A method for predicting the life of a screw shaft

Abstract

This invention belongs to the field of predictive maintenance technology for equipment, and provides a method for predicting the lifespan of a helical shaft. The method includes: segmenting real-time helical shaft data using a sliding time window; obtaining feature parameter vectors for each segment to form a real-time feature matrix; inputting this matrix into a pre-trained dynamic health benchmark model; and outputting the minimum reconstruction error value corresponding to each segment. If the minimum reconstruction error values ​​corresponding to several consecutive segments of real-time helical shaft data are all less than a normal error threshold, a health index is calculated based on the minimum reconstruction error values, and a health degradation trajectory is fitted. When the slope of the health degradation trajectory exceeds a threshold, the health degradation trajectory is predicted for the future to obtain the remaining lifespan. This invention triggers degradation degree monitoring of the degradation trajectory when the minimum reconstruction error corresponding to consecutive real-time data is less than the normal error threshold, enabling intelligent monitoring of actual performance degradation and improving the timeliness of fault detection.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology, specifically to a method for predicting the lifespan of a screw shaft. Background Technology

[0002] Screw presses, as a highly efficient and energy-saving sludge dewatering device, are widely used in various fields such as municipal wastewater, industrial wastewater, and food processing wastewater. The screw shaft, as the core component of the screw press, has a crucial impact on the overall performance and service life of the equipment through its design, manufacturing, application, and maintenance.

[0003] As a key component in screw presses that withstands complex torsional loads and wear, the performance degradation of the screw shaft is often gradual and insidious, typically manifesting as the evolution of weak temporal characteristics embedded in complex nonlinear vibration signals. These characteristics have low energy and poor signal-to-noise ratio in the early stages and are easily masked by strong background noise and multiple harmonic components during equipment operation. This makes it difficult for traditional threshold- or spectrum-based analysis methods to effectively identify and accurately locate these characteristics, let alone accurately predict remaining lifespan.

[0004] Currently, the common approach for predicting the remaining life of spiral shafts is to directly extract local spatial correlation features from equipment operation data using convolutional and pooling layers, and then input these extracted spatial features into a hybrid long short-term memory (LSTM) network for temporal dependency modeling. However, this method has significant limitations. Convolutional layers are highly dependent on the dimension and size of the input data, resulting in insufficient flexibility in feature extraction when dealing with multi-source, heterogeneous, and dynamically changing equipment monitoring data. Furthermore, the spatial features are input into the LSTM network without filtering out redundant information, leading to a large amount of invalid information that increases the computational complexity of the model and reduces the efficiency and accuracy of temporal modeling. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting the lifespan of a helical shaft, thereby solving the problems of high estimation complexity and performance degradation caused by the ambiguity in the definition of remaining service life during normal operation.

[0006] This invention provides a method for predicting the life of a spiral shaft, comprising: The real-time data of the spiral axis is segmented using a sliding time window. After obtaining the feature parameter vector for each segment, a real-time feature matrix is ​​formed. The real-time feature matrix is ​​input into a pre-trained dynamic health benchmark model, and the minimum reconstruction error value corresponding to each segment is output. If the minimum reconstruction error value corresponding to several consecutive segments is less than the normal error threshold, then the health index is calculated based on the minimum reconstruction error value, and the health degradation trajectory of the health index over time is obtained by fitting. When the slope of the health degradation trajectory exceeds a threshold, the health degradation trajectory is predicted to obtain the remaining service life; the remaining service life is determined based on the intersection of the health degradation trajectory and the failure health index threshold.

[0007] As can be seen from the above technical solution, the spiral shaft life prediction method provided by the present invention segments the real-time data of the spiral shaft and inputs each segment into a trained dynamic health benchmark model. When the minimum reconstruction error corresponding to the continuous real-time data is less than the normal error threshold, the degradation degree (slope) monitoring of the degradation trajectory is triggered. When it exceeds the threshold, the health degradation trajectory is extended in the future direction to obtain the remaining service life. Based on this, the actual performance degradation can be intelligently monitored, improving the timeliness of fault detection.

[0008] Optionally, the training methods for the dynamic health benchmark model include: The PRONOSTIA dataset is preprocessed to obtain a clean and normal dataset. The PRONOSTIA dataset includes full-lifetime operation data of multiple identical spiral shafts under different working conditions. The clean normal dataset is segmented based on the sliding time window. Feature parameter vectors are obtained for each segment and stacked into a two-dimensional feature matrix according to the window index. The two-dimensional feature matrix is ​​then standardized to obtain a standardized feature matrix. The standardized feature matrix is ​​divided into working condition sub-data sets by a clustering algorithm; each working condition sub-data set is associated with and stored with the corresponding health baseline value under the working condition. Each working condition subset is divided into a training set and a test set. The training set is used to train the dynamic health benchmark model. Then, the test set is input into the trained dynamic health benchmark model to obtain the normal error threshold corresponding to each working condition subset.

[0009] Optionally, the feature parameter vector includes root mean square value, peak value, kurtosis, spectral centroid, and specific frequency band energy.

[0010] Optionally, the preprocessing of the PRONOSTIA dataset to obtain a clean, normal dataset includes: The running data from the start time point to the preset time point in the entire life cycle of multiple screw shafts under various working conditions is extracted as health sample data. The healthy sample data of multiple spiral shafts under the same operating conditions are merged to obtain a clean and normal dataset.

[0011] Optionally, the step of segmenting the clean, normal dataset based on the sliding time window, obtaining feature parameter vectors for each segment and stacking them into a two-dimensional feature matrix according to the window index, and then standardizing the two-dimensional feature matrix to obtain a standardized feature matrix includes: The sliding time window is used to segment multiple continuous data segments in the clean and normal dataset. For the data within each segment, obtain the feature parameter vector; The feature parameter vectors are stacked according to the window index to form a two-dimensional feature matrix; each row of data in the two-dimensional feature matrix represents the feature parameter vector of a certain window. The two-dimensional feature matrix is ​​standardized column-wise to obtain the standardized feature matrix.

[0012] Optionally, the step of re-dividing the clean, normal dataset into working condition subsets using a clustering algorithm includes: The optimal number of clusters is determined by the profile coefficient, an internal effectiveness index, and the number of known operating state categories of the shaft system. Iteratively optimize the location of the cluster center, minimize the sum of squared Euclidean distances from the sample to its cluster center, and complete the cluster partitioning and association of sample labels; Based on the sample labels, the standardized feature matrix is ​​inverted and reorganized to divide it into sub-datasets for each working condition.

[0013] Optionally, the real-time data of the spiral axis is segmented using a sliding time window. After obtaining feature parameter vectors for each segment, a real-time feature matrix is ​​formed. This real-time feature matrix is ​​input into a pre-trained dynamic health benchmark model, which outputs the minimum reconstruction error value corresponding to each segment, including: After obtaining the feature parameter vectors for each segment, a real-time feature matrix is ​​formed. The real-time feature matrix is ​​input into multiple pre-established dynamic health benchmark models to obtain multiple reconstruction error values ​​for each segment. The dynamic health benchmark model is an autoencoder-long short-term memory hybrid network model. The minimum value among the multiple reconstruction error values ​​obtained by the segment is taken as the minimum reconstruction error value of the segment; the dynamic health benchmark model corresponding to the minimum reconstruction error value is the dynamic health benchmark model corresponding to the segment.

[0014] Optionally, the step of predicting the future trajectory of health degradation to obtain the remaining lifespan includes: Historical data segments with slopes exceeding a threshold are extracted from the health degradation trajectory, and the curve of health index changing over time is obtained by fitting an exponential decay model; the fitting objective is to minimize the mean square error. By extending the time axis towards the future, we obtain the future health index prediction curve. Solve for the time coordinates of the intersection point between the predicted future health index curve and the preset failure health threshold; The difference between the current time point and the time coordinate is the remaining service life.

[0015] By adopting the above technical solution, this application has the following beneficial effects: This invention utilizes only historical monitoring data from the normal operation of the helical shaft. It establishes a multi-state health dynamic benchmark by constructing an unsupervised hybrid model that integrates the feature reconstruction capabilities of an autoencoder and the temporal modeling capabilities of a long short-term memory hybrid network. During monitoring, when the helical shaft experiences initial wear, subtle time-frequency distortions implicit in its vibration signal will significantly increase the model reconstruction error. By continuously tracking the temporal evolution of this error, the system can accurately capture the anomaly initiation pattern and output the precise timestamp of the anomaly's onset, thereby achieving early warning of the degradation initiation stage and providing a critical time window for pre-maintenance decisions. This invention, by establishing the anomaly initiation time, can construct a complete degradation trajectory from the initial anomaly to functional failure, and use this trajectory as a basis to train a high-precision life prediction model, significantly improving the engineering reliability of the prediction results. Thus, a seamless connection is formed from anomaly detection and degradation tracking to life prediction, constructing a complete screw shaft health management process, providing systematic technical support for predictive maintenance, spare parts optimization management, and operational safety assurance of screw presses. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 A flowchart of a helical shaft life prediction method provided by an embodiment of the present invention is shown; Figure 2 A framework diagram of a helical shaft life prediction method provided by an embodiment of the present invention is shown. Detailed Implementation

[0018] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention. It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0019] In one embodiment, such as Figure 1-2 As shown, a method for predicting the life of a helical shaft is provided, including: S110. The real-time data of the spiral axis is segmented using a sliding time window. The feature parameter vectors of each segment are obtained to form a real-time feature matrix. The real-time feature matrix is ​​input into the pre-trained dynamic health benchmark model, and the minimum reconstruction error value corresponding to each segment is output.

[0020] The vibration acceleration sensor continuously acquires equipment vibration signals at a fixed sampling rate and transmits the digital signal stream to the data buffer of the monitoring server in real time through the data acquisition system. The latest signal is extracted from the buffer at fixed intervals, and the real-time data is truncated using a sliding time window with the same parameters as the preprocessing stage. The standardized time-series sample of the current moment is output, and the real-time data sample is normalized and standardized in the same way as in the training stage.

[0021] S111. After obtaining the feature parameter vectors of each segment, a real-time feature matrix is ​​formed. The real-time feature matrix is ​​input into multiple pre-established dynamic health benchmark models to obtain multiple reconstruction error values ​​of each segment. The dynamic health benchmark model is an autoencoder-long short-term memory hybrid network model.

[0022] The preprocessed samples are simultaneously input into multiple trained dynamic health benchmark models. Each model independently encodes and decodes concurrently, resulting in multiple reconstruction error values ​​associated with each segment, forming a real-time reconstruction error set. The real-time reconstruction error set reflects the degree of matching deviation between the real-time data of the spiral shaft and the health feature models of the condition-specific sub-models in the multimodal dynamic health benchmark model.

[0023] S112. The minimum value among the multiple reconstruction error values ​​obtained by the segment is taken as the minimum reconstruction error value of the segment; the dynamic health benchmark model corresponding to the minimum reconstruction error value is the dynamic health benchmark model corresponding to the segment.

[0024] The minimum value is found from the real-time reconstruction error set, and the model that produced the minimum value is recorded. The current status label of the device is the health sub-state represented by the model. The pre-calibrated normal error threshold bound to the model is called. If the error value exceeds the threshold, an abnormal warning is triggered; otherwise, the state is determined to be normal.

[0025] (1) Formula (1) is mainly used for calculating the real-time reconstruction error, where, Let be the real-time reconstruction error at time t, where t is the monitoring timestamp; D is the feature dimension of a single sample. Let d be the original eigenvalue of the d-th dimension at time t. Let be the d-th dimension reconstructed eigenvalue at time t.

[0026] S120. If the minimum reconstruction error value corresponding to the segments of real-time data of several consecutive spiral axes is less than the normal error threshold, then the health index is calculated based on the minimum reconstruction error value, and the health degradation trajectory of the health index over time is obtained by fitting.

[0027] S130. When the slope of the health degradation trajectory is detected to exceed the threshold, the health degradation trajectory is predicted in the future to obtain the remaining service life; the remaining service life is determined based on the intersection of the health degradation trajectory and the failure health index threshold.

[0028] The historical data, real-time data and corresponding operating parameters of the current helical shaft are input into the pre-trained dynamic health benchmark model in chronological order, and the corresponding real-time reconstruction error value is output. Based on the nonlinear decay formula, the reconstruction error is converted into a health index, and the health degradation trajectory is plotted based on the calculation results. (2) In formula (2), To reconstruct errors in real time, To match the normal error threshold of the working condition, k is the attenuation coefficient (value 3.0); the health index HI ranges from (0,1], and the closer the value is to 1, the better the health status.

[0029] Using time as the horizontal axis and the health index as the vertical axis, the HI values ​​at each timestamp are plotted in chronological order to form a health degradation trajectory curve. Rotating shaft machinery degrades slowly during normal operation, but the degradation rate accelerates significantly as it approaches failure. Its health status exhibits typical nonlinear accelerated degradation characteristics over operating time, highly consistent with the variation law of the exponential decay model. Therefore, using an exponential function can more accurately fit the actual degradation process of equipment from health to failure. This newly generated health degradation trajectory curve is analyzed to detect whether it exhibits a statistically significant monotonically decreasing trend. If a significant decreasing trend is detected, and the current health index is below the preset warning threshold, the system automatically triggers the life prediction program. The core of remaining life expectancy prediction lies in mathematical modeling and forward-looking extrapolation of identified health degradation trajectories. Once prediction is triggered, a nonlinear regression fitting is first performed on significant downward trend segments in the historical health index sequence. An exponential decay model is employed to capture the dynamic characteristics of the degradation process. The fitting process determines the model parameters by minimizing the prediction error, thereby obtaining an analytical curve that accurately describes the past degradation patterns.

[0030] Specifically, step S130 includes: S131. Extract historical data segments with slopes exceeding a threshold from the health degradation trajectory, and use an exponential decay model to fit the curve of the health index changing over time; the goal of the fitting is to minimize the mean square error between the calculated value and the measured value of the historical health index.

[0031] S132. The analysis curve is extended along the time axis in the future direction to obtain the future health index prediction curve; (3) In formula (3), a, b, and c are fitting parameters (b>0, representing decay rate), and t is a historical timestamp; S133. Solve for the time coordinate of the intersection point between the predicted future health index curve and the preset failure health threshold. This intersection point is the time point of equipment functional failure inferred by the algorithm. The difference between the current time and this future time point is the predicted remaining service life. The preset failure health index is calibrated based on the objective judgment standard of equipment functional failure. Collect the full life cycle monitoring data of the same type of equipment, locate the sensor data corresponding to the time of equipment functional failure, input the sensor data at the time of failure into the trained and converged autoencoder-long short-term memory hybrid network model, and calculate the reconstruction error at the time of failure. Substitute it into formula (8) to obtain the preset failure health index threshold of equipment i. (4) Formula (4) is used to calculate the health index value at the time of failure of each device, where i is the i-th failed device and k is the attenuation coefficient. This is the normal error threshold for the corresponding operating condition.

[0032] Statistical analysis was performed on the failure health index thresholds of multiple devices. After removing outliers, the statistical mean was taken as the final failure health index threshold. For devices operating under multiple conditions, the failure health index thresholds for each condition needed to be calibrated by grouping them according to load and speed, and then bound to the dynamic health benchmark model for the corresponding operating condition. Outliers could be determined using percentage thresholds, and all outliers were calculated. The preliminary average value, when a certain device corresponds to Values ​​deviating from the preliminary average by more than ±30% are considered outliers and must be removed. To quantify the uncertainty of the prediction, the module typically employs Bayesian inference techniques to generate a probability distribution or confidence interval for the remaining useful life.

[0033] S134. Obtain the difference between the current time point and the time coordinate, which is the remaining service life.

[0034] In one embodiment, the method further includes: automatically synthesizing a structured prediction report, specifically including: The structured prediction report not only includes point and interval estimates of remaining useful life, but also covers key indicators such as the predicted specific failure date, the instantaneous rate of decline of the current health index, the real-time status assessment level based on the health index, and model fit goodness, providing multi-dimensional quantitative basis for maintenance decisions. Combining preset warning thresholds and failure thresholds for the health index, if the warning threshold is lower than the failure threshold, the health index is classified according to its current value range. I. A health index above the warning threshold indicates a healthy state; II. A health index below the warning threshold and above the failure threshold indicates a mild degradation state; III. A health index greater than 0.5 times the failure threshold and less than or equal to the failure threshold indicates a moderate degradation state. IV. A health index greater than 0.1 times the failure threshold and less than or equal to 0.5 times the failure threshold indicates a severely degraded state. V. A health index less than or equal to 0.1 times the failure threshold indicates a failure state.

[0035] Model goodness of fit is determined by comparing the degree of fit between the measured trajectory of the health index and the fitted curve. A higher degree of overlap and smaller deviation indicates a better goodness of fit, and the life expectancy prediction derived from this fitted curve is more reliable. Conversely, a large deviation indicates a poor goodness of fit, requiring further optimization of the health index fitting model to ensure prediction accuracy. Warning signs, status labels, anomaly scores, and the first anomaly timestamp are packaged into a complete monitoring record, stored in a database, and a real-time health status query interface is provided to the upper-level system.

[0036] As the helical shaft continues to operate, each time new operational data is acquired, the fitting of the degradation trajectory is optimized using the latest data points, and future curve prediction is performed again. Consistent with the core principle of extending the previous healthy degradation trajectory, both methods calculate the remaining service life based on the extension of the health index degradation fitting curve into the future. This process allows the prediction model to dynamically adapt to the latest performance of equipment degradation, thereby gradually correcting prediction biases and achieving an adaptive optimization closed loop where prediction accuracy continuously improves with data accumulation.

[0037] In one embodiment, the training method for the dynamic health benchmark model used in step S110 includes: S210. Preprocess the PRONOSTIA dataset to obtain a clean and normal dataset; the PRONOSTIA dataset includes full-lifetime operation data of multiple identical spiral shafts under different working conditions.

[0038] The PRONOSTIA dataset contains full-life operational data for multiple identical screw shafts under various load and speed combinations. For each screw shaft, the test maintained a preset constant load and speed until complete failure, with the full life cycle duration for a single screw shaft ranging from 1 hour to 7 hours. For the fixed operating conditions, two accelerometer sensors were deployed at 90° angles to measure vertical and horizontal vibration directions, corresponding to the X and Y directions respectively, with a sampling frequency of 25.6 kHz, to acquire the vibration acceleration signals during the screw shaft's operation in real time. The vibration acceleration signals represent the physical response measurement of the screw shaft under the corresponding load-speed conditions.

[0039] The above mainly introduces the experimental conditions, sensor arrangement, and data acquisition method of the PRONOSTIA screw shaft fault dataset, which is used to illustrate the source and physical meaning of the data used in this embodiment. Subsequent data preprocessing, lifecycle segment selection, and model training are all based on this dataset, assuming that its data size can meet the training and validation requirements of the proposed method.

[0040] S211. Extract the operating data from the start time point to the preset time point throughout the entire life cycle of multiple screw shafts under various operating conditions, and use it as health sample data. In a specific example, extract the data for the first 30% of the entire life cycle as health sample data.

[0041] S212. Merge health sample data from multiple spiral axes under the same operating condition to obtain a clean, normal dataset. Under the premise of consistent healthy operating status, by merging health samples from multiple spiral axes, a dataset containing normal operating conditions is constructed to generate a data segment representing the absolute health status.

[0042] S220. The clean normal dataset is segmented based on the sliding time window. Feature parameter vectors are obtained for each segment and stacked into a two-dimensional feature matrix according to the window index. The two-dimensional feature matrix is ​​then standardized to obtain a standardized feature matrix. The feature parameter vectors include root mean square value, peak value, kurtosis, spectral centroid and specific frequency band energy. The calculation method is shown in Table 1.

[0043] Table 1

[0044] The calculation of specific frequency band energy is based on the unified delineation of a specific frequency band. All data segments follow the same delineation rules for their specific frequency bands to ensure the consistency and effectiveness of feature extraction. This frequency band is determined based on the failure evolution law of the helical shaft and the needs of lifetime prediction, and adopts a delineation method based on the characteristic frequencies of helical shaft failures. Narrow frequency bands are delineated centered on the failure characteristic frequencies of the outer race (BPFO), inner race (BPFI), rolling elements (BSF), and cage (FTF) of the helical shaft to capture the energy change characteristics of the corresponding failures. Physically speaking, when the helical shaft fails and gradually deteriorates, the signal energy within its fault-related frequency band shows a significant and regular increase. Therefore, the specific frequency band energy can serve as the core quantitative indicator for helical shaft failure evolution characterization and remaining lifetime prediction, and it is also the key basis for this scheme to use it as the core frequency domain statistic.

[0045] Specifically, S220 includes: S221. Divide multiple consecutive data segments in a clean, normal dataset into segments using a sliding time window.

[0046] S222. For the data in each segment, obtain the feature parameter vector.

[0047] S223. Stack the feature parameter vectors according to the window index to form a two-dimensional feature matrix; each row of data in the two-dimensional feature matrix represents the feature parameter vector of a certain window.

[0048] S224. Perform standardization operations on the two-dimensional feature matrix column by column to obtain the standardized feature matrix.

[0049] Using the time-domain and frequency-domain statistics system shown in the table above, the original waveform of each window can be converted into a 5-dimensional low-dimensional feature vector. This low-dimensional feature vector is formed by combining the root mean square value, peak value, kurtosis, spectral centroid, and specific frequency band energy in a fixed order. By stacking the feature vectors corresponding to all data windows row by row, a two-dimensional feature matrix F∈R can be constructed. W×5 (W represents the total number of data windows, and 5 represents the feature vector dimension). Perform column-wise standardization on the two-dimensional feature matrix F, specifically according to formula (5); (5) in, Let be the mean of the j-th column of the two-dimensional feature matrix. Let be the standard deviation of the j-th column of the two-dimensional feature matrix. This operation unifies the dimensions of each column corresponding to the same feature dimension, eliminating numerical differences between different features. After standardization, a standardized feature space is obtained, and each independent feature vector in this standardized feature space serves as the input data for the clustering algorithm.

[0050] Steps S221-S224 are described below by way of specific examples, but are not intended to limit the scope of protection of the present invention.

[0051] A specific example is as follows: The completed PRONOSTIA spiral shaft clean and normal dataset is segmented using a sliding time window. The window length is set to 1024 sampling points, with adjacent windows overlapping by 512 sampling points. Window data is sequentially extracted from the clean and normal dataset, resulting in a total of 136,718 window samples. For the vibration signal of each window, five feature parameters are calculated in a fixed order: root mean square value → peak value → kurtosis → spectral centroid → specific frequency band energy. This constructs a 1×5 low-dimensional feature vector. The feature calculation results for the first three windows are as follows: The root mean square value of window 1 (at time t1) is 0.023 m / s. 2 The peak value is 0.089 m / s 2 The kurtosis is 3.12, the spectral centroid is 1250Hz, and the specific frequency band energy (a narrow band centered at BPFO=180Hz) is 0.0015m. 2 / s 4 The corresponding feature vectors are [0.023, 0.089, 3.12, 1250, 0.0015]. The root mean square value of window 2 (at time t2) is 0.021 m / s. 2 The peak value is 0.085 m / s 2 The kurtosis is 3.08, the spectral centroid is 1248Hz, and the specific frequency band energy is 0.0014m. 2 / s 4 The corresponding feature vector is [0.021, 0.085, 3.08, 1248, 0.0014]. The root mean square value of window 3 (time t3) is 0.024 m / s. 2 The peak value is 0.091 m / s 2 The kurtosis is 3.15, the spectral centroid is 1252Hz, and the specific frequency band energy is 0.0016m. 2 / s 4 The corresponding feature vector is [0.024, 0.091, 3.15, 1252, 0.0016].

[0052] Stack all 136718 1×5 low-dimensional feature vectors in rows to obtain a 136718×5 two-dimensional feature matrix, where the rows of the matrix correspond to the number of window samples and the columns correspond to the 5-dimensional feature parameters.

[0053] The two-dimensional feature matrix is ​​processed by the standardization method of formula (5). The mean of the five-dimensional features is close to 0 and the standard deviation is 1, forming a standardized feature space. The standardized feature space is still a matrix of 136718×5. The standardized feature vectors of the first three windows are respectively: window 1 standardized feature vector [0.12, 0.09, 0.07, 0.05, 0.11], window 2 standardized feature vector [-0.21, -0.15, -0.08, -0.12, -0.09], and window 3 standardized feature vector [0.25, 0.18, 0.15, 0.18, 0.23].

[0054] In one example, the normalized feature matrix can be represented as .

[0055] S230. The standardized feature matrix is ​​divided into working condition subsets using a clustering algorithm, forming working condition subset datasets. Each working condition subset dataset is associated with and stores the corresponding health baseline value under the working condition. The elements in each working condition subset dataset are sub-matrices resulting from the working condition division of the standardized feature matrix. Specifically, S230 includes: S231. Determine the optimal number of clusters using the internal validity index silhouette coefficient and the number of known operating state categories of the shaft system. A key feature of the PRONOSTIA dataset is that it contains operational data for the entire lifecycle of the spiral shaft, and the dataset itself predefines the number of basic operating state categories of the spiral shaft, eliminating the need for additional algorithmic calculations. (6) Formula (6) silhouette coefficient is a commonly used internal clustering effectiveness index used to evaluate the quality of clustering results. Here, a is the average distance from a sample to all other samples within the same cluster (intra-cluster compactness), and b is the average distance from a sample to all samples in the nearest other cluster (inter-cluster separation). It measures the similarity of a sample to its own cluster and its dissimilarity to the nearest other clusters, with a value between -1 and 1.

[0056] S232. Iteratively optimize the location of the cluster center, minimizing the sum of squared Euclidean distances from the sample to its cluster center, thus completing the cluster partitioning and the association of sample labels. Based on this, feature parameter vectors with highly similar vibration spectra and time-domain statistical characteristics are automatically aggregated into the same cluster, forming several sub-matrices. Each final cluster represents an operating mode that is tightly clustered in the feature space and clearly separated from others, and can be regarded as a stable and healthy intrinsic sub-state, thereby completing the data-driven, preliminary semantically unlabeled state partitioning. S233. Based on the sample labels, perform reverse indexing and reorganization of the standardized feature matrix to divide it into sub-datasets for each working condition.

[0057] All feature parameter vectors belonging to the same cluster are mapped and extracted back to their corresponding original time-series data segments based on their original timestamps or sequence indices. These grouped time-series data segments constitute the basic data units for subsequent sub-state specificity analysis. A specific example of an original time-series data segment is shown below. Its core consists of a data point sequence, timestamps, and optional metadata, used to support the inverse indexing after clustering and subsequent sub-state specificity analysis. The data point sequence is the amplitude of the radial vibration acceleration signal of the helical shaft collected in chronological order, using a 10kHz sampling frequency. Each segment of time-series data contains 10 consecutive sampling points, corresponding to a collection duration of 1ms. The specific data point sequence is {0.023g, 0.021g, 0.024g, 0.022g, 0.025g, 0.023g, 0.022g, 0.024g, 0.023g, 0.023g, 0.024g, 0.023g, 0.023g, 0.024g, 0.023g, 0.023g, 0.024g, 0.023g, 0.024g, 0.025g, 0.023g, 0.02 ...5g, 0.025g, 0.023g, 0.022g, 0.024g, 0.025g, 0.025g, 0.025g, 0.022g, 0.025g, 0.023g, 0.024g 0.021g}, all data points are calibrated by sensors and the unit is gravitational acceleration (g), directly reflecting the real-time dynamic response and health status of the screw shaft; the timestamp is marked in a format accurate to the millisecond level, clearly defining the acquisition time context of the data segment and ensuring that the clustering results can be accurately traced back to the original data; optional metadata synchronously records the equipment operating parameters during the acquisition period, specifically the sampling frequency of 10kHz, screw shaft shell temperature of 38℃, equipment operating load of 500N, and screw shaft speed of 1500r / min. This type of metadata is used for subsequent sub-state specific analysis and fault mode identification to improve the accuracy of life prediction.

[0058] S240. Divide each working condition subset into a training set and a test set. Use the training set to train the dynamic health benchmark model. Then, input the test set into the trained dynamic health benchmark model to obtain the normal error threshold corresponding to each working condition subset.

[0059] For a specific subset of the work condition dataset, the training set is preprocessed by denoising, slicing, and normalization, and then loaded in mini-batches into the dynamic health benchmark model (autoencoder-long short-term memory hybrid network model). The encoder-long short-term memory hybrid network maps this data to a low-dimensional context vector, which the decoder-long short-term memory hybrid network uses to reconstruct the original sequence. The dynamic health benchmark model outputs the reconstruction error. It should be clarified that the original sequence is a subset of the preprocessed time-series data from the training set, and the reconstructed sequence is the vibration signal sequence of the same dimension output by the decoder. The reconstruction loss of the autoencoder is obtained by calculating the mean square error between the reconstructed sequence and the original sequence. This value objectively quantifies the model's reconstruction deviation of the normal pattern under the current parameters; (7) Formula (7) is mainly used to calculate the reconstruction loss, where N is the batch size. Let i be the original feature sequence of the i-th healthy sample. This is the reconstructed feature sequence output by the model.

[0060] Based on the reconstructed loss function, the backpropagation algorithm is used to optimize the trainable parameters of the autoencoder-long short-term memory hybrid network model corresponding to each cluster. Based on the chain rule, starting from the output of the loss function, the specific influence of the loss value on each trainable parameter in the model is calculated in reverse, i.e., the gradient is calculated. The trainable parameters of the autoencoder and long short-term memory network algorithm models are parameters that need to be iteratively optimized during the training phase. These mainly include the weight matrix and bias vector of the autoencoder and long short-term memory hybrid network. Considering the helical shaft vibration time-series data processing scenario of this scheme, the specific settings and meanings of each parameter are as follows: For the autoencoder (AE) part, the weight matrix W from the encoder input layer to the hidden layer... enc The dimensions are set to (10, 64), where the input dimension 10 corresponds to the 10 sampling points of the original time-series data segment mentioned above, and the hidden layer dimension is set to 64; the encoder hidden layer bias vector b is optimized in conjunction with this weight matrix. enc The dimension is (64,1). In the decoder part, the weight matrix W from the encoder output to the decoder hidden layer... dec1 The dimension is (64, 32), with a corresponding bias vector b. dec1 The dimension is (32,1); the weight matrix W from the hidden layer to the output layer of the decoder. dec2 The dimension is (32, 10), and the output dimension is consistent with the input dimension to achieve sequence reconstruction, with a corresponding output layer bias vector b. dec2 The dimension is (10,1); In the hybrid long short-term memory network, the parameter settings are uniform for each time step t. The input gate, forget gate, output gate, and cell state update mechanism all correspond to independent trainable parameters, and the dimensionality of each parameter remains consistent. The input dimension of 10 corresponds to 10 sampling points of the time-series data. The LSTM hidden layer dimension is 128, meaning the input weight matrix dimension is (10, 128), the recurrent weight matrix dimension is (128, 128), and the bias vector dimension is (128, 1). Specifically, the parameters corresponding to the input gate are the input weight matrix W. xi Circular weight matrix W hi and bias vector b i The forget gate corresponds to the input weight matrix W. xf Circular weight matrix W hf and bias vector b f The output gate's corresponding parameter is the input weight matrix W. xo Circular weight matrix W ho and bias vector b o The parameter corresponding to the cell state update is the input weight matrix W. xc Circular weight matrix W hcand bias vector b c All of the above parameters were iteratively optimized during the training phase using the backpropagation algorithm; The optimizer uses this gradient information to make targeted adjustments to all parameters. Gradient calculations penetrate backward from the loss output to each layer of the model. After obtaining the gradient values ​​of all parameters, the optimizer adjusts the parameters according to the magnitude and direction of the gradients. Parameters with large absolute gradient values, meaning they contribute significantly to the error, are significantly corrected, while parameters with small absolute gradient values, meaning they have a smaller impact, are fine-tuned. This process clarifies the direction and magnitude of each parameter's contribution to the total loss, resulting in smaller reconstruction errors when the model encounters similar input data. It completes the learning of the spiral axis health status features, and after training, outputs the health baseline values ​​and error thresholds for each cluster, providing quantitative support for remaining lifetime prediction and fault warning. This process determines the direction and magnitude of each parameter's contribution to the total loss, enabling the model to produce smaller reconstruction errors when encountering the same or similar inputs. The aforementioned forward and backward propagation processes, from the input layer to the output loss value, are performed cyclically across all training data batches. Each complete traversal of the dataset is called a training epoch. In the early stages of training, the loss value typically decreases rapidly with each epoch, indicating that the model is quickly learning the basic patterns of normal data. As training progresses, the rate of loss decrease gradually slows. When the loss value on the validation set decreases by less than a preset convergence threshold over several consecutive epochs, or fluctuates randomly without trend within a preset allowable fluctuation range centered on the mean, the model training is considered to have reached convergence. At this point, the model parameters have been optimized to a stable point, and the learned intrinsic patterns of the data no longer undergo fundamental improvement with additional training; the training process terminates. The final set of model parameters defines a fixed equipment health benchmark that can be used for online monitoring. The test set not used in model training is input into the converged dynamic health benchmark model. The reconstruction error value of the model for each sample in the test set is recorded, forming a set of error values. Statistical analysis is performed on the set to calculate the high quantile of its distribution. This high quantile value is set as the normal error threshold. (8) Formula (8) is mainly used to calculate the reconstruction error value, where is the reconstruction error value of a single test sample, and D is the feature dimension of a single sample (time-frequency domain features in both X and Y directions). These are the original feature values ​​of the d-th dimension of the test sample. Let d be the reconstructed feature value of the d-th dimension of the test sample.

[0061] Each different original data grouped according to label is trained with the above steps to train an autoencoder-long short-term memory hybrid network model. The reconstruction error is reduced by iterative optimization to obtain the sub-model of the autoencoder-long short-term memory hybrid network model corresponding to each label group. Each sub-model has a normal error threshold. By integrating the parameters of all sub-models, the health baseline values ​​for each operating condition, and the normal error threshold, a multimodal dynamic health baseline model is obtained. The health baseline values ​​for each operating condition are derived from the standardized feature mean vector of the corresponding operating condition data subset. For each operating condition subset, the 5-dimensional standardized feature mean of all window samples is calculated separately, and this mean vector is the health baseline value for the corresponding operating condition.

[0062] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for the purpose of helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A method for predicting the life of a screw shaft, characterized in that, include: The real-time data of the spiral axis is segmented using a sliding time window. After obtaining the feature parameter vector for each segment, a real-time feature matrix is ​​formed. The real-time feature matrix is ​​input into a pre-trained dynamic health benchmark model, and the minimum reconstruction error value corresponding to each segment is output. If the minimum reconstruction error value corresponding to the segments of real-time data of several consecutive spiral axes is less than the normal error threshold, then the health index is calculated based on the minimum reconstruction error value, and the health degradation trajectory of the health index over time is obtained by fitting. When the slope of the health degradation trajectory exceeds a threshold, the health degradation trajectory is predicted to obtain the remaining service life; the remaining service life is determined based on the intersection of the health degradation trajectory and the failure health index threshold.

2. The method according to claim 1, characterized in that, Training methods for dynamic health benchmark models include: The PRONOSTIA dataset is preprocessed to obtain a clean and normal dataset. The PRONOSTIA dataset includes full-lifetime operation data of multiple identical spiral shafts under different working conditions. The clean normal dataset is segmented based on the sliding time window. Feature parameter vectors are obtained for each segment and stacked into a two-dimensional feature matrix according to the window index. The two-dimensional feature matrix is ​​then standardized to obtain a standardized feature matrix. The standardized feature matrix is ​​divided into working condition sub-data sets by a clustering algorithm; each working condition sub-data set is associated with and stored with the corresponding health baseline value under the working condition. Each working condition subset is divided into a training set and a test set. The training set is used to train the dynamic health benchmark model. Then, the test set is input into the trained dynamic health benchmark model to obtain the normal error threshold corresponding to each working condition subset.

3. The method according to claim 1, characterized in that, The feature parameter vector includes root mean square value, peak value, kurtosis, spectral centroid, and specific frequency band energy.

4. The method according to claim 2, characterized in that, The preprocessing of the PRONOSTIA dataset to obtain a clean, normal dataset includes: The running data from the start time point to the preset time point in the entire life cycle of multiple screw shafts under various working conditions is extracted as health sample data. The healthy sample data of multiple spiral shafts under the same operating conditions are merged to obtain a clean and normal dataset.

5. The method according to claim 2, characterized in that, The process of segmenting the clean, normal dataset based on the sliding time window, obtaining feature parameter vectors for each segment and stacking them into a two-dimensional feature matrix according to the window index, and then standardizing the two-dimensional feature matrix to obtain a standardized feature matrix includes: The sliding time window is used to segment multiple continuous data segments in the clean and normal dataset. For the data within each segment, obtain the feature parameter vector; The feature parameter vectors are stacked according to the window index to form a two-dimensional feature matrix; each row of data in the two-dimensional feature matrix represents the feature parameter vector of a certain window. The two-dimensional feature matrix is ​​standardized column-wise to obtain the standardized feature matrix.

6. The method according to claim 5, characterized in that, The step of re-dividing the clean, normal dataset into working condition subsets using a clustering algorithm includes: The optimal number of clusters is determined by the profile coefficient, an internal effectiveness index, and the number of known operating state categories of the shaft system. Iteratively optimize the location of the cluster center, minimize the sum of squared Euclidean distances from the sample to its cluster center, and complete the cluster partitioning and association of sample labels; Based on the sample labels, the standardized feature matrix is ​​inverted and reorganized to divide it into sub-datasets for each working condition.

7. The method according to claim 1, characterized in that, The real-time data of the spiral axis is segmented using a sliding time window. Feature parameter vectors are obtained for each segment to form a real-time feature matrix. This real-time feature matrix is ​​input into a pre-trained dynamic health benchmark model, which outputs the minimum reconstruction error value corresponding to each segment, including: After obtaining the feature parameter vectors for each segment, a real-time feature matrix is ​​formed. The real-time feature matrix is ​​input into multiple pre-established dynamic health benchmark models to obtain multiple reconstruction error values ​​for each segment. The dynamic health benchmark model is an autoencoder-long short-term memory hybrid network model. The minimum value among the multiple reconstruction error values ​​obtained by the segment is taken as the minimum reconstruction error value of the segment; the dynamic health benchmark model corresponding to the minimum reconstruction error value is the dynamic health benchmark model corresponding to the segment.

8. The method according to claim 7, characterized in that, The method of predicting the future trajectory of health degradation to obtain the remaining lifespan includes: Historical data segments with slopes exceeding a threshold are extracted from the health degradation trajectory, and the curve of health index changing over time is obtained by fitting an exponential decay model; the fitting objective is to minimize the mean square error. By extending the time axis towards the future, we obtain the future health index prediction curve. Solve for the time coordinates of the intersection point between the predicted future health index curve and the preset failure health threshold; The difference between the current time point and the time coordinate is the remaining service life.

Images