actual diagnostic experience, gradually improving the accuracy and applicability of the rules.

[0060] In the specific implementation process, the parameter preprocessing of the input activation module is crucial. For the bearing health status probability value, the output result of the probabilistic neural network model needs to be received directly and range-checked to ensure that it is between 0 and 1. For the temperature change gradient, differential calculation needs to be performed on the real-time data of the temperature sensor, for example, by calculating the change per unit time using the temperature values ​​of two adjacent sampling points, and then smoothing is performed to remove noise interference. For the bearing running time, it needs to be accumulated using the system clock and bearing start / stop signals, and the current duration needs to be saved when the system is powered off to ensure data continuity.

[0061] The construction of a rule base must adhere to the principles of completeness and consistency. Completeness requires the rule base to cover all possible input combinations to avoid diagnostic blind spots; consistency requires that there are no contradictory conclusions between rules—for example, the same input combination cannot simultaneously correspond to different warning levels. To meet these requirements, orthogonal experimental design can be used to generate input combinations during rule base construction. Domain experts can then be invited to evaluate the diagnostic conclusions corresponding to each combination, ensuring the comprehensiveness and rationality of the rules.

[0062] The computational process of the evidence fusion module needs to consider both computational efficiency and numerical stability. Since the computational complexity of DS evidence theory is high when synthesizing multiple pieces of evidence, a recursive synthesis method can be used to gradually merge basic probability assignments, reducing the computational load. Simultaneously, the input basic probability assignments need to be normalized to avoid deviations in the synthesis results due to numerical errors. Furthermore, by setting a confidence threshold, when the confidence level of a certain warning level in the comprehensive confidence distribution exceeds the threshold, that level can be directly output as the diagnostic result, improving diagnostic efficiency.

[0063] The vibration spectrum feature templates in the fault mode knowledge base need to be established through a large amount of experimental data and fault cases. For example, for pitting faults, pitting defects are artificially created on the bearing, vibration signals are collected under different speeds and loads, and spectrum analysis is performed to determine its characteristic frequency as 1.5 times the rotational frequency. The manifestation of this feature under different operating conditions is recorded. For crack and wear faults, a similar method is used to obtain the characteristic frequency and spectrum features. These templates can serve as auxiliary basis in the diagnostic process. When the inference results of the confidence rule base are uncertain, correlation matching between the real-time acquired time-frequency feature matrix and the template can further assist in determining the fault type.

[0064] During system integration, the confidence rule base inference engine needs to interact with the feature extraction module, state assessment module, and optimization decision module. The time-frequency feature matrix generated by the feature extraction module provides input data for the fault mode knowledge base. The bearing health state probability value output by the state assessment module serves as one of the parameters input to the activation module. The optimal feature combination data selected by the optimization decision module is used to improve the effectiveness of the input parameters and the diagnostic accuracy. Through the collaborative work between these modules, a complete bearing operating condition detection and diagnosis process is formed.

[0065] Example 4: In the data preprocessing process of the multi-channel sensing module, appropriate algorithms need to be used for different types of sensor data to extract effective features and suppress noise, providing a reliable data foundation for subsequent state assessment and diagnosis. The following describes in detail the preprocessing implementation methods for each sensor data using specific examples:

[0066] The signals acquired by vibration acceleration sensors typically contain rich fault characteristics, but are also easily interfered with by background noise from mechanical vibrations. Therefore, envelope demodulation algorithms are needed to extract the frequency components of the fault characteristics. Taking bearing pitting faults as an example, suppose a vibration acceleration time-domain signal is acquired at a certain moment. Its original waveform contains high-frequency noise and low-frequency mechanical vibration interference. The preprocessing steps are as follows: Low-frequency interference is removed using a high-pass filter. The cutoff frequency of the high-pass filter is set to 500Hz to filter out noise generated by low-frequency mechanical vibrations such as gear meshing, retaining high-frequency signals that may contain bearing fault characteristics (such as impact vibration signals from bearing components). Although the filtered signal still contains high-frequency noise, the frequency components related to the bearing fault have been preliminarily separated.

[0067] The filtered signal is then demodulated using Hilbert envelope demodulation. The Hilbert transform converts a real signal into an analytic signal, extracting its envelope and highlighting periodic impact components. For example, pitting faults cause periodic impact vibrations in a bearing during rotation, and its envelope signal exhibits periodic fluctuations related to the fault's characteristic frequency. By calculating the power spectral density of the envelope signal, a distinct peak frequency can be observed in the spectrum, corresponding to the characteristic frequency of the bearing pitting fault (e.g., 1.5 times the rotational speed). This process clearly extracts fault features hidden within complex vibration signals, facilitating subsequent feature analysis and fault diagnosis.

[0068] A current sensor monitors the current signal in the motor input circuit in real time, and its waveform changes can reflect changes in bearing load and motor operating status. Taking the case of increased load due to bearing wear as an example, the acquired current signal may contain fundamental and higher harmonic components, with an increase in higher harmonics often indicating mechanical faults. The preprocessing steps are as follows: The current time-domain signal is converted into a frequency-domain signal using a Fast Fourier Transform (FFT) to generate a spectrum. The number of FFT points is set to 4096 to ensure that the frequency resolution meets diagnostic requirements (frequency resolution = sampling frequency / number of FFT points). Assuming a sampling frequency of 10kHz, the frequency resolution is approximately 2.44Hz, which can clearly distinguish the harmonic components in the current signal (such as the fundamental 50Hz, the 3rd harmonic 150Hz, the 5th harmonic 250Hz, etc.).

[0069] Background noise is suppressed through adaptive threshold filtering. The threshold is calculated based on the average power of the noise segment in the spectrum. Specifically, a frequency band with no obvious signal components (such as a high-frequency noise segment) is selected in the spectrum, and the average power spectral density within that band is calculated. 1.5 times this average value is used as the cutoff threshold. Frequency components below the threshold are considered background noise and set to zero; frequency components above the threshold are retained as effective features reflecting the bearing condition. For example, under normal operating conditions, the harmonic components of the current signal are mainly the fundamental frequency, with lower power for higher harmonics. When bearing wear leads to uneven load, significant higher harmonics (such as the 3rd and 5th harmonics) appear in the current signal, and their power spectral density is higher than the threshold. Filtering can highlight these features, providing a basis for judging abnormal bearing load.

[0070] Temperature sensors (such as PT100 RTDs) are embedded in the bearing lubrication cavity to monitor the bearing's operating temperature. Due to sensor measurement errors and environmental fluctuations, the raw temperature data may exhibit random fluctuations, requiring the use of an exponentially weighted moving average (EWMA) algorithm to eliminate these fluctuations. Taking the temperature monitoring of a bearing during continuous operation as an example, assuming temperature data is collected once per minute, a series of temperature values ​​are obtained. .

[0071] The EWMA algorithm generates a smoothed temperature value by weighting the current measurement value and the filtered result from the previous time step. Specifically, the filtered value at the current time step equals the smoothing coefficient multiplied by the current measurement value, plus (1 - smoothing coefficient) multiplied by the filtered value from the previous time step. The smoothing coefficient is typically set to 0.2, giving more weight to recent measurements, enabling faster response to temperature trends while suppressing the impact of short-term fluctuations. For example, if the current measurement value is... The measured temperature in minutes is , No. The filter value per minute Then the first The filter value per minute In this way, random noise in the temperature data is effectively smoothed, and the temperature change trend (such as slow rise or fall) is clearly presented, making it easier to analyze whether the thermal state of the bearing is normal.

[0072] The speed encoder is rigidly connected to the drive shaft via an incremental code disk, outputting pulse signals to reflect the bearing speed. Due to mechanical transmission backlash and measurement errors, the directly acquired pulse signals may exhibit jitter. Therefore, a Kalman filter algorithm is used to fuse angular velocity and angular acceleration data to output the instantaneous bearing speed and speed fluctuation rate. Taking the speed change of a gearbox bearing during gear shifting as an example, assuming the speed encoder's sampling period is 10ms, the angular velocity (unit: rad / s) and angular acceleration (unit: rad / s²) can be calculated within each sampling period.

[0073] The Kalman filter algorithm uses angular velocity and angular acceleration as state variables to construct state equations and observation equations. The state equations describe the changes of state variables over time; for example, the derivative of angular velocity is angular acceleration. The observation equations describe the relationship between observed values ​​(such as angular velocity calculated through pulse counting) and state variables. By recursively calculating state estimates and covariance matrices, Kalman filtering can effectively fuse measurement data from multiple time points, suppress noise interference, and output smooth instantaneous speed and speed fluctuation rates. For example, during the transition of a bearing from high speed (2000 rpm) to low speed (1000 rpm), the original pulse signal may experience brief jitter due to the shift shock. After Kalman filtering, the instantaneous speed curve can transition smoothly, and the speed fluctuation rate (such as the standard deviation of speed change per second) can accurately reflect the severity of the speed change, providing data support for judging whether the bearing's operating condition transition is smooth.

[0074] In practical applications, data from various sensors in a multi-channel sensing module needs to be processed collaboratively to ensure data consistency and reliability. For example, vibration acceleration data and speed encoder data are strongly correlated, and changes in bearing speed directly affect the frequency components of the vibration signal (such as rotational frequency and its harmonics). Therefore, during preprocessing, signal anomalies can be detected by calculating the correlation coefficients (such as the Spearman correlation coefficient) of the data from each channel. If the correlation coefficient between a channel's data and related channels is lower than a set threshold (such as 0.5), it is considered that the data in that channel may be abnormal (such as sensor failure or loose wiring), and data reconstruction needs to be initiated. Data reconstruction can utilize the correlation between historical data and adjacent channel data to generate alternative data through linear interpolation or regression models, ensuring the continuity of subsequent analysis.

[0075] Taking a vibration accelerometer channel anomaly as an example: Suppose that at a certain moment, the correlation coefficient of the vibration signal suddenly drops to 0.3, while the speed, current, and temperature data are all normal, then it is determined that the vibration sensor may be faulty. At this time, the system automatically retrieves the vibration and speed data from the previous 10 minutes, establishes a linear regression model (such as the linear relationship between the root mean square value of vibration and speed), and predicts the estimated value of the vibration signal based on the current speed value, replacing the abnormal data. In this way, even if a single sensor fails, the system can still continue to operate based on data from other sensors, avoiding diagnostic interruptions due to data loss.

[0076] The data preprocessing of the multi-channel sensing module effectively improves data quality and feature identifiability through algorithm design tailored to the characteristics of different sensors. Envelope demodulation of vibration data, spectral analysis of current data, smoothing of temperature data, and Kalman filtering of rotational speed data extract key information about bearing operation from different dimensions, while the correlation verification and reconstruction mechanism of multi-sensor data further enhances the robustness of the system. These preprocessing steps lay a solid foundation for subsequent feature extraction, condition assessment, and fault diagnosis, ensuring accurate capture of bearing operating state changes under active / passive switching conditions, enabling early fault warning and precise maintenance.

[0077] Example 5: In the implementation of multi-level diagnostic strategies generated in the analysis layer of the hierarchical diagnostic control model, it is necessary to combine the bearing health state probability value, state transition law, and trend prediction requirements to construct a diagnostic system covering different risk levels. The implementation method is described in detail below with specific working condition examples:

[0078] I. Health Level Classification and Response Mechanism

[0079] The bearing health status is divided into three levels: "Normal," "Warning," and "Fault," with each level corresponding to different probability thresholds and response strategies. For example, in the Normal level, the health status probability value is ≥0.7, indicating that the bearing is in good operating condition. At this level, periodic monitoring is triggered, with a monitoring cycle set at 30 minutes. The system only records trend data for key parameters such as vibration energy entropy and current harmonic distortion rate, without performing real-time in-depth analysis. For instance, if a bearing operates under high-speed, light-load conditions with a health probability value of 0.85, the system collects and stores data every 30 minutes for long-term trend analysis.

[0080] Warning Level: A health status probability value between 0.4 and 0.7 indicates a possible early abnormality in the bearing. At this point, feature tracking monitoring is triggered, the monitoring cycle is shortened to 5 minutes, and a multi-feature joint analysis process is initiated. For example, when the bearing health probability value drops to 0.6, the system automatically increases the analysis frequency of features such as the envelope spectrum of the vibration signal and temperature change gradient, comparing historical data to determine whether the abnormal trend continues.

[0081] Fault Level: A health probability value ≤ 0.4 indicates that the bearing has experienced a significant fault or is about to fail. In this case, the shutdown protection protocol is triggered, sending a shutdown command to the transmission controller via the CAN bus to prevent the fault from escalating. For example, if the health probability value suddenly drops to 0.2, the system immediately cuts off the motor power output to prevent the bearing from seizing and causing serious damage to the transmission.

[0082] II. State Transition Matrix and Upgrade Diagnostic Process

[0083] A state transition matrix is ​​constructed to describe the changes in health level over time. Matrix elements are defined based on historical data statistics or expert experience. For example, a state transition matrix can be represented as: The first row indicates the probability of transitioning from the "normal" level to "normal", "warning", and "fault" is 90%, 9%, and 1%, respectively; the second row indicates the probability of transitioning from the "warning" level to each state is 10%, 80%, and 10%, respectively; and the third row indicates the probability of transitioning from the "fault" level to each state is 0%, 5%, and 95%, respectively.

[0084] When the health level declines for three consecutive sampling periods (assuming a sampling period of 1 second) (e.g., "normal → warning → fault"), an upgraded diagnostic process is triggered. For example, if a bearing's health level is "normal" at time t1, drops to "warning" at time t2, and further drops to "fault" at time t3, the system determines that the condition is continuously deteriorating and automatically activates a backup sensor channel (e.g., a redundant vibration acceleration sensor) to verify the data. It compares the vibration data from the main channel and the backup channel to rule out misjudgments caused by a single sensor failure. If the verification confirms a consistent downward trend in the health level, fault level response measures are implemented; if the data are contradictory, a sensor anomaly warning is issued, and hardware repair is prioritized.

[0085] III. Trend Prediction Mechanism and Active Testing Mode

[0086] A trend prediction mechanism is introduced. When the rate of change of the bearing health status probability value exceeds a preset gradient threshold (e.g., an absolute value of 0.05 / second), an active testing mode is activated. For example, if a bearing's health probability value suddenly drops at a rate of 0.06 / second during normal operation, the system determines that a rapidly developing fault (e.g., bearing crack propagation) may exist and immediately triggers the active testing mode: applying a step load excitation (e.g., a sudden increase from 50% to 80% of the rated load) through the gearbox controller, collecting the bearing's response data (e.g., vibration signals, current signals) under dynamic operating conditions, and performing in-depth time-frequency analysis and fault feature matching.

[0087] In active testing mode, the system can capture fault characteristics that are not obvious under static conditions. For example, an early-stage crack in a bearing may only cause slight vibration under static light load, but under a step load, the crack deformation will produce significant impact vibration, and the kurtosis and characteristic frequency components (such as twice the rotational speed) of the vibration signal will increase significantly. By comparing the changes in data before and after excitation, the fault type and severity can be determined more accurately, providing a basis for maintenance strategy formulation.

[0088] IV. Constraints of the Multi-Parameter State Evaluation Model

[0089] Multi-parameter state assessment models must meet the following constraints to ensure engineering practicality: Feature dimension constraint: The maximum dimension of the feature vector is limited to no more than the upper limit threshold of computational resource allocation (e.g., 12 dimensions). For example, the original features extracted from data such as vibration, current, temperature, and rotational speed may exceed 20 dimensions (e.g., root mean square, kurtosis, and margin indices of vibration, and harmonic energy of current). Principal component analysis (PCA) or quantum genetic optimization algorithms are needed to select the most representative 12-dimensional features (e.g., vibration energy entropy, current harmonic distortion rate, and temperature gradient) to avoid excessive computational complexity due to too many features, which could affect the real-time performance of the system.

[0090] Operating condition coverage constraints: The adaptability of characteristic combinations is tested by simulating four typical operating conditions: high speed, low speed, heavy load, and light load. For example, under high-speed, light-load conditions, bearings mainly exhibit high-frequency vibration, and the characteristic combination must include parameters that can reflect high-frequency impacts (such as the high-frequency energy of the vibration signal). Under low-speed, heavy-load conditions, bearing temperature changes and current fluctuations are more significant, and the characteristic combination needs to focus on the stability of the temperature gradient and the fundamental current component. Through multi-condition testing, it is ensured that the characteristic combination can effectively distinguish between healthy and faulty states under different operating conditions.

[0091] Real-time constraint: The maximum latency of a single diagnostic process is limited to no more than 20% of the control cycle (e.g., if the control cycle is 100ms, the diagnostic latency should be ≤20ms). To meet this requirement, the algorithm execution efficiency needs to be optimized. For example, a parallel computing architecture can be adopted for the variational mode decomposition algorithm, and hardware acceleration can be performed on the probabilistic neural network model to ensure that the entire process from data acquisition to output of diagnostic results is completed within the specified time, avoiding delays in fault response.

[0092] V. Signal Verification Module and Data Reconstruction

[0093] A signal verification module is constructed to detect signal anomalies by calculating the correlation coefficients of multi-sensor data. For example, vibration acceleration data and rotational speed data usually have a strong correlation (e.g., the vibration characteristic frequency and rotational frequency are multiples of each other). If the Spearman correlation coefficient between vibration data and rotational speed data in a certain channel is lower than 0.5 (a preset threshold), the vibration data is determined to be abnormal. At this time, data reconstruction is initiated, utilizing the linear relationship between historical vibration data and rotational speed data (e.g., the root mean square value of vibration increases with increasing rotational speed) to generate alternative data through linear interpolation.

[0094] Taking a speed encoder failure as an example: If the speed data suddenly jumps to zero, while the vibration, current, and temperature data are all normal, the system detects a sharp drop in the correlation between the vibration data and the speed data, and determines that the speed data is abnormal. At this time, the system estimates the current speed value and reconstructs the data based on the speed trend of the previous 5 seconds (such as a uniform decrease) and the current motor current value (which is negatively correlated with the speed), maintaining the continuity of the diagnostic process. At the same time, it issues a speed sensor fault warning, prompting maintenance personnel to perform timely repairs.

[0095] VI. Example of Diagnostic Procedures under Typical Operating Conditions

[0096] Taking the transition of an AMT transmission bearing from active operating condition (motor drive) to passive operating condition (coasting) as an example, the practical application of the hierarchical diagnostic control model is illustrated: Active operating condition stage: The bearing rotates at high speed under motor drive, with a health probability value of 0.75 (lower limit of the warning level). The system triggers feature tracking monitoring, analyzing the time-frequency characteristics of the vibration signal every 5 minutes. It detects energy concentration at 0.8 times the rotational frequency in the vibration signal (corresponding to wear fault characteristics), but the health probability value has not yet exceeded the fault threshold. The system maintains the warning level and records the trend of characteristic changes.

[0097] During the condition switching phase: After the gearbox shifts gears, it enters a passive condition, the bearing speed decreases, and the health probability value suddenly drops to 0.38 (fault level) at a rate of 0.07 / second. Since the rate of change exceeds the threshold, the system starts the active test mode, applies a short-term load excitation, and the 2x frequency component in the collected vibration signal increases significantly (corresponding to crack fault characteristics). Combined with the high transition probability (10%) of "warning → fault" in the state transition matrix, it is determined that the bearing has a risk of crack propagation.

[0098] Fault Response Phase: The system immediately triggers the shutdown protection protocol and simultaneously outputs a Level III warning signal and a "stop immediately for maintenance" recommendation via the confidence rule base inference engine. After maintenance, a deep crack was found in the inner ring of the bearing, confirming the accuracy of the diagnosis.

[0099] The hierarchical diagnostic control model achieves dynamic tracking and precise response to bearing conditions through multi-level diagnostic strategies, state transition analysis, trend prediction, and constraint management. Differential monitoring of health levels, continuous judgment of state transitions, fault feature mining through active testing, and feature optimization under multiple constraints collectively ensure the system's reliability and real-time performance under active-passive switching conditions. Signal verification and data reconstruction mechanisms further enhance the system's fault tolerance, enabling stable operation in complex industrial environments and providing effective technical support for the full lifecycle health management of AMT transmission bearings.

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the operating status of an AMT bearing under active / passive switching conditions, characterized in that, include: Real-time operating data of the gearbox bearing is collected through a multi-channel sensing module, which includes a vibration acceleration sensor, a current sensor, a temperature sensor, and a speed encoder. The intrinsic mode components of the real-time operating data are extracted based on the variational mode decomposition algorithm, and a time-frequency feature matrix is ​​generated by Hilbert transform. The time-frequency feature matrix is ​​then input into a pre-trained probabilistic neural network model, which adopts a sliding time window mechanism to adjust the activation threshold of hidden layer nodes based on real-time operating data and outputs the bearing health status probability value. A multi-parameter condition assessment model is constructed based on the bearing health status probability value. The multi-parameter condition assessment model aims to maximize fault sensitivity and minimize feature redundancy. A quantum genetic optimization algorithm is used to perform a global search on the feature subset. The quantum genetic optimization algorithm introduces adaptive rotation angle and dynamic crossover probability. The optimal feature combination data is output based on the multi-parameter condition assessment model. A hierarchical diagnostic control model is established based on the optimal feature combination data. The hierarchical diagnostic control model includes a data acquisition layer, an analysis layer, and an execution layer. The data acquisition layer updates the operating condition mode based on the time-frequency feature matrix. The analysis layer generates a multi-level diagnostic strategy based on the bearing health status probability value. The execution layer maps diagnostic instructions into early warning levels and maintenance suggestion signals through a confidence rule base inference engine. A probabilistic network structure is constructed, which includes vibration feature nodes, current feature nodes, and health state nodes, wherein the vibration feature nodes and the current feature nodes are parent nodes, and the health state nodes are child nodes. The probability density function of the vibration feature node is updated based on real-time acquired data, and the frequency band energy of the current feature node is analyzed by wavelet packet decomposition algorithm to obtain the posterior probability of the feature frequency band energy distribution. A dynamic threshold adjustment mechanism is introduced to dynamically adjust the inference step size of the probability network according to the bearing load change. The width of the time window is negatively correlated with the bearing speed. The expectation-maximization algorithm is used to update the parameters of the probabilistic neural network online, and the weight parameters of the hidden layer nodes are optimized by maximizing the likelihood function. A genetic optimization objective function is constructed, which includes a feature sensitivity term and a feature independence term. The feature sensitivity term is calculated by the inter-class distance of the fault samples, and the feature independence term is calculated by the correlation coefficient between feature vectors. An adaptive rotation angle update rule is designed, wherein the rotation angle changes with an exponential decay curve as the number of generations increases, maintaining a large search range in the initial stage and converging to the neighborhood of the optimal solution in the later stage. A dynamic crossover probability adjustment mechanism is designed, wherein the crossover probability is dynamically adjusted according to the population fitness variance. When the population variance is higher than a threshold, the frequency of crossover operation is increased, and vice versa, the frequency of mutation operation is increased. An input activation module is constructed, wherein the inputs include the bearing health status probability value, temperature change gradient and bearing running time, and a Gaussian membership function is used to assign confidence to the input variables. Construct a rule base containing confidence rules; The design of the evidence fusion module uses the DS evidence theory to fuse multi-source rule outputs into a comprehensive confidence distribution. The output includes a warning level code and a maintenance suggestion code. A rule weight correction mechanism is introduced to dynamically update the rule confidence coefficient based on historical diagnostic accuracy and maintenance records; The envelope demodulation algorithm is used to extract the fault characteristic frequency components from the vibration acceleration data; The current sensor data is processed by Fast Fourier Transform to generate a spectrum, and background noise is suppressed by adaptive threshold filtering. An exponentially weighted moving average algorithm is used to eliminate measurement fluctuations in temperature sensor data; The Kalman filter algorithm is used to fuse the angular velocity and angular acceleration data of the speed encoder data to output the instantaneous speed and speed fluctuation rate of the bearing. The health level is divided into three levels: normal, warning, and fault. The normal level triggers periodic monitoring, the warning level triggers feature tracking monitoring, and the fault level triggers a shutdown protection protocol. A state transition matrix is ​​constructed, which transitions the state based on continuous changes in health level. If the health level decreases for three consecutive sampling periods, an upgrade diagnosis process is triggered. A trend prediction mechanism is introduced, and an active testing mode is activated when the rate of change of the bearing health status probability value exceeds a preset gradient threshold. A signal verification module is constructed to detect signal anomalies through multi-sensor data correlation analysis. If the correlation coefficient between a certain channel's data and the associated channel is lower than a set threshold, data reconstruction is initiated.

2. The method for detecting the operating status of an AMT bearing under active / passive switching conditions according to claim 1, characterized in that, The rule construction of the confidence rule base inferencer includes: establishing a fault mode knowledge base, which contains vibration spectrum feature templates for three typical faults: bearing pitting, cracking, and wear.

3. The method for detecting the operating status of an AMT bearing under active / passive switching conditions according to claim 1, characterized in that, The constraints of the multi-parameter state evaluation model include: Feature dimension constraint: the maximum dimension of the feature vector is limited to no more than the upper limit threshold of computing resource allocation; Operating condition coverage constraints are used to verify the adaptability of feature combinations through typical operating condition simulation tests. Real-time constraints limit the maximum delay of a single diagnostic process to no more than 20% of the control cycle.

4. A system for detecting the operating status of an AMT bearing under active / passive switching conditions, characterized in that, include: Multi-channel sensing module: used to collect real-time operating data of gearbox bearings, including vibration acceleration sensor, current sensor, temperature sensor and speed encoder; The envelope demodulation algorithm is used to extract the fault characteristic frequency components from the vibration acceleration data; The current sensor data is processed by Fast Fourier Transform to generate a spectrum, and background noise is suppressed by adaptive threshold filtering. An exponentially weighted moving average algorithm is used to eliminate measurement fluctuations in temperature sensor data; The Kalman filter algorithm is used to fuse the angular velocity and angular acceleration data of the speed encoder data to output the instantaneous speed and speed fluctuation rate of the bearing. Feature extraction module: Based on the variational mode decomposition algorithm, the intrinsic mode components of the real-time operating data collected by the multi-channel sensing module are extracted, and the time-frequency feature matrix is ​​generated by Hilbert transform; State assessment module: Inputs the time-frequency feature matrix into a pre-trained probabilistic neural network model. This probabilistic neural network model employs a sliding time window mechanism, adjusts the activation threshold of hidden layer nodes based on real-time operating data, and outputs a bearing health state probability value. The sliding time window mechanism employed in the probabilistic neural network model includes: Construct a probabilistic network structure, which includes vibration feature nodes, current feature nodes, and health state nodes. Vibration feature nodes and current feature nodes are parent nodes, and health state nodes are child nodes. The probability density function of the vibration feature node is updated based on real-time acquired data. The frequency band energy of the current feature node is analyzed by wavelet packet decomposition algorithm to obtain the posterior probability of the feature frequency band energy distribution. A dynamic threshold adjustment mechanism is introduced to dynamically adjust the inference step size of the probability network according to the bearing load change. The width of the time window is negatively correlated with the bearing speed. The expectation-maximization algorithm is used to update the parameters of the probabilistic neural network online, and the weight parameters of the hidden layer nodes are optimized by maximizing the likelihood function. Optimization Decision Module: Based on the bearing health status probability value, a multi-parameter status assessment model is constructed. With the highest fault sensitivity and lowest feature redundancy as the optimization objectives, a quantum genetic optimization algorithm that introduces adaptive rotation angle and dynamic crossover probability is used to perform a global search on the feature subset and output the optimal feature combination data. The quantum genetic optimization algorithm introduces adaptive rotation angle and dynamic crossover probability, including: Construct a genetic optimization objective function, which includes a feature sensitivity term and a feature independence term. The feature sensitivity term is calculated by the inter-class distance of the fault samples, and the feature independence term is calculated by the correlation coefficient between the feature vectors. An adaptive rotation angle update rule is designed, in which the rotation angle changes exponentially with the number of generations, maintaining a large search range in the initial stage and converging to the neighborhood of the optimal solution in the later stage. Design a dynamic crossover probability adjustment mechanism. The crossover probability is dynamically adjusted according to the population fitness variance. When the population variance is higher than the threshold, the frequency of crossover operation is increased, and vice versa. Diagnostic execution module: Based on the optimal feature combination data, it includes an acquisition layer, an analysis layer and an execution layer. The acquisition layer updates the operating mode based on the time-frequency feature matrix. The analysis layer generates a multi-level diagnostic strategy based on the bearing health status probability value and divides the health level into three levels: normal, warning and fault. The normal level triggers periodic monitoring, the warning level triggers feature tracking monitoring, and the fault level triggers the shutdown protection protocol. Construct a state transition matrix and jump to a state based on continuous changes in health level. If the health level decreases for three consecutive sampling periods, trigger the upgrade diagnosis process. A trend prediction mechanism is introduced, and an active testing mode is activated when the rate of change of the bearing health status probability value exceeds a preset gradient threshold. The execution layer maps diagnostic commands into warning levels and maintenance suggestion signals through a confidence rule base inference engine, constructs an input activation module, and the inputs include the bearing health status probability value, temperature change gradient and bearing running time. Gaussian membership function is used to assign confidence to the input variables. Build a rule base that includes confidence rules; The evidence fusion module is designed, and the DS evidence theory is used to fuse the multi-source rule outputs into a comprehensive confidence distribution. The output includes warning level codes and maintenance suggestion codes. A rule weight correction mechanism is introduced to dynamically update the rule confidence coefficient based on historical diagnostic accuracy and maintenance records; Signal verification module: Detects signal anomalies through multi-sensor data correlation analysis. If the correlation coefficient between a certain channel's data and the associated channel is lower than a set threshold, data reconstruction is initiated.

Images