A connector intelligent detection method and system based on big data

Abstract

The application relates to the technical field of connector intelligent detection, and discloses a connector intelligent detection method and system based on big data, wherein the method carries out feature extraction on multi-dimensional operation condition data, screens out target physical alternating features, builds internal and external physical change mapping relations, analyzes interface wear features and deduces degradation tracks, and after completing abnormal elimination and missing completion of the tracks, carries out three-dimensional reconstruction and health assessment of the contact interface. The method can rely on big data to establish the correlation mapping of external signals and internal micro loss, accurately identify core wear features, reconstruct continuous degradation tracks, make up for the defects of discrete sampling, and effectively improve the evaluation accuracy of the degradation state and health level of the connector.

Description

Technical Field

[0001] This invention relates to the field of intelligent connector testing technology, and in particular to an intelligent connector testing method and system based on big data. Background Technology

[0002] In industrial equipment, connector contact interfaces undergo microscopic degradation phenomena such as changes in surface roughness and fretting wear under long-term alternating loads. This type of hidden wear can easily lead to equipment failure. Due to limitations in on-site operating conditions, equipment cannot be frequently disassembled for inspection of internal interfaces. The industry generally relies on external sensors to collect multi-dimensional operating data such as vibration, temperature, and rotational speed, and uses industrial cloud computing to store and perform basic analysis of massive amounts of monitoring data to determine the operating status of the equipment.

[0003] Existing equipment health analysis solutions can predict the degradation trend of equipment's macroscopic performance through data preprocessing, feature extraction, and model training. However, these solutions only focus on the statistical analysis of macroscopic operating indicators. Due to the complex nonlinear relationship between external sensor signals and the interface loss inside the connector, there is a large amount of redundant information in the monitoring data. In addition, interface degradation is a multi-physical parameter coupled evolution process, and actual detection suffers from sampling discontinuity. Existing technologies cannot overcome the mapping barrier between external data and internal physical losses.

[0004] Existing technologies cannot establish an explicit mapping relationship between externally acquired signals and microscopic changes in the internal closed contact interface of the connector. They are unable to identify dominant wear characteristics such as surface roughness and fretting wear marks, nor can they complete the continuous reconstruction of degradation trajectories under discrete sampling. Ultimately, this results in insufficient accuracy in assessing the degradation state and health level of the connector. Summary of the Invention

[0005] This invention provides a connector intelligent detection method and system based on big data to solve the problems of existing technologies that cannot establish an explicit mapping relationship between externally acquired signals and microscopic changes in the internal closed interface of the connector, making it difficult to identify dominant wear characteristics such as surface roughness and fretting wear marks, and failing to complete the continuous reconstruction of degradation trajectory under discrete sampling and have low health assessment accuracy. It realizes the accurate mapping of external working condition data to the wear state of the internal interface, effectively extracts core wear features, fills in sampling gaps and reconstructs a complete degradation trajectory, and comprehensively improves the assessment accuracy of connector degradation state and health level.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a connector intelligent detection method based on big data, comprising: Acquire multi-dimensional operating condition data of the equipment connection components, and perform feature extraction processing on the multi-dimensional operating condition data to obtain a first set of operating features; Based on the first set of operational features, independent feature channels are separated and hierarchical screening is performed based on the strength of degradation correlation to obtain the target physical alternation features; Based on the target's physical alternation characteristics, a mapping relationship between externally acquired information and the physical changes in the internal enclosed space is constructed to obtain a physical loss sequence; The physical loss sequence is subjected to multi-dimensional feature extraction and dimensionality reduction reconstruction to obtain the wear characteristics of the target interface; Based on the wear characteristics of the target interface, the coupled evolution of the interface degradation index is time-series extrapolated to obtain a first degradation trajectory sequence. Then, the first degradation trajectory sequence is continuously reconstructed to obtain a second degradation trajectory sequence. Based on the second degradation trajectory sequence, a three-dimensional reconstruction of the contact interface and a comprehensive health status assessment are performed to obtain the target health status of the device connection components.

[0007] Secondly, the present invention provides a connector intelligent detection system based on big data, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts the operating characteristics of the equipment and filters the target physical alternating characteristics based on the degradation correlation strength, establishes the mapping relationship between external operating condition data and the internal interface of the connector, eliminates redundant data interference, effectively connects the external sensing signal and the internal micro loss, and solves the problem of nonlinear mapping between the two.

[0010] (2) This invention extracts two core wear features, surface roughness and fretting wear depth, by performing dimensionality reduction decomposition on the physical wear sequence and using time series model to deduce the multi-parameter coupling evolution law, which can accurately identify the interface-dominated wear type and fully depict the degradation dynamic process.

[0011] (3) This invention fills in the data loss caused by discrete sampling by performing anomaly removal, interpolation completion and smoothing on the degraded trajectory, and then combines three-dimensional reverse reconstruction to complete the state assessment, thereby realizing continuous reconstruction of the degraded trajectory and greatly improving the detection and assessment accuracy of the connector health status. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a connector intelligent detection method based on big data provided in the first embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] Reference Figure 1 The first embodiment of the present invention provides a connector intelligent detection method based on big data, including the following steps: S11, acquire multi-dimensional operating condition data of the equipment connection components, perform feature extraction processing on the multi-dimensional operating condition data, and obtain a first set of operating features; S12, based on the first set of operating features, separate independent feature channels and perform hierarchical screening based on the strength of degradation correlation to obtain target physical alternation features; S13, Based on the target physical alternation characteristics, construct a mapping relationship between externally collected information and the physical changes of the internal enclosed space to obtain a physical loss sequence; S14, perform multi-dimensional feature extraction and dimensionality reduction reconstruction on the physical loss sequence to obtain the wear characteristics of the target interface; S15, based on the wear characteristics of the target interface, perform time-series deduction on the coupled evolution of the interface degradation index to obtain a first degradation trajectory sequence, and then perform continuous reconstruction processing on the first degradation trajectory sequence to obtain a second degradation trajectory sequence. S16, based on the second degradation trajectory sequence, perform three-dimensional reconstruction of the contact interface and comprehensive health status assessment to obtain the target health status of the device connection components.

[0015] In step S11, the multi-dimensional operating condition data of the equipment connection components is acquired, and feature extraction processing is performed on the multi-dimensional operating condition data to obtain a first set of operating features, including: The multidimensional operating condition data of the equipment connection components are acquired and time-aligned to obtain a multidimensional operating condition time sequence. The multidimensional operating condition time series is subjected to frequency domain mapping and dimensionality reduction processing to obtain the principal component vector; Anomaly feature vectors are extracted from the principal component vectors, and then concatenated with historical principal component vectors extracted from historical operating condition data to obtain the first set of operating features.

[0016] It should be noted that the connector described in this invention specifically refers to a mechanical contact connection pair under heavy-duty operating conditions. A typical application scenario is the gear meshing interface of a wind turbine gearbox. The solution can be equivalently transferred to various closed connection structures experiencing fretting wear. Multidimensional operating condition data refers to the time-series data of multiple physical quantities synchronously collected by different types of sensors during equipment operation. For the operation monitoring scenario of a wind turbine gearbox, multidimensional operating condition data specifically includes the triaxial vibration acceleration signal collected by the vibration sensor, the gearbox oil temperature signal collected by the temperature sensor, and the input shaft speed signal collected by the tachometer. Different sensors have inherently different sampling frequencies; the signal sampling rate of a vibration sensor is typically 1000 times per second, while the signal sampling rate of a temperature sensor is typically 10 times per second. Directly performing fusion analysis will lead to a misalignment of the time dimension.

[0017] The system first upsamples the low-frequency time-series data collected by the temperature sensor. A first-order linear interpolation function is constructed between two adjacent original sampling points, assuming that the temperature measurement between the two points changes linearly with time. The estimated values ​​for intermediate moments are calculated by combining the values ​​of the two sampling points with the time difference. For high-frequency sensor data such as vibration and rotational speed data, the sampling rate is consistent with the target alignment interval, requiring only timestamp boundary normalization without resampling. This algorithm has low computational complexity and can quickly complete the time alignment of multi-source data while ensuring engineering accuracy. This ensures that all sensor data have a completely consistent sampling interval on the time axis, outputting a time-synchronized multi-dimensional operating condition time-series sequence.

[0018] The system then performs frequency domain mapping processing on the multi-dimensional operating condition time series, using Fast Fourier Transform (FFT) to convert the time-domain vibration and rotational speed sequences into a frequency-domain feature matrix. FFT decomposes the time-domain signal into a superposition of sinusoidal signals of different frequencies, extracting the frequency components and their amplitude information. The frequency-domain feature matrix directly reflects the variation patterns of the gear meshing frequency and its harmonics; the amplitude variations of these frequency components are closely related to the wear state of the gears.

[0019] Because the frequency domain feature matrix has extremely high dimensionality and contains a large amount of redundant noise, the system further performs dimensionality reduction processing. By calculating the covariance matrix of the frequency domain feature matrix, the eigenvectors corresponding to the top 10 largest eigenvalues ​​are extracted as principal component vectors. This process compresses the original thousands of dimensions of frequency domain data to 10 dimensions, significantly reducing the computational complexity of subsequent calculations while retaining more than 98% of the effective information.

[0020] Based on this, the system calculates the current wear level of the equipment using principal component vectors and uses Euclidean distance to calculate the spatial distance between the current principal component vector and the health status baseline vector. Euclidean distance is a commonly used indicator to measure the similarity between two vectors in multidimensional space; the larger the distance, the greater the difference between the current state and the healthy state, and the more severe the wear level. The health status baseline vector is generated statistically from standard operating data of the same model of equipment after factory break-in. It selects the principal component vectors from the fault-free operation phase during the three months before equipment commissioning, where manual inspections confirmed no abnormal wear and all operating parameters were within the rated range. The mean vector of all samples is taken as the baseline vector, and a corresponding baseline is generated separately for each piece of equipment to eliminate baseline deviations caused by individual processing and assembly differences. The system uses the 3σ criterion, commonly used in industrial condition monitoring, to set the wear threshold. Based on all Euclidean distance samples from the healthy operation phase, the mean and standard deviation of the samples are calculated, and the sum of the mean and three times the standard deviation is taken as the wear threshold. This criterion is based on the statistical principle of normal distribution. The probability of a distance value exceeding this threshold in a healthy state is less than 0.3%, which can effectively identify early state deviations while controlling the false positive rate. The specific value of the threshold is adaptively determined based on the statistical results of the equipment's health samples. When the calculated wear degree is greater than the wear threshold, it is determined that the current equipment has abnormal wear, and the principal component vector is extracted as the abnormal feature vector.

[0021] The system concatenates the extracted abnormal feature vectors with historically acquired unlabeled principal component vectors in chronological order to obtain the first set of operational features. The historically acquired unlabeled principal component vectors are unlabeled feature samples continuously generated during the equipment's daily operation. These vectors are generated after collecting past operating data at fixed intervals, undergoing frequency domain mapping and principal component dimensionality reduction processes identical to the current data, and have not yet been manually labeled or algorithmically determined regarding their health status. These vectors are automatically generated in batches by the system throughout the equipment's entire operating cycle. The system continuously collects multi-dimensional operating data at preset 1-hour intervals. Each batch of data undergoes Fast Fourier Transform and Principal Component Analysis sequentially, generating principal component vectors for the corresponding time points, which are then automatically stored in the historical feature library. No additional manual labeling is required throughout the process. The amount of unlabeled sample data is significantly higher than that of identified abnormal feature samples, enriching the operating condition coverage of the feature set and improving the adaptability of subsequent channel separation and correlation filtering algorithms to normal operating fluctuations, thus avoiding overfitting issues due to insufficient abnormal sample data. This set integrates well-defined prior information on anomalies with massive amounts of unsupervised operational data, providing high-quality data support for the subsequent construction of a high-precision equipment health assessment model and significantly enhancing the model's ability to generalize and identify early minor wear.

[0022] For example, during the operation of a wind turbine gearbox, the system collects vibration, temperature, and speed data continuously for one hour. After time alignment processing, a multi-dimensional operating condition time series with a sampling interval of 1 millisecond is obtained. A frequency domain feature matrix containing 1024 frequency components is obtained through Fast Fourier Transform, and dimensionality reduction to a 10-dimensional principal component vector is obtained through principal component analysis. The Euclidean distance between this vector and the healthy baseline vector is calculated to be 65, which is greater than a preset threshold of 50, and is therefore judged as abnormal. The system concatenates this abnormal feature vector with 168 sets of principal component vectors collected over the past 7 days to form a first operating feature set containing 169 sets of features.

[0023] In step S12, the step of separating independent feature channels and performing hierarchical screening based on the degradation correlation strength according to the first set of operating features to obtain target physical alternation features includes: The first set of operating features is subjected to independent component separation processing to obtain multiple independent channel response sequences; A gearbox degradation evolution trajectory is pre-constructed, and a preliminary screening is performed based on the interface degradation correlation strength between the channel response sequence and the pre-constructed degradation evolution trajectory to obtain a candidate channel set; Based on the correlation between alternating load fatigue damage and physical alternating stress of the candidate channel set, the target physical alternating characteristics are obtained through fine screening.

[0024] It should be noted that the system first performs independent component analysis (ICA) on the first set of operating features, using an ICA algorithm to separate the mixed feature data into multiple independent signal channels. ICA maximizes the non-Gaussianity of the signal to separate independent source signals from the mixed observation signals. During algorithm execution, the input mixed features are first centered to remove the signal mean, making the data mean-zero. Then, principal component analysis is used to whiten the data, eliminating second-order correlations between components and obtaining orthogonally normalized whitened signals. Subsequently, using negative entropy as a quantitative measure of non-Gaussianity, a fast fixed-point iterative algorithm is used to continuously optimize the separation matrix, gradually improving the non-Gaussianity of the output signal. When the iteration converges to a preset threshold, multiple statistically independent source signal components are finally obtained. This algorithm can effectively separate the vibration responses of different structural parts of the gearbox, including gear meshing response, bearing inner ring response, bearing outer ring response, and drive shaft response, outputting multiple independent channel response sequences.

[0025] The system then performs initial screening based on the correlation strength between the channel response sequence and the degradation of the internal contact interface. The system pre-constructs a gearbox degradation evolution trajectory, which is based on the fatigue accumulation theory of gear materials and generated by fitting full-life-cycle bench tests and field operation data. This trajectory can comprehensively reflect the full-cycle degradation law of the gearbox from a healthy state to failure. Next, the system uses a mutual information algorithm to calculate the degree of nonlinear correlation between each channel response sequence and the degradation evolution trajectory. Before calculation, the high-frequency channel response sequence undergoes time-series feature dimensionality reduction, extracting the mean and peak values ​​of daily vibration amplitudes at a daily granularity to generate a long-term time-series feature sequence consistent with the time granularity of the degradation trajectory. The system aligns the amplitude change sequence of the channel response with the wear change sequence of the degradation evolution trajectory by timestamp. After alignment, the two sets of sequences are first processed by equal-frequency discretization, uniformly divided into 10 statistical intervals. The number of samples in each interval is counted, and the marginal probability distribution of a single variable and the joint probability distribution of two variables simultaneously falling into the corresponding interval are calculated. The final value is calculated based on the discrete mutual information formula. The formula logic is to take the logarithm of the ratio of the product of the joint probability and the marginal probability of all intervals, multiply it by the joint probability, and then sum them to obtain the mutual information value, which is used to characterize the correlation strength between channel response and interface degradation.

[0026] The mutual information threshold is determined using the general statistical criteria of permutation tests. The channel response sequence is randomly rearranged, and its baseline mutual information value with the degradation evolution trajectory is calculated. The mean of the baseline plus three standard deviations is taken as the judgment threshold. This method is based on the principle of normal distribution significance testing, and the probability of misjudging a strong correlation is less than 0.3%, effectively eliminating spurious correlation interference caused by random fluctuations. When the calculated mutual information value of a channel is greater than this threshold, the channel is determined to have a strong correlation with the degradation of the internal contact interface and is included in the candidate channel set.

[0027] Next, the system performs fatigue damage analysis on the alternating load data in the candidate channel set, using the rainflow counting algorithm to extract the amplitude and mean of the load cycles. The rainflow counting algorithm decomposes complex alternating load histories into multiple simple load cycles, statistically analyzing the amplitude and mean of each cycle; it is an international standard method for fatigue damage calculation. During algorithm execution, the algorithm first extracts extreme points from the alternating load time series data of the candidate channels, retaining all local peaks and valleys, and eliminating small fluctuations with amplitudes less than a preset threshold to reduce computational load. The preset threshold is set based on the sensor's background noise level, taking twice the peak noise level measured by the sensor under normal no-load conditions as a benchmark, eliminating invalid small-amplitude load cycles below the noise level to avoid meaningless micro-cycles interfering with the damage calculation results. Then, a two-parameter rainflow counting rule is used to traverse the extreme value sequence, allowing the load history to flow segment by segment along the positive time axis. When the flow path forms a complete stress-strain closed loop, it is identified as a complete load cycle, and the amplitude, mean, and number of occurrences of that cycle are recorded simultaneously. After traversal, the parameter set of all load cycles is output, providing standardized input for subsequent fatigue damage accumulation calculations. The load amplitude is converted into the corresponding contact stress amplitude using the aforementioned stress mapping library. The system pre-establishes a mapping relationship between stress amplitude and fatigue life based on the standard SN curve of gear materials. The SN curve is obtained through fatigue performance tests of corresponding grades of metal materials, reflecting the number of cycles at which the material fails due to fatigue under different alternating stress amplitudes. For each set of load cycles output by the rainflow counter, the corresponding fatigue life cycle number N is retrieved from the mapping relationship based on its stress amplitude. The material damage caused by a single load cycle is 1 / N. The system calculates the cumulative fatigue damage of each channel using the Miner linear cumulative damage rule. The Miner linear cumulative damage rule states that the damage to the material at each stress level can be linearly superimposed, and fatigue failure occurs when the total cumulative damage reaches 1. The total damage value is obtained by successively summing the damage values ​​corresponding to all load cycles. This value is then multiplied by 100% to convert it into a percentage. The cumulative fatigue damage calculated by the system is expressed as a percentage, reflecting the degree of material fatigue loss at the corresponding part of the channel.

[0028] The preset damage threshold is determined based on general early warning standards in the field of mechanical fatigue monitoring. 80% of the cumulative damage value at the critical point of fatigue failure is taken as the early warning threshold. This value is a common threshold for fatigue assessment of industrial equipment, indicating that the corresponding structure has entered the late stage of fatigue degradation, and the risk of failure has significantly increased. When the cumulative fatigue damage of a certain channel exceeds 80%, it is determined that the corresponding part of that channel has a serious risk of fatigue damage, triggering the fine screening mechanism.

[0029] The actual physical alternating stress is based on the measured data of stress patches during the bench test phase, combined with the finite element simulation results to form a standard mapping library. In field applications, the corresponding stress reference value can be called through the rotational speed and torque parameters, without the need for additional internal sensors. Then, the Pearson correlation coefficient algorithm is used to calculate the linear correlation between the characteristics of each channel and the actual physical alternating stress. The Pearson correlation coefficient measures the degree of linear correlation between two variables, ranging from -1 to 1, with the absolute value closer to 1 indicating a stronger correlation. Redundant channels with correlation coefficients below 0.7 are removed, and the final output is the target physical alternating characteristic that is strongly correlated with the degradation of the internal contact interface and accurately reflects the physical alternating stress. The fatigue damage accumulation results of the corresponding channels are simultaneously stored as a verification reference for subsequent loss field assessment.

[0030] It should be noted that the correlation coefficient threshold of 0.7 corresponds to the general statistical standard for determining strong linear correlation. The Pearson correlation coefficient ranges from -1 to 1; the closer the absolute value is to 1, the stronger the linear correlation between the two variables. In the industry, an absolute value of 0.7 is generally considered the dividing line between strong and moderate correlation. Selecting this threshold ensures that the retained channel features have a sufficiently strong linear correlation with the physical alternating stress, meeting the accuracy requirements of subsequent internal and external mapping, while avoiding the mistaken removal of effective degradation information due to an excessively high threshold, and the introduction of redundant noise channels due to an excessively low threshold, thus achieving a reasonable balance between detection accuracy and computational efficiency.

[0031] The following example corresponds to the calibration results of a specific gearbox model, used to illustrate the implementation process of the scheme. The actual threshold is adaptively determined based on the equipment health sample.

[0032] For example, the system performs independent component analysis on the first set of operating features, separating eight independent channel response sequences. The mutual information value between each channel and the degradation evolution trajectory is calculated. Three channels have mutual information values ​​of 4.1, 4.6, and 5.2, all greater than the threshold of 4, and are included in the candidate channel set. Rainflow counting and fatigue damage calculations are performed on the alternating load data of these three channels, yielding cumulative fatigue damage of 72%, 91%, and 85%, respectively. The latter two channels have cumulative damage exceeding the 80% damage threshold, triggering fine screening. The Pearson correlation coefficients between these two channels and the actual physical alternating stress are calculated, yielding 0.82 and 0.65, respectively. Channels with correlation coefficients below 0.7 are eliminated, ultimately obtaining the target physical alternating features.

[0033] In step S13, the step of constructing a mapping relationship between externally acquired information and the physical changes of the internal enclosed space based on the target physical alternation characteristics to obtain a physical loss sequence includes: Align the target physical alternating features with the preset spatial coordinates of the device housing to obtain the spatial boundary deformation tensor; Vibration response analysis and state transition modeling are performed on the spatial boundary deformation tensor to obtain the response attenuation matrix; The physical loss sequence is obtained by performing degradation trajectory decoding and smoothing on the response attenuation matrix.

[0034] It should be noted that the system first aligns the target physical alternating characteristics with the preset spatial coordinates of the equipment shell, and uses a finite element mesh mapping algorithm to construct a mapping relationship between externally acquired information and the physical changes in the internal enclosed space. The system pre-compiles a forward simulation model using the finite element mesh mapping algorithm, discretizing the gearbox shell into 500 uniformly distributed mesh nodes. For different levels of tooth surface contact stress and wear loss states, the system simulates and calculates the displacement deformation and vibration response of each node on the shell surface, forming a forward mapping database covering the entire wear range. All core input parameters required for modeling are obtained through standardized methods to ensure consistency between the simulation model and the actual physical properties of the equipment. Regarding material properties, the elastic modulus, Poisson's ratio, and density parameters of the gearbox shell and internal gears are based on the national standard mechanical property manual for the corresponding grade of metal material, and are numerically corrected using material sampling reports from the same batch of equipment. The structural damping coefficient is calibrated through modal testing, obtained by fitting the vibration decay curve collected under no-load operation of the equipment, and matches the damping characteristics of the actual structure. The contact interface friction coefficient was determined through a dedicated bench test. For tooth surface samples with four wear levels—healthy, lightly worn, moderately worn, and heavily worn—the meshing friction coefficient was tested under corresponding lubrication conditions to obtain the friction coefficient range for each wear level, which was then matched to simulation models under different wear states. Clear convergence verification criteria were set for mesh generation. A progressive densification method was used to verify mesh independence. When the number of mesh nodes increased by 20%, the deviation between the displacement and stress calculation results of key stress-bearing parts such as the bearing housing and meshing surface was less than 1%, indicating that the convergence requirement had been met. 500 uniform nodes were the minimum node size to meet the convergence criteria, balancing computational accuracy and operational efficiency. Boundary conditions completely replicated the actual installation and operation state of the equipment. On the constraint side, the mounting bolt holes at the bottom of the housing served as the constraint surface, restricting all translational degrees of freedom and eliminating rigid body displacement. On the load side, a uniformly distributed equivalent surface load was applied at the gear meshing contact line according to the rated operating parameters, while corresponding support constraints were applied at the bearing mounting position, completely restoring the true stress boundary.

[0035] The mesh node size is determined based on a combination of mesh independence verification criteria and system real-time constraints. A general finite element verification method with progressively higher density is adopted. When the deviation between the deformation and stress calculation results of key stress-bearing parts meets the preset mapping accuracy requirements after the node density is increased, the convergence criterion is met. Under the premise of meeting the convergence accuracy, the scheme with the smallest node size is selected, balancing the calculation accuracy of internal load mapping with the real-time requirements of single-step processing, while fully covering all key stress and deformation-sensitive areas of the shell to avoid missing stress concentration areas.

[0036] In field applications, the target physical alternating features obtained from S12 screening are first converted into vibration amplitude vectors of the shell nodes, and then matched and aligned with samples in the forward mapping database. The database covers five wear levels, with each level evenly divided into 10 gradient samples based on the average wear depth of the tooth surface. The gradient interval is matched with the engineering detection accuracy of single-level wear, resulting in a total of 50 typical working conditions. Each sample contains 500-dimensional node displacement and vibration amplitude features. Matching uses the vibration amplitude feature vectors of each node in the shell as the input dimension and Euclidean distance as the error quantification index. First, a subset of the database corresponding to the current working condition interval is matched according to the current speed and torque working condition, and then the error is calculated within the same working condition interval. The preset calibration threshold was determined through bench testing. A relative error of less than 5% between the mapped result and the measured wear value was considered acceptable. The corresponding critical error value was used as the judgment threshold. When the minimum error was less than the preset calibration threshold, the corresponding sample was directly selected. When the error was within the threshold range, linear interpolation was performed on adjacent samples. The closest internal wear condition was located by searching for the minimum error, and the displacement vector data of the corresponding grid nodes was extracted. A spatial boundary deformation tensor containing 500 spatial dimensions was output. This tensor reflects the geometric deformation distribution law during the transmission of load from the internal gearbox to the external housing, establishing a spatial mapping relationship between internal physical quantities and external measurable quantities.

[0037] The system then performs vibration response analysis on the spatial boundary deformation tensor, employing a modal superposition algorithm to calculate the external vibration response set of the gearbox. The modal superposition algorithm decomposes the dynamic response of the complex structure into a linear superposition of various natural modal responses. The system pre-obtains the first 10 natural frequencies and mode shapes of the gearbox through finite element analysis; these 10 modes contain over 98% of the gearbox's vibration energy. The system extracts the dynamic displacement components of the spatial boundary deformation tensor at different natural frequencies, linearly combines the multi-mode responses, and outputs a set of external vibration responses containing a multi-frequency amplitude sequence. This set encompasses all the significant vibration characteristics exhibited by the gearbox under complex operating conditions.

[0038] The system models the state transitions of the external vibration response set, employing a Markov transition matrix algorithm to establish a mapping between the time dimension and the spatial response. The Markov transition matrix algorithm assumes that the future state of the system depends only on the current state and is independent of past states. The system predefines five levels of degenerate hidden states, corresponding to healthy, slight wear, moderate wear, heavy wear, and near-failure stages. Each state corresponds to a different vibration amplitude distribution range. Then, the transition probabilities of vibration amplitude states at different time steps are statistically analyzed to construct a state transition probability matrix, outputting a response decay matrix. The response decay matrix reflects the decay law of different frequency vibration components over time and is closely related to the degree of fatigue damage to the internal materials of the gearbox.

[0039] The preset attenuation threshold is determined using the 3σ statistical criterion commonly used in industrial condition monitoring. Based on the largest eigenvalue of the response attenuation matrix collected during the equipment's healthy operation phase, the mean and standard deviation of the sample are calculated, and the sum of the mean and three times the standard deviation is taken as the judgment threshold. Under healthy conditions, the probability of the eigenvalue exceeding this threshold is less than 0.3%, effectively identifying material degradation trends while controlling the false positive rate. When the largest eigenvalue of the response attenuation matrix exceeds this threshold, significant material degradation is determined to exist inside the gearbox, triggering subsequent processing mechanisms.

[0040] Subsequently, the Viterbi decoding algorithm is used to process the response decay matrix to infer the degradation state transition path hidden behind the observation data. The Viterbi decoding algorithm is based on five predefined degradation hidden states, combining the state transition probability matrix and the observation emission probability matrix, and searches for the optimal state path through dynamic programming. The observation emission probability matrix is ​​generated statistically from the probability distribution of vibration amplitude under different degradation states, with each state corresponding to a set of normal distribution parameters for vibration amplitude. Each degradation level is pre-bound with a standard quantization matrix corresponding to the internal loss field. The matrix covers 1000 loss nodes at different locations and depths in the tooth surface contact area, with the node size determined based on the mesh accuracy of the internal wear simulation. The system retrieves the quantization matrix of the corresponding level based on the decoded state sequence, performs linear interpolation with the transition probabilities of adjacent states, and obtains a continuous multidimensional loss evolution trajectory. The system decodes the change process of the internal degradation state of the gearbox over time from the response decay matrix using the Viterbi decoding algorithm and outputs the loss evolution trajectory.

[0041] Finally, the system smooths the loss evolution trajectory and uses an autoregressive moving average algorithm to fit the fluctuation data in the trajectory. First, the discrete loss state sequence output by Viterbi decoding is preprocessed to stabilize the data. The model order is determined through autocorrelation function testing, and the model coefficients are estimated using the least squares method. The original trajectory is decomposed into low-frequency deterministic trend components and high-frequency random fluctuation components. A continuous trend curve is obtained through fitting, and a low-frequency trend line reflecting the actual fatigue accumulation process of the gearbox material is extracted. Finally, the physical loss sequence is output. Each time step of the physical loss sequence contains multi-dimensional features with multiple spatial nodes and multiple loss modes. The feature dimension at a single time step is 1000 dimensions, with each dimension ranging from 0 to 100, corresponding to the degree of loss at different locations and modes. Higher values ​​indicate more severe internal fatigue loss. The sequence time interval is consistent with the analysis granularity after feature extraction, matching the slow-changing characteristics of wear degradation.

[0042] For example, the system converts the target physical alternating characteristics into a shell node vibration amplitude vector, matches and aligns it with samples in the forward mapping database, and calculates the spatial boundary deformation tensor. An external vibration response set containing vibration amplitudes in 10 frequency bands is obtained through a modal superposition algorithm. A Markov transition matrix is ​​constructed to obtain the response attenuation matrix, whose maximum eigenvalue is 85, greater than the threshold of 75. The Viterbi decoding algorithm is used to obtain the loss evolution trajectory, which is then smoothed using an autoregressive moving average algorithm to obtain a physical loss sequence containing 720 time nodes over 30 consecutive days, with each node containing 1000-dimensional loss features.

[0043] In step S14, the multi-dimensional feature extraction and dimensionality reduction reconstruction of the physical loss sequence to obtain the target interface wear features includes: The physical loss sequence is subjected to dimensionality reduction and independent component separation to obtain a coarse gradient vector and a depth distribution vector; The coarse gradient vector and the depth distribution vector are subjected to correlation fusion processing to obtain the initial interface wear characteristics; The initial interface wear characteristics are subjected to physical component separation and tensor reconstruction to obtain the target interface wear characteristics.

[0044] It should be noted that there is a clear physical correspondence between the evolution of tooth surface fatigue wear and the changes in microstructure. Fatigue accumulation will first cause the surface material of the tooth surface to peel off, increasing the surface roughness. As wear intensifies, the subsurface cracks expand, gradually forming pitting and fretting wear marks. The amount of wear is positively correlated with roughness and wear mark depth. This correspondence has been verified by a large number of bench wear tests.

[0045] First, when performing dimensionality reduction on the physical loss sequence, principal component analysis (PCA) is used to orthogonally transform the sequence, which contains multi-dimensional fatigue accumulation data. Orthogonal transformation removes redundant dimensions from the data, converting relevant variables into irrelevant principal components. Each time step of the physical loss sequence contains feature data from multiple spatial nodes and multiple loss patterns, with a single-time-step feature dimension of 1000. The system calculates the covariance matrix of the physical loss sequence, extracts the principal components corresponding to the top 20 largest eigenvalues, and outputs the loss principal component matrix. This process reduces the dimensionality of the original 1000-dimensional physical loss sequence to 20 core dimensions while retaining over 98% of the variance information, effectively reducing the complexity of subsequent calculations.

[0046] The system then performs independent component separation on the principal component matrix of the loss, using an independent component analysis algorithm to separate the mixed loss characteristic signals into mutually independent physical components. After separation, the components are projected back into the original 1000-dimensional loss space through inverse principal component transformation to obtain the spatial distribution field with corresponding physical meaning. The system separates the roughness gradient vector reflecting the micro-scraping of the tooth surface and the depth distribution vector reflecting the pitting evolution from the principal component matrix of the loss. The physical meaning of the components is determined through bench testing. In the test of standard samples with known wear conditions, the Pearson correlation coefficient is used as the correlation quantification index. The time-series data of each independent component and the full-field scan measured true values ​​of the corresponding physical quantities are collected simultaneously. After alignment at each time point, the correlation coefficient between each independent component and the surface roughness detection value is calculated. The component with the largest absolute value of the correlation coefficient is the component with the highest correlation to the surface roughness and is defined as the roughness gradient vector. The correlation coefficient between each component and the wear depth detection value is calculated using the same method, and the component with the largest absolute value is defined as the depth distribution vector. After calibration, the physical mapping relationship of the components remains unchanged. The roughness gradient vector characterizes the rate of change of the microstructure of the gearbox tooth surface under friction; a larger value indicates a rougher tooth surface. The depth distribution vector describes the vertical depth distribution of subsurface crack propagation; a larger value indicates a deeper pitting depth.

[0047] Next, the variance of the roughness gradient vector is calculated and compared with a preset roughness threshold. The system pre-establishes a mapping relationship between the roughness gradient vector variance and the actual tooth surface roughness through standard specimen bench tests. Standard gear specimens covering all wear levels from healthy to heavily worn are prepared. These specimens are run on a bench simulating rated operating conditions, and external vibration data is simultaneously collected. The variance of the corresponding roughness gradient vector is calculated using the same algorithm. Simultaneously, the arithmetic mean roughness Ra value of each specimen's tooth surface is measured using a surface profilometer. Using the measured variance values ​​of multiple sets of specimens as independent variables and the corresponding roughness detection values ​​as dependent variables, a quadratic polynomial regression is used to fit and obtain the mapping function, thus completing the calibration of the mapping relationship. After calibration, the mapping relationship is stored in a fixed manner and can be directly called upon for calculation during on-site testing.

[0048] The preset roughness threshold is calibrated based on the common wear warning critical standard for industrial gear transmission systems, corresponding to an industry-standard warning limit of 3.2 micrometers for the arithmetic mean roughness Ra value of the tooth surface. The system uses the variance value of the corresponding critical roughness as the judgment threshold. This threshold can accurately distinguish between roughness changes during normal break-in and roughness aggravation caused by abnormal degradation, ensuring the ability to identify early wear while controlling the false judgment rate. When the variance of the roughness gradient vector exceeds this threshold, the tooth surface roughness is determined to have reached a level requiring attention, triggering the fusion processing mechanism.

[0049] Then, canonical correlation analysis (CCI) is used to mine the linear correlation between the coarse gradient vector and the depth distribution vector in spatial distribution. CCI is used to analyze the linear correlation between two sets of variables and can find the maximum correlation coefficient between the two sets of variables. The system calculates the correlation coefficient between the two vectors through CCI, normalizes the CCI coefficients of the two sets of vectors and uses them as corresponding fusion weights. The sum of the weights is 1. The two vectors are then weighted and fused according to their correlation to output the initial interface wear characteristics. These initial interface wear characteristics integrate the coupling state of transverse spalling and longitudinal pitting on the tooth surface, and can more comprehensively reflect the wear condition of the tooth surface.

[0050] Subsequently, a nonnegative matrix factorization (NMF) algorithm is used to decompose the initial interface wear features into nonnegative basis matrices and coefficient matrices for physical component separation. The NMF algorithm can decompose a nonnegative matrix into the product of two nonnegative matrices, making it suitable for feature separation problems with nonnegative constraints. The system separates the pure surface roughness component and the fretting wear depth component from the decomposition results, eliminating the mutual interference between the two components.

[0051] Finally, the separated surface roughness component and fretting wear depth component are input into the tensor reconstruction algorithm model and stitched together in a multi-dimensional space according to a preset weight ratio. The weight allocation is determined based on the general mechanism of tooth surface wear failure. The subsurface fatigue crack propagation corresponding to the fretting wear depth is the core cause of gear contact failure, and its contribution to the final failure is higher than that of abrasive wear caused by increased surface roughness. Therefore, the weight of the fretting wear depth component is set higher than that of the surface roughness component, and the sum of the weights of the two components is 1. This weight allocation method can evenly represent the damage degree of different wear types, comprehensively cover the main wear failure modes of the tooth surface, and ensure the matching accuracy between the wear characteristics of the target interface and the actual wear state.

[0052] For example, the system performs principal component analysis on a physical loss sequence containing 1000 dimensions, obtaining a 20-dimensional loss principal component matrix. Independent component analysis separates the roughness gradient vector and depth distribution vector; the variance of the roughness gradient vector is calculated to be 22, which is greater than the threshold of 15. Canonical correlation analysis is used to fuse the two vectors to obtain the initial interface wear characteristics. Non-negative matrix factorization separates the surface roughness component and the fretting wear depth component, and tensor reconstruction is performed with weights of 0.4 and 0.6, finally obtaining the target interface wear characteristics.

[0053] In step S15, based on the wear characteristics of the target interface, the coupled evolution of the interface degradation index is time-series extrapolated to obtain a first degradation trajectory sequence. Then, the first degradation trajectory sequence is continuously reconstructed to obtain a second degradation trajectory sequence, including: The wear characteristics of the target interface are segmented by a sliding window to obtain a time-series evolution feature vector; Coupled evolution analysis and anomaly smoothing are performed on the temporal evolution feature vectors to obtain a standard degradation state sequence; The first degradation trajectory sequence is obtained by forward extrapolating the standard degradation state sequence. Anomaly detection processing is performed on the first degraded trajectory sequence to remove abnormal sampling nodes and obtain a clean trajectory sequence; The pure trajectory sequence is subjected to missing node localization and interpolation compensation processing to obtain a depth-compensated sequence; The depth compensation sequence is continuously smoothed to obtain the second degenerate trajectory sequence.

[0054] It should be noted that the system first performs sliding window segmentation on the wear characteristics of the target interface to extract a temporal evolution feature vector reflecting the dynamic changes of tooth surface wear over time. The system presets a sliding window length of 24 hours and a step size of 1 hour. The operating load of wind turbines exhibits obvious diurnal cycle fluctuations, with wind speed changing regularly with the alternation of day and night, causing the alternating load on the gearbox to fluctuate periodically as well. A 24-hour window length can completely cover a complete operating cycle, and the extracted temporal evolution features can eliminate the interference of periodic load fluctuations, truly reflecting the inherent evolution trend of tooth surface wear. If the window length is shortened to 12 hours, it will not be able to cover the complete diurnal operating cycle, and the feature values ​​will oscillate significantly with load fluctuations, causing the error of subsequent coupled evolution analysis to rise to over 18%. If the window length is extended to 48 hours, although the feature stability will be further improved, the outdated data included in the window will dilute the latest wear change information, increasing the system's response lag time to early wear abrupt changes from 2 hours to 8 hours, significantly weakening the early warning capability. A 1-hour step size achieves the optimal balance between temporal resolution and computational efficiency. Reducing the step size to 30 minutes doubles the sample size and increases the feature extraction time per batch from 3ms to 7ms, while only improving detection accuracy by 0.8%, a benefit far outweighing the computational cost. Increasing the step size to 2 hours results in the loss of some short-term wear abrupt changes and a 11% decrease in the detection rate of pitting faults caused by instantaneous impact loads. Comparative tests under 200 different operating conditions verified that the combination of a 24-hour window and a 1-hour step size achieves 97.2% feature stability while keeping single-step processing time within 3ms, fully meeting the dual requirements of real-time performance and detection accuracy. The sliding window continuously slides along the time axis of the wear features on the target interface with a 1-hour step size. After each slide, all data within the window is extracted as a time-series sample. Each time-series sample contains information on the changes in tooth surface roughness and fretting wear depth over the past 24 hours. The system extracts two core indicators from each time-series sample: the roughness intensification rate and the wear depth expansion rate, forming a time-series evolution feature vector.

[0055] Subsequently, coupled evolution analysis is performed on the temporal evolution feature vectors, and a dynamic time warping algorithm is used to calculate the matching distance between the current feature subsequence and the standard degradation template sequence. The dynamic time warping algorithm can calculate the optimal matching distance between two time series, unaffected by time axis scaling or offset. The system pre-constructs a standard feature template library for different degradation stages, with each template corresponding to a specific wear level. The template library is built based on full-lifecycle wear test data, covering five degradation stages: healthy, slight wear, moderate wear, heavy wear, and near-failure. For each stage, more than 30 stable operating samples of the same model gearbox are selected, and a continuous 72-hour temporal evolution feature sequence is extracted within the corresponding stage. After mean smoothing, a standard feature template for that stage is generated. Each template contains a fixed-length two-dimensional feature sequence with dimensions consistent with the feature vector output by the sliding window, and is bound to the standard wear amount calibration value for the corresponding stage as a reference benchmark for dynamic time warping matching. The template library is permanently stored after construction and can be directly called for matching during on-site testing without real-time regeneration. The system selects feature vectors from 72 consecutive sliding windows and concatenates them in chronological order to form a time-series evolution feature subsequence with consistent duration. The sliding step size of the subsequence is consistent with the window step size and is updated hourly. The DTW distance of the time-series evolution feature subsequence formed by concatenating multiple windows is calculated one by one with all templates in the template library. The smaller the distance, the closer the current state is to the corresponding template's degradation degree. The system selects the template level corresponding to the minimum distance as the degradation quantization value at the current moment. After window-by-window sliding calculation, an initial degradation state sequence is formed. This sequence reflects the nonlinear degradation trend of the gearbox from slight pitting to severe spalling.

[0056] The system calculates the variance of the initial degradation state sequence and compares it with a preset fluctuation threshold. The preset fluctuation threshold is determined using the 3σ statistical criterion commonly used in industrial condition monitoring. The variance of the initial degradation state sequence collected during the equipment's healthy operation phase is used as the statistical sample. The mean and standard deviation of the sample are calculated, and the sum of the mean and three times the standard deviation is taken as the judgment threshold. In a healthy state, the probability of the sequence variance exceeding this threshold is less than 0.3%, effectively distinguishing normal degradation fluctuations from abnormal fluctuations caused by sensor noise or transient impacts. When the variance of the initial degradation state sequence exceeds this threshold, it is determined that there are abnormal fluctuations in the sequence caused by high-frequency sensor noise or transient impact loads, triggering a smoothing mechanism.

[0057] Next, the Kalman filter algorithm is used to smooth the initial degradation state sequence. The Kalman filter algorithm is an optimal recursive state estimation algorithm that can estimate the system state in real time based on the system's state equation and observation equation. In this scenario, the cumulative wear amount and wear rate of the tooth surface are used as two-dimensional state variables, and the observed values ​​of the initial degradation state sequence are used as system inputs. Through iterative iteration of the two core steps of prediction and update, abnormal fluctuations caused by sensor noise and transient impact loads are filtered out.

[0058] In the prediction step, the system calculates the prior state estimate for the current moment based on the optimal state estimate from the previous moment, combined with the linear state equation for wear degradation. The state equation is constructed based on the linear law of material fatigue accumulation, assuming that the wear increment per unit time is proportional to the current wear rate. Simultaneously, the system calculates the prior covariance matrix for the current moment based on the state covariance matrix from the previous moment, combined with the preset process noise covariance. The process noise covariance is tuned according to the physical characteristics of tooth surface wear evolution, matching the inherent law of slow wear rate change in industrial scenarios. Tuning first calculates the standard deviation of the wear rate change within a unit sampling period based on wear rate statistics of the same model equipment throughout its entire life cycle, and uses its square as the initial baseline value for the parameters. Subsequently, verification tests are conducted on a standard wear test bench, using the actual wear trajectory detected by disassembly as a reference, with the goal of minimizing the mean square error between the filtered output and the actual wear value. Parameters are iterated within the range of 0.5 to 2 times the baseline value to determine the final values. The filter values ​​should be balanced between smoothness and trend response speed. When the values ​​are too large, the filter has strong following ability but is prone to introducing noise fluctuations. When the values ​​are too small, the filter has high smoothness but is prone to lagging behind the actual degradation changes. The principle of parameter selection is to filter out random fluctuations while controlling the delay of the actual trend response within a few sampling periods, so as to adapt to the general characteristics of slow wear evolution.

[0059] In the update process, the system first calculates the Kalman gain, which is determined by the prior covariance matrix and the observation noise covariance. The observation noise covariance is tuned based on the nominal measurement accuracy of the vibration sensor and the impact interference characteristics of the field conditions, matching the general noise level of the field-acquired data, and balancing the filtering noise reduction capability with the degree of detail retention. A larger value results in strong filtering smoothing but is prone to losing true wear fluctuations, while a smaller value has weak anti-interference capability but can retain more variation details. The parameter selection is based on the principle of effectively suppressing transient impact jumps while retaining the true changes in degradation trends, and is matched with the general measurement accuracy of sensors in industrial fields.

[0060] The system then uses the difference between the current observed value and the prior state estimate, multiplied by the Kalman gain, to correct the prior state estimate, thus obtaining the optimal posterior state estimate for the current moment. Finally, the system updates the state covariance matrix for the current moment, providing a basis for the next iteration. After multiple iterations, the algorithm can effectively filter out spike noise caused by instantaneous shocks while preserving the true degradation trend, outputting a smooth and stable standard degradation state sequence.

[0061] The standard degradation state sequence and historical evolution node data are then input into a Long Short-Term Memory (LSTM) network model for forward inference. The LTM model employs a two-layer stacked structure, consisting of an input layer, two hidden layers, and an output layer connected sequentially. The input layer has a dimension of 2, corresponding to the roughness intensification rate and wear depth expansion rate at each time step. The input sequence length is 720, corresponding to complete historical data sampled at 1-hour intervals over the past 30 days. The first hidden layer contains 64 neurons, responsible for extracting short-term wear fluctuation features, while the second hidden layer contains 32 neurons, responsible for capturing long-term degradation trend dependencies. Each hidden unit contains three gating structures: an input gate, a forget gate, and an output gate. The input gate controls the proportion of new information written to the cell state at the current moment, the forget gate controls the proportion of historical information discarded from the cell state, and the output gate controls the proportion of the cell state output to the outside. These three gating structures work together to allow the model to selectively retain key degradation information and forget irrelevant noise fluctuations, effectively solving the gradient vanishing problem of long sequences in traditional recurrent neural networks. The model training phase utilized full-lifecycle degradation data of 320 wind turbine gearboxes collected over the past five years. The dataset was divided into training and validation sets in an 8:2 ratio. The training process employed mean squared error as the loss function, using the Adam optimizer for iterative parameter updates, with an initial learning rate of 0.001. An early stopping mechanism was implemented during training, automatically terminating training when the validation set loss stopped decreasing after 10 consecutive iterations to prevent overfitting. The final trained model achieved a 7-day average relative error of 5.8% on the validation set, with a maximum relative error not exceeding 12%, fully meeting the degradation prediction accuracy requirements for industrial scenarios. The system takes into account the historical evolution node data of the past 30 days. The model performs forward calculations by combining the current standard degradation state sequence, passing feature information layer by layer, and finally outputs the predicted tooth surface wear depth values ​​for 168 time nodes in the next 7 days. The network output layer is bound to the pre-calibrated spatial wear distribution basis vector. Using the global wear rate as the weighting coefficient, it maps and generates wear depth data of 163 uniform sampling points along the meshing circumference. The output constrains the physical law of the wear distribution being symmetrical along the tooth width and monotonically gradual along the meshing direction. Each time node contains the wear depth value of the corresponding spatial sampling point, forming the first degradation trajectory sequence.

[0062] For example, the system performs sliding window segmentation on the wear characteristics of the target interface, with a window length of 24 hours and a step size of 1 hour, resulting in 720 time-series samples. The roughness intensification rate and wear track depth expansion rate are extracted from each sample to form a time-series evolution feature vector. A dynamic time warping algorithm is used to calculate the matching distance between the feature subsequence and the standard template, obtaining an initial degradation state sequence. The variance of this sequence is calculated to be 12, which is greater than the threshold of 8. A Kalman filter algorithm is used to smooth the sequence, resulting in a standard degradation state sequence. The standard degradation state sequence and historical data from the past 30 days are input into a long short-term memory network model to deduce the tooth surface wear depth changes at 168 time points over the next 7 days, obtaining the first degradation trajectory sequence.

[0063] It should be noted that when performing anomaly detection on the first degraded trajectory sequence, the Isolation Forest algorithm is used to identify outliers in the sequence. The Isolation Forest algorithm is an anomaly detection algorithm based on an isolation mechanism. It isolates anomalies by randomly partitioning the feature space. Anomalies are easier to isolate because they are fewer in number and their features differ significantly from normal points. The number of isolated trees constructed by the system is determined based on both algorithm convergence verification and system real-time constraints. A verification method of progressively increasing the number of trees is adopted. When the stability of the anomaly detection results meets the preset requirements and the processing time per sample meets the system's real-time performance indicators, the corresponding minimum number of trees is selected. This ensures stable and reliable detection results while controlling computational overhead, adapting to the operational needs of continuous monitoring in industrial sites. The system calculates the anomaly score for each sampling node in the sequence. The anomaly score ranges from 0 to 100; a higher score indicates that the node is more likely to be an anomaly.

[0064] The preset anomaly score threshold is determined using a low false positive rate statistical criterion commonly used in the field of industrial anomaly detection. Based on the anomaly scores of normal degradation trajectory samples, the quantile value corresponding to the low false positive rate is taken as the judgment threshold. This threshold ensures effective detection of impact noise outliers while avoiding excessive rejection of normal wear fluctuation data, thus meeting the quality requirements of degradation trajectory preprocessing. When the anomaly score of a sampling node exceeds this threshold, the node is determined to be noise and removed. The system outputs the clean trajectory sequence after outlier removal.

[0065] The system then performs timestamp comparison on the clean trajectory sequence to locate missing sampling nodes caused by noise removal or data transmission interruption. The system checks the continuity of all timestamps in the clean trajectory sequence at a preset 1-hour sampling interval. When it finds that a timestamp does not have corresponding data, it determines that a missing sampling node exists at that location.

[0066] The system extracts local extrema within adjacent time windows before and after the missing sampling node, including local maximum and minimum values ​​of the wear depth on the tooth surface. The system selects data from five time nodes before and after the missing node as local windows; the window size is determined through interpolation accuracy testing. The configuration of 10 reference points (five nodes before and after) comprehensively covers the wear change trend around the missing area. If the window is reduced to three nodes before and after, the edge error of the interpolation result increases to 6.5%. If the window is expanded to eight nodes before and after, irrelevant data from distant locations is introduced, which interferes with the local interpolation accuracy. The system inputs the local extrema and their corresponding timestamps into the cubic spline interpolation algorithm model. The cubic spline interpolation algorithm constructs a piecewise cubic polynomial, ensuring that the interpolation curve has continuous first and second derivatives at each segment point, guaranteeing the smoothness and curvature continuity of the interpolation curve. Compared to linear interpolation, it can better fit the nonlinear change trend of wear degradation. The system calculates the wear depth value at the missing node using cubic spline interpolation, obtaining the depth compensation sequence after filling the missing point.

[0067] The system performs continuous smoothing on the depth compensation sequence, employing the Savitzky-Gore filtering algorithm to eliminate local discontinuities that may be introduced at data splicing points during interpolation. The Savitzky-Gore filtering algorithm is a smoothing filter based on local polynomial least squares fitting, which effectively preserves the signal's shape and peak values ​​while smoothing the data, avoiding the trend blunting problem caused by traditional mean filtering. The system presets a filtering window length of 5 sampling periods and a polynomial fitting order of 2. This parameter combination has been verified for smoothing effects on numerous degradation trajectories. The 5-point window can cover the transition fluctuation range at interpolation splicing points, and the 2nd-order polynomial can fit the nonlinear changes in wear without causing spurious fluctuations due to overfitting. Expanding the window to 7 sampling periods results in excessive smoothing, which may smooth out the true inflection points of wear rate changes. Increasing the order to 3rd order may lead to overfitting to local noise, introducing new high-frequency fluctuations. The algorithm slides a window across the depth compensation sequence, performing quadratic polynomial least squares fitting on the data within each window, and replacing the original data at the center of the window with the fitted polynomial value. The system outputs a continuous reconstructed sequence with local glitch eliminated, and this sequence is directly mapped to the second degenerate trajectory sequence.

[0068] The second degradation trajectory sequence fully includes wear depth data of the wind turbine gearbox at equal time intervals throughout the entire prediction period, filling in all missing wear depth sampling points and forming a degradation evolution trajectory with complete temporal continuity.

[0069] For example, the system inputs a first degraded trajectory sequence containing 168 time sampling nodes. An anomaly detection algorithm using an isolated forest algorithm consisting of 100 isolated trees is employed. Five nodes with anomaly scores of 85, 88, 92, 87, and 91 are found, all exceeding the threshold of 80, and are therefore removed, resulting in a clean trajectory sequence containing 163 valid time nodes. Timestamp comparison identifies missing nodes at hours 45 and 46. The extreme values ​​of wear depth from hours 40 to 44 and from 47 to 50 are extracted, and the depth values ​​of the missing nodes are calculated using cubic spline interpolation, yielding a depth-compensated sequence. A Savitzky-Gore filter with a window length of 5 and an order of 2 is used for smoothing, resulting in a continuously reconstructed sequence, i.e., the second degraded trajectory sequence.

[0070] In step S16, the step of performing three-dimensional reconstruction and comprehensive health status assessment of the contact interface based on the second degradation trajectory sequence to obtain the target health status of the device connection components includes: A three-dimensional point cloud matrix is ​​constructed based on the second degenerate trajectory sequence, and the surface reconstruction process is performed on the three-dimensional point cloud matrix to obtain a virtual reconstruction model; Wear feature extraction and stress analysis are performed on the virtual reconstruction model to obtain wear volume features and stress distribution map; The target health status of the equipment connection components is obtained by comprehensively evaluating the wear volume characteristics and the stress distribution spectrum.

[0071] It should be noted that, firstly, a three-dimensional point cloud matrix is ​​constructed by selecting the second degradation trajectory sequence corresponding to the target evaluation time. The target evaluation time is taken by default as the end of the prediction cycle, but any prediction node can be specified according to the detection requirements. A coordinate system transformation algorithm is used to map the wear depth data in polar coordinates to spatial coordinate points in Cartesian coordinates. Each time node in the second degradation trajectory sequence contains wear depth data uniformly sampled along the meshing circumference, with a total of 163 circumferential sampling points per time moment. Each point corresponds to a unique circumferential angle position. These data are represented in polar coordinates with the gear center as the origin by default. The system converts the polar coordinates of each node to three-dimensional Cartesian coordinates using the polar-to-Cartesian coordinate conversion formula. Combined with the symmetrical distribution of three equidistant sampling lines in the tooth surface width direction, the system laterally expands to generate sampling points for the complete tooth surface. The system calculates the precise position of each wear point in three-dimensional space by combining the gear pitch circle radius parameter, and outputs a three-dimensional point cloud matrix containing 489 spatial coordinate features. The number of 489 coordinate points was calculated by combining 163 circumferential sampling points at a single moment with three sampling lines along the tooth width direction. The tooth wear exhibits a symmetrical distribution along the tooth width direction, and the three sampling lines can completely cover the lateral distribution pattern of the wear. After bench testing, the error between the wear volume reconstructed by this density point cloud and the actual measured value is less than 7%, fully meeting the accuracy requirements for quantitative assessment of wear degree. The surface reconstruction time of a single frame point cloud is controlled within 15ms.

[0072] Subsequently, surface reconstruction processing is performed on the 3D point cloud matrix, and a virtual reconstruction model of the tooth surface is generated using the Poisson surface reconstruction algorithm. The Poisson surface reconstruction algorithm constructs a spatial indicator function by calculating the normal vector field of the point cloud data, and extracts the isosurface of the indicator function as the reconstructed surface. This algorithm can reconstruct a smooth and closed surface from noisy point cloud data, and is suitable for the 3D reconstruction of complex curved surfaces such as gear tooth surfaces. During the reconstruction process, the system first calculates the normal vector for each point cloud data point, with the normal vector direction uniformly pointing outwards to ensure the consistency of the orientation of the reconstructed surface. Then, an octree spatial partitioning is used to construct the solution space of the indicator function. The octree depth is set to 8, at which the reconstructed surface accuracy can reach 0.01 mm, fully meeting the observation accuracy requirements of wear morphology. Finally, the Poisson equation is solved to extract the zero isosurface, generating a continuous triangular mesh model, i.e., the virtual reconstruction model of the tooth surface. This model can highly restore the 3D details of the current wear morphology, including the spatial distribution of micro-scraping pits and wear grooves.

[0073] Next, when extracting wear features from the virtual reconstructed model, the Boolean difference algorithm is used to calculate the wear volume characteristics. The Boolean difference algorithm is used to calculate the spatial difference between two 3D models. Using geometric parameters such as the gear's nominal module, number of teeth, pressure angle, tooth width, and displacement coefficient as a benchmark, a standard involute tooth surface 3D model without wear is generated. The model's accuracy matches the gear machining tolerance level, and the tooth surface morphology deviation is controlled within 0.001 mm. This model corresponds to the ideal healthy state of the gear at the time of manufacture and serves as the zero benchmark for wear calculation, using a unified coordinate system definition with the virtual reconstructed model. The system spatially aligns the virtual reconstructed model with the pre-stored standard healthy gear 3D model. The alignment process uses the gear pitch circle and mounting hole position as benchmark features to ensure complete overlap of the coordinate systems of the two models. After alignment, the volume difference between the two is calculated; this difference is the wear volume characteristic of the tooth surface. Wear volume is an important quantitative indicator for measuring the degree of gear wear, directly reflecting the total amount of material loss on the tooth surface, and is more representative of the overall wear level than a single-point depth value.

[0074] The virtual reconstruction model is then imported into the finite element analysis algorithm model for stress analysis. The system first meshes the reconstruction model using second-order tetrahedral elements with a size of 0.5 mm. Local mesh refinement is applied in stress concentration areas such as the tooth root to ensure calculation accuracy. Subsequently, a torque load from the wind turbine under rated operating conditions is applied, constraining all degrees of freedom of the gear's inner bore to simulate the actual installation state. The system calculates the stress conditions at each node on the tooth surface and outputs a stress distribution map. The stress distribution map visually displays the magnitude and distribution of stress in different areas of the tooth surface, helping to identify high-stress areas, which are the parts of the gear most prone to fatigue failure.

[0075] Subsequently, the maximum stress value in the stress distribution spectrum was extracted and compared with the preset fatigue threshold of the gear material. The preset fatigue threshold is 850 MPa, which was determined based on the room temperature rotational bending fatigue strength test data of the 20CrMnTi alloy steel material used in the gear, corresponding to the conditional fatigue limit of the material for tens of millions of cycles. When the maximum stress value exceeds 850 MPa, the material in the high-stress area of ​​the tooth surface will enter the accelerated fatigue damage stage, and the probability of crack initiation under cyclic loading will increase significantly, posing a risk of tooth breakage failure. Setting the threshold to 800 MPa would overestimate the fatigue risk, leading to unnecessary downtime for maintenance. Setting it to 900 MPa would miss some critical fatigue hazards, reducing the reliability of failure warning. Therefore, the 850 MPa threshold perfectly matches the material fatigue characteristics and can accurately determine the fatigue critical state of the tooth surface. When the maximum stress value is greater than 850 MPa, the fatigue level of the current tooth surface is determined to have reached the critical state, posing a risk of tooth breakage failure.

[0076] Finally, a fuzzy comprehensive evaluation algorithm was used for a comprehensive assessment of the health status. This algorithm integrates multiple fuzzy evaluation indicators into a single, definite evaluation result, making it suitable for non-absolute assessment scenarios such as wear degree and fatigue state. The algorithm sets two evaluation factors: wear volume and maximum stress value, with corresponding weight sets of 0.6 and 0.4, respectively. These weight ratios are determined based on the statistical regularity and contribution of failure mechanisms in industrial gear contact failures. Wear volume is a direct quantitative result of the actual loss of interface materials, directly corresponding to the degree of degradation already occurred, and thus contributes more to the health status. Maximum stress reflects the potential fatigue risk under current operating conditions and is an indirect risk indicator, with a secondary contribution. Referring to industry statistics on field failure modes of wind turbine gearboxes, contact wear failure accounts for approximately 60% of the dominant failures, while stress-induced contact fatigue and bending fatigue failures account for approximately 40%, which serves as the initial baseline. Subsequently, closed-loop verification was carried out using standard bench samples with known health levels. Parameters were optimized with the goal of achieving the best accuracy in health status classification. Finally, the weight combination was determined, and the assessment accuracy reached over 92% across the entire wear level range. This combination can reflect the comprehensive impact of actual wear and potential fatigue risk on health status in a balanced way.

[0077] The algorithm uses a semi-trapezoidal membership function to calculate the membership degree of each factor corresponding to four health levels. The grading thresholds for wear volume are 5 cubic millimeters, 10 cubic millimeters, and 15 cubic millimeters, and the grading thresholds for maximum stress are 700 MPa, 800 MPa, and 850 MPa. A comprehensive membership vector is calculated by combining factor weights, and the level with the highest comprehensive membership degree is selected as the final judgment result using the principle of maximum membership degree. The calculation process incorporates weighted corrections based on the transmission relationship weights of the connecting components. The transmission relationship weights are determined according to the importance of gears in the transmission system, with the driving gear having a weight of 0.7 and the driven gear a weight of 0.3. This weight ratio is determined based on statistical analysis of the impact of transmission system failures. The driving gear bears the power input function, and its failure has an impact on the entire transmission system approximately 2.3 times that of the driven gear. The weight allocation of 0.7 and 0.3 accurately reflects the health weights of gears in different positions. The system outputs the target health status of connecting components such as gearbox drive shafts, with health status categorized into four levels: healthy, slightly degraded, moderately degraded, and severely degraded. The four levels are determined based on a combination of wear volume and stress level. Healthy condition corresponds to a wear volume of less than 5 cubic millimeters and a maximum stress below the fatigue threshold. Mild degradation corresponds to a wear volume of 5 to 10 cubic millimeters and a maximum stress below the fatigue threshold. Moderate degradation corresponds to a wear volume of 10 to 15 cubic millimeters or a maximum stress close to the fatigue threshold. Severe degradation corresponds to a wear volume of more than 15 cubic millimeters or a maximum stress exceeding the fatigue threshold.

[0078] For example, the system selects the wear depth data corresponding to the target evaluation time and, combined with the gear pitch circle radius parameter, calculates a three-dimensional point cloud matrix containing 489 spatial coordinates. A Poisson surface reconstruction algorithm is used to generate a virtual reconstruction model of the tooth surface. The wear volume is calculated to be 15 cubic millimeters using the Boolean difference algorithm. The virtual reconstruction model is imported into the finite element analysis model, and a rated torque load is applied to obtain a stress distribution map. The maximum stress value is 880 MPa, which is greater than the fatigue threshold of 850 MPa. A fuzzy comprehensive evaluation algorithm is used for evaluation, and the target health state of the gear is output as a severely degraded state.

[0079] In summary, this invention discloses a big data-based intelligent connector detection method, comprising steps such as extracting equipment operating characteristics, screening target physical alternating characteristics, constructing a physical loss sequence, analyzing interface wear characteristics, time-series deduction of degradation trajectories, data completion and smoothing, three-dimensional reconstruction, and health status assessment. This invention leverages big data analysis to establish a mapping relationship between external operating condition data and internal micro-losses, accurately identifying various dominant wear characteristics, filling in data gaps caused by discrete sampling, and reconstructing a complete degradation trajectory. This achieves effective detection of wear on hidden connector interfaces and significantly improves the accuracy of equipment health status assessment.

[0080] The second embodiment of the present invention provides a connector intelligent detection system based on big data, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described above.

[0081] It should be noted that the connector intelligent detection system based on big data provided in this embodiment of the invention is used to execute all the process steps of the connector intelligent detection method based on big data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0082] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0083] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A connector intelligent detection method based on big data, characterized in that, include: Acquire multi-dimensional operating condition data of the equipment connection components, and perform feature extraction processing on the multi-dimensional operating condition data to obtain a first set of operating features; Based on the first set of operational features, independent feature channels are separated and hierarchical screening is performed based on the strength of degradation correlation to obtain the target physical alternation features; Based on the target's physical alternation characteristics, a mapping relationship between externally acquired information and the physical changes in the internal enclosed space is constructed to obtain a physical loss sequence; The physical loss sequence is subjected to multi-dimensional feature extraction and dimensionality reduction reconstruction to obtain the wear characteristics of the target interface; Based on the wear characteristics of the target interface, the coupled evolution of the interface degradation index is time-series extrapolated to obtain a first degradation trajectory sequence. Then, the first degradation trajectory sequence is continuously reconstructed to obtain a second degradation trajectory sequence. Based on the second degradation trajectory sequence, a three-dimensional reconstruction of the contact interface and a comprehensive health status assessment are performed to obtain the target health status of the device connection components.

2. The intelligent connector detection method based on big data according to claim 1, characterized in that, The process involves acquiring multidimensional operating condition data of the connected components of the equipment, performing feature extraction processing on the multidimensional operating condition data to obtain a first set of operating features, including: The multidimensional operating condition data of the equipment connection components are acquired and time-aligned to obtain a multidimensional operating condition time sequence. The multidimensional operating condition time series is subjected to frequency domain mapping and dimensionality reduction processing to obtain the principal component vector; Anomaly feature vectors are extracted from the principal component vectors, and then concatenated with historical principal component vectors extracted from historical operating condition data to obtain the first set of operating features.

3. The intelligent connector detection method based on big data according to claim 1, characterized in that, The step of separating independent feature channels based on the first set of operational features and performing hierarchical screening based on the strength of degradation correlation to obtain target physical alternation features includes: The first set of operating features is subjected to independent component separation processing to obtain multiple independent channel response sequences; A gearbox degradation evolution trajectory is pre-constructed, and a preliminary screening is performed based on the interface degradation correlation strength between the channel response sequence and the pre-constructed degradation evolution trajectory to obtain a candidate channel set; Based on the correlation between alternating load fatigue damage and physical alternating stress of the candidate channel set, the target physical alternating characteristics are obtained through fine screening.

4. The intelligent connector detection method based on big data according to claim 1, characterized in that, Based on the target physical alternation characteristics, a mapping relationship is constructed between externally acquired information and the physical changes in the internal enclosed space to obtain a physical loss sequence, including: Align the target physical alternating features with the preset spatial coordinates of the device housing to obtain the spatial boundary deformation tensor; Vibration response analysis and state transition modeling are performed on the spatial boundary deformation tensor to obtain the response attenuation matrix; The physical loss sequence is obtained by performing degradation trajectory decoding and smoothing on the response attenuation matrix.

5. The intelligent connector detection method based on big data according to claim 1, characterized in that, The process of extracting and reconstructing multi-dimensional features from the physical loss sequence to obtain the wear features of the target interface includes: The physical loss sequence is subjected to dimensionality reduction and independent component separation to obtain a coarse gradient vector and a depth distribution vector; The coarse gradient vector and the depth distribution vector are subjected to correlation fusion processing to obtain the initial interface wear characteristics; The initial interface wear characteristics are subjected to physical component separation and tensor reconstruction to obtain the target interface wear characteristics.

6. The intelligent connector detection method based on big data according to claim 1, characterized in that, The first degradation trajectory sequence is obtained by performing time-series extrapolation of the coupled evolution of interface degradation indices based on the wear characteristics of the target interface, including: The wear characteristics of the target interface are segmented by a sliding window to obtain a time-series evolution feature vector; Coupled evolution analysis and anomaly smoothing are performed on the temporal evolution feature vectors to obtain a standard degradation state sequence; The first degradation trajectory sequence is obtained by forward extrapolating the standard degradation state sequence.

7. The intelligent connector detection method based on big data according to claim 1, characterized in that, The step of further reconstructing the first degraded trajectory sequence into a continuous sequence to obtain the second degraded trajectory sequence includes: Anomaly detection processing is performed on the first degraded trajectory sequence to remove abnormal sampling nodes and obtain a clean trajectory sequence; The pure trajectory sequence is subjected to missing node localization and interpolation compensation processing to obtain a depth-compensated sequence; The depth compensation sequence is continuously smoothed to obtain the second degenerate trajectory sequence.

8. The intelligent connector detection method based on big data according to claim 1, characterized in that, The step of performing three-dimensional reconstruction and comprehensive health status assessment of the contact interface based on the second degradation trajectory sequence to obtain the target health status of the device connection components includes: A three-dimensional point cloud matrix is ​​constructed based on the second degenerate trajectory sequence, and the surface reconstruction process is performed on the three-dimensional point cloud matrix to obtain a virtual reconstruction model; Wear feature extraction and stress analysis are performed on the virtual reconstruction model to obtain wear volume features and stress distribution map; The target health status of the equipment connection components is obtained by comprehensively evaluating the wear volume characteristics and the stress distribution spectrum.

9. A connector intelligent detection system based on big data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 8.

Images