Valve durability test method and system based on vibration environment

Abstract

The invention provides a valve durability testing method and system based on a vibration environment, and relates to the technical field of automatic testing, and the method comprises the steps: employing a multi-degree-of-freedom vibration excitation mechanism to carry out the vibration loading of a valve, employing a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping, and carrying out the testing of the durability of the valve. Multi-direction vibration superposition is generated by controlling a multi-degree-of-freedom vibration excitation mechanism, and the vibration direction and vibration parameters are dynamically adjusted in combination with multi-target self-adaptive optimization; a three-layer cascade attention network structure comprising an axial attention layer, a working condition attention layer and a time sequence attention layer and a double-flow feature fusion network are adopted to process the acquired response data under the vibration working condition, and the durability state of the valve is predicted; and based on a durability prediction result, performing optimization adjustment on a composite vibration mode generation algorithm to form closed-loop control of durability testing.

Description

Technical Field

[0001] The present invention relates to the technical field of automated testing, and particularly to a method and system for testing the durability of a valve based on a vibration environment. Background Art

[0002] Traditional methods for testing the durability of valves mainly use single-axis vibration excitation, making it difficult to truly simulate the complex vibration conditions in the actual working environment. Although some studies have started to use multi-axis vibration testing, there is a lack of an effective cooperative control mechanism between the vibration units of each axis, resulting in insufficient vibration loading accuracy.

[0003] Existing vibration mode generation algorithms often use preset fixed parameters, without considering the inter-axis coupling effect and the characteristics of working condition conversion. It is difficult to accurately reflect the vibration characteristics of the valve under different working conditions. At the same time, the lack of the ability to adaptively adjust the vibration direction and vibration parameters reduces the effectiveness of the test. The existing evaluation of the durability state of valves mainly relies on simple signal processing methods and statistical analysis, and cannot effectively process multi-source heterogeneous vibration response data.

[0004] These problems restrict the accuracy and scientific nature of valve durability testing, and there is an urgent need to develop new testing methods. Summary of the Invention

[0005] Embodiments of the present invention provide a method and system for testing the durability of a valve based on a vibration environment, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] A method for testing the durability of a valve based on a vibration environment is provided, including:

[0008] Vibrating and loading the valve by using a multi-degree-of-freedom vibration excitation mechanism, where the multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit, and a Z-axis vibration unit under cooperative control;

[0009] Based on the multi-degree-of-freedom vibration excitation mechanism, a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping is used to generate multi-directional vibration superposition by controlling the X-axis vibration unit, the Y-axis vibration unit, and the Z-axis vibration unit, and the vibration direction and vibration parameters are dynamically adjusted by combining multi-objective adaptive optimization;

[0010] The response data of the valve during vibration is collected in real time through a sensor array, and the three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network are used to process the response data to predict the durability state of the valve; based on the durability prediction result, the composite vibration mode generation algorithm is optimized and adjusted to form a closed-loop control for durability testing.

[0011] In an alternative embodiment,

[0012] the multi-degree-of-freedom vibration excitation mechanism includes:

[0013] The X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit all include a driver, a displacement sensor, and an acceleration sensor, and vibration state data is obtained in real time through a multi-channel data acquisition module;

[0014] The X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit adopt a hierarchical control architecture, including bottom-layer execution control, middle-layer collaborative control, and upper-layer motion planning. Among them, the bottom-layer execution control uses an independent PID controller, the middle-layer collaborative control calculates the mutual interference amount between axes according to the current vibration state data and compensates it in real time, and the upper-layer motion planning generates a vibration trajectory according to the test requirements;

[0015] Vibration precision control is performed using a phase synchronization control and adaptive amplitude control algorithm based on phase-locked loop technology. Among them, the phase synchronization control calculates the error between the target phase and the actual phase, and adjusts the control amount in real time according to the proportional-integral control strategy; the adaptive amplitude control iteratively updates the output amplitude by comparing the ratio of the target amplitude to the measured amplitude.

[0016] In an alternative embodiment,

[0017] A composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping is adopted. By controlling the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit to generate multi-directional vibration superposition, the vibration direction and vibration parameters are dynamically adjusted by combining multi-objective adaptive optimization, including:

[0018] Establish a composite vibration mode generation algorithm. By performing frequency domain analysis and principal component extraction, a vibration working condition library for the valve usage environment is constructed. Calculate the vibration correlation matrix of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit, and introduce a coupling correction coefficient to correct the vibration transfer function. The coupling correction coefficient characterizes the ratio relationship between the actual amplitude and the theoretical amplitude between different axial vibration units;

[0019] The vibration characteristic parameters of the valve in normal operation, start-stop transition, extreme environment, and fault simulation in the actual environment are mapped to the vibration frequencies of each axis of the multi-degree-of-freedom vibration excitation mechanism through a working condition conversion matrix. The working condition conversion matrix includes vibration type, frequency parameters, and duration; according to the vibration frequencies of each axis, the synthetic acceleration amplitude and direction vector are calculated through spatial vibration synthesis, and a vibration energy distribution algorithm is used to determine the vibration energy of each axis;

[0020] Construct a multi-objective optimization function including a working condition simulation accuracy term, an energy consumption term, and a system stability term based on the synthetic acceleration amplitude and direction vector, and use an adaptive weight coefficient to dynamically balance each optimization term; the working condition simulation accuracy term is evaluated through time-domain signal similarity, frequency-domain error, and statistical feature comparison; based on the optimization results of the multi-objective optimization function, dynamically update the vibration frequencies, amplitudes, and phase parameters of each vibration unit to achieve adaptive adjustment of the vibration direction and vibration parameters.

[0021] In an alternative embodiment,

[0022] The calculation of the adaptive weight coefficient includes:

[0023] Calculate the vibration energy proportion of each vibration unit, where the vibration energy proportion is the ratio of the vibration energy of each vibration unit within a preset time window to the total vibration energy, and calculate the vibration energy imbalance degree based on the vibration energy proportion, where the vibration energy imbalance degree is the maximum value of the difference in vibration energy proportions between any two vibration units;

[0024] Divide the vibration working conditions into an equilibrium section, a transition section, and a non-equilibrium section according to the vibration energy imbalance degree; in the equilibrium section, positively adjust the weight of the working condition simulation accuracy term, negatively adjust the weight of the energy consumption term, and keep the weight of the system stability term unchanged; in the transition section, perform exponential decay adjustment on the weight of the working condition simulation accuracy term, keep the weight of the energy consumption term unchanged, and set the weight of the system stability term to a compensation value with a sum of weights equal to one; in the non-equilibrium section, perform exponential adjustment on the weights of the working condition simulation accuracy term and the system stability weight term, and set the weight of the energy consumption term to a compensation value with a sum of weights equal to one; use a state transition function based on the vibration energy imbalance degree to smoothly transition the weights under each vibration working condition to obtain the adaptive weight coefficient.

[0025] In an alternative embodiment,

[0026] Collect the response data of the valve during vibration in real time through a sensor array, and use a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network to process the response data. The prediction of the durability state of the valve includes:

[0027] Construct a three - layer cascaded attention network structure. The three - layer cascaded attention network structure includes an axial attention layer, a working condition attention layer, and a time - series attention layer. The axial attention layer calculates the feature weights of the X - axis vibration unit, Y - axis vibration unit, and Z - axis vibration unit based on the vibration correlation matrix. The working condition attention layer calculates the feature weights of the equilibrium section, transition section, and non - equilibrium section based on the vibration energy distribution. The time - series attention layer calculates the time - series feature weights based on the historical state information. Input the response data into the three - layer cascaded attention network structure, and perform weighted calculations with the corresponding feature weights respectively to obtain weighted feature vectors.

[0028] Establish a two - stream feature fusion network. The two - stream feature fusion network includes a vibration feature stream and a durability evaluation stream. Input the weighted feature vectors into the two - stream feature fusion network. Among them, the vibration feature stream extracts the vibration state characterization vector, and the durability evaluation stream extracts the performance degradation characterization vector. Use a feature fusion matrix to fuse the vibration state characterization vector and the performance degradation characterization vector to obtain a fused feature vector.

[0029] Based on the fused feature vector, establish a loss function. The loss function includes a durability state prediction loss term, a fault feature loss term, and a performance degradation trend loss term. Obtain the durability state prediction result of the valve by minimizing the loss function.

[0030] In an alternative embodiment,

[0031] The calculation of the feature weights in the three - layer cascaded attention network structure includes:

[0032] Calculate the eigen - decomposition of the vibration correlation matrix, obtain the principal eigen - vector and the corresponding eigenvalue, construct an axial dynamic weight function based on the eigenvalue, and weight the axial dynamic weight function with the original features of the X - axis vibration unit, Y - axis vibration unit, and Z - axis vibration unit to obtain the feature weights of the axial attention layer.

[0033] Use a sliding time window to calculate the vibration energy change rate, extract the energy gradient feature at the working condition switching moment, construct a working condition transition probability matrix based on the energy gradient feature, and map the working condition transition probability matrix to the feature weight coefficients of the equilibrium section, transition section, and non - equilibrium section to obtain the feature weights of the working condition attention layer.

[0034] Calculate the attenuation coefficient based on the time interval between the historical state information and the current moment, and perform weighted accumulation of the attenuation coefficient and the historical state information to obtain the feature weights of the time - series attention layer.

[0035] In an alternative embodiment,

[0036] Establish a dual-stream feature fusion network. Inputting the weighted feature vector into the dual-stream feature fusion network to obtain a fused feature vector includes:

[0037] Construct a vibration feature stream using a temporal convolutional network with three-layer residual structure and convolutional kernel size increasing with the number of layers. Perform feature transformation on the weighted feature vector through multi-layer convolutional operations, and calculate attention weights based on the frequency-domain energy distribution to perform weighted combination on the feature transformation results, obtaining a vibration state representation vector;

[0038] Construct a durability assessment stream using a long short-term memory network with a bidirectional structure and a skip connection mechanism. Perform temporal modeling on the weighted feature vector based on the forget gate state and the input gate state to obtain a hidden state, and calculate temporal weights based on the temporal change trend of the hidden state to perform weighted accumulation on the hidden state, obtaining a performance degradation representation vector;

[0039] Perform a concatenation operation on the vibration state representation vector and the performance degradation representation vector. Use the attention mechanism to calculate the inter-channel correlation of the concatenated features to obtain a feature fusion matrix. Calculate the feature fusion importance score through the feature fusion matrix, construct a dynamic fusion weight based on the feature fusion importance score, perform a weighted operation on the dynamic fusion weight and the feature fusion matrix, and introduce a residual connection to obtain a fused feature vector.

[0040] In the second aspect of the embodiments of the present invention,

[0041] Provide a valve durability test system based on the vibration environment, including:

[0042] The first unit is used to perform vibration loading on the valve using a multi-degree-of-freedom vibration excitation mechanism, and the multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit, and a Z-axis vibration unit that are collaboratively controlled;

[0043] The second unit is used to, based on the multi-degree-of-freedom vibration excitation mechanism, adopt a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping, generate multi-directional vibration superposition by controlling the X-axis vibration unit, the Y-axis vibration unit, and the Z-axis vibration unit, and dynamically adjust the vibration direction and vibration parameters through multi-objective adaptive optimization;

[0044] The third unit is used to collect the response data of the valve during vibration in real time through a sensor array, process the response data using a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a temporal attention layer and a dual-stream feature fusion network, and predict the durability state of the valve; based on the durability prediction result, optimize and adjust the composite vibration mode generation algorithm to form a closed-loop control for durability testing.

[0045] In the third aspect of the embodiments of the present invention,

[0046] Provided is an electronic device, comprising:

[0047] a processor;

[0048] a memory for storing instructions executable by the processor;

[0049] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0050] In a fourth aspect of the embodiments of the present invention,

[0051] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0052] The beneficial effects of the present invention are as follows:

[0053] The proposed three-level cascaded attention network structure realizes multi-dimensional feature extraction of vibration response data through the collaborative action of the axial attention layer, the working condition attention layer, and the time series attention layer. Among them, the axial attention layer captures the correlation between different vibration units, the working condition attention layer adaptively adjusts the feature weights under different working conditions, and the time series attention layer effectively models the historical state information, significantly improving the accuracy and reliability of durability state prediction.

[0054] The designed composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping corrects the vibration transfer function by introducing a coupling correction coefficient, and realizes the dynamic mapping of vibration characteristic parameters by using a working condition conversion matrix. Combining a multi-objective adaptive optimization method to adjust the vibration direction and vibration parameters in real time, effectively solving the coupling interference problem in the multi-axis vibration superposition process, and significantly improving the accuracy of vibration loading.

[0055] A two-stream feature fusion network including a vibration feature stream and a durability evaluation stream is constructed. The vibration state representation vector and the performance degradation representation vector are respectively extracted through a time series convolutional network and a long short-term memory network, and an attention mechanism is used to realize the adaptive fusion of features. This network structure makes full use of the time series correlation and multi-scale features of vibration data, provides a complete closed-loop control scheme for the durability test of the valve, and effectively improves the test efficiency and reliability. Description of the Drawings

[0056] Figure 1 It is a schematic flowchart of the valve durability test method based on the vibration environment in the embodiments of the present invention;

[0057] Figure 2 Adaptive weight adjustment process diagram based on vibration energy imbalance;

[0058] Figure 3 It is a comparison chart of the prediction accuracies of different methods under typical working conditions. Specific implementation manners

[0059] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0061] Figure 1 It is a schematic flowchart of a method for testing the durability of a valve based on a vibration environment in an embodiment of the present invention. As Figure 1 shown, the method includes:

[0062] Vibration loading is applied to the valve by a multi-degree-of-freedom vibration excitation mechanism, and the multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit and a Z-axis vibration unit under coordinated control;

[0063] Based on the multi-degree-of-freedom vibration excitation mechanism, a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping is adopted, and multi-directional vibration superposition is generated by controlling the X-axis vibration unit, the Y-axis vibration unit and the Z-axis vibration unit, and the vibration direction and vibration parameters are dynamically adjusted by combining multi-objective adaptive optimization;

[0064] Response data of the valve during vibration is collected in real time by a sensor array, and the response data is processed by a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer and a time series attention layer and a dual-stream feature fusion network to predict the durability state of the valve; based on the durability prediction result, the composite vibration mode generation algorithm is optimized and adjusted to form a closed-loop control for durability testing.

[0065] In an alternative implementation manner,

[0066] The multi-degree-of-freedom vibration excitation mechanism includes:

[0067] The X-axis vibration unit, the Y-axis vibration unit and the Z-axis vibration unit all include a driver, a displacement sensor and an acceleration sensor, and vibration state data is obtained in real time through a multi-channel data acquisition module;

[0068] The X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit adopt a hierarchical control architecture, including bottom-layer execution control, middle-layer collaborative control, and upper-layer motion planning. Among them, the bottom-layer execution control uses independent PID controllers, the middle-layer collaborative control calculates the mutual interference amount between axes based on the current vibration state data and compensates in real time, and the upper-layer motion planning generates vibration trajectories according to test requirements;

[0069] Vibration precision control is carried out using a phase synchronization control and an adaptive amplitude control algorithm based on phase-locked loop technology. Among them, the phase synchronization control calculates the error between the target phase and the actual phase, and adjusts the control amount in real time according to the proportional-integral control strategy; the adaptive amplitude control iteratively updates the output amplitude by comparing the ratio of the target amplitude to the measured amplitude.

[0070] Exemplarily, the multi-degree-of-freedom vibration excitation mechanism consists of three vibration units arranged orthogonally, corresponding to the X-axis, Y-axis, and Z-axis directions respectively. Each axis vibration unit is composed of the same basic components. Taking the X-axis vibration unit as an example, this unit includes a DC servo motor as a driver, an optical encoder as a displacement sensor, and a piezoelectric acceleration sensor. The vibration state data is obtained in real time through a 24-bit multi-channel data acquisition module.

[0071] The control system of the vibration unit adopts a hierarchical control architecture, divided into three levels:

[0072] The bottom-layer execution control layer uses independent PID controllers, which are configured separately for each vibration unit. Taking the X-axis vibration unit as an example, the PID parameters are configured as: proportional coefficient Kp = 2.5, integral coefficient Ki = 0.8, derivative coefficient Kd = 0.3, and the control period is 1 millisecond. This controller receives the real-time feedback from the displacement sensor and the acceleration sensor, calculates the control deviation after comparing it with the target value, and outputs a PWM control signal to drive the motor to achieve closed-loop control.

[0073] The middle-layer collaborative control mainly solves the problem of mutual interference between axes. For example, when the X-axis vibration unit operates at 85 Hz with an amplitude of 2 mm, it will generate an interference amplitude of about 0.3 mm on the Y-axis. In response to this situation, the middle-layer collaborative control performs real-time compensation through an interference matrix. The elements in the interference matrix represent the mutual influence coefficients between different axes. For example, the interference coefficient of the X-axis on the Y-axis is 0.15, and the interference coefficient of the Y-axis on the Z-axis is 0.12. By measuring the actual vibration state, calculating the interference amount, and then correcting the control signal in real time.

[0074] The upper - layer motion planning is responsible for generating vibration trajectories according to test requirements. The upper - layer motion planner has built - in various typical vibration condition templates, including sine sweep, random vibration, shock response, etc. The planner selects an appropriate template according to the test requirements and then generates corresponding vibration parameters such as frequency, amplitude, and phase. For example, for the test of the valve start - stop process, a logarithmic sweep mode with a frequency range of 20 - 200 Hz is adopted, the sweep time is 60 seconds, the amplitude gradually increases to 1.5 mm and then remains for 30 seconds, and then gradually decreases to 0 within 20 seconds. This preset trajectory can simulate the frequent start - stop process of the valve in the actual working environment.

[0075] Adopt a phase - synchronous control based on phase - locked loop technology and an adaptive amplitude control algorithm. The implementation process of phase - synchronous control is as follows: Collect real - time vibration signals from the acceleration sensor, calculate the actual phase angle through FFT; compare with the target phase angle and calculate the phase error; calculate the frequency adjustment amount through a proportional - integral control strategy, with the proportional coefficient set to 0.5 and the integral coefficient set to 0.2; update the frequency of the vibration control signal according to the frequency adjustment amount. The adaptive amplitude control algorithm ensures that the actual vibration amplitude is consistent with the target by dynamically adjusting the amplitude of the control signal. The specific implementation is to calculate the ratio of the target amplitude to the measured amplitude, multiply by a correction factor (usually set to 0.8) as the adjustment factor, and then multiply by the current output amplitude to obtain the new output amplitude. This process is iterated every 50 milliseconds until the error between the actual amplitude and the target amplitude is less than the preset threshold (usually 5%). For example, when the target amplitude is 2 mm and the measured value is 1.8 mm, the adjustment factor is (2 / 1.8)×0.8 = 0.889, and the new output amplitude is the current output amplitude×0.889.

[0076] The multi - degree - of - freedom vibration excitation mechanism of the present invention realizes omnidirectional vibration loading with three - axis linkage, improves the vibration accuracy through a hierarchical control architecture and phase - locked loop technology, effectively simulates the complex vibration conditions in the actual working environment of the valve, and can expose potential durability problems more quickly compared with traditional single - axis tests, greatly improving the test efficiency and accuracy.

[0077] In an alternative embodiment,

[0078] Adopt a composite vibration mode generation algorithm based on inter - axis coupling correction and working condition conversion mapping. By controlling the X - axis vibration unit, Y - axis vibration unit, and Z - axis vibration unit to generate multi - direction vibration superposition, combined with multi - objective adaptive optimization to dynamically adjust the vibration direction and vibration parameters, including:

[0079] Establish a composite vibration mode generation algorithm. Construct a vibration working condition library for the valve usage environment through frequency domain analysis and principal component extraction. Calculate the vibration correlation matrix of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit, and introduce a coupling correction coefficient to correct the vibration transfer function. The coupling correction coefficient characterizes the ratio relationship between the actual amplitude and the theoretical amplitude among different axial vibration units;

[0080] Map the vibration characteristic parameters of the valve in normal operation, start-stop transition, extreme environment, and fault simulation in the actual environment to the vibration frequencies of each axis of the multi-degree-of-freedom vibration excitation mechanism through a working condition conversion matrix. The working condition conversion matrix includes vibration type, frequency parameters, and duration; Calculate the synthetic acceleration amplitude and direction vector through spatial vibration synthesis according to the vibration frequencies of each axis, and use the vibration energy distribution algorithm to determine the vibration energy of each axis;

[0081] Construct a multi-objective optimization function including a working condition simulation accuracy term, an energy consumption term, and a system stability term based on the synthetic acceleration amplitude and direction vector, and use an adaptive weight coefficient to dynamically balance each optimization term; The working condition simulation accuracy term is evaluated through time-domain signal similarity, frequency-domain error, and statistical feature comparison; Based on the optimization results of the multi-objective optimization function, dynamically update the vibration frequencies, amplitudes, and phase parameters of each vibration unit to achieve adaptive adjustment of the vibration direction and vibration parameters.

[0082] Exemplarily, the composite vibration mode generation algorithm constructs a vibration working condition library for the valve usage environment through frequency domain analysis and principal component extraction. Use an acceleration sensor to collect vibration data. After the collected raw data is filtered by a band-pass filter to remove noise, perform a fast Fourier transform to obtain spectral features. Perform dimensionality reduction processing on the spectral features through principal component analysis, extract the principal component features with a contribution rate exceeding 85%, and construct a working condition library containing multiple typical vibration modes. Each mode includes three types of parameters: frequency distribution, amplitude range, and duration.

[0083] The vibration correlation matrix is obtained by measuring the degree of mutual influence of the vibration units on each axis at different excitation frequencies. The test process is as follows: The X-axis vibration unit is excited alone, with the frequency ranging from 10 Hz to 500 Hz in steps of 10 Hz, and maintained for 30 seconds at each frequency point, while recording the outputs of the triaxial acceleration sensors. The above process is repeated for the Y-axis and Z-axis vibration units respectively. For example, at the frequency point of 120 Hz, the influence coefficient of the X-axis on the Y-axis is 0.18, and on the Z-axis is 0.12; the influence coefficient of the Y-axis on the X-axis is 0.22, and on the Z-axis is 0.16; the influence coefficient of the Z-axis on the X-axis is 0.14, and on the Y-axis is 0.19. These data form a 3×3 vibration correlation matrix. The coupling correction coefficient is calculated by comparing the ratio of the theoretical amplitude to the measured amplitude. For example, when the three axes work simultaneously, the set amplitude of the X-axis is 2 mm, and the actually measured amplitude is 1.75 mm, with a correction coefficient of 1.14; the set amplitude of the Y-axis is 1.5 mm, and the actually measured amplitude is 1.35 mm, with a correction coefficient of 1.11; the set amplitude of the Z-axis is 1 mm, and the actually measured amplitude is 0.92 mm, with a correction coefficient of 1.09. These correction coefficients are applied to the vibration transfer function between axes to pre-correct the control commands, making the actual output closer to the theoretical expectation.

[0084] The condition conversion mapping is a key technology for converting the vibration characteristic parameters in the actual environment into the control parameters of the multi-degree-of-freedom vibration excitation mechanism. The implementation process is to map the vibration characteristics collected from the site into the specific parameters of the vibration units on each axis through the condition conversion matrix. For example, the vibration characteristics of a certain pressure regulating valve under specific conditions are: main frequency 85 Hz, additional frequencies 45 Hz and 125 Hz, and peak acceleration 3.2g. Through the condition conversion matrix, these characteristics are mapped to: X-axis vibration frequency 85 Hz, amplitude 1.2 mm; Y-axis vibration frequency 45 Hz, amplitude 0.8 mm; Z-axis vibration frequency 125 Hz, amplitude 0.6 mm; duration 240 seconds. The condition conversion matrix is a mapping table containing vibration types, frequency parameters and durations. This table is constructed based on experimental data and covers four major types of typical conditions: normal operation, start-stop transition, extreme environment and fault simulation. Taking the start-stop transition condition as an example, its characteristic is that the frequency changes rapidly from low to high, and the amplitude first increases and then decreases. In the condition conversion matrix, this condition is decomposed into: the X-axis performs a frequency sweep from 20 Hz to 150 Hz for 15 seconds, and the amplitude increases from 0.5 mm to 2 mm and then drops; the Y-axis performs a fixed frequency of 80 Hz, and the amplitude increases from 0.3 mm to 1.5 mm and then remains; the Z-axis performs a frequency sweep from 30 Hz to 100 Hz, and the amplitude is fixed at 0.8 mm.

[0085] Spatial vibration synthesis is a technology that synthesizes vibration parameters of each axis into a spatial vibration vector. The vibrations of each axis are regarded as vector components, and the total acceleration amplitude and direction are calculated through vector synthesis. For example, when the acceleration of the X-axis is 2g, the acceleration of the Y-axis is 1.5g, and the acceleration of the Z-axis is 1g, the synthesized acceleration amplitude is 2.69g, and the direction vector is (0.74, 0.56, 0.37). The vibration energy distribution algorithm calculates the vibration energy that should be distributed to each axis in reverse according to the target synthesized vibration.

[0086] The multi-objective optimization function includes three evaluation items: the working condition simulation accuracy item, the energy consumption item, and the system stability item. The working condition simulation accuracy item is evaluated by calculating the similarity between the actual vibration and the target vibration, including the correlation coefficient of the time-domain waveform (the target value is greater than 0.85), the root mean square error of the frequency-domain characteristics (the target value is less than 10%), and the deviation of the statistical characteristics (peak value, mean value, standard deviation) (the target value is less than 15%). The energy consumption item considers the power consumption of each vibration unit and is calculated by measuring the input power of the driving motor. The system stability item is evaluated based on the fluctuation range of the displacement and acceleration of each axis, and the stability index value should usually be controlled within 0.2. The adaptive weight coefficient dynamically adjusts the weights of each optimization item according to the current working condition.

[0087] Based on the optimization results of the multi-objective optimization function, the system dynamically updates the vibration parameters once every preset time (such as 200 milliseconds), including the frequency of each axis (the adjustment accuracy is 0.1 Hz), the amplitude (the adjustment accuracy is 0.05 mm), and the phase (the adjustment accuracy is 1 degree).

[0088] The composite vibration mode generation algorithm of the present invention accurately simulates the complex vibration conditions in the actual working environment of the valve through inter-axis coupling correction and working condition conversion mapping technology. It can calculate the vibration correlation matrix in real time and apply the coupling correction coefficient, significantly reducing the vibration error and improving the simulation accuracy; the multi-objective adaptive optimization realizes the dynamic balance of the working condition simulation accuracy, energy consumption, and system stability, reducing energy consumption while ensuring the test effect and improving the test efficiency; the constructed working condition library of various typical vibration modes covers the usage scenarios of the entire life cycle of the valve, significantly enhancing the comprehensiveness and reliability of the test.

[0089] In an alternative embodiment,

[0090] The calculation of the adaptive weight coefficient includes:

[0091] Calculate the vibration energy proportion of each vibration unit. The vibration energy proportion is the ratio of the vibration energy of each vibration unit within a preset time window to the total vibration energy, and calculate the vibration energy imbalance degree based on the vibration energy proportion. The vibration energy imbalance degree is the maximum value of the difference in the vibration energy proportion between any two vibration units;

[0092] The vibration conditions are divided into an equilibrium section, a transition section, and a non-equilibrium section according to the vibration energy imbalance; in the equilibrium section, the weight of the working condition simulation accuracy term is adjusted positively, the weight of the energy consumption term is adjusted negatively, and the weight of the system stability term remains unchanged; in the transition section, the weight of the working condition simulation accuracy term is adjusted with exponential decay, the weight of the energy consumption term remains unchanged, and the weight of the system stability term is set to a compensation value with a sum of weights equal to one; in the non-equilibrium section, the weights of the working condition simulation accuracy term and the system stability weight term are adjusted exponentially, and the weight of the energy consumption term is set to a compensation value with a sum of weights equal to one; a state transition function based on the vibration energy imbalance is used to smoothly transition the weights under each vibration condition to obtain the adaptive weight coefficients.

[0093] Exemplarily, calculate the vibration energy proportion of each vibration unit, collect the acceleration data of the vibration units on the X-axis, Y-axis, and Z-axis, square and sum the collected acceleration signals to obtain the vibration energy of each axis within the time window. For example, in a certain test, the vibration energy of the X-axis is 42 joules, the vibration energy of the Y-axis is 36 joules, the vibration energy of the Z-axis is 22 joules, and the total vibration energy is 100 joules. Accordingly, calculate the vibration energy proportion of each axis: 0.42 for the X-axis, 0.36 for the Y-axis, and 0.22 for the Z-axis. Calculate the vibration energy imbalance based on the vibration energy proportion. The vibration energy imbalance is defined as the maximum value of the difference in the vibration energy proportion between any two vibration units. In the above case, the difference between the X-axis and the Y-axis is 0.06, the difference between the X-axis and the Z-axis is 0.20, and the difference between the Y-axis and the Z-axis is 0.14. Therefore, the vibration energy imbalance is 0.20.

[0094] As Figure 2 As shown in the adaptive weight adjustment process diagram based on the vibration energy imbalance, the vibration conditions are divided into three types according to the vibration energy imbalance: an equilibrium section, a transition section, and a non-equilibrium section. For example, when the vibration energy imbalance is less than 0.15, it is the equilibrium section; when the vibration energy imbalance is between 0.15 and 0.30, it is the transition section; when the vibration energy imbalance is greater than 0.30, it is the non-equilibrium section.

[0095] In the equilibrium section, the energy distribution of each vibration unit is relatively uniform, and the system operation state is stable. At this time, priority is given to ensuring the working condition simulation accuracy, and at the same time, the energy consumption is appropriately reduced. The adjustment strategy is: positively adjust the weight of the working condition simulation accuracy term, and the adjustment amplitude is 1.2 times the initial weight, but not exceeding 0.7; negatively adjust the weight of the energy consumption term, and the adjustment amplitude is 0.8 times the initial weight, but not less than 0.1; keep the weight of the system stability term unchanged. For example, if the initial weight configuration is 0.

[0096] In the transition section, the vibration conditions are changing, and it is necessary to enhance the system stability control. The adjustment strategy is as follows: exponentially decay the weight of the working condition simulation accuracy term, with the decay factor set to 0.85; keep the weight of the energy consumption term unchanged; set the weight of the system stability term to a compensation value with a sum of weights equal to one. Taking the initial weight configuration of 0.5 for the working condition simulation accuracy term, 0.3 for the energy consumption term, and 0.2 for the system stability term as an example, after adjustment, the weight of the working condition simulation accuracy term is 0.5×0.85 = 0.425, the weight of the energy consumption term remains 0.3, and the weight of the system stability term is 1 - 0.425 - 0.3 = 0.275.

[0097] In the unbalanced section, the vibration energy distribution is extremely uneven. It is necessary to prioritize ensuring system stability while reducing energy consumption. The adjustment strategy is as follows: exponentially adjust the weights of the working condition simulation accuracy term and the system stability weight term. The decay factor of 0.7 is used for the working condition simulation accuracy term, and the growth factor of 1.6 is used for the system stability term; set the weight of the energy consumption term to a compensation value with a sum of weights equal to one. Taking the initial weight configuration of 0.5 for the working condition simulation accuracy term, 0.3 for the energy consumption term, and 0.2 for the system stability term as an example, after adjustment, the weight of the working condition simulation accuracy term is 0.5×0.7 = 0.35, the weight of the system stability term is 0.2×1.6 = 0.32, and the weight of the energy consumption term is 1 - 0.35 - 0.32 = 0.33.

[0098] In addition, a state transition function based on the vibration energy imbalance is also used to smoothly transition the weights under each vibration condition. This transition function adopts the Sigmoid form, and this smooth transition mechanism effectively avoids the instability problem caused by sudden weight changes.

[0099] The adaptive weight coefficient calculation method of the present invention can intelligently adjust and optimize the target weights according to the real-time changes of vibration conditions, realize the adaptive control of the test system, accurately divide the working conditions into the balanced section, the transition section, and the unbalanced section through the evaluation of the vibration energy ratio and imbalance degree, and adopt a differentiated weight adjustment strategy; introduce a state transition function to achieve smooth weight transition, effectively avoiding the impact on the system caused by sudden changes; significantly improve the stability of the system under complex working conditions, shorten the response time to sudden vibrations, improve the vibration simulation accuracy, and enhance the safety and effectiveness of the valve durability test.

[0100] In an alternative embodiment,

[0101] By using a sensor array to collect the response data of the valve during vibration in real time, and processing the response data by using a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network, predicting the durability state of the valve includes:

[0102] Construct a three - layer cascaded attention network structure, which includes an axial attention layer, a working condition attention layer, and a time - series attention layer. The axial attention layer calculates the feature weights of the X - axis vibration unit, Y - axis vibration unit, and Z - axis vibration unit based on the vibration correlation matrix. The working condition attention layer calculates the feature weights of the equilibrium section, transition section, and non - equilibrium section based on the vibration energy distribution. The time - series attention layer calculates the time - series feature weights based on the historical state information; Input the response data into the three - layer cascaded attention network structure, and perform weighted calculations with the corresponding feature weights respectively to obtain a weighted feature vector;

[0103] Establish a two - stream feature fusion network, which includes a vibration feature stream and a durability evaluation stream. Input the weighted feature vector into the two - stream feature fusion network, where the vibration feature stream extracts the vibration state representation vector, and the durability evaluation stream extracts the performance degradation representation vector; Use a feature fusion matrix to fuse the vibration state representation vector and the performance degradation representation vector to obtain a fused feature vector;

[0104] Based on the fused feature vector, establish a loss function, which includes a durability state prediction loss term, a fault feature loss term, and a performance degradation trend loss term. Obtain the durability state prediction result of the valve by minimizing the loss function.

[0105] Exemplarily, the raw data collected by the sensor needs to be pre - processed, including filtering and denoising, feature extraction, and data standardization. Construct a three - layer cascaded attention network structure to focus on information in different dimensions.

[0106] The axial attention layer calculates the feature weights of the X - axis, Y - axis, and Z - axis vibration units based on the vibration correlation matrix. The working condition attention layer calculates the feature weights of the equilibrium section, transition section, and non - equilibrium section based on the vibration energy distribution. The time - series attention layer calculates the time - series feature weights based on the historical state information.

[0107] After inputting the pre - processed data into the three - layer cascaded attention network, perform three - layer weighted calculations in sequence: first, use the axial feature weights to weight the data of each axis; then use the working condition feature weights to weight the data of different working conditions; finally, use the time - series feature weights (based on the attenuation coefficient) to weight the time series. After three - layer weighting, reduce the dimension through a fully - connected layer to obtain a weighted feature vector.

[0108] A two-stream feature fusion network including a vibration feature stream and a durability evaluation stream is established, and the weighted feature vector is input into the two-stream feature fusion network to obtain a fused feature vector. A loss function is established based on the fused feature vector to predict the durability state of the valve. The loss function consists of three components: a durability state prediction loss term that measures the difference between the predicted remaining service life and the actual value through the mean square error; a fault feature loss term that measures the accuracy of the fault type prediction through cross-entropy; and a performance degradation trend loss term that measures the accuracy of the performance degradation rate prediction through the mean absolute error. The weights of the three loss terms are 0.5, 0.3, and 0.2 respectively, which can be adjusted according to actual applications.

[0109] The loss function is minimized by the gradient descent method, and the Adam optimizer is used to update the model parameters. The initial learning rate is 0.001, and the learning rate decays to 0.9 times the original value every 50 training epochs. Mini-batch training with a batch size of 32 is adopted, and the training is carried out for 200 epochs. Finally, three prediction results can be output: the estimated remaining service life of the valve (in hours), the probability distribution of potential fault types (including categories such as wear, corrosion, loosening, etc.) and their occurrence probabilities, and the performance degradation rate (the percentage of the performance index decrease per hour), which comprehensively constitute the prediction result of the valve's durability state.

[0110] Based on the durability prediction results, the energy distribution of the corresponding frequency bands in the working condition conversion matrix is adjusted; when the decline rate of the remaining life prediction value is abnormal, the vibration intensity is automatically reduced; the test cycle is adjusted according to the performance degradation rate. After each standard test cycle is completed, the vibration parameters are automatically updated to form a closed-loop control mechanism.

[0111] The existing valve durability test uses a single-layer attention mechanism and a simple neural network to process vibration data, which can only focus on single-dimensional features, has limited prediction accuracy, and lacks adaptability to different working conditions. The present invention innovatively proposes a three-level cascaded attention network structure that calculates feature weights from three dimensions: axial, working condition, and time series, enabling the system to adaptively focus on the most predictive information. At the same time, a two-stream feature fusion network is introduced to extract vibration features and performance degradation features respectively, and weighted fusion is performed through a feature fusion matrix, solving the problem that traditional single-stream networks cannot simultaneously focus on vibration states and degradation trends. By introducing a three-level cascaded attention mechanism and a two-stream feature fusion network, the accuracy and reliability of valve durability prediction are significantly improved. These improvements provide more reliable technical support for the predictive maintenance of valves, reduce maintenance costs, and improve equipment safety.

[0112] In an alternative embodiment,

[0113] The calculation of feature weights in the three-level cascaded attention network structure includes:

[0114] Calculate the eigenvalue decomposition of the vibration correlation matrix, obtain the principal eigenvector and the corresponding eigenvalue, construct an axial dynamic weight function based on the eigenvalue, and weight the axial dynamic weight function with the original features of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit to obtain the feature weights of the axial attention layer;

[0115] Calculate the vibration energy change rate using a sliding time window, extract the energy gradient feature at the moment of working condition switching, construct a working condition transition probability matrix based on the energy gradient feature, map the working condition transition probability matrix to the feature weight coefficients of the equilibrium section, transition section, and non-equilibrium section to obtain the feature weights of the working condition attention layer;

[0116] Calculate the decay coefficient based on the time interval between the historical state information and the current moment, and perform weighted accumulation of the decay coefficient and the historical state information to obtain the feature weights of the time series attention layer.

[0117] Exemplarily, perform eigenvalue decomposition on the vibration correlation matrix to obtain the principal eigenvector and the corresponding eigenvalue. For example, use the Jacobi iterative method to solve, and the obtained eigenvalues are 2.15, 0.52, and 0.33 respectively, and the corresponding eigenvectors are [0.62, 0.65, 0.44], [-0.56, 0.12, 0.82], and [0.55, -0.75, 0.37] respectively. Construct an axial dynamic weight function based on the eigenvalue. First, normalize the eigenvalue to obtain the weight coefficients 0.72, 0.17, and 0.11. Then multiply these coefficients by the eigenvector to calculate the weight contribution of each axis. Finally, the weight of the X-axis is 0.32, the weight of the Y-axis is 0.32, and the weight of the Z-axis is 0.36. Weight the axial dynamic weight function with the original features of each axis. Multiply the original features of the X-axis, Y-axis, and Z-axis vibration units by the corresponding weight values to obtain the feature weights of the axial attention layer. These weighted features can better reflect the influence degree of the vibration of each axis on the durability of the valve.

[0118] The working condition attention layer calculates the feature weights of the equilibrium section, transition section, and non-equilibrium section based on the vibration energy distribution. The system uses a sliding time window (window length is 2 seconds, overlap rate is 50%) to calculate the vibration energy change rate. When the change rate exceeds the preset threshold (usually 30%), it is marked as a working condition switching point. Extract the energy gradient feature at the switching point and construct a 3×3 working condition transition probability matrix. This matrix represents the conversion probability between each working condition. For example, the probability of transitioning from the equilibrium section to the transition section is 0.25, and the probability of transitioning from the transition section to the non-equilibrium section is 0.30. According to this matrix, calculate the feature weight of the equilibrium section to be 0.50, the feature weight of the transition section to be 0.30, and the feature weight of the non-equilibrium section to be 0.20.

[0119] The temporal attention layer calculates the feature weights based on the relationship between the historical state information and the current state. The system maintains a historical queue containing the state information of the past 10 time points, with a time interval of 1 second between time points, covering the data of the past 10 seconds. The decay coefficient is calculated based on the time interval between the current moment and the historical moment. The exponential decay function is used: when the time interval is t seconds, the decay coefficient is equal to 0.95 to the power of t. For example, when the time interval is 1 second, the decay coefficient is 0.95; when the time interval is 2 seconds, the decay coefficient is 0.9025; and so on. The decay coefficient is weighted and accumulated with the historical state information to obtain the feature weights of the temporal attention layer. For each state in the historical queue, the product of its feature vector and the decay coefficient is calculated, and then all the weighted vectors are added and normalized. This mechanism enables the system to adaptively focus on information at different time scales, taking into account both the immediacy of recent states and the long-term impact of distant states.

[0120] Through the calculation of the feature weights of the three-layer cascaded attention structure, the present invention can simultaneously focus on the spatial distribution characteristics, operating condition change characteristics, and temporal evolution characteristics of vibrations, comprehensively capture the key information of the valve durability state, and provide a solid foundation for accurate prediction.

[0121] In an alternative embodiment,

[0122] Establishing a two-stream feature fusion network and inputting the weighted feature vector into the two-stream feature fusion network to obtain a fusion feature vector includes:

[0123] Using a temporal convolutional network with three-layer residual structure and increasing convolutional kernel size with the number of layers to construct a vibration feature stream, performing feature transformation on the weighted feature vector through multi-layer convolutional operations, and calculating attention weights based on the frequency domain energy distribution to perform weighted combination on the feature transformation results to obtain a vibration state representation vector;

[0124] Using a long short-term memory network with a bidirectional structure and a skip connection mechanism to construct a durability evaluation stream, performing temporal modeling on the weighted feature vector based on the forget gate state and the input gate state to obtain a hidden state, calculating temporal weights based on the temporal change trend of the hidden state to perform weighted accumulation on the hidden state to obtain a performance degradation representation vector;

[0125] Performing a concatenation operation on the vibration state representation vector and the performance degradation representation vector, using the attention mechanism to calculate the inter-channel correlation degree of the concatenated features to obtain a feature fusion matrix, calculating a feature fusion importance score based on the feature fusion matrix, constructing a dynamic fusion weight based on the feature fusion importance score, performing a weighted operation on the dynamic fusion weight and the feature fusion matrix, and introducing a residual connection to obtain a fusion feature vector.

[0126] Exemplarily, as one of the core technologies of the present invention, the dual-stream feature fusion network aims to simultaneously extract the short-term dynamic features of vibration signals and the long-term evolution trend of durability, and adaptively fuse the two types of features to achieve accurate prediction of durability states.

[0127] The vibration feature stream is designed using a temporal convolutional network structure to extract the frequency-domain and time-domain features of vibration signals. This network contains three residual structures, and the convolutional kernel size increases with the number of layers to capture vibration patterns at different time scales. The first residual structure uses 32 convolutional kernels of size 3, and the input is a 128-dimensional weighted feature vector. After convolutional operations, a feature map with 32 channels is obtained. The stride of the convolutional operation is 1, and the padding is 1 to ensure that the output dimension is the same as the input. Then, non-linearity is introduced through the ReLU activation function, and at the same time, Batch Normalization is used for feature normalization to help the network converge faster. The output of the first residual structure serves as the input to the second residual structure, and at the same time, it is added to the output of the first layer through a skip connection to form a residual connection. The second residual structure uses 64 convolutional kernels of size 5, with a stride of 1 and a padding of 2. After convolution, ReLU activation and Batch Normalization are performed to obtain a feature map with 64 channels. Similar to the first layer, the output of the second layer is also added to its input through a residual connection. The third residual structure uses 128 convolutional kernels of size 7, with a stride of 1 and a padding of 3. After the same activation and normalization operations, a feature map with 128 channels is obtained. The design of the three residual structures enables the network to learn vibration features at different scales while alleviating the problem of vanishing gradients.

[0128] After the feature transformation is completed, attention weights are calculated based on the frequency-domain energy distribution. The feature map output by the temporal convolutional network is subjected to a frequency-domain transformation to calculate the energy proportion of each frequency component. For a certain test sample, the energy distribution of the detected main frequency components is as follows: 15% for 5 - 10 Hz, 35% for 10 - 20 Hz, 40% for 20 - 50 Hz, and 10% for 50 - 100 Hz. According to the energy distribution, the attention weights for each channel are calculated, and the weight values are proportional to the energy proportion of the frequency components. The attention weights are combined with the feature transformation results through weighted combination to obtain a vibration state representation vector. The dimension of this vector is 256, which contains the main feature information of the vibration signal. In practical applications, the first 32 dimensions of this vector mainly represent low-frequency vibration characteristics, the middle 128 dimensions represent medium-frequency vibration characteristics, and the last 96 dimensions represent high-frequency vibration characteristics.

[0129] The durability assessment process uses a long short-term memory (LSTM) network with a bidirectional architecture and skip connections to capture the long-term evolution of vibration characteristics. The network input is a 128-dimensional weighted feature vector processed by a three-layer cascaded attention network. This vector is first mapped to a 64-dimensional vector through a linear projection layer, which serves as the input to the LSTM network. The LSTM network consists of two layers, each with 128 hidden units. The first layer is a bidirectional structure with 64 hidden units each in the forward and reverse directions; the second layer is a unidirectional structure with 128 hidden units. Within the LSTM unit, the system calculates the forget gate, input gate, and output gate states based on the current input and the previous hidden state. The forget gate determines which historical information should be forgotten, the input gate determines which new information should be remembered, and the output gate determines which information should be output. For example, in one test, when a sudden increase in vibration energy was detected, the input gate state value increased to above 0.85, indicating that the system was paying close attention to the current information. During stable operation, the forget gate state value remained around 0.35, indicating that the system retained most historical information. The skip connection mechanism of the LSTM network involves directly adding the output of the first layer to the output of the second layer. This design effectively alleviates the vanishing gradient problem while preserving both short-term and long-term temporal information. The network ultimately outputs a 128-dimensional sequence of hidden states, one hidden state at each time step. A temporal weight is calculated based on the temporal trend of the hidden state. This trend is measured by the difference between the hidden states at adjacent time steps, with larger differences indicating more drastic state changes. In a valve test, the system detected an average hidden state change rate of 0.08 during normal performance periods, while the average change rate during performance degradation periods was as high as 0.27. Based on this trend, a temporal weight was calculated for each time step, with time steps with more drastic changes receiving higher weights. The temporal weights are then weighted and accumulated with the hidden states to produce a performance degradation representation vector. This vector, with a dimension of 128, contains information about the long-term evolution of the valve's durability state.

[0130] The feature fusion process first concatenates the vibration state representation vector and the performance degradation representation vector to produce a 384-dimensional joint feature vector. An attention mechanism is used to calculate the inter-channel correlation of the concatenated features. The joint feature vector is then processed through a multi-layer perceptron to produce a 384×384 feature fusion matrix. Each element of this matrix represents the strength of the correlation between the two feature dimensions. For example, in one test, the correlation between the resonant frequency in the vibration feature and the friction coefficient growth rate in the performance degradation feature reached a high of 0.78, indicating a strong correlation between the two.

[0131] Calculate the feature fusion importance score through the feature fusion matrix. Sum each row of the feature fusion matrix to obtain 384 importance scores, which reflect the importance of each dimension feature in the fusion process. Construct the dynamic fusion weight based on the feature fusion importance score. Process the importance score through Softmax to obtain the weight vector, where the weight values are between 0 and 1 and the sum is 1. For example, in the dynamic fusion weight obtained from a certain test, the average weight of the vibration state characterization part is 0.65, and the average weight of the performance degradation characterization part is 0.35, indicating that the vibration feature contributes more to the prediction result in this sample. Perform a weighted operation on the dynamic fusion weight and the feature fusion matrix to obtain the weighted fusion result. At the same time, introduce a residual connection, add the original spliced feature vector and the weighted fusion result to obtain the final fused feature vector.

[0132] Existing technologies usually use neural networks with a single structure for feature extraction, and fail to simultaneously focus on the short-term dynamic characteristics of vibration signals and the long-term evolution law of durability states. Figure 3 It is a comparison chart of the prediction accuracy of different methods under typical working conditions. It can be seen from the figure that the dual-stream feature fusion network proposed by the present invention has the optimal prediction performance under various working conditions: under normal operating conditions, the prediction accuracy of the present invention reaches 95%, which is 13% and 8% higher than that of CNN + weighted average fusion and LSTM + attention fusion respectively; under start-stop transition conditions, the prediction accuracy of the present invention is 92%, and the improvement ranges are 17% and 11% respectively; under extreme environments, the prediction accuracy is 88%, and the improvement range is 20% and 15%; under fault conditions, the prediction accuracy is 90%, and the improvement ranges are 19% and 14%. The results show that the present invention effectively extracts temporal and spatial features through a dual-stream architecture and adopts an innovative feature fusion strategy, significantly improving the prediction accuracy of the system under various working conditions, especially showing stronger robustness under complex working conditions. The dual-stream feature fusion network proposed in this application extracts short-term dynamic features and long-term evolution trends through the vibration feature stream and the durability evaluation stream respectively, and realizes adaptive fusion through the feature fusion matrix, realizing the efficient utilization of multi-dimensional vibration data and the collaborative analysis of multi-scale features, overcoming the limitation of insufficient prediction accuracy of traditional methods under complex working conditions.

[0133] In the second aspect of the embodiments of the present invention,

[0134] Provide a valve durability test system based on a vibration environment, including:

[0135] The first unit is used to perform vibration loading on the valve by using a multi-degree-of-freedom vibration excitation mechanism, and the multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit, and a Z-axis vibration unit that are cooperatively controlled;

[0136] The second unit is configured to generate a multi-directional vibration superposition by controlling the X-axis vibration unit, the Y-axis vibration unit, and the Z-axis vibration unit based on the multi-degree-of-freedom vibration excitation mechanism, adopting a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping, and dynamically adjusting the vibration direction and vibration parameters through multi-objective adaptive optimization in combination with the multi-directional vibration superposition;

[0137] The third unit is configured to collect the response data of the valve during vibration in real time through a sensor array, and process the response data by using a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network to predict the durability state of the valve; based on the durability prediction result, optimize and adjust the composite vibration mode generation algorithm to form a closed-loop control for durability testing.

[0138] In the third aspect of the embodiments of the present invention,

[0139] A kind of electronic device is provided, including:

[0140] A processor;

[0141] A memory for storing instructions executable by the processor;

[0142] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0143] In the fourth aspect of the embodiments of the present invention,

[0144] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0145] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0146] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A durability test method for a valve based on a vibration environment, characterized in that Including: A multi-degree-of-freedom vibration excitation mechanism is used to apply vibration loading to the valve. The multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit, and a Z-axis vibration unit that are coordinately controlled; Based on the multi-degree-of-freedom vibration excitation mechanism, a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping is adopted. By controlling the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit to generate multi-directional vibration superposition, the vibration direction and vibration parameters are dynamically adjusted by combining multi-objective adaptive optimization; The response data of the valve during vibration is collected in real time through a sensor array. A three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network are used to process the response data to predict the durability state of the valve; Based on the durability prediction result, the composite vibration mode generation algorithm is optimized and adjusted to form a closed-loop control for durability testing.

2. The method according to claim 1, wherein The multi-degree-of-freedom vibration excitation mechanism includes: The X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit all include a driver, a displacement sensor, and an acceleration sensor, and vibration state data is obtained in real time through a multi-channel data acquisition module; The X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit adopt a hierarchical control architecture, including bottom-layer execution control, middle-layer coordinated control, and upper-layer motion planning. Among them, the bottom-layer execution control uses an independent PID controller, and the middle-layer coordinated control calculates the mutual interference amount between each axis based on the current vibration state data and compensates it in real time. The upper-layer motion planning generates a vibration trajectory according to the test requirements; A phase synchronization control and adaptive amplitude control algorithm based on phase-locked loop technology is used for vibration precision control. Among them, the phase synchronization control calculates the error between the target phase and the actual phase, and adjusts the control amount in real time according to the proportional-integral control strategy; The adaptive amplitude control iteratively updates the output amplitude by comparing the ratio of the target amplitude to the measured amplitude.

3. The method according to claim 1, characterized in that, Adopting a composite vibration mode generation algorithm based on inter-axis coupling correction and working condition conversion mapping, by controlling the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit to generate multi-directional vibration superposition, and dynamically adjusting the vibration direction and vibration parameters by combining multi-objective adaptive optimization includes: Establish a composite vibration mode generation algorithm. By frequency domain analysis and principal component extraction, a vibration working condition library of the valve usage environment is constructed. Calculate the vibration correlation matrix of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit, and introduce a coupling correction coefficient to correct the vibration transfer function. The coupling correction coefficient characterizes the ratio relationship between the actual amplitude and the theoretical amplitude between different axial vibration units; The vibration characteristic parameters of the valve in normal operation, start-stop transition, extreme environment, and fault simulation in the actual environment are mapped to the vibration frequencies of each axis of the multi-degree-of-freedom vibration excitation mechanism through a working condition conversion matrix. The working condition conversion matrix includes vibration type, frequency parameters, and duration; According to the vibration frequencies of each axis, the synthetic acceleration amplitude and direction vector are calculated through spatial vibration synthesis, and a vibration energy distribution algorithm is used to determine the vibration energy of each axis; Construct a multi-objective optimization function that includes a working condition simulation accuracy term, an energy consumption term, and a system stability term based on the synthetic acceleration amplitude and direction vector, and use an adaptive weight coefficient to dynamically balance each optimization term; the working condition simulation accuracy term is evaluated through time-domain signal similarity, frequency-domain error, and statistical feature comparison; based on the optimization results of the multi-objective optimization function, dynamically update the vibration frequencies, amplitudes, and phase parameters of each vibration unit to achieve adaptive adjustment of the vibration direction and vibration parameters.

4. The method according to claim 3, characterized in that, The calculation of the adaptive weight coefficient includes: Calculate the vibration energy proportion of each vibration unit, where the vibration energy proportion is the ratio of the vibration energy of each vibration unit within a preset time window to the total vibration energy, and calculate the vibration energy imbalance degree based on the vibration energy proportion. The vibration energy imbalance degree is the maximum value of the difference in vibration energy proportions between any two vibration units; Divide the vibration working conditions into an equilibrium section, a transition section, and a non-equilibrium section according to the vibration energy imbalance degree; in the equilibrium section, positively adjust the weight of the working condition simulation accuracy term, negatively adjust the weight of the energy consumption term, and keep the weight of the system stability term unchanged; in the transition section, exponentially decay the weight of the working condition simulation accuracy term, keep the weight of the energy consumption term unchanged, and set the weight of the system stability term to a compensation value with a sum of weights of one; in the non-equilibrium section, exponentially adjust the weights of the working condition simulation accuracy term and the system stability weight term, and set the weight of the energy consumption term to a compensation value with a sum of weights of one; use a state transition function based on the vibration energy imbalance degree to smoothly transition the weights under each vibration working condition to obtain the adaptive weight coefficient.

5. The method according to claim 4, wherein Real-time collect the response data of the valve during vibration through a sensor array, and use a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network to process the response data. The prediction of the durability state of the valve includes: Construct a three-level cascaded attention network structure, which includes an axial attention layer, a working condition attention layer, and a time series attention layer. The axial attention layer calculates the feature weights of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit based on the vibration correlation matrix. The working condition attention layer calculates the feature weights of the equilibrium section, transition section, and non-equilibrium section based on the vibration energy distribution. The time series attention layer calculates the time series feature weights based on the historical state information; input the response data into the three-level cascaded attention network structure, and calculate the weighted feature vectors by weighted calculation with the corresponding feature weights respectively; Establish a dual-stream feature fusion network, which includes a vibration feature stream and a durability evaluation stream. Input the weighted feature vector into the dual-stream feature fusion network, where the vibration feature stream extracts the vibration state characterization vector, and the durability evaluation stream extracts the performance degradation characterization vector; use a feature fusion matrix to fuse the vibration state characterization vector and the performance degradation characterization vector to obtain a fused feature vector; A loss function is established based on the fused feature vector. The loss function includes a durability state prediction loss term, a fault feature loss term, and a performance degradation trend loss term. The durability state prediction result of the valve is obtained by minimizing the loss function.

6. The method according to claim 5, characterized in that, The calculation of the feature weights in the three-level cascaded attention network structure includes: Calculating the eigen-decomposition of the vibration correlation matrix to obtain the principal eigenvector and the corresponding eigenvalue, constructing an axial dynamic weight function based on the eigenvalue, and weighting the original features of the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit with the axial dynamic weight function to obtain the feature weights of the axial attention layer; Calculating the vibration energy change rate using a sliding time window, extracting the energy gradient feature at the moment of working condition switching, constructing a working condition transition probability matrix based on the energy gradient feature, and mapping the working condition transition probability matrix to the feature weight coefficients of the equilibrium section, transition section, and non-equilibrium section to obtain the feature weights of the working condition attention layer; Calculating the decay coefficient based on the time interval between the historical state information and the current moment, and performing weighted accumulation of the decay coefficient and the historical state information to obtain the feature weights of the temporal attention layer.

7. The method according to claim 5, wherein Establishing a two-stream feature fusion network, and inputting the weighted feature vector into the two-stream feature fusion network to obtain the fused feature vector includes: Constructing a vibration feature stream using a temporal convolutional network with three-layer residual structure and the convolutional kernel size increasing with the number of layers, performing feature transformation on the weighted feature vector through multi-layer convolutional operations, and calculating the attention weight based on the frequency domain energy distribution to perform weighted combination on the feature transformation result to obtain the vibration state representation vector; Constructing a durability evaluation stream using a long short-term memory network with a bidirectional structure and a skip connection mechanism, performing temporal modeling on the weighted feature vector based on the forget gate state and the input gate state to obtain the hidden state, and calculating the temporal weight based on the temporal change trend of the hidden state to perform weighted accumulation on the hidden state to obtain the performance degradation representation vector; Performing a concatenation operation on the vibration state representation vector and the performance degradation representation vector, calculating the inter-channel correlation of the concatenated features using the attention mechanism to obtain the feature fusion matrix, calculating the feature fusion importance score through the feature fusion matrix, constructing a dynamic fusion weight based on the feature fusion importance score, performing weighted operation on the feature fusion matrix with the dynamic fusion weight, and introducing a residual connection to obtain the fused feature vector.

8. A valve durability test system based on a vibration environment, which is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that, Including: The first unit is used to perform vibration loading on the valve using a multi-degree-of-freedom vibration excitation mechanism, and the multi-degree-of-freedom vibration excitation mechanism includes an X-axis vibration unit, a Y-axis vibration unit, and a Z-axis vibration unit under collaborative control; The second unit is used to, based on the multi-degree-of-freedom vibration excitation mechanism, adopt a composite vibration mode generation algorithm based on axial coupling correction and working condition conversion mapping, generate multi-directional vibration superposition by controlling the X-axis vibration unit, Y-axis vibration unit, and Z-axis vibration unit, and dynamically adjust the vibration direction and vibration parameters in combination with multi-objective adaptive optimization. The third unit collects the response data of the valve during vibration in real time through a sensor array, and processes the response data by using a three-level cascaded attention network structure including an axial attention layer, a working condition attention layer, and a time series attention layer and a dual-stream feature fusion network to predict the durability state of the valve; based on the durability prediction result, optimize and adjust the composite vibration mode generation algorithm to form a closed-loop control for durability testing.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images