Mechanical transmission system fault trend prediction system based on dynamic feature recognition

Abstract

The invention discloses a mechanical transmission system fault trend prediction system based on dynamic feature recognition, and relates to the technical field of mechanical state monitoring. Comprising the following steps: synchronously acquiring a load torque signal and a lubrication state parameter signal of a transmission system and vibration acceleration signals of a plurality of measuring points through a signal acquisition module; the working condition decoupling characteristic generation module carries out time-frequency analysis on the vibration signal, calls a pre-stored load disturbance spectrum template according to a load torque signal to carry out adaptive differential processing so as to eliminate load fluctuation interference, and calls a correction rule set according to a lubrication state parameter signal to carry out form recombination on the signal so as to compensate the lubrication state influence; and finally outputting a working condition decoupling feature representing the health state of the mechanical part. And the trend prediction module calculates and obtains fault development trend and residual life estimation data through a pre-trained fault prediction model. According to the method, the dynamic characteristics representing the essential degradation of the part are effectively extracted, and the accuracy and reliability of fault trend prediction of the mechanical transmission system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical state monitoring, in particular to a mechanical transmission system fault trend prediction system based on dynamic feature recognition. BACKGROUND

[0002] Mechanical transmission system fault prediction and health management is a core technology in the field of industrial intelligent operation and maintenance, aiming to identify fault signs and predict remaining life in advance through real-time monitoring and data analysis of equipment operating state, so as to realize predictive maintenance and ensure production safety and efficiency. In recent years, with the progress of vibration analysis technology and artificial intelligence algorithm, fault prediction methods are developing towards high precision and self-adaptation.

[0003] In the prior art, most schemes collect vibration signals of the transmission system, and combine basic working condition parameters such as load and speed to perform feature extraction and trend analysis. These methods can realize a certain degree of early warning for common faults under stable working conditions by establishing a statistical relationship between vibration features and equipment degradation, and have considerable practical value. However, in actual operation of the mechanical transmission system, its dynamic response is a complex product of the mutual coupling of various working condition factors such as load torque and lubrication state and the health state of components. The existing methods often regard load fluctuation as the main interference for simple filtering or working condition segmentation processing, but generally ignore the nonlinear modulation effect of dynamic changes in lubrication state on the form of vibration features. For example, poor lubrication can change the contact stiffness and damping characteristics of gears and bearings, causing changes in the energy distribution of the vibration frequency spectrum, and this change is highly overlapped with the fault features caused by component wear in the frequency spectrum.

[0004] Therefore, in the existing technology, in the industrial scene where the lubrication condition fluctuates greatly, the prediction model is prone to false positives and false negatives, and its generalization ability and prediction accuracy face a bottleneck. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a mechanical transmission system fault trend prediction system based on dynamic feature recognition.

[0006] In order to achieve the above-mentioned purpose, the technical scheme of the present application is as follows:

[0007] The present application discloses a mechanical transmission system fault trend prediction system based on dynamic feature recognition, comprising:

[0008] A signal acquisition module is used to synchronously acquire load torque signals and lubrication state parameter signals of the mechanical transmission system and vibration acceleration signals of multiple measuring points on the transmission chain;

[0009] A working condition decoupling feature generation module is used to perform the following operations:

[0010] generate time-frequency distribution data based on the vibration acceleration signal;

[0011] According to the load torque signal, a corresponding load disturbance spectrum template is called from a pre-stored load disturbance spectrum mapping relationship, and adaptive differential processing is performed on the time-frequency distribution data to generate processed time-frequency distribution data.

[0012] According to the lubrication state parameter signal, a corresponding correction rule set is called to perform morphological reorganization on the processed time-frequency distribution data, and working condition decoupling features representing the health state of the mechanical component are output.

[0013] A trend prediction module is configured to calculate the fault development trend data and the remaining life estimation data of the mechanical transmission system based on the working condition decoupling features through a pre-trained fault prediction model.

[0014] Compared with the prior art, the present application has the following advantages:

[0015] 1. The present application can selectively separate the linear disturbance component related to the load fluctuation from the complex vibration response. This processing makes the time-frequency data relied on subsequent analysis no longer change dramatically in energy characteristics with the load fluctuation, laying a foundation for extracting stable fault features under variable load conditions.

[0016] 2. The present application can compensate or suppress the nonlinear modulation effect of the lubrication condition change on the vibration signal spectrum form, so that the final output working condition decoupling features can more purely reflect the essential state change of the mechanical component caused by wear, fatigue and other degradation processes, effectively avoiding the feature confusion and misjudgment caused by lubrication fluctuation.

[0017] 3. The present application can more accurately track the fault development trend of the mechanical transmission system and provide reliable remaining life estimation under the complex industrial scene of changing load and lubrication state, improving the accuracy and practicality of the predictive maintenance strategy. BRIEF DESCRIPTION OF DRAWINGS

[0018] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. In the drawings, the same reference numerals are used to refer to the same parts. Among them:

[0019] Figure 1 is a system module connection diagram of the present application;

[0020] Figure 2 is a system module flowchart of the present application;

[0021] Figure 3 is a construction process flowchart of the lubrication feature form mapping relationship of the present application;

[0022] Figure 4 A step flow chart for the cycle drift identification module of the present application;

[0023] Figure 5 A step flow chart for the conduction delay difference extraction module of the present application. DETAILED DESCRIPTION

[0024] It is easy to understand that, according to the technical solution of the present application, a person skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical solution of the present application.

[0025] In the prior art, most solutions can combine basic working condition parameters such as load, but generally lack a fine compensation mechanism for dynamic changes in lubrication state. When the load fluctuates or the lubrication state changes, the vibration frequency spectrum characteristics of the traditional method are easily disturbed, resulting in inaccurate fault feature extraction. The existing solution lacks a fine compensation mechanism for dynamic changes in working conditions, especially under variable load and variable lubrication conditions. A single signal processing model will have systematic deviations, making it difficult to meet the demand for high-precision predictive maintenance.

[0026] In order to solve the above problems, it is found that there is a stable mapping relationship between the load torque and the disturbance mode of the vibration frequency spectrum, and the lubrication state has a nonlinear modulation effect on the frequency spectrum form. By establishing a load disturbance spectrum and a lubrication correction rule set, the working condition interference can be peeled off layer by layer. It is further found that adaptive difference based on the load interval can effectively suppress linear disturbances, and the form reorganization of the lubrication state can compensate for nonlinear distortion. Therefore, the idea of generating working condition decoupling features through two-level processing of "difference-reorganization" is proposed. Further, through online monitoring and model updating, the load disturbance spectrum mapping relationship and the lubrication feature form mapping relationship are incorporated into a dynamic optimization mechanism to form a continuously adaptive feature extraction system.

[0027] After introducing the basic concept of the present application, the embodiments of the present application will be specifically introduced with reference to the drawings.

[0028] Embodiment:

[0029] As shown in Figure 1 The mechanical transmission system fault trend prediction system based on dynamic feature identification comprises:

[0030] A signal acquisition module is configured to synchronously acquire a load torque signal and a lubrication state parameter signal of the mechanical transmission system, and vibration acceleration signals of a plurality of measuring points on the transmission chain.

[0031] The working condition decoupling feature generation module is configured to perform the following operations:

[0032] Generate time-frequency distribution data based on the vibration acceleration signal;

[0033] According to the load torque signal, a corresponding load disturbance spectrum template is called from a pre-stored load disturbance spectrum mapping relationship, and adaptive differential processing is performed on the time-frequency distribution data to generate processed time-frequency distribution data;

[0034] According to the lubrication state parameter signal, a corresponding correction rule set is called to perform morphological reorganization on the processed time-frequency distribution data, and working condition decoupling features representing the health state of the mechanical component are output.

[0035] The trend prediction module is configured to calculate the fault development trend data and the remaining life estimation data of the mechanical transmission system based on the working condition decoupling features through a pre-trained fault prediction model.

[0036] The training of the fault prediction model is completed through historical full life cycle data. Complete data sequences from healthy state operation to complete failure are collected from no less than 50 mechanical transmission systems of the same type. Each data sequence includes working condition decoupling features, auxiliary feature vectors, and corresponding remaining life labels. Before training, the feature data is standardized to eliminate the dimension effect. The model uses a deep fully connected network structure, including an input layer, three hidden layers, and an output layer. The number of hidden layer neurons is 256, 128, and 64, respectively, and the ReLU activation function is used. During training, the Adam optimizer is used, the initial learning rate is set to 0.001, the batch size is 32, and the number of training rounds is 200. The loss function uses a weighted combination of mean square error and cross-entropy loss, corresponding to the remaining life regression and fault classification tasks. Overfitting is prevented by early stopping, and training is terminated when the validation set loss does not decrease for 10 consecutive rounds. The final model can output fault development trend scores and remaining life estimates.

[0037] As shown in Figure 2 The working principle of the present application is as follows: The signal acquisition module is composed of a group of physical sensors arranged at key nodes of the mechanical transmission chain. Specifically, it includes torque sensors installed on the input / output shafts for real-time measurement of load torque signals; oil condition sensors integrated in the gearboxes or bearing seats for detecting the dielectric constant or viscosity parameters of the lubricating oil; and three-axis ICP type acceleration sensors distributed on the bearing seats along the transmission path for collecting vibration acceleration signals. All sensors are time-synchronized to the microsecond level through a synchronous data acquisition card to ensure the timing alignment of multi-source signals. The collected raw signals are transmitted to the edge computing node via industrial Ethernet for subsequent processing.

[0038] The operating condition decoupling feature generation module is deployed in an industrial computer with parallel computing capabilities, and its data processing flow is as follows:

[0039] 1. Time-Frequency Distribution Generation: The vibration acceleration signal is analyzed using a short-time Fourier transform algorithm to generate time-frequency distribution data containing three-dimensional information of time, frequency, and energy. This process involves truncating signal segments using a sliding window, applying a Hanning window, and then performing an FFT transform to ultimately form the time-frequency spectrum matrix.

[0040] 2. Adaptive Differential Processing: Based on the real-time load torque signal, the system retrieves the reference template for the corresponding torque range from a pre-stored load disturbance spectrum database. This load disturbance spectrum database is stored in the form of a lookup table and interacts with the processing unit at high speed via a PCIe interface. Differential processing employs weighted spectral subtraction, dynamically calculating the attenuation factor based on the energy ratio of the current spectrum and the template spectrum within the characteristic frequency band. The proportionally scaled template component is subtracted from the original time-frequency distribution, effectively suppressing spectral disturbances caused by load fluctuations.

[0041] 3. Morphological Reconstruction Processing: To address the impact of lubrication conditions, the system invokes a set of correction rules based on a lubrication condition parameter signal-spectral response relationship model. This set of rules is stored in a configuration file and contains frequency band gain coefficients corresponding to different lubrication conditions. The processing unit determines the specific gain vector based on the real-time lubrication condition parameter signal using an interpolation algorithm and performs element-wise multiplication with the differentially processed time-frequency data. This achieves energy compensation or suppression of the lubrication-affected frequency bands, ultimately outputting a clean decoupling characteristic that reflects the component's health status.

[0042] The trend prediction module runs on a server platform, and its core is a fault prediction model built on a deep learning framework. After receiving the decoupled features of the operating conditions, the module first performs feature standardization to eliminate the influence of dimensions. Then, it fuses the processed feature vectors with auxiliary feature vectors obtained from other analysis paths at the feature level to form a comprehensive feature description. The fused features are input into a pre-trained deep neural network model. This network acquires fault mode recognition capabilities through training on historical fault data and ultimately outputs in parallel fault development trend data representing the changing trend of system health, as well as remaining life estimation data expressed in time units.

[0043] Through the above technical solutions, this application combines multi-source sensor data synchronous acquisition, step-by-step decoupling of operating condition interference, and intelligent prediction models to enable the system to effectively extract characteristic information strongly correlated with the inherent degradation of mechanical components under varying operating conditions, thereby providing accurate condition assessment and life prediction basis for predictive maintenance of mechanical transmission systems.

[0044] This application further proposes that, in the step of calling the corresponding load perturbation spectrum template from the pre-stored load perturbation spectrum mapping relationship, the process of constructing and updating the load perturbation spectrum mapping relationship includes:

[0045] The construction and updating of the load disturbance spectrum mapping relationship in this application is achieved by establishing a benchmark reference under a healthy system condition and continuously optimizing it during operation. Specifically, after the mechanical transmission system is installed and commissioned but before formal operation, it is determined to be in a stable operating phase under healthy conditions. The system uses torque and vibration acceleration sensors in the signal acquisition module to collect vibration acceleration signals under different load torque signals. This includes setting sampling points at intervals of 5%-10% of the rated load within the 20%-100% rated load range, continuously collecting stable operating data for at least 300 seconds at each load point. During the data acquisition process, it is necessary to ensure that the lubrication system is working normally and its parameters are within the rated range, which serves as the basis for determining the equipment's healthy condition.

[0046] After obtaining the basic data, the data processing unit performs time-frequency analysis on the vibration acceleration signals for each load interval. Using the same short-time Fourier transform parameters as in the real-time processing stage (e.g., a window length of 1024 points and an overlap rate of 75%), the time-domain signal is converted into time-frequency distribution data. The time-frequency distribution data of the vibration acceleration signal is extracted, and the arithmetic mean of all time-frequency distribution data within each load interval is calculated along the time dimension and stored as a load perturbation spectrum template. It is a baseline spectrum established under healthy conditions, stored in memory as an array.

[0047] Load perturbation spectrum template The calculation formula is:

[0048]

[0049] Where f is the frequency variable, corresponding to the frequency resolution of the time-frequency distribution data;

[0050] This represents the number of sampling frames within this load range;

[0051] This represents the time-frequency distribution data of the i-th frame.

[0052] The calculated average time-frequency distribution data corresponding to each load interval are stored in non-volatile memory in the form of key-value pairs of <load interval identifier, perturbation spectrum template>, forming the initial load perturbation spectrum mapping relationship.

[0053] To maintain the accuracy of the mapping relationship during online monitoring, a dynamic update mechanism is introduced. When the system is in a quasi-steady-state condition where the load fluctuation rate is below 15% and the total vibration energy fluctuation rate is below 20%, the current health status data is determined to be usable for updating. The health status data meets the following conditions:

[0054] 1. Load stability: Within the current time window (e.g., 60 seconds), the fluctuation rate of the load torque signal is lower than a preset threshold (e.g., 15%).

[0055] 2. Vibration stability: Within the same time window, the fluctuation rate of the total energy (or characteristic frequency band energy) of the vibration acceleration signal is lower than a preset threshold (e.g., 20%).

[0056] 3. Normal lubrication condition: Lubrication parameters (such as viscosity and dielectric constant) are within the normal operating range specified by the equipment manufacturer;

[0057] 4. No alarm history: Within the current time window and a previous period (e.g., 10 minutes), the system has not triggered any fault alarms.

[0058] The load disturbance spectrum mapping relationship is dynamically updated using a dynamic update algorithm only when data from M consecutive time points (e.g., M=5) are all determined to be in a healthy state. The dynamic update algorithm employs an exponentially weighted moving average algorithm, with the specific formula as follows:

[0059]

[0060] in, For the updated load perturbation spectrum template;

[0061] The original load perturbation spectrum template;

[0062] This is the average spectrum over a given period of time.

[0063] γ is the forgetting factor, which is usually taken between 0.85 and 0.95 depending on the stability requirements of the application scenario.

[0064] The index-weighted moving average algorithm is executed by a real-time computing engine on an embedded system, ensuring that the mapping relationship can gradually adapt to the slow changes in the device over time, while avoiding template distortion caused by a single abnormal data point.

[0065] Through the aforementioned construction and updating methods, this application enables the load disturbance spectrum mapping relationship to provide an accurate benchmark reference during the initial stage of system commissioning, while also adapting to the slow changes in equipment status during long-term operation. This dynamic maintenance mechanism ensures that adaptive differential processing can always effectively separate load disturbances from fault characteristics, thereby maintaining the accuracy of fault prediction throughout the entire system lifecycle.

[0066] This application further proposes that, based on the load torque signal, the specific steps for calling the corresponding load disturbance spectrum template from the pre-stored load disturbance spectrum mapping relationship and performing adaptive differential processing on the time-frequency distribution data include:

[0067] The system reads the load torque signal from the torque sensor in real time, maps the load torque signal to multiple preset load ranges, and calls the load disturbance spectrum template corresponding to each load range from the load disturbance spectrum mapping relationship. ,in The frequency is indicated. These load ranges are pre-defined based on the actual operating characteristics of the transmission system. For example, 0%-30% of the rated load can be classified as a light load range, 30%-70% as a medium load range, and 70%-100% as a heavy load range. The specific range boundary values ​​can be adjusted according to the system's rated parameters. The mapping process uses a nearest neighbor matching algorithm to assign the current load torque value to the nearest load range.

[0068] Subsequently, the system enters the core adaptive scaling factor calculation phase. This phase requires calculating two key energy parameters: one is the current time-frequency distribution data. In the characteristic frequency band Total energy within Secondly, the template energy of the load perturbation spectrum template P(f) within the same characteristic frequency band. .

[0069] The selection of the characteristic frequency band is determined based on the equipment characteristics. For gear transmission systems, the band containing the meshing frequency and its harmonics is typically selected, while for bearings, the band near the fault characteristic frequency is selected. Energy calculation employs a numerical integration method, specifically:

[0070] Total Energy Calculated by numerical integration:

[0071]

[0072] Template energy The same calculation method is used:

[0073]

[0074] in, , These are the upper and lower limits of the characteristic frequency band, respectively.

[0075] Based on the above energy values, the system calculates the adaptive scaling factor. The calculation formula is as follows:

[0076]

[0077] in, The preset gain parameter typically ranges from 0.5 to 1.2. The specific value of this coefficient is determined through experimental calibration. Under healthy system conditions, a series of step load changes covering the entire operating range are applied, and the results are calculated for different... The energy volatility of the decoupled operating conditions after value processing. Select the value that minimizes this volatility. The value is used as the final calibration value. The purpose of this process is to adjust the intensity of the differential processing to avoid overcompensation. Excessive size leads to feature distortion) or insufficient compensation ( (Too small a value can lead to residual load interference).

[0078] After calculating the scaling factor, the system performs adaptive differencing on the time-frequency distribution data to generate processed time-frequency distribution data. The calculation formula is:

[0079]

[0080] Adaptive differential operations are performed in parallel at each frequency point in the time-frequency domain, and the computation process is accelerated by the vector operation unit in the embedded processor. To address potential negative values ​​in the processing results, the system employs a zero-setting process, setting all negative values ​​to zero to ensure the physical validity of the time-frequency distribution data.

[0081] Through the aforementioned adaptive differential processing method, this application can dynamically adjust the load disturbance compensation intensity according to the current vibration energy level, avoiding the inadequacy of fixed coefficient compensation under varying operating conditions. This processing method retains the benchmark reference provided by the load disturbance spectrum template and achieves precise control of the compensation amount through real-time energy comparison, thereby effectively separating the vibration characteristics related to component degradation under varying load conditions.

[0082] like Figure 3 As shown, this application further proposes to invoke the corresponding correction rule set based on the lubrication state parameter signal. The correction rule set includes a lubrication characteristic morphology mapping relationship, aiming to quantify the nonlinear modulation effect of lubrication state on vibration spectrum morphology and provide a quantitative correction basis for subsequent morphology reconstruction processing. This mapping relationship is constructed by collecting corresponding data of lubrication state and vibration response during actual operation and based on statistical learning methods.

[0083] The process of establishing the mapping relationship between lubrication feature morphology includes:

[0084] During the mapping relationship establishment phase, the system continuously monitors and collects lubrication status parameter signals and synchronous vibration acceleration signals during the operation of the mechanical transmission system. The lubrication status parameters are obtained through online oil sensors, including but not limited to oil viscosity (unit: cSt), oil dielectric constant (dimensionless), and oil temperature (unit: ℃).

[0085] To ensure that the collected samples accurately reflect the essential correlation between lubrication status and vibration characteristics, the system sets strict sample screening conditions: when the fluctuation rate of the load torque signal within a preset third time window is lower than the first preset threshold and the total energy fluctuation rate of the vibration acceleration signal within the same time window is lower than the second preset threshold, the equipment is determined to be in a stable and healthy operating state, and only the data collected under these conditions can be used as valid samples.

[0086] The first and second preset thresholds are empirical values ​​determined based on the inherent fluctuation characteristics of the transmission system under steady-state operating conditions. The first preset threshold ranges from 10% to 20%, preferably 15%, and the second preset threshold ranges from 10% to 30%, preferably 15%. The preset third time window is typically set to 10-60 seconds to cover several complete equipment operating cycles. This screening mechanism effectively eliminates abnormal data caused by sudden load changes or external shocks, ensuring the quality of the training samples.

[0087] The system combines the current lubrication state parameter signal with the energy distribution data of time-frequency distribution data generated based on synchronized vibration acceleration signals across different frequency bands. For each valid sample point, the system records the current multidimensional lubrication state parameter vector L and simultaneously extracts the energy distribution vector E of the time-frequency distribution data generated based on vibration acceleration signals across different frequency bands. These are recorded and stored as sample pairs. The frequency band division is determined based on the characteristic frequencies of the transmission system, typically including the 1st to 3rd harmonics of the meshing frequency and its sidebands, thus forming a feature vector describing the spectral morphology. The lubrication state parameter vector L and the energy distribution vector E together constitute a sample pair (L, E) and are stored in the training sample database.

[0088] Using accumulated sample pairs, a quantitative prediction model for the relationship between lubrication state and spectral energy distribution is established through a multiple linear regression algorithm. This quantitative prediction model characterizes the expected changes in spectral morphology caused by variations in lubrication conditions under ideal health conditions. The mathematical expression of the quantitative prediction model is:

[0089]

[0090] in, Let be the predicted energy distribution vector, W be the regression coefficient matrix, and b be the bias vector. The model parameters of the quantitative prediction model are solved by minimizing the mean squared error between the predicted and observed values.

[0091]

[0092] Where N is the number of training samples. and Let represent the energy distribution vector and lubrication state parameter vector of the i-th sample, respectively.

[0093] The parameters (W, b) of the quantitative prediction model are stored and updated to represent the lubrication characteristic morphology mapping relationship; the correction rule set consists of the gain coefficients predicted after inputting the current lubrication state parameter signal into the quantitative prediction model. .

[0094] In the set of correction rules for practical applications, the lubrication state parameter signals currently collected in real time are included. Input the quantitative prediction model to calculate the predicted energy distribution. Furthermore, by comparing the gain coefficient with the original reference energy distribution, the gain coefficient for each frequency point can be derived. It is used to compensate and correct the vibration spectrum morphology.

[0095] Through the above technical solutions, this application can quantitatively evaluate the modulation effect of lubrication state changes on vibration spectrum, effectively reduce the interference of lubrication condition fluctuations on fault feature extraction, improve the ability of operating condition decoupling features to characterize the true health state of components, and provide a technical basis for accurate fault trend prediction under variable lubrication conditions.

[0096] This application further proposes that, based on the lubrication state parameter signal, the corresponding correction rule set is invoked to perform morphological reconstruction on the processed time-frequency distribution data. This is a crucial step in compensating for spectral distortion caused by lubrication state after load disturbance suppression. This processing, based on a pre-established lubrication characteristic morphological mapping relationship, achieves refined correction of the vibration spectrum through frequency domain gain adjustment. Specific steps include:

[0097] The system first classifies the real-time acquired lubrication status parameter signals. These parameters include oil viscosity, dielectric constant, and oil temperature. The lubrication status parameter signals are compared with the boundary values ​​of multiple preset lubrication status intervals to determine the current lubrication status category. The lubrication status intervals are divided according to the equipment manufacturer's technical specifications and practical experience, for example, into three categories: "excellent lubrication," "normal lubrication," and "poor lubrication," with each category corresponding to a set of specific parameter thresholds.

[0098] In practical implementation, taking oil viscosity as an example, a viscosity value higher than 80% of the standard value for new oil is considered "excellent lubrication," between 60% and 80% is considered "normal lubrication," and below 60% is considered "poor lubrication." The system uses a multi-parameter joint criterion; when the state categories indicated by the various parameters are inconsistent, the final state category is determined according to the worst-case principle.

[0099] Based on the lubrication condition category, the gain coefficients for different frequency bands are retrieved from the stored lubrication characteristic morphology mapping relationship. Gain coefficient The quantitative prediction model is generated with dimensions consistent with the number of frequency points in the time-frequency distribution data. The specific value of the gain coefficient reflects the degree of compensation required for each frequency band under different lubrication conditions. In frequency bands significantly affected by lubrication (usually the low-frequency range of 200-800Hz), it may exhibit a gain value greater than 1, while in the relatively stable high-frequency range, it is close to 1.

[0100] After the gain coefficient is called, the system performs a spectrum morphology adjustment operation, converting the time-frequency distribution data after load differential processing... With gain coefficient Multiply by frequency points, adjust the spectral shape, and output the decoupling characteristics of the operating conditions. The calculation formula is:

[0101] ;

[0102] in, It represents the Hadamard product, used to compensate for or suppress frequency band energy that is significantly affected by lubrication conditions.

[0103] This operation is performed independently in the frequency dimension, applying a corresponding gain adjustment to the amplitude at each frequency point f, thereby achieving precise reshaping of the spectrum shape.

[0104] Through the aforementioned morphological recombination process, this application can effectively compensate for spectral distortion caused by changes in lubrication conditions, ensuring that the final decoupled operating condition features no longer contain feature variations caused by lubrication conditions. This targeted spectral correction, combined with the aforementioned load disturbance suppression, forms a complete operating condition decoupling processing chain, providing subsequent fault trend prediction with feature inputs that purely reflect the health status of mechanical components, thereby maintaining the accuracy of the fault prediction model under varying lubrication conditions.

[0105] This application further proposes that the working condition decoupling feature generation module is also used for:

[0106] Low-energy residual components are extracted from the time-frequency distribution data after morphological reconstruction. Further mining is conducted on those subtle features with low energy but fault early warning value, and the feature characterization ability is enhanced by fusing information from multiple measurement points.

[0107] In practical implementation, the system extracts low-energy residual components by setting a dynamic energy threshold. This threshold is set to 5%-10% of the global energy of the current time-frequency matrix, a proportion determined based on statistical analysis of the energy levels of early fault characteristics in numerous fault cases. The extraction method employs soft-threshold filtering to retain weak components below the threshold but above the noise baseline, forming an independent low-energy residual time-frequency distribution for each measurement point. , where i represents the measurement point number.

[0108] After obtaining the low-energy residual components at each measuring point, the low-energy residual components extracted from different measuring points on the transmission chain are time-aligned and energy-normalized. Time alignment is based on the physical layout of the transmission chain and the signal propagation speed. By calculating and compensating for the fixed time delay between each measuring point, it is ensured that the signals of the same vibration event at different measuring points are synchronized in time. Energy normalization adopts the minimum-maximum normalization method to map the residual component energy of each measuring point to a unified range of [0,1], eliminating the inconsistency of basic energy levels caused by differences in sensor sensitivity or installation position.

[0109] When generating the focused residual feature vector, the system introduces an adaptive weight allocation mechanism based on historical fault sensitivity. Based on the historical fault sensitivity of each measurement point, adaptive weights are assigned to the corresponding low-energy residual components and then weighted and superimposed to generate the focused residual feature vector.

[0110] Weight of each measuring point The early warning capability for specific fault types is calculated based on historical data from this measurement point. Specifically:

[0111]

[0112] in, This represents the early detection probability of the i-th measurement point in historical fault cases;

[0113] This indicates the false alarm probability at that measurement point;

[0114] It is a tiny positive number (usually 0.01) set to prevent division by zero.

[0115] Weight calculation relies on a fault case library accumulated during system operation and is reassessed and updated every six months. The weighted summation process is implemented using the following formula:

[0116]

[0117] Where N is the total number of measuring points. This refers to the residual feature vector generated after focusing.

[0118] The trend prediction module, based on the decoupled features of operating conditions and the focused residual feature vectors, calculates fault development trend data and remaining life estimation data through the fault prediction model. This fusion approach allows the prediction model to utilize both the obvious fault information in the main features and the early, subtle signs in the residual features.

[0119] Through the above technical solution, this application can effectively enhance the detection sensitivity of early faults, especially for weak fault features that are easily masked by the main vibration components in conventional vibration analysis. The multi-measurement point weighted fusion strategy makes full use of monitoring information at different locations on the transmission chain. By assigning higher weights to fault-sensitive areas, it optimizes the feature representation efficiency, thereby improving the early warning capability and prediction accuracy of the fault prediction system for potential faults.

[0120] like Figure 4 As shown, in one embodiment, this application further proposes that the system also includes a periodic drift identification module, which aims to obtain dynamic features reflecting changes in the state of mechanical components such as clearance and wear by analyzing the temporal stability of periodic events in the transmission system. This periodic drift identification module operates in parallel with the decoupled feature generation for operating conditions, together forming a multi-dimensional feature extraction system.

[0121] In practical implementation, the periodic drift identification module is used for:

[0122] This study identifies event time sequence of periodic meshing or impact events from vibration acceleration signals. A peak detection algorithm based on wavelet transform is employed. A Morlet wavelet basis function is designed centered on the inherent meshing frequency of the transmission system. By identifying the modulus maxima of the wavelet coefficients, the event occurrence time is accurately located, forming the original event time sequence. , where i=1,2,...,M represents the event sequence number.

[0123] Synchronously, the system acquires the rotational speed signal of the mechanical transmission system through an encoder mounted on the input shaft. This rotational speed signal records the rotor angle change in the form of a pulse sequence. By mapping the event time point sequence to the same rotor cycle reference system through the rotational speed signal, a standardized event time point sequence is obtained.

[0124] Specifically, a unified time base is constructed using rotational speed signals, mapping the original event time sequence to the angle domain:

[0125]

[0126] in, This represents the standardized perspective of the event. The reference start time for an analysis window. This represents the rotor rotation period. This mapping process eliminates the impact of speed fluctuations on event time series analysis, resulting in a standardized event time point sequence. .

[0127] Calculate the time interval sequence between adjacent periods based on the standardized event time point sequence. The calculation formula is as follows:

[0128]

[0129] Where j and j+1 represent the current event time point and the next event time point, respectively;

[0130] Event time interval sequence This reflects the stability of periodic events in the angular domain.

[0131] The rate of change within a preset sliding window of the event time interval sequence is calculated as the drift rate sequence. The system uses a sliding window of 20-30 periods to calculate the rate of change, i.e., the drift rate, of the interval sequence.

[0132]

[0133] in, This represents the drift rate of the k-th window. The sliding window length is typically set to 20 to 30 event cycles. This length is determined based on the following: it should be much larger than the inherent fluctuation period of the transmission system's rotational speed (usually <5 cycles) to ensure smoothing of random fluctuations; at the same time, it should be much smaller than the typical timescale of cycle drift caused by fault development (usually >50 cycles) to ensure sensitivity in detecting early faults.

[0134] Based on the calculated drift rate sequence The system constructs a periodic drift feature vector as an auxiliary feature vector. This vector contains statistical characteristics of the drift rate: mean, standard deviation, skewness, and the proportion exceeding a threshold (usually set to 0.5%), forming a four-dimensional feature vector. These features describe the stability characteristics of periodic events from different perspectives. The mean reflects the average drift trend, the standard deviation characterizes the degree of volatility, the skewness indicates the asymmetry of the distribution, and the proportion exceeding the threshold reflects the degree of anomaly.

[0135] The trend prediction module, based on the decoupling characteristics of operating conditions and the periodic drift eigenvector, calculates fault development trend data and remaining life estimation data through a fault prediction model. This fusion allows the prediction system to not only focus on changes in vibration energy but also consider the evolution of the dynamic temporal characteristics of the mechanical system.

[0136] By introducing periodic drift feature analysis, this application can capture subtle temporal variations caused by changes in mechanical state, such as increased gear clearance and changes in bearing clearance. These temporal features complement the vibration energy features, providing additional evidence for fault prediction. This is particularly helpful in the early identification of progressive faults that primarily manifest as changes in fit rather than an increase in vibration energy, thereby improving the comprehensiveness and reliability of the prediction system.

[0137] like Figure 5 As shown, in one embodiment, this application further proposes that the system also includes a transmission delay difference extraction module, which aims to capture the changes in dynamic characteristics caused by structural stiffness degradation, loose connections or component wear by analyzing the changes in the propagation characteristics of vibration signals in the transmission chain, and provide supplementary features reflecting the wave propagation characteristics for fault prediction.

[0138] The conduction delay difference extraction module is used for:

[0139] Short-time cross-correlation analysis was performed on the synchronous vibration acceleration signals at different measuring points on the transmission chain to obtain the transmission delay difference sequence between the signals.

[0140] In practical implementation, the conduction delay difference extraction module first performs short-time cross-correlation analysis on the synchronous vibration acceleration signals at different measuring points on the transmission chain. A sliding time window processing method is adopted, with the window length set to include the duration of 2-3 major impact events (typically 100-200 milliseconds), and the window overlap rate set to 50% to ensure the continuity of the analysis. For the signals of any two adjacent measuring points i and j... and During the time window , The cross-correlation function within the range is calculated as follows:

[0141]

[0142] Where τ is the time delay variable.

[0143] The value of τ corresponding to the maximum value of the cross-correlation function of the conduction delay difference:

[0144]

[0145] By repeating this process over a continuous time window, a conduction delay difference sequence is obtained. .

[0146] Considering the impact of load variations on wave propagation speed, load scaling correction is applied to the conduction delay difference sequence based on the load torque signal to eliminate the influence of load variations on conduction speed. The relationship between load and conduction speed was established through previous calibration experiments, and the correction formula is as follows:

[0147]

[0148] in, This is the corrected conduction delay difference. To measure the conduction delay difference, The current load torque, Where N is the rated load torque, and k is the load influence coefficient (determined experimentally, with a typical value of 0.001-0.005 N·m). This correction eliminates the systematic impact of load fluctuations on the conduction delay difference, allowing subsequent analysis to focus on delay variations caused by equipment status.

[0149] A conduction delay difference feature vector is constructed as an auxiliary feature vector based on the corrected conduction delay difference sequence. This feature vector includes several statistical characteristics such as the mean, standard deviation, coefficient of variation, and the proportion of delay differences exceeding a threshold (usually set to 15% of the baseline value). Furthermore, the decay time of the autocorrelation function of the delay difference sequence is calculated to assess the persistence of delay variations. These features collectively constitute a multidimensional feature vector describing the stability of wave propagation characteristics.

[0150] The trend prediction module calculates fault development trend data and remaining life estimation data based on the decoupling characteristics of operating conditions and the characteristic vector of conduction delay difference through the fault prediction model.

[0151] Through the aforementioned process of extracting the transmission delay difference feature, this application can effectively monitor the characteristic changes of the wave propagation path in a mechanical transmission system. These changes are often closely related to the structural connection state, component fit clearance, and material stiffness characteristics. By fusing the transmission delay difference feature vector as an auxiliary feature with the decoupled feature of the main operating condition, unique information reflecting the wave propagation characteristics is provided for the fault prediction model. This enhances the early identification capability of faults such as structural loosening, bearing wear, and abnormal gear clearance. In particular, it provides supplementary diagnostic basis that is difficult to obtain through traditional vibration analysis when predicting fault types related to the dynamic force transmission path.

[0152] This application further proposes to calculate the fault development trend data of the mechanical transmission system through a pre-trained fault prediction model, specifically including the following steps:

[0153] The system caches the decoupling features of the most recent N time points in chronological order, generating a time series matrix of decoupling features. In implementation, the system first establishes a time series analysis framework for the decoupling features. The decoupling features of the most recent 128 sampling time points are cached in chronological order. This value is determined based on the time scale of typical fault development in the transmission system, covering operating data from several hours to several days. These features are then arranged in chronological order to form the time series matrix of decoupling features. , where d is the feature dimension.

[0154] Principal component analysis was performed along the time dimension on the time series matrix of the decoupling features of the operating conditions to extract the first principal component score sequence containing the maximum variance information. , ,..., This reflects the most significant change pattern in equipment status. A linear fit is then performed on the first principal component score sequence:

[0155]

[0156] Where a is the intercept and b is the slope obtained from the fitting. This represents the residual term. The slope b obtained from the fitting is used as the global trend strength factor. A positive value indicates that the state is deteriorating, while a negative value indicates that the state is recovering. The absolute value reflects the rate of change.

[0157] The decoupled features of the current operating condition are input into a pre-trained multi-class neural network. The training process of the pre-trained multi-class neural network and the fault prediction model is based on historical full lifecycle data. The training dataset comes from complete data records of at least 30 identical mechanical transmission systems from healthy operation to failure, containing more than 500 sample sequences covering different loads and lubrication conditions. In the data preprocessing stage, the original vibration signals are subjected to time-frequency transformation, and the Z-score normalization method is used to eliminate the influence of dimensions. The multi-class neural network adopts a deep fully connected architecture, containing 3 hidden layers, each with 256 neurons, using the ReLU activation function, and is trained using historical full lifecycle data from operation to failure. During training, the Adam optimizer is used, the learning rate is set to 0.001, and the batch size is 32.

[0158] The probability distribution of the output mechanical transmission system transitioning from its current healthy state to multiple predefined degenerate states within a future preset first time window (e.g., a 24-hour time window). , where m is the number of state categories (usually set to 4-6 ordered degenerate states). A state deterioration tendency index is calculated based on this probability distribution. :

[0159]

[0160] in, The weight coefficient for the i-th state is set incrementally according to the severity of the state.

[0161] Finally, the global trend intensity factor and the state deterioration tendency index are weighted and fused to generate a comprehensive fault development trend score, which serves as the main component of the fault development trend data. The weight allocation is based on the historical predictive performance of each indicator; typically, the trend intensity factor is weighted at 0.6, and the state deterioration tendency index at 0.4, resulting in the comprehensive fault development trend score.

[0162]

[0163] in, and The normalization coefficient ensures that the two indicators have the same dimensions.

[0164] By employing the aforementioned fault development trend scoring method, this application can simultaneously consider both the long-term, slow changing trend of equipment status and the short-term risk of state transitions. It reflects both the gradual degradation of the status and the possibility of sudden changes. This dual-perspective assessment mechanism provides a more comprehensive basis for predictive maintenance decisions, enabling maintenance personnel to adjust maintenance strategies promptly based on score changes, thereby optimizing the allocation of maintenance resources while ensuring the safe operation of equipment.

[0165] This application further proposes that the specific steps for calculating the fault development trend data and remaining life estimation data of the mechanical transmission system through a pre-trained fault prediction model include:

[0166] The pre-stored fault feature template is invoked to calculate the feature degradation distance D between the current operating condition decoupling feature and the fault feature template. The fault feature template is pre-stored in the equipment database and is derived from the feature states of historical operation up to complete failure. It is established by collecting the feature average of the last 8 hours of operating data of multiple similar equipment before the failure occurred.

[0167] The feature degradation distance D is calculated using Mahalanobis distance to account for the correlation between the feature dimensions:

[0168]

[0169] Where D is the feature degradation distance. This represents the decoupling feature vector of the current operating condition. Σ is the fault feature template vector, and Σ is the covariance matrix of the feature distribution, which is estimated through historical normal operation data.

[0170] Simultaneously, the system calculates the slope K of the change trend of the decoupling characteristics of the operating conditions within a preset second time window (usually set to 8-24 hours, preferably 24 hours). This calculation uses the weighted least squares method to linearly fit the feature sequence of the most recent 24 sampling points, with more recent data points being given higher weights to more sensitively reflect the recent state change trend.

[0171] The feature degradation distance D and the slope of the change trend are input into a pre-trained nonlinear function to calculate the remaining lifetime estimate. This nonlinear function is trained by collecting no less than 50 sets of historical full life cycle data from operation to failure, and the Levenberg-Marquardt optimization algorithm is used to solve the model parameters.

[0172] The nonlinear function is trained using full lifecycle data from historical operation to failure, and its expression is:

[0173]

[0174] Where RUL is the estimated remaining lifetime (in hours), D is the feature degradation distance, K is the slope of the trend, A represents the initial remaining lifetime baseline value of the equipment in a healthy state, B controls the influence of the feature degradation distance on lifetime estimation, and C adjusts the weight of the state change trend. The model parameters A, B, and C are determined through the training process, with typical value ranges of A∈[200,500], B∈[0.1,0.5], and C∈[50,150], respectively. The specific values ​​vary depending on the equipment type and operating conditions.

[0175] During training, the objective function is set to minimize the root mean square error between the predicted remaining lifetime and the actual remaining lifetime. The training data needs to cover the complete degradation process from a healthy state to a complete failure, including diverse degradation paths under different operating conditions and load conditions.

[0176] Using the aforementioned remaining life estimation method, the system can dynamically assess the remaining lifespan of equipment based on the relative position of the current state and the fault state, as well as the rate of state change. It considers both the absolute degree of state degradation and the trend information of state changes, enabling the lifespan estimation results to reflect the accelerated degradation or recovery of equipment status in a timely manner. This provides a more accurate and adaptive time reference for maintenance planning, exhibiting better predictive adaptability, especially when facing nonlinear degradation processes.

[0177] The following is a specific embodiment of a fault trend prediction system for mechanical transmission systems based on dynamic feature recognition:

[0178] The main conveyor belt gear transmission system (model ZLYJ250, rated power 160kW) of a large open-pit mine operates under conditions of high dust concentration and frequent load fluctuations, with the rated load switching range covering 50% to 100%. The lubrication system is susceptible to particulate contamination, leading to dynamic changes in its condition. Traditional fault prediction methods often result in false alarms due to sudden load increases or missed potential faults due to lubrication contamination masking wear characteristics. However, by applying a fault trend prediction system based on dynamic feature recognition, accurate monitoring and early warning under complex operating conditions have been achieved.

[0179] The system employs a multi-source sensor array for synchronous data acquisition: an HBM T40B torque sensor (100Hz sampling rate) is installed on the input shaft to acquire load torque signals in real time; a viscosity sensor (2-20cSt measurement range) and a dielectric constant sensor (50Hz sampling rate) are integrated into the gearbox lubrication circuit to monitor lubrication status; three PCB 352C65 triaxial accelerometers (2048Hz sampling rate) are distributed in the transmission chain bearing housing to acquire vibration acceleration signals. All sensors achieve hardware-level synchronization via the IEEE 1588 PTP protocol, with time errors controlled within 1ms.

[0180] The vibration signal is converted into time-frequency distribution data S(f,t) by short-time Fourier transform (window length 1024 points, Hanning window, overlap rate 75%). When the real-time load torque is 80% of the rated load, the system calls the pre-stored load disturbance spectrum template P(f) in the mid-load range; and calculates the total energy of the current time-frequency distribution in the gear meshing frequency band (250Hz±50Hz). (Relative energy unit), template energy (Relative energy unit), preset gain parameter =0.8, adaptive scaling factor ; obtained through adaptive differential processing This effectively eliminates load fluctuation interference. For the lubrication condition (current viscosity 18 cSt, dielectric constant 1.8), the gain coefficient G(f) is generated by calling the lubrication characteristic morphology mapping relationship—based on a multiple linear regression model to predict energy distribution. =[1.36,2.39], compared to the baseline energy distribution In comparison =1.13、 =1.195; Morphological recombination is achieved through the Hadamard product. It compensates for spectral distortion caused by lubrication contamination and decouples the output operating conditions.

[0181] The periodic drift identification module identifies the time sequence of gear meshing events from the vibration signal, maps it to the rotor periodic reference system in conjunction with the speed encoder signal (1024 pulses / revolution), and calculates the interval between adjacent periodic events. Sliding window drift rate The periodic drift feature vector is constructed; the conduction delay difference extraction module performs short-time cross-correlation analysis on vibration signals at different measuring points, and obtains the conduction delay difference sequence after load scaling correction, and extracts statistical features to form an auxiliary feature vector.

[0182] The fault prediction model receives decoupling features and auxiliary features of operating conditions, caches the decoupling features of the most recent 128 time points to generate a time series matrix, extracts the first principal component score sequence through principal component analysis and obtains a global trend intensity factor b=0.03 through linear fitting; the multi-class neural network outputs the probability distribution of degradation state in the next 24 hours. =[0.6,0.3,0.1], calculate the deterioration tendency index. =0.6 + 10.3 + 20.1 = 0.5; Overall fault development trend score =100.03 + 0.50.5 = 0.55, indicating a slow degradation trend in the system. The remaining lifetime is estimated by calculating the characteristic degradation distance D = 1.2 (Mahanobis distance) and the trend slope K = 0.04, and then substituting these values ​​into a nonlinear function. =300 exp(-0.41.2)+500.04≈187 hours, providing a precise basis for the maintenance plan.

[0183] Traditional monitoring systems often trigger false alarms under these mine conditions due to sudden load increases (such as increased vibration amplitude during heavy-load startup being misjudged as a fault) or miss alarms due to lubrication contamination masking early wear characteristics. After applying this system, in a case where the load jumped from 60% to 90%, adaptive differential processing successfully isolated the load disturbance and did not generate false alarms. When the dielectric constant of the lubricating oil increased abnormally (indicating contamination), morphological recombination compensation effectively captured the early wear characteristics of the gear tooth surface, issuing an early warning. Maintenance personnel replaced the lubricating oil and worn gears in time, avoiding mine production interruptions caused by conveyor belt shutdowns and improving the reliability and economy of equipment operation and maintenance.

[0184] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A fault trend prediction system for mechanical transmission systems based on dynamic feature recognition, characterized in that: include: The signal acquisition module is used to synchronously acquire the load torque signal and lubrication status parameter signal of the mechanical transmission system, as well as the vibration acceleration signal of multiple measuring points on the transmission chain, and send the acquired multi-source signals to the working condition decoupling feature generation module. The operating condition decoupling feature generation module, connected to the signal acquisition module, is used to receive the multi-source signals and perform the following operations: Time-frequency distribution data is generated based on the vibration acceleration signal; Based on the load torque signal, the corresponding load disturbance spectrum template is retrieved from the pre-stored load disturbance spectrum mapping relationship, and the time-frequency distribution data is subjected to adaptive differential processing to generate processed time-frequency distribution data. Based on the lubrication state parameter signal, the corresponding correction rule set is invoked to reshape the processed time-frequency distribution data, output the working condition decoupling feature that characterizes the health status of the mechanical component, and send the working condition decoupling feature to the trend prediction module. The trend prediction module, connected to the working condition decoupling feature generation module, is used to calculate the fault development trend data and remaining life estimation data of the mechanical transmission system based on the working condition decoupling features and through a pre-trained fault prediction model. The fault development trend data of the mechanical transmission system is calculated using a pre-trained fault prediction model, specifically including the following steps: The operating condition decoupling features of the most recent N time points are cached in chronological order to generate a time series matrix of operating condition decoupling features. Principal component analysis is performed on the time series matrix of the decoupling feature of the working condition along the time dimension. The first principal component score sequence is extracted and linearly fitted. The slope obtained from the fitting is used as the global trend intensity factor. The decoupling features of the current working condition are input into a pre-trained multi-class neural network, which outputs the probability distribution of the mechanical transmission system transitioning from the current healthy state to multiple predefined degenerate states within a preset first time window in the future, and calculates the state deterioration tendency index. The global trend intensity factor and the state deterioration tendency index are weighted and fused to generate a comprehensive fault development trend score, which is used as a component of the fault development trend data.

2. The mechanical transmission system fault trend prediction system based on dynamic feature recognition according to claim 1, characterized in that: The step of calling the corresponding load perturbation spectrum template from the pre-stored load perturbation spectrum mapping relationship includes the following process for constructing and updating the load perturbation spectrum mapping relationship: During the stable operation phase when the mechanical transmission system is determined to be in a healthy state, vibration acceleration signals under different load torque signals are collected. Extract the time-frequency distribution data of the vibration acceleration signal, calculate the average time-frequency distribution data corresponding to each load torque signal, and store it as the load disturbance spectrum template; During online monitoring, the load disturbance spectrum mapping relationship is dynamically updated based on newly collected health status data using a dynamic update algorithm.

3. The mechanical transmission system fault trend prediction system based on dynamic feature recognition according to claim 1, characterized in that: The specific steps for retrieving the corresponding load disturbance spectrum template from the pre-stored load disturbance spectrum mapping relationship based on the load torque signal and performing adaptive differential processing on the time-frequency distribution data include: The load torque signal is mapped to multiple preset load ranges, and the load perturbation spectrum template corresponding to each load range is called from the load perturbation spectrum mapping relationship. ,in Indicates frequency; Calculate the current time-frequency distribution data Total energy within the characteristic frequency band and the load perturbation spectrum template Template energy within the same characteristic frequency band Calculate the adaptive scaling factor : ,in These are the preset gain parameters; Adaptive difference operations are performed on the time-frequency distribution data to generate the processed time-frequency distribution data. The calculation formula is: .

4. The mechanical transmission system fault trend prediction system based on dynamic feature recognition according to claim 3, characterized in that: The corresponding correction rule set is invoked based on the lubrication state parameter signal; wherein, the correction rule set includes a lubrication feature morphology mapping relationship, and the process of establishing the lubrication feature morphology mapping relationship includes: During the operation of the mechanical transmission system, the lubrication status parameter signals and synchronous vibration acceleration signals are continuously monitored and collected; When the fluctuation rate of the load torque signal is lower than the first preset threshold and the total energy fluctuation rate of the vibration acceleration signal is lower than the second preset threshold, the equipment is determined to be in a stable and healthy operating state. The lubrication status parameter signal at the current moment and the energy distribution of the time-frequency distribution data generated based on the synchronized vibration acceleration signal in different frequency bands are recorded and stored as sample pairs. Using accumulated sample pairs, a quantitative prediction model is fitted using a multiple linear regression algorithm; The parameters of the quantitative prediction model are stored and updated to the lubrication characteristic morphology mapping relationship; wherein, the correction rule set is the gain coefficient predicted after inputting the current lubrication state parameter signal into the quantitative prediction model. .

5. The mechanical transmission system fault trend prediction system based on dynamic feature recognition according to claim 4, characterized in that: The specific steps for morphological reconstruction of the processed time-frequency distribution data based on the lubrication state parameter signal and the corresponding correction rule set include: The lubrication state parameter signal is compared with the boundary values ​​of multiple preset lubrication state intervals to determine the current lubrication state category. Based on the lubrication state category, the gain coefficients for different frequency bands are retrieved from the lubrication characteristic morphology mapping relationship. ; The processed time-frequency distribution data With the gain coefficient Multiply by frequency points, adjust the spectral shape, and output the decoupling characteristics of the operating condition. The calculation formula is: ； in, It represents the Hadamard product, used to compensate for or suppress frequency band energy that is significantly affected by lubrication conditions.

6. The mechanical transmission system fault trend prediction system based on dynamic feature recognition according to claim 1, characterized in that: The specific steps for calculating the fault development trend data and remaining life estimation data of the mechanical transmission system using a pre-trained fault prediction model include: Call the pre-stored fault feature template and calculate the feature degradation distance between the current operating condition decoupling feature and the fault feature template; The feature degradation distance and the slope of the change trend of the operating condition decoupling feature within a preset second time window are input into a pre-trained nonlinear function to calculate the remaining lifetime estimate. The nonlinear function is obtained by training on the entire lifecycle data from historical operation to failure, and its expression is: Where RUL is the estimated remaining lifetime, D is the feature degradation distance, K is the slope of the change trend, and A, B, and C are model parameters determined through training.

Images