Transfer learning prediction method and system for residual life of hydraulic system

Abstract

The invention provides a transfer learning prediction method and system for the residual life of a hydraulic system, and the method comprises the following steps: obtaining the measured hydraulic data of a field device and the full-life-cycle data of a test bench, and extracting the power spectrum density characteristics of the measured hydraulic data and the full-life-cycle data; and calculating a mapping conversion matrix by comparing spectrum features of a specific frequency window, aligning source domain features to a target domain, and generating a migration feature set. And in an evolution state space formed by fluid thermodynamic entropy increase and vibration energy level, constructing a standard degradation track by using the migration feature set. And mapping real-time data to the space, and deducing a time distance to a failure boundary in combination with a tropism weight of the real-time data to obtain a life prediction result. According to the method, the technical problem that a prediction model based on full-life data training is difficult to be directly applied to field equipment due to the difference between the laboratory working condition and the field actual working condition is solved.

Description

Technical Field

[0001] This invention belongs to the field of equipment management technology, specifically relating to a transfer learning prediction method and system for the remaining life of a hydraulic system. Background Technology

[0002] In extreme marine engineering operations such as deep-sea saturation diving, the heave compensation hydraulic system is a critical system for ensuring the safety of personnel and equipment. The core power component of this system, the hydraulic pump, operates under unstable, highly dynamic load conditions for extended periods, and the reliability of its performance directly impacts the success or failure of the entire operation. To avoid catastrophic accidents caused by sudden hydraulic pump failures, predictive maintenance is necessary. Currently, predictive maintenance primarily employs remaining life prediction methods. Specifically, this involves conducting accelerated aging tests on hydraulic pumps of the same model on a laboratory hydraulic test bench, covering their entire lifecycle from new to failure. Complete degradation data is collected, and based on this data, a deep learning or machine learning model is trained to characterize the evolution of the pump's health status.

[0003] However, the load spectrum applied in accelerated aging tests in the laboratory is usually constant or periodically varying, while the hydraulic pumps on a saturation diving support vessel are subjected to random and non-stationary loads under real harsh sea conditions. The spectral characteristics of actual pressure pulsations and vibration signals differ significantly from laboratory data. This inconsistency in data distribution caused by differences in operating conditions can lead to a sharp decline in the performance of models trained in the laboratory when applied to real-world conditions, resulting in huge prediction biases and even completely erroneous conclusions. Summary of the Invention

[0004] This invention provides a transfer learning prediction method and system for the remaining life of a hydraulic system to solve the above-mentioned technical problems.

[0005] In a first aspect, the present invention provides a transfer learning prediction method for the remaining life of a hydraulic system, the method comprising the following steps: Real-time hydraulic measurement data of the target domain is read from the control system of the on-site hydraulic equipment; Hydraulic test data for the entire life cycle of the source domain is acquired using a hydraulic test bench. The hydraulic measured data and hydraulic test data are then converted into two sets of power spectral density characteristic sequences by time-frequency conversion. The hydraulic test bench has the same hydraulic circuit design as the field hydraulic equipment. The characteristic frequency window is divided according to the mechanical structure characteristics of the hydraulic components in the hydraulic equipment on site. The spectral characteristics of the two sets of power spectral density characteristic sequences are compared based on the characteristic frequency window and the mapping transformation matrix is ​​calculated. The mapping transformation matrix is ​​applied to the power spectral density characteristic sequence of the hydraulic test data to generate a migration feature set aligned with the target domain distribution. Based on hydraulic test data, fluid thermodynamic entropy increase index and vibration energy level index are calculated. An evolutionary state space is constructed with fluid thermodynamic entropy increase index and vibration energy level index as the reference dimensions. The migration feature set is mapped to the evolutionary state space to form a standard degradation trajectory. Risk singularity regions representing different failure modes are pre-set in the evolutionary state space. Hydraulic measured data is mapped into the evolutionary state space and the position and evolutionary trend vector of the target data point in the evolutionary state space at the current moment are calculated in real time. The directional weight is calculated based on the degree to which the evolutionary trend vector points to the risk singularity region. After using the power spectral density feature sequence of hydraulic measured data weighted by directional weights, the time distance from the target data point to the failure state region in the evolution state space is deduced, and the life prediction result of the hydraulic equipment in the field is generated based on the deduction result.

[0006] Optionally, the step of comparing the spectral features of two sets of power spectral density feature sequences based on a feature frequency window and calculating a mapping transformation matrix, and then applying the mapping transformation matrix to the power spectral density feature sequences of the hydraulic test data to generate a migration feature set aligned with the target domain distribution, includes the following steps: Source domain data segments that match the system load pressure level under stable operating conditions in the target domain are selected from the hydraulic test data. Time-frequency transformations were performed on the source domain data segments and the measured hydraulic data under steady-state conditions in the target domain, and the source domain power spectrum vector and the target domain power spectrum vector within the characteristic frequency window were extracted. By combining the power spectral vectors of the source domain and the target domain, a mapping optimization equation is constructed. The gradient descent method is used to solve the mapping optimization equation, and a mapping transformation matrix representing the mapping relationship from the feature space of the source domain to the feature space of the target domain is obtained. Multiplying the mapping transformation matrix on the left by the power spectral density feature sequence of the hydraulic test data yields a migration feature set that conforms to the data distribution characteristics of the target domain.

[0007] Optionally, the step of filtering source domain data segments from hydraulic test data that are consistent with the system load pressure level under steady-state conditions in the target domain includes the following steps: By monitoring the system pressure and flow parameters of the hydraulic measured data within a preset time window, the pressure fluctuation variance and flow fluctuation variance are calculated respectively. When both the pressure fluctuation variance and the flow fluctuation variance are less than the preset stability judgment threshold, the preset time window is marked as the stable operating condition window, and the average system pressure value and average flow value calculated within the stable operating condition window are used as the operating condition matching benchmark. Calculate the Euclidean distance between the source domain pressure and source domain flow values ​​at each sampling time of the hydraulic test data and the operating condition matching benchmark; The hydraulic test data were sorted in ascending order of Euclidean distance, and the first few data segments with the smallest Euclidean distance were selected as the source domain data segments.

[0008] Optionally, the step of performing time-frequency transformation on the source domain data segment and the measured hydraulic data under steady-state conditions in the target domain, and extracting the source domain power spectrum vector and the target domain power spectrum vector within the characteristic frequency window, includes the following steps: Calculate the spectral energy concentration of the source domain data segment within the initial characteristic frequency window; Using the gradient descent search algorithm and aiming to maximize the spectral energy concentration, the left and right boundary frequencies of the feature frequency window are iteratively adjusted to determine the optimal feature frequency window; Within the optimal characteristic frequency window, the hydraulic measured data within the source domain data segment and the steady-state operating condition window are discretized and sampled. The discretized spectrum amplitude points are arranged in frequency order to generate the source domain power spectrum vector and the target domain power spectrum vector.

[0009] Optionally, the hydraulic test data includes inlet oil temperature, outlet oil temperature, inlet pressure, outlet pressure, and flow parameters. The steps of calculating the fluid thermodynamic entropy increase index and vibration energy level index based on the hydraulic test data, constructing an evolutionary state space using the fluid thermodynamic entropy increase index and vibration energy level index as reference dimensions, and mapping the migration feature set to the evolutionary state space to form a standard degradation trajectory include the following steps: Based on hydraulic test data and the enthalpy difference between the inlet and outlet of the hydraulic test bench calculated according to the fluid thermodynamic equation, the total power loss of the hydraulic test bench is calculated based on the flow parameters. The fluid thermodynamic entropy production rate is obtained by dividing the total power loss by the average thermodynamic temperature of the hydraulic test bench, and the fluid thermodynamic entropy production rate is normalized to obtain the fluid thermodynamic entropy increase index. The vibration energy level index is calculated by integrating the power spectral density feature sequence in the migration feature set across the entire frequency band. A two-dimensional coordinate system is established with the fluid thermodynamic entropy increase index as the horizontal axis and the vibrational energy level index as the vertical axis as the evolution state space. The migration feature set is projected as coordinate points onto the evolutionary state space according to the sampling time of the hydraulic test data; The projected coordinate points are smoothed using a polynomial fitting algorithm to generate a standard degenerate trajectory that connects the healthy state region and the failed state region in the evolutionary state space.

[0010] Optionally, the step of mapping hydraulic measured data to the evolutionary state space and calculating the position and evolutionary trend vector of the target data point in the evolutionary state space at the current moment in real time, and calculating the directional weight based on the degree to which the evolutionary trend vector points to the risk singularity region includes the following steps: Based on the hydraulic measured data, the fluid thermodynamic entropy increase index and vibration energy level index at each sampling moment in the target domain are calculated, and the coordinate positions of the corresponding data points in the evolution state space at each sampling moment in the target domain are determined. Based on the change in coordinate position between the current time and the previous time in the target domain, calculate the instantaneous evolution velocity vector of the target data point corresponding to the current time in the target domain; Construct a risk pointing vector that connects the coordinates of the target data point with the coordinates of the risk singularity in the risk singularity region; Calculate the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector; An exponential gravity function is constructed with cosine similarity and Euclidean distance from the target data point to the risk singularity as independent variables, and the output value of the exponential gravity function is used as the directional weight.

[0011] Optionally, the step of using the power spectral density feature sequence of the hydraulic measured data with directional weights to deduce the time distance from the target data point to the failure state region in the evolution state space includes the following steps: The power spectral density feature sequence in the migration feature set is multiplied by the directional weight to generate a weighted fault-sensitive feature vector. The fault-sensitive feature vector is input into the preset lifetime prediction model, and the degradation-dependent features on the time series are extracted through the long short-term memory network layer in the lifetime prediction model. The degradation-dependent features are input into the fully connected regression layer in the lifespan prediction model, and the predicted health value is output through the fully connected regression layer. Obtain the health threshold of the standard degradation trajectory in the failure state region of the evolution state space, and calculate the time distance required for the health value to drop to the health threshold by linear extrapolation.

[0012] Optionally, the construction of an exponential gravity function with cosine similarity and the Euclidean distance from the target data point to the risk singularity as independent variables, and the use of the output value of the exponential gravity function as the directional weight, includes the following steps: Set the gravitational constant parameter, distance decay factor parameter, and direction sensitivity parameter; Calculate the reciprocal of the Euclidean distance from the target data point to the risk singularity, and use the distance decay factor parameter to perform a power operation on the reciprocal of the Euclidean distance to obtain the distance influence term; The direction influence term is obtained by nonlinearly amplifying the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector using the direction sensitivity parameter; Multiply the distance effect term and the direction effect term, and then multiply by the gravitational constant parameter to obtain the fundamental gravitational value; The Sigmoid activation function is introduced to map the basic gravity value to a preset weight range, and the value mapped to the weight range is used as the directional weight.

[0013] In a second aspect, the present invention also provides a transfer learning prediction system for the remaining life of a hydraulic system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the transfer learning prediction method for the remaining life of a hydraulic system as described in any one of the first aspects.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing instructions, characterized in that, when executed by a processor, the instructions cause the processor to be configured to perform a transfer learning prediction method for the remaining life of a hydraulic system according to any one of the first aspects.

[0015] The beneficial effects of this invention are: This invention establishes a measurement benchmark for the health degradation process of hydraulic pumps by constructing an evolutionary state space based on fluid thermodynamic entropy increase and vibration energy level indices. Then, using characteristic frequency windows divided based on the mechanical structural characteristics of hydraulic components, laboratory test data and field measurement data are compared in the frequency domain, and a mapping transformation matrix is ​​calculated. This mapping transformation matrix effectively eliminates the inconsistency in characteristic distribution caused by the difference between ideal laboratory conditions and real deep-sea operation conditions, thus successfully transferring and applying the complete degradation laws obtained in the laboratory to specific field equipment. Furthermore, by introducing directional weights pointing to the singularity region of specific failure mode risk, the prediction model can be dynamically adjusted according to the current degradation trend during the prediction process, thereby significantly improving the life prediction accuracy of the model trained in the laboratory under actual operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a transfer learning prediction method for the remaining life of a hydraulic system in one embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the evolutionary state space in one embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0020] Figure 1 This is a flowchart illustrating a transfer learning prediction method for the remaining life of a hydraulic system in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 1 As shown, the transfer learning prediction method for the remaining life of a hydraulic system disclosed in this invention specifically includes the following steps: S101. Read the measured hydraulic data of the target domain in real time from the control system of the on-site hydraulic equipment.

[0021] Among these features, the high-frequency data acquisition interface deployed in the field hydraulic equipment control system can continuously capture multi-dimensional time-series signals, including hydraulic oil pressure, flow rate, temperature, and actuator vibration acceleration. These signals accurately reflect the operating status of the equipment under complex working conditions. Simultaneously, using a hydraulic test bench with a physical structure and hydraulic circuit design identical to the field equipment, accelerated degradation tests covering the entire life cycle from a healthy state to a complete failure state are conducted to obtain comprehensive source-domain hydraulic test data.

[0022] S102. Use a hydraulic test bench to acquire hydraulic test data for the entire life cycle of the source domain, and perform time-frequency conversion on the hydraulic measured data and hydraulic test data respectively to obtain two sets of power spectral density characteristic sequences.

[0023] To uncover the periodic degradation patterns hidden beneath the chaotic time-domain waveforms, time-frequency domain transformations need to be performed on the two sets of acquired raw time-series data. Typically, Fast Fourier Transform (FFT) or Wavelet Transform (WFT) techniques are used to deconstruct the time-domain signal into its energy distribution in the frequency domain, calculating the power spectral density (PSD) characteristic sequence. PSD quantifies the power intensity of the signal at different frequency components, effectively revealing specific frequency band energy changes caused by faults such as hydraulic pump wear, valve core jamming, or seal failure.

[0024] S103. Divide the characteristic frequency window according to the mechanical structure characteristics of the hydraulic components in the on-site hydraulic equipment. Based on the characteristic frequency window, compare the spectral characteristics of the two sets of power spectral density characteristic sequences and calculate the mapping transformation matrix. Apply the mapping transformation matrix to the power spectral density characteristic sequence of the hydraulic test data to generate a migration feature set aligned with the target domain distribution.

[0025] After acquiring the basic frequency domain features, based on the inherent mechanical structural characteristics of hydraulic components, such as the number of pump plungers, rotational speed frequency, and their harmonics, a feature frequency window containing key fault information is defined. Within this window, the spectral morphology differences between two sets of power spectral density feature sequences are compared and analyzed. In the specific implementation process, firstly, the target domain stable operating condition data and corresponding source domain data segments are screened by monitoring the fluctuation variance of system pressure and flow. Then, the frequency window boundary is dynamically adjusted using a gradient descent algorithm with the goal of maximizing spectral energy concentration. Based on this, a mapping optimization equation is constructed and solved to obtain a mapping transformation matrix that maps the source domain feature space to the target domain feature space. The working mechanism of this matrix is ​​similar to a linear transformation operator; multiplying it on the left by the power spectral density feature sequence of the hydraulic test data generates a transfer feature set. This not only achieves data alignment at the feature level, enabling the use of full-lifecycle labeled data acquired in a laboratory environment to train predictive models suitable for field equipment, but also largely preserves the degradation trend information in the original data, effectively overcoming the problem of scarce samples due to the lack of full-lifecycle fault data for field equipment, and significantly improving the matching accuracy of cross-domain features.

[0026] S104. Based on hydraulic test data, the fluid thermodynamic entropy increase index and vibration energy level index are calculated. The evolutionary state space is constructed using the fluid thermodynamic entropy increase index and vibration energy level index as the reference dimensions, and the migration feature set is mapped to the evolutionary state space to form a standard degradation trajectory.

[0027] In this study, the fluid thermodynamic entropy increase index and the vibration energy level index were selected as two orthogonal reference dimensions. Fluid thermodynamic entropy increase reflects irreversible energy dissipation and efficiency reduction caused by leakage and friction within the system, while the vibration energy level directly reflects the structural loosening and surface wear of mechanical components. In the specific calculations, based on the fluid thermodynamic equations, the fluid thermodynamic entropy production rate was calculated using the total power loss and average thermodynamic temperature of the hydraulic test bench, and the entropy increase index was obtained after normalization. Simultaneously, the vibration energy level index was obtained by full-band integration of the power spectral density of the migration feature set. (Refer to...) Figure 2 A two-dimensional coordinate system is constructed with the fluid thermodynamic entropy increase index as the abscissa and the vibration energy level index as the ordinate. The migration feature set of the whole life cycle is projected into it in chronological order, and the scattered points are smoothed by a polynomial fitting algorithm to form a standard degradation trajectory connecting the healthy region to the failure region.

[0028] S105. Map the hydraulic measured data to the evolutionary state space and calculate the position of the target data point in the evolutionary state space and the evolutionary trend vector at the current moment in real time. Calculate the directional weight based on the degree to which the evolutionary trend vector points to the risk singularity region.

[0029] This process involves mapping real-time, pre-processed hydraulic measurement data onto an evolutionary state space to determine the precise coordinates of the target data point at the current moment. The focus shifts beyond isolated point locations to analyzing the dynamic changes of the data point over time. By calculating the difference between the current and previous coordinate positions, an instantaneous evolution velocity vector characterizing the rate and direction of equipment performance degradation is derived. Simultaneously, a risk pointing vector is constructed connecting the current target data point to the center of a pre-defined risk singularity region in space. Vector analysis is then used to calculate the cosine of the angle between the instantaneous evolution velocity vector and the risk pointing vector, quantifying the degree to which the current degradation trend points to a specific failure mode. To ensure the life prediction model focuses on the most pressing failure risk, directional weights are calculated based on the evolution trend, and features are weighted accordingly. The specific calculation logic is as follows: First, the Euclidean distance from the target data point to the risk singularity is calculated, and the cosine similarity between the evolution velocity vector and the risk pointing vector is obtained. A weight calculation formula is then constructed using a distance decay factor and a direction sensitivity parameter. The larger the calculated trend weight value, the closer the current state of the equipment is to a certain failure mode and the more consistent the evolution trend is. Subsequently, this weight is multiplied by the power spectral density feature sequence of the hydraulic measured data to generate a weighted fault-sensitive feature vector.

[0030] S106. After using the power spectral density feature sequence of hydraulic measured data with directional weights, the time distance from the target data point to the failure state region in the evolution state space is deduced, and the life prediction result of the hydraulic equipment in the field is generated based on the deduction result.

[0031] The fault-sensitive feature vector, after being weighted by directional factors, is input into a pre-trained life prediction model. This model includes a Long Short-Term Memory (LSTM) network layer, which effectively extracts long-term degradation dependencies from time-series features and remembers the impact of historical states on the current state. The degradation dependency features extracted by the LSTM layer then enter a fully connected regression layer, directly outputting the predicted health value of the equipment at the current moment. To obtain the specific remaining lifespan, it is necessary to combine the health threshold of the standard degradation trajectory in the failure state region in the evolutionary state space. Using linear extrapolation or gradient-based extrapolation algorithms, the time span required for the health curve to decline to the threshold is extrapolated using the current predicted health value and the historical degradation rate, ultimately generating the life prediction result for the hydraulic equipment in the field.

[0032] In one embodiment, the process of comparing the spectral features of two sets of power spectral density feature sequences based on a characteristic frequency window and calculating a mapping transformation matrix, and then applying the mapping transformation matrix to the power spectral density feature sequences of hydraulic test data to generate a migration feature set aligned with the target domain distribution, includes the following steps: Source domain data segments that match the system load pressure level under stable operating conditions in the target domain are selected from the hydraulic test data. Time-frequency transformations were performed on the source domain data segments and the measured hydraulic data under steady-state conditions in the target domain, and the source domain power spectrum vector and the target domain power spectrum vector within the characteristic frequency window were extracted. By combining the power spectral vectors of the source domain and the target domain, a mapping optimization equation is constructed. The gradient descent method is used to solve the mapping optimization equation, and a mapping transformation matrix representing the mapping relationship from the feature space of the source domain to the feature space of the target domain is obtained. Multiplying the mapping transformation matrix on the left by the power spectral density feature sequence of the hydraulic test data yields a migration feature set that conforms to the data distribution characteristics of the target domain.

[0033] In this embodiment, a sliding time window is set in the measured data stream of the hydraulic equipment in operation, and the system pressure and flow parameters within the window are monitored in real time. The degree of data fluctuation is quantified by calculating the variance index. When both the pressure variance and the flow variance are lower than the preset stability judgment threshold, the window is considered to be in a stable operating condition, and the average system pressure within the window is calculated. and average flow This serves as the benchmark for operating condition matching. Subsequently, the hydraulic test data throughout the entire life cycle of the source domain is traversed, and the corresponding source domain pressure value is extracted for each sampling moment. and source domain traffic value Using the Euclidean distance formula Calculate the difference between the source domain operating conditions at each time point and the field matching benchmark. Sort the source domain data in ascending order based on the calculated distance values, and select the first few data segments with the smallest distance as source domain data segments that are consistent with the target domain operating conditions.

[0034] During operation, wear or failure of internal components of a hydraulic system, such as piston pumps and hydraulic valves, is often accompanied by changes in vibration energy at specific frequencies. Time-frequency transformation can deconstruct the time-domain signal into a frequency distribution, thereby highlighting these fault characteristics. In practice, a Fast Fourier Transform (FFT) is used to process source domain data segments and target domain steady-state operating condition data. A characteristic frequency window is pre-defined based on the kinematic parameters of the hydraulic components (such as rotational speed and number of pistons). This window covers the fault characteristic frequencies and their main harmonic components. From the transformed spectrum, amplitude data within this characteristic frequency window is extracted to construct source domain power spectrum vectors. and target domain power spectrum vector These two vectors directly quantify the energy distribution within the key frequency band, eliminating background noise interference across the entire frequency band.

[0035] In this embodiment, a mapping optimization equation is constructed by combining the source domain power spectral vector and the target domain power spectral vector. The mapping optimization equation is then solved using the gradient descent method to obtain the mapping transformation matrix that characterizes the mapping relationship from the source domain feature space to the target domain feature space. The steps include the following: A k-nearest neighbor similarity graph is constructed for the source domain power spectral vector set. The adjacency weight matrix and degree matrix of the k-nearest neighbor similarity graph are calculated. The Laplacian matrix of the spectral graph is obtained by subtracting the adjacency weight matrix from the degree matrix. A characteristic frequency sensitivity diagonal matrix is ​​constructed based on the distribution of the characteristic frequency window. The diagonal elements corresponding to the frequency points within the characteristic frequency window are assigned high sensitivity weight values, and the diagonal elements corresponding to the frequency points outside the characteristic frequency window are assigned low background weight values. A graph regularized weighted mapping optimization equation is constructed, which includes a reconstruction error term, a manifold consistency constraint term, and a sparsity regularization term. The specific mathematical expression of the graph regularized weighted mapping optimization equation is as follows:

[0036] In the formula, W is the mapping transformation matrix to be solved; This indicates that the objective function is minimized for matrix W; The source domain data matrix is ​​composed of source domain power spectrum vectors; is the target domain data matrix composed of target domain power spectrum vectors; A is the characteristic frequency sensitivity diagonal matrix, used to weight the reconstruction error in terms of frequency dimension; L is the spectral Laplacian matrix, used to constrain the consistency of the manifold structure of the source domain data before and after mapping; Represents the trace operation of a matrix; Denotes the Frobenius norm of a matrix; For the regularization parameter of the manifold consistency constraint term; is the regularization parameter for the sparsity regularization term.

[0037] In the graph-regularized weighted mapping optimization equation, the reconstruction error between the predicted value (after multiplying the source domain power spectral vector by the mapping transformation matrix) and the target domain power spectral vector is weighted using a characteristic frequency sensitivity diagonal matrix to form a reconstruction error term. The local geometric structure difference of the source domain power spectral vector before and after the mapping transformation matrix transformation is calculated using the spectral graph Laplacian matrix to form a manifold consistency constraint term, ensuring that the neighborhood topology of the source domain data remains unchanged during the migration process. The graph-regularized weighted mapping optimization equation is solved using the alternating direction multiplier method or gradient descent algorithm to obtain the mapping transformation matrix that minimizes the weighted sum of the reconstruction error term, the manifold consistency constraint term, and the sparsity regularization term.

[0038] Finally, the complete power spectral density feature sequence matrix obtained after time-frequency transformation of the source domain's full-life-cycle hydraulic test data is extracted. Following matrix multiplication rules, the mapping transformation matrix is ​​multiplied by the source domain feature sequence matrix. Specifically, left or right multiplication is performed depending on the data dimension arrangement to meet linear transformation requirements. The result is a transfer feature set aligned with the target domain distribution. Thus, data originally belonging to different distribution spaces are unified under the same metric system, enabling the rich remaining life label information carried in the source domain data to be directly transferred to field equipment lacking failure samples.

[0039] In one implementation, selecting source domain data segments from hydraulic test data that correspond to the system load pressure level under steady-state conditions in the target domain includes the following steps: By monitoring the system pressure and flow parameters of the hydraulic measured data within a preset time window, the pressure fluctuation variance and flow fluctuation variance are calculated respectively. When both the pressure fluctuation variance and the flow fluctuation variance are less than the preset stability judgment threshold, the preset time window is marked as the stable operating condition window, and the average system pressure value and average flow value calculated within the stable operating condition window are used as the operating condition matching benchmark. Calculate the Euclidean distance between the source domain pressure and source domain flow values ​​at each sampling time of the hydraulic test data and the operating condition matching benchmark; The hydraulic test data were sorted in ascending order of Euclidean distance, and the first few data segments with the smallest Euclidean distance were selected as the source domain data segments.

[0040] In this embodiment, the latest pressure and flow data segments are captured by sliding a preset time window along the time axis. For each captured time window, the dispersion of the data within it, i.e., the variance of the fluctuation, is calculated using statistical methods. Specifically, for the N sampling points contained within the window, the pressure measurement sequence P(i) and the flow measurement sequence Q(i) are extracted respectively. First, the arithmetic mean of the data within the window is calculated as the central tendency. Then, the mean of the sum of squares of the deviations of each sampling point from the mean is calculated, thereby obtaining the pressure fluctuation variance. and flow fluctuation variance The calculation formula can be expressed as follows: ,in This represents the average pressure within the window. By updating these two variance indices in real time, the complex hydraulic dynamics can be transformed into quantifiable stability values.

[0041] The pressure fluctuation variance and flow fluctuation variance are respectively compared with the pre-set stability judgment threshold. and A comparison is performed. These two thresholds are determined based on the rated parameters of the hydraulic system and the allowable fluctuation tolerance range, representing the maximum permissible fluctuation limit of the system under steady-state operation. Only when the conditions are met... and When both parameters are simultaneously valid—meaning their fluctuations are limited to a small, permissible range—the current time window is officially designated as a stable operating condition window. Once the window is confirmed to be in a stable state, it indicates that the system is in a constant load or steady-state operation phase, at which point the statistical characteristics of the data within the window are highly representative. Subsequently, all sampled data within this stable operating condition window are averaged to calculate the average system pressure value. and average flow value .

[0042] Since the source-domain hydraulic test bench may experience various load levels and flow settings during accelerated life testing, it is necessary to evaluate the similarity between the source-domain data and the target-domain benchmark through traversal calculations. For each sampling time k in the hydraulic test data, the corresponding source-domain pressure value is extracted. and source domain traffic value To quantify the differences between the two source domain parameters and the target domain operating condition benchmark, Euclidean distance is used as the metric. A two-dimensional feature space is constructed with pressure and flow rate as orthogonal coordinate axes, and the straight-line distance between the source domain data points and the target domain benchmark points is calculated. The calculation formula is expressed as follows: This formula integrates pressure and flow differences into a single scalar indicator, intuitively reflecting the deviation between the operating conditions of the source domain at a certain moment and the actual operating conditions on site. Since the calculated Euclidean distance set directly reflects the degree of matching between the data from the source domain at each moment and the current operating conditions of the target domain, the smaller the distance value, the closer the corresponding source domain data is to the actual operating state of the equipment on site in terms of physical conditions. Therefore, the sequence containing distance values ​​at all sampling moments is sorted in ascending order, that is, the source domain data index is reorganized according to the order of distance from smallest to largest. Based on the sorting result, several data items at the beginning of the sequence are extracted, and the continuous data segments associated with the time points corresponding to these minimum distances are extracted as the finally selected source domain data segments.

[0043] In one embodiment, performing time-frequency transformation on the source domain data segment and the measured hydraulic data under steady-state conditions in the target domain, and extracting the source domain power spectrum vector and the target domain power spectrum vector within the characteristic frequency window, includes the following steps: Calculate the spectral energy concentration of the source domain data segment within the initial characteristic frequency window; Using the gradient descent search algorithm and aiming to maximize the spectral energy concentration, the left and right boundary frequencies of the feature frequency window are iteratively adjusted to determine the optimal feature frequency window; Within the optimal characteristic frequency window, the hydraulic measured data within the source domain data segment and the steady-state operating condition window are discretized and sampled. The discretized spectrum amplitude points are arranged in frequency order to generate the source domain power spectrum vector and the target domain power spectrum vector.

[0044] In this embodiment, because the rotational frequency of mechanical components in hydraulic equipment often drifts due to load fluctuations or changes in hydraulic oil viscosity during actual operation, the initial window defined solely by theoretical formulas may not accurately cover the peak values ​​of fault characteristics. Therefore, high-resolution power spectral density analysis is first performed on the selected source domain data segments to convert the time-domain vibration or pressure signal into a frequency-domain energy distribution map. Based on this, initial upper and lower frequency limits are set according to the mechanical kinematic parameters of the hydraulic pump or motor, and the power spectral density curve is integrated within this range to obtain the total energy value within the window. To eliminate the influence of the window width itself on the total energy value, the integration result needs to be divided by the bandwidth of the frequency window to obtain the average energy density per unit frequency, i.e., the spectral energy concentration. Let S(f) be the power spectral density function of the source domain data, and the lower limit frequency of the initial window be... The upper limit frequency is The formula for calculating the spectral energy concentration is defined as follows: A high concentration value means that the window not only captures the main fault energy but also has a high signal-to-noise ratio; conversely, a low concentration value indicates that the window position is off or contains too much background noise.

[0045] After determining that maximizing the spectral energy concentration is the optimization objective, the left and right boundaries of the characteristic frequency window are considered as independent variables to be optimized. Since the energy concentration function typically exhibits nonlinear convex characteristics with respect to the frequency boundaries, a gradient descent algorithm is suitable for solving it. During algorithm initialization, starting from the theoretically calculated initial boundaries, the partial derivatives of the energy concentration function with respect to the left and right boundaries are calculated respectively. These partial derivatives indicate the direction of the fastest increase in energy density. Based on the calculated gradient direction and the preset learning rate, the values ​​of the left and right boundaries are gradually updated, causing the window to move or expand towards the more energy-dense region. The boundary update formula can be expressed as: ,in This is a parameter vector containing the frequencies of the left and right boundaries. To control the learning rate of the iteration step, This represents the gradient operator. As the number of iterations increases, the window boundary gradually approaches the edge of the actual fault characteristic frequency band until the change in energy concentration is less than the set convergence threshold. Within the locked optimal frequency window, the source domain data segment and the measured target domain data under stable operating conditions are simultaneously resampled. N equally spaced frequency points are selected within the window, and the power spectral density amplitudes corresponding to these frequency points are extracted respectively. Since the source domain and target domain data may have different original sampling rates or data lengths, this equally spaced sampling within a unified physical frequency band forces the two sets of data to maintain strict consistency in the feature dimension. The extracted N amplitude points are arranged strictly in ascending order of frequency to construct a feature representation in the form of a column vector.

[0046] In one embodiment, the hydraulic test data includes inlet oil temperature, outlet oil temperature, inlet pressure, outlet pressure, and flow parameters. Based on the hydraulic test data, fluid thermodynamic entropy increase index and vibration energy level index are calculated. An evolutionary state space is constructed using the fluid thermodynamic entropy increase index and vibration energy level index as the baseline dimensions. Mapping the migration feature set into the evolutionary state space to form a standard degradation trajectory includes the following steps: Based on hydraulic test data and the enthalpy difference between the inlet and outlet of the hydraulic test bench calculated according to the fluid thermodynamic equation, the total power loss of the hydraulic test bench is calculated based on the flow parameters. The fluid thermodynamic entropy production rate is obtained by dividing the total power loss by the average thermodynamic temperature of the hydraulic test bench, and the fluid thermodynamic entropy production rate is normalized to obtain the fluid thermodynamic entropy increase index. The vibration energy level index is calculated by integrating the power spectral density feature sequence in the migration feature set across the entire frequency band. A two-dimensional coordinate system is established with the fluid thermodynamic entropy increase index as the horizontal axis and the vibrational energy level index as the vertical axis as the evolution state space. The migration feature set is projected as coordinate points onto the evolutionary state space according to the sampling time of the hydraulic test data; The projected coordinate points are smoothed using a polynomial fitting algorithm to generate a standard degenerate trajectory that connects the healthy state region and the failed state region in the evolutionary state space.

[0047] In this embodiment, high-precision sensors deployed at the inlet and outlet of the hydraulic test bench are used to collect the oil pressure in real time. and temperature Data. Based on the fluid thermodynamic equation of state, combined with the specific heat capacity of the oil... Using physical properties such as the coefficient of thermal expansion, the specific enthalpy difference between the inlet and outlet is calculated. The specific enthalpy difference reflects the change in energy state of a unit mass of fluid flowing through the measured element. The calculation must consider the combined effect of pressure work and internal energy change. The formula can be approximated as follows: ,in Let be the fluid density. Then, the calculated specific enthalpy difference is multiplied by the real-time monitored mass flow rate to obtain the total power loss of the system at that moment, thus realizing the transformation of the invisible mechanical wear and volumetric efficiency reduction inside the hydraulic system into a directly measurable macroscopic thermodynamic energy index.

[0048] According to the second law of thermodynamics, any irreversible process will lead to the generation of entropy, and the entropy production rate is directly related to the degradation rate of the system. This is in addition to obtaining the total power loss. Next, the reference temperature of the system needs to be determined. Usually, the arithmetic mean of the inlet and outlet temperatures is taken as the average thermodynamic temperature. Using formulas The thermodynamic entropy production rate of a fluid is calculated, which quantifies the increase in entropy of a system per unit time due to irreversible losses. However, due to the significant differences in rated power and flow rate among different specifications of hydraulic components, it is difficult to make cross-equipment comparisons using the entropy production rate directly. Therefore, normalization is necessary. The calculated entropy production rate is divided by the maximum theoretical entropy production rate under rated operating conditions or the baseline entropy production rate under initial healthy conditions to obtain a dimensionless thermodynamic entropy increase index, thus confining its numerical range to [0,1].

[0049] Mechanical faults in hydraulic systems, such as bearing spalling, gear pitting, or rotor imbalance, directly manifest as increased vibration intensity. In the preceding steps, a transfer feature set aligned with the target domain distribution has been obtained through transfer learning, containing a power spectral density feature sequence over the entire lifecycle. The area under the power spectral density curve represents the total power of the signal, i.e., the vibration energy. Therefore, numerical integration is performed on each power spectral density vector in the transfer feature set across the entire frequency band. Let the power spectral density function be P(f), and the frequency range be... The formula for calculating the vibration energy level index is: In the case of discrete data, this integral is transformed into the summation of the amplitudes at each frequency point, thereby aggregating fault information from various frequency bands. Whether the fault characteristics appear in the low-frequency fluid pulsation region or the high-frequency mechanical meshing region, they can be captured by this index.

[0050] Next, a fluid thermodynamic entropy increase index is selected as the horizontal axis to represent the efficiency degradation dimension of the system; a vibration energy level index is selected as the vertical axis to represent the mechanical vibration degradation dimension of the system. A two-dimensional Cartesian coordinate system is constructed using these two indices, which are physically independent but interconnected in fault evolution. In this space, the area near the origin typically corresponds to the healthy state of the equipment (low entropy production, low vibration), while regions far from the origin correspond to varying degrees of degradation. Finally, according to the timestamp order of the hydraulic test data, the fluid thermodynamic entropy increase index and vibration energy level index corresponding to each sampling moment are extracted sequentially, combined into two-dimensional coordinate points, and plotted in the evolutionary state space. Since the original measurement data is often affected by random noise, the directly projected points are usually discrete and jittery, making them difficult to use directly for trend prediction. Therefore, a polynomial fitting algorithm is used to smooth these scattered points. This algorithm finds an optimal curve function that minimizes the sum of squared distances between the curve and all data points. The smoothed curve obtained by fitting connects the region representing the initial healthy state and the region representing the final failure state, which is the standard degradation trajectory. This trajectory reveals the average patterns and main trends in the evolution of performance indicators of this type of hydraulic equipment over its entire life cycle.

[0051] In one implementation, mapping hydraulic measurement data into an evolutionary state space and calculating the position and evolutionary trend vector of the target data point in the evolutionary state space at the current moment in real time, and calculating the directional weight based on the degree to which the evolutionary trend vector points to the risk singularity region, includes the following steps: Based on the hydraulic measured data, the fluid thermodynamic entropy increase index and vibration energy level index at each sampling moment in the target domain are calculated, and the coordinate positions of the corresponding data points in the evolution state space at each sampling moment in the target domain are determined. Based on the change in coordinate position between the current time and the previous time in the target domain, calculate the instantaneous evolution velocity vector of the target data point corresponding to the current time in the target domain; Construct a risk pointing vector that connects the coordinates of the target data point with the coordinates of the risk singularity in the risk singularity region; Calculate the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector; An exponential gravity function is constructed with cosine similarity and Euclidean distance from the target data point to the risk singularity as independent variables, and the output value of the exponential gravity function is used as the directional weight.

[0052] In this embodiment, the current fluid thermodynamic entropy increase index is calculated based on the inlet and outlet pressure, temperature, and flow rate data collected on-site. Simultaneously, time-frequency conversion and full-frequency integration are performed on the on-site vibration or pressure pulsation signals to calculate the vibration energy level index. These two calculated values ​​constitute the state coordinates at the current time t. The coordinate point is then mapped to a pre-constructed two-dimensional evolutionary state space. This process is essentially a real-time positioning operation for the equipment's health status; each coordinate point precisely reflects the equipment's efficiency level and mechanical integrity at the current instant. By continuously calculating and mapping these coordinate points, an actual operating trajectory of the field equipment can be drawn in the state space. Then, the coordinate position of the target domain at the current time t is extracted. Coordinates of the previous sampling time t-1 The instantaneous evolution velocity vector is defined as the displacement vector between these two points, and the calculation formula is: .

[0053] This vector contains not only information about the magnitude of the velocity, indicating the severity of degradation, but also directional information, indicating the trend of changes in the equipment's performance parameters. For example, a vector pointing to the upper right corner means that vibration and entropy production are increasing simultaneously, and the equipment is deteriorating comprehensively; if it points horizontally, it may only indicate a decrease in efficiency while the mechanical structure is temporarily stable. In the evolutionary state space, several risk singularity regions representing different failure modes (such as pump shaft breakage or complete seal failure) are pre-defined, each region consisting of a central coordinate point. Identification. For the target data point at the current time. $, to build a connection The vector to the risk singularity is called the risk pointing vector. The calculation formula is: This vector represents the shortest path and direction that a device must take to reach a specific catastrophic failure state from its current state. Multiple risk pointing vectors can be constructed for various risk singularities in space (corresponding to different failure modes).

[0054] After obtaining the instantaneous evolution velocity vector and risk pointing vector, the cosine similarity formula from vector algebra is used to evaluate the directional consistency between the two. Let... Let be the angle between two vectors, and the formula for calculating cosine similarity is: The value ranges from [-1, 1]. When When the value is close to 1, it indicates that the direction of the evolution velocity vector highly coincides with the direction of the risk pointing vector, meaning that the equipment is rapidly evolving along the direction pointing to the fault singularity, indicating an extremely high risk; when When the value is close to 0 or negative, it indicates that the current evolution direction of the device is unrelated to the fault mode, or even deviates from the fault region.

[0055] Next, we need to construct an exponential gravity function, designed using cosine similarity and the Euclidean distance from the target data point to the risk singularity. Let be the independent variable. Construct an exponential gravitational function of the following form: ,in Based on the fundamental gravitational constant, This is the direction sensitivity coefficient, used to amplify the effect of directional consistency. This is the distance decay exponent (usually a positive value, indicating that the closer the distance, the greater the weight). To prevent small values ​​with a denominator of zero, the calculated output value is the trend weight. This weight directly reflects the urgency with which the current equipment state evolves towards a specific failure mode. Using it as an input weighting factor in subsequent lifetime prediction models forces the models to focus on feature components highly correlated with the current high-risk failure mode, while suppressing irrelevant features.

[0056] In one embodiment, after using the power spectral density feature sequence of hydraulic measured data weighted by directional weights, the time distance from the target data point to the failure state region in the evolution state space is deduced through the following steps: The power spectral density feature sequence in the migration feature set is multiplied by the directional weight to generate a weighted fault-sensitive feature vector. The fault-sensitive feature vector is input into the preset lifetime prediction model, and the degradation-dependent features on the time series are extracted through the long short-term memory network layer in the lifetime prediction model. The degradation-dependent features are input into the fully connected regression layer in the lifespan prediction model, and the predicted health value is output through the fully connected regression layer. Obtain the health threshold of the standard degradation trajectory in the failure state region of the evolution state space, and calculate the time distance required for the health value to drop to the health threshold by linear extrapolation.

[0057] In this embodiment, although the original power spectral density feature sequence contains rich frequency domain information, it lacks the ability to directly represent the degree of fault risk. By applying directional weights to the feature sequence, a signal enhancement process is actually performed on the original physical features: when the equipment is in a high-risk evolution stage, the weight values ​​increase, and the amplitude in the feature vector is amplified as a whole, thereby activating the subsequent neural network model at the numerical level, making it highly attentive to the current input; conversely, when the equipment is in a healthy or stable state, lower weights suppress the feature amplitude and reduce the interference of noise on the model prediction. Specifically, scalar multiplication is used to multiply the scalar directional weights with the power spectral density feature sequence in the vector form of the migration feature set. The wear and fatigue accumulation of hydraulic components is a long-term and historically memory-based nonlinear process. The current health state depends not only on the current vibration characteristics but also on the previous historical state. By introducing gating mechanisms (forget gate, input gate, and output gate), LSTM networks can effectively maintain and update state information over long time spans, overcoming the gradient vanishing problem of traditional recurrent neural networks. During implementation, the fault-sensitive feature vector sequence arranged in chronological order is input into the LSTM layer one by one. At each time step, the forget gate determines how much information to retain from the cell state at the previous time step, the input gate determines how much new information to write into the cell state at the current time step, and finally the output gate combines the updated cell state to calculate the current hidden layer state vector, which is the output degradation dependency feature vector.

[0058] While the feature vectors output by LSTM layers contain rich temporal information, their dimensionality is typically high (e.g., 64 or 128 dimensions), making them difficult for maintenance personnel to understand directly. A fully connected regression layer acts as a comprehensive evaluator, compressing high-dimensional features into a specific scalar value—health—through linear weighted combination. This health value is usually normalized to between 0 and 1, where 1 represents complete health and 0 represents complete failure. In implementation, the fully connected layer contains a weight matrix and a bias term. A matrix multiplication operation is performed between the d-dimensional degradation dependency feature vector and the weight matrix, and the bias term is added. Finally, an activation function is applied to obtain the final result. Next, the threshold of the standard degradation trajectory is obtained, and linear extrapolation is performed to calculate the remaining lifetime. In the evolutionary state space, the standard degradation trajectory connects the healthy region and the failure region; the health value corresponding to the failure state region is the system's failure threshold. The linear extrapolation algorithm is based on the principle of inertia, assuming that the degradation rate of the device will remain consistent with the current time in the near future. In specific implementation, the predicted health value output by the model at the current time is first obtained. Simultaneously, using historical health predictions from the previous sliding window, the current degradation rate (slope) is fitted using the least squares method. Subsequently, the amount of change required for the health to drop from the current value to the failure threshold is calculated and divided by the degradation rate to derive the remaining time distance.

[0059] In one implementation, constructing an exponential gravity function with cosine similarity and the Euclidean distance from the target data point to the risk singularity as independent variables, and using the output value of the exponential gravity function as the directional weight, includes the following steps: Set the gravitational constant parameter, distance decay factor parameter, and direction sensitivity parameter; Calculate the reciprocal of the Euclidean distance from the target data point to the risk singularity, and use the distance decay factor parameter to perform a power operation on the reciprocal of the Euclidean distance to obtain the distance influence term; The direction influence term is obtained by nonlinearly amplifying the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector using the direction sensitivity parameter; Multiply the distance effect term and the direction effect term, and then multiply by the gravitational constant parameter to obtain the fundamental gravitational value; The Sigmoid activation function is introduced to map the basic gravity value to a preset weight range, and the value mapped to the weight range is used as the directional weight.

[0060] In this implementation, within the framework of physics field simulation, three key control variables need to be defined: the gravitational constant parameter, the distance decay factor parameter, and the direction sensitivity parameter. The gravitational constant serves as a benchmark for adjusting the overall weight magnitude, ensuring that the calculated gravitational value is within a reasonable order of magnitude. The distance decay factor controls how quickly gravity decays with increasing distance; a larger value means that significant weight is only generated when the device is extremely close to the fault point, and vice versa. The direction sensitivity parameter is used to adjust the nonlinear contribution of cosine similarity to the final weight. Higher direction sensitivity ensures that the model only responds strongly to cases where the evolution direction precisely points to the fault point, thereby enhancing the identification of specific fault modes. Then, the coordinates of the target data point in the evolutionary state space and the coordinates of the preset risk singularity are extracted, and the straight-line distance between them is calculated using the Euclidean distance formula. Considering that gravity and distance are negatively correlated, the reciprocal of the distance is taken as the base quantity. Next, the pre-set distance decay factor parameter is introduced, and the reciprocal of the distance is exponentially calculated. Power operations are not merely simple numerical transformations; they actually define the potential energy distribution pattern of the risk field. When the distance decay factor parameter is greater than 1, the distance influence term increases exponentially and rapidly as the distance decreases.

[0061] Simply being close in distance might only indicate that the device is in a critical steady state, while a clear direction of evolution is the sign of escalating risk. The cosine similarity between the instantaneous evolution velocity vector and the risk direction vector, obtained through vector analysis, ranges from -1 to 1. Directly using linear cosine values ​​is often insufficient to distinguish ambiguous trends; therefore, a nonlinear amplification of the cosine similarity is performed using a direction sensitivity parameter, typically in the form of an exponential function. Next, the distance and direction influence terms are coupled through multiplication, and the calculated result is the base gravity value. Because the base gravity value includes the reciprocal of the distance and exponential operations, its dynamic range can be extremely large. Directly using it as a weight could lead to instability or gradient explosion in subsequent neural network calculations. Therefore, the Sigmoid activation function is introduced as a mapping tool. The Sigmoid function has the characteristic of compressing any real number input to an open interval of (0,1) and has good nonlinear smoothing capabilities. The final directional weights are calculated, ensuring that the output weights always remain within the acceptable stable range of the model. Meanwhile, the saturation characteristic of the Sigmoid function allows the weights to remain relatively stable in both low-risk and high-risk areas, while exhibiting high sensitivity in the transition region of risk abrupt changes. This aligns with the need for monitoring state abrupt change points in fault monitoring, enabling the final weights to serve as both an amplification factor for features and a direct indicator of risk level.

[0062] The present invention also discloses a transfer learning prediction system for the remaining life of a hydraulic system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the transfer learning prediction method for the remaining life of the hydraulic system as described in any one of the above.

[0063] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0064] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0065] The present invention also discloses a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the transfer learning prediction method for the remaining life of a hydraulic system as described in any of the above embodiments.

[0066] The computer program can be stored in a machine-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The machine-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the machine-readable medium includes, but is not limited to, the above-mentioned components.

[0067] The transfer learning prediction method for the remaining life of the hydraulic system in the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0068] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0069] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A transfer learning prediction method for the remaining life of a hydraulic system, characterized in that, The method includes the following steps: Real-time hydraulic measurement data of the target domain is read from the control system of the on-site hydraulic equipment; Hydraulic test data for the entire life cycle of the source domain is acquired using a hydraulic test bench. The hydraulic measured data and hydraulic test data are then converted into two sets of power spectral density characteristic sequences by time-frequency conversion. The hydraulic test bench has the same hydraulic circuit design as the field hydraulic equipment. The characteristic frequency window is divided according to the mechanical structure characteristics of the hydraulic components in the hydraulic equipment on site. The spectral characteristics of the two sets of power spectral density characteristic sequences are compared based on the characteristic frequency window and the mapping transformation matrix is ​​calculated. The mapping transformation matrix is ​​applied to the power spectral density characteristic sequence of the hydraulic test data to generate a migration feature set aligned with the target domain distribution. Based on hydraulic test data, fluid thermodynamic entropy increase index and vibration energy level index are calculated. An evolutionary state space is constructed with fluid thermodynamic entropy increase index and vibration energy level index as the reference dimensions. The migration feature set is mapped to the evolutionary state space to form a standard degradation trajectory. Risk singularity regions representing different failure modes are pre-set in the evolutionary state space. Hydraulic measured data is mapped into the evolutionary state space and the position and evolutionary trend vector of the target data point in the evolutionary state space at the current moment are calculated in real time. The directional weight is calculated based on the degree to which the evolutionary trend vector points to the risk singularity region. After using the power spectral density feature sequence of hydraulic measured data weighted by directional weights, the time distance from the target data point to the failure state region in the evolution state space is deduced, and the life prediction result of the hydraulic equipment in the field is generated based on the deduction result.

2. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 1, characterized in that, The process of comparing the spectral features of two sets of power spectral density feature sequences based on a feature frequency window and calculating a mapping transformation matrix, then applying the mapping transformation matrix to the power spectral density feature sequences of hydraulic test data to generate a migration feature set aligned with the target domain distribution includes the following steps: Source domain data segments that match the system load pressure level under stable operating conditions in the target domain are selected from the hydraulic test data. Time-frequency transformations were performed on the source domain data segments and the measured hydraulic data under steady-state conditions in the target domain, and the source domain power spectrum vector and the target domain power spectrum vector within the characteristic frequency window were extracted. By combining the power spectral vectors of the source domain and the target domain, a mapping optimization equation is constructed. The gradient descent method is used to solve the mapping optimization equation, and a mapping transformation matrix representing the mapping relationship from the feature space of the source domain to the feature space of the target domain is obtained. Multiplying the mapping transformation matrix on the left by the power spectral density feature sequence of the hydraulic test data yields a migration feature set that conforms to the data distribution characteristics of the target domain.

3. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 2, characterized in that, The process of selecting source domain data segments from hydraulic test data that match the system load pressure level under stable operating conditions in the target domain includes the following steps: By monitoring the system pressure and flow parameters of the hydraulic measured data within a preset time window, the pressure fluctuation variance and flow fluctuation variance are calculated respectively. When both the pressure fluctuation variance and the flow fluctuation variance are less than the preset stability judgment threshold, the preset time window is marked as the stable operating condition window, and the average system pressure value and average flow value calculated within the stable operating condition window are used as the operating condition matching benchmark. Calculate the Euclidean distance between the source domain pressure and source domain flow values ​​at each sampling time of the hydraulic test data and the operating condition matching benchmark; The hydraulic test data were sorted in ascending order of Euclidean distance, and the first few data segments with the smallest Euclidean distance were selected as the source domain data segments.

4. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 3, characterized in that, The step of performing time-frequency transformation on the source domain data segment and the hydraulic measured data under steady-state conditions in the target domain, and extracting the source domain power spectral vector and the target domain power spectral vector within the characteristic frequency window, includes the following steps: Calculate the spectral energy concentration of the source domain data segment within the initial characteristic frequency window; Using the gradient descent search algorithm and aiming to maximize the spectral energy concentration, the left and right boundary frequencies of the feature frequency window are iteratively adjusted to determine the optimal feature frequency window; Within the optimal characteristic frequency window, the hydraulic measured data within the source domain data segment and the steady-state operating condition window are discretized and sampled. The discretized spectrum amplitude points are arranged in frequency order to generate the source domain power spectrum vector and the target domain power spectrum vector.

5. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 1, characterized in that, The hydraulic test data includes inlet oil temperature, outlet oil temperature, inlet pressure, outlet pressure, and flow parameters. The process of calculating fluid thermodynamic entropy increase and vibration energy level indices based on the hydraulic test data, constructing an evolutionary state space using these indices as baseline dimensions, and mapping the migration feature set to the evolutionary state space to form a standard degradation trajectory includes the following steps: Based on hydraulic test data and the enthalpy difference between the inlet and outlet of the hydraulic test bench calculated according to the fluid thermodynamic equation, the total power loss of the hydraulic test bench is calculated based on the flow parameters. The fluid thermodynamic entropy production rate is obtained by dividing the total power loss by the average thermodynamic temperature of the hydraulic test bench, and the fluid thermodynamic entropy production rate is normalized to obtain the fluid thermodynamic entropy increase index. The vibration energy level index is calculated by integrating the power spectral density feature sequence in the migration feature set across the entire frequency band. A two-dimensional coordinate system is established with the fluid thermodynamic entropy increase index as the horizontal axis and the vibrational energy level index as the vertical axis as the evolution state space. The migration feature set is projected as coordinate points onto the evolutionary state space according to the sampling time of the hydraulic test data; The projected coordinate points are smoothed using a polynomial fitting algorithm to generate a standard degenerate trajectory that connects the healthy state region and the failed state region in the evolutionary state space.

6. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 1, characterized in that, The process of mapping hydraulic measured data into the evolutionary state space and calculating the position and evolutionary trend vector of the target data point in the evolutionary state space at the current moment in real time, and calculating the directional weight based on the degree to which the evolutionary trend vector points to the risk singularity region, includes the following steps: Based on the hydraulic measured data, the fluid thermodynamic entropy increase index and vibration energy level index at each sampling moment in the target domain are calculated, and the coordinate positions of the corresponding data points in the evolution state space at each sampling moment in the target domain are determined. Based on the change in coordinate position between the current time and the previous time in the target domain, calculate the instantaneous evolution velocity vector of the target data point corresponding to the current time in the target domain; Construct a risk pointing vector that connects the coordinates of the target data point with the coordinates of the risk singularity in the risk singularity region; Calculate the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector; An exponential gravity function is constructed with cosine similarity and Euclidean distance from the target data point to the risk singularity as independent variables, and the output value of the exponential gravity function is used as the directional weight.

7. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 6, characterized in that, After utilizing the power spectral density feature sequence of the hydraulic measured data with directional weighting, the time distance from the target data point to the failure state region in the evolution state space is deduced through the following steps: The power spectral density feature sequence in the migration feature set is multiplied by the directional weight to generate a weighted fault-sensitive feature vector. The fault-sensitive feature vector is input into the preset lifetime prediction model, and the degradation-dependent features on the time series are extracted through the long short-term memory network layer in the lifetime prediction model. The degradation-dependent features are input into the fully connected regression layer in the lifespan prediction model, and the predicted health value is output through the fully connected regression layer. Obtain the health threshold of the standard degradation trajectory in the failure state region of the evolution state space, and calculate the time distance required for the health value to drop to the health threshold by linear extrapolation.

8. The transfer learning prediction method for the remaining life of a hydraulic system according to claim 6, characterized in that, The construction of an exponential gravity function with cosine similarity and the Euclidean distance from the target data point to the risk singularity as independent variables, and the use of the output value of the exponential gravity function as the directional weight, includes the following steps: Set the gravitational constant parameter, distance decay factor parameter, and direction sensitivity parameter; Calculate the reciprocal of the Euclidean distance from the target data point to the risk singularity, and use the distance decay factor parameter to perform a power operation on the reciprocal of the Euclidean distance to obtain the distance influence term; The directional influence term is obtained by nonlinearly amplifying the cosine similarity between the instantaneous evolution velocity vector and the risk pointing vector using the directional sensitivity parameter; Multiply the distance effect term and the direction effect term by the gravitational constant parameter to obtain the fundamental gravitational value; The Sigmoid activation function is introduced to map the basic gravity value to a preset weight range, and the value mapped to the weight range is used as the directional weight.

9. A transfer learning prediction system for the remaining life of a hydraulic system, 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 computer program, it implements the transfer learning prediction method for the remaining life of the hydraulic system as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform a transfer learning prediction method for the remaining life of a hydraulic system according to any one of claims 1 to 8.

Images