Residual life prediction method and system for engineering mechanical structural member
By combining multi-source sensor data acquisition with physical memory neural networks, the problem of loose model fusion in the life prediction of mechanical structural components is solved, achieving accurate damage monitoring and life prediction under complex working conditions, and it remains effective even in small sample situations.
Patent Information
- Application Number
- CN202511706703.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-04-03
AI Technical Summary
In the current technology for life prediction of mechanical structural components, the integration of physical models and data-driven methods is not close enough, resulting in inaccurate capture of damage evolution characteristics under complex working conditions and difficulty in accurately predicting the remaining life.
Multi-source sensors are used to collect multi-dimensional data. Time-frequency domain features are extracted by combining natural frequency and fault characteristic frequency. Wavelet packet analysis is used to obtain a high-dimensional degradation feature set. Adaptive dimensionality reduction optimization is performed by physical memory neural network CP-NMN. Remaining lifetime is predicted by combining long short-term memory network. Physical constraints are introduced to improve prediction accuracy.
It enables precise monitoring of damage to mechanical structural components and accurate life prediction under complex working conditions, reducing prediction errors, and effectively assessing health status even with small sample sizes.
Smart Images
Figure CN121786612A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical data processing technology, and in particular to a method and system for predicting the remaining life of engineered mechanical structural components. Background Technology
[0002] With the development of modern industry, the reliability and safety of various mechanical equipment in industrial production are directly related to production efficiency, economic benefits, and personnel safety. As a core component of mechanical equipment, the health status monitoring and remaining life prediction of mechanical structural components have become a research focus. Although the integration of physical models and data-driven methods has become a research trend, most current integration methods simply combine the two, failing to fully utilize the constraints and guidance of physical knowledge and the flexibility of data-driven methods. During model training, the integration of physical knowledge and data features is not tight enough, resulting in inaccurate capture of the damage evolution characteristics of mechanical structural components under complex working conditions, making it difficult to accurately predict remaining life.
[0003] Currently, the methods for predicting the remaining life of mechanical structural components mainly rely on fixed algorithm models. These methods establish mathematical models of mechanical structural components and calculate parameters such as stress and strain under given working conditions based on the mechanical properties of materials and the dynamic characteristics of the structure, thereby inferring their lifespan. However, with the increasing complexity of mechanical equipment, the limitations of fixed algorithm models become more and more apparent. When dealing with complex mechanical structural components, traditional methods based on fixed algorithm models face great difficulty in accurately establishing physical models due to the diversity and uncertainty of actual working conditions. These models are difficult to describe precisely with simple mathematical formulas, resulting in a large deviation between the calculation results based on ideal models and the actual situation. Therefore, this paper proposes a method and system for predicting the remaining life of engineered mechanical structural components. Summary of the Invention
[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: A method for predicting the remaining service life of engineered mechanical structural components, comprising: S1: Collect multi-dimensional raw datasets of mechanical structural components using multi-source sensors; S2: Time-frequency domain feature extraction is performed by matching the multi-dimensional original dataset with the inherent frequency and the fault characteristic frequency, and a high-dimensional degradation feature set is obtained by wavelet packet analysis and combining it with the characterization of damage evolution. S3: Adaptive dimensionality reduction optimization of high-dimensional feature sets is performed using the pre-trained physical memory neural network fusion model CP-NMN to obtain the same-dimensional health index across working conditions. Set the high-dimensional degradation features as As input to the CP-NMN model, the features are transformed by combining physical constraints with the various layers of the neural network, achieving dimensionality reduction and constraint fusion. The structure of the physical memory neural network fusion model CP-NMN is represented as follows: ; in, Indicates the first The output of the layer, and These are the weight matrix and bias vector of the l-th layer, respectively, and σ is the activation function. This refers to the physical memory constraint term of the l-th layer; S4: Based on the same-dimensional health indicators, a long short-term memory network is used to dynamically predict the remaining lifespan, and a preset lifespan threshold is used to achieve engineering early warning.
[0005] The time-frequency domain features include spectral peak value, center frequency, and spectral entropy.
[0006] The process for obtaining the spectrum peak value is as follows: For the original multidimensional raw dataset Perform a Fourier transform and set the frequency resolution to 10Hz. Based on each frequency point k after the discrete Fourier transform, calculate the spectral amplitude obtained from the discrete Fourier transform. Its corresponding frequency value is Iterate through the values of k, when When the frequency falls within the range of 100-500Hz or 2000-5000Hz, record the spectral amplitude. The maximum value is taken as the peak value of the spectrum. .
[0007] The process for obtaining the center frequency is as follows: Determine the set of frequency points to be used in the calculation Where M is the number of frequency points, for each frequency point Corresponding spectral amplitude Introducing the vibration transmission coefficient at the damaged site For gear meshing, the transmission coefficient For the bearing raceway, the transmission coefficient ; Assign a corresponding transmission coefficient based on the damage location corresponding to frequency point k. Value to obtain center frequency Center frequency The calculation formula is: .
[0008] The process of obtaining the spectral entropy is as follows: For each frequency point k, calculate its corresponding spectral amplitude. Then calculate the energy at that frequency point. And obtain the total energy of all 512 frequency points. ; Let the fault characteristic frequency be For each frequency point k, determine Whether the condition is met or not, if it is met, the energy proportion at that frequency point is weighted and enhanced with an enhancement coefficient of 1.2; otherwise, the energy proportion is calculated normally. The calculation formula is: ; Based on energy percentage Calculate the spectral entropy : , where i is the index.
[0009] The process of obtaining a high-dimensional degenerative feature set representing damage evolution is as follows: Time-frequency domain features The original signal is subjected to wavelet packet decomposition, and energy features are extracted from each sub-band signal obtained by decomposition. Let the signal of the i-th sub-band in the j-th layer be... and obtain time-frequency domain features. discrete-time signal energy The calculation formula is: The extracted time-frequency domain features High-dimensional energy features obtained from wavelet packet decomposition To integrate the features, suppose there are m time-frequency domain features, and wavelet packet decomposition yields n sub-band energy features. Then the integrated high-dimensional degenerate feature set... Represented as: in, This is a transpose.
[0010] The training process of the physical memory neural network fusion model is as follows: Physical memory constraint Including the constraints of the physical laws governing damage evolution and structural dynamics constraints Physical memory constraint terms It is embedded into the structure and training process of the neural network in the form of weight bias initialization and loss function regularization, by minimizing the physical loss function of the model. Optimize the network's weight matrix and bias vector 7. Output low-dimensional cross-condition same-dimensional health index H; Theoretical loss function Based on the fatigue Miner's rule for fatigue damage accumulation, let the physical quantity output by the model be... The physical quantity output according to the theoretical relationship of Miner's rule is The physical loss is then expressed as: m is the number of physical constraints. The physical quantity output by the physical memory constraint term model; After a predetermined number of training rounds, the CP-NMN model can adaptively process high-dimensional degenerate features, transforming multi-dimensional high-dimensional degenerate feature sets. The table is transformed into a low-dimensional, same-dimensional health indicator H.
[0011] The physical laws governing damage evolution constrain The implementation process is as follows: Based on Miner's law of fatigue, the cumulative fatigue damage satisfies... ,in, This represents the actual number of loops. For fatigue life, the physical laws governing damage evolution are constrained. satisfy: ; in, This indicates the damage characteristics corresponding to the current loop count.
[0012] The structural dynamics constraints The implementation process is as follows: Determine the natural frequency of mechanical structural components And calculate the natural angular frequency. The formula is: ; For the l-th layer of the CP-NMN model, to ensure that the frequency-related features obtained after processing this layer do not produce resonant errors with the natural frequency, structural dynamics constraints are introduced. Requirements for structural dynamic characteristics satisfy: .
[0013] A system for predicting the remaining life of engineered mechanical structural components, comprising: Data acquisition module: Collects multi-dimensional raw datasets of mechanical structural components through multi-source sensors; High-dimensional acquisition module: Based on the matching of multi-dimensional original datasets, time-frequency domain features are extracted by combining inherent frequency and fault feature frequency, and high-dimensional degradation feature set is obtained by wavelet packet analysis and combined with characterization of damage evolution; Same-dimensional decision module: The pre-trained physical memory neural network fusion model CP-NMN is used to adaptively reduce the dimensionality of the high-dimensional feature set to obtain same-dimensional health indicators across working conditions. Lifespan prediction module: Based on the same-dimensional health indicators, it uses a long short-term memory network to dynamically predict the remaining lifespan and presets a lifespan threshold to achieve engineering early warning.
[0014] The present invention has the following beneficial effects: In this invention, firstly, by matching and combining the inherent frequency and the fault characteristic frequency through a multi-dimensional raw dataset, time-frequency domain feature extraction can be performed, which can accurately locate the key frequency information related to the damage of mechanical structural components. Then, by wavelet packet analysis and combining it with the characterization of damage evolution, a high-dimensional degradation feature set is obtained, which comprehensively extracts the features of the signal in different frequency bands. This not only covers the entire stage evolution of damage from very early small defects to mid-term extended damage and then to late-term critical failure, but also reveals the inherent evolution logic of damage from local defects to global degradation, providing a rich information foundation for subsequent remaining life prediction. Secondly, the physical knowledge of mechanical structural components is integrated into the neural network. During the training process, the physical loss term is used to ensure that the model output conforms to the actual damage evolution law. In fatigue damage prediction, the physical loss function is constructed based on the Miner rule of fatigue damage accumulation, so that the model follows the physical mechanism of fatigue damage while learning data features, which effectively improves the accuracy of predicting the remaining life of mechanical structural components. Compared with the simple data-driven model, the prediction error is greatly reduced. Finally, even with small sample sizes, the CP-NMN model can leverage prior physical information to compensate for insufficient data, still enabling relatively accurate health status assessments and lifespan predictions. For example, in the early monitoring phase of new mechanical structural components, traditional data-driven models struggle to function effectively due to limited operational data, while the CP-NMN model, constrained by physical knowledge, can achieve effective monitoring of structural component health status and lifespan prediction. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the method and system for predicting the remaining life of engineered mechanical structural components proposed in this invention.
[0016] Figure 2 This invention presents a system block diagram of a method and system for predicting the remaining life of engineered mechanical structural components. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Example 1: As Figure 1 As shown, the present invention proposes a method for predicting the remaining life of engineered mechanical structural components, comprising: S1: Collect multi-dimensional raw datasets of mechanical structural components using multi-source sensors; Deploy multi-source sensors to collect multi-dimensional raw data. Multi-source sensors include vibration sensors, temperature sensors, pressure sensors, and acoustic emission sensors, etc. The data acquisition time interval is set according to the operating frequency of the mechanical components and the monitoring requirements. For high-speed rotating mechanical components, the time interval is set to the millisecond level, and for components that run smoothly and change slowly, the time interval is set to the second or minute level. By continuously collecting this multi-source data, a multi-dimensional raw dataset is constructed. Let the multi-dimensional raw dataset of the mechanical structural component at time t be... Where n is the number of sensors. This represents the raw data collected by the i-th sensor at time t.
[0019] S2: Time-frequency domain feature extraction is performed by matching the multi-dimensional original dataset with the inherent frequency and the fault characteristic frequency, and a high-dimensional degradation feature set is obtained by wavelet packet analysis and combining it with the characterization of damage evolution. The process of extracting time-frequency domain features by combining the natural frequency and the fault characteristic frequency is as follows: Time-frequency domain characteristics include spectral peak value, center frequency, and spectral entropy; When performing Fourier transform on the original time-domain signal, the frequency resolution is set to 10Hz, the meshing frequency of the structural gear is 480Hz, and its sideband interval is 10Hz. This resolution can accurately capture sideband changes. Only the peak values within 100-500Hz (gear meshing frequency band) and 2000-5000Hz (bearing impact frequency band) are extracted, while irrelevant frequency band interference is filtered out; For the original multidimensional raw dataset To perform a Fourier transform and set the frequency resolution to 10Hz, first clarify the relationship between the sampling frequency f and the number of sampling points N. Then, based on the definition of frequency resolution, the frequency resolution... This refers to the interval between adjacent frequency points after the Fourier transform, and its calculation formula is: ; Spectrum peak It refers to the maximum value of the spectral amplitude within a specific frequency band, extracting the peak values within the 100-500Hz gear meshing frequency band and the 2000-5000Hz bearing impact frequency band, and filtering out interference from irrelevant frequency bands; First, based on each frequency point k after the Discrete Fourier Transform, the spectral amplitude obtained from the Discrete Fourier Transform... Its corresponding frequency value is Iterate through the values of k, when When the frequency falls within the range of 100-500Hz or 2000-5000Hz, record the spectral amplitude. The maximum value is taken as the peak value of the spectrum. ; Center frequency The index reflects the main frequency range of signal energy concentration. When calculating it, the vibration transmission coefficient of the damaged part of the structure needs to be considered so that the energy weight is more in line with the actual damage energy distribution. First, determine the set of frequency points to be used in the calculation. Where M is the number of frequency points, for each frequency point Corresponding spectral amplitude Introducing the vibration transmission coefficient at the damaged site For gear meshing, the transmission coefficient For the bearing raceway, the transmission coefficient ; Assign a corresponding transmission coefficient based on the damaged location (gear or bearing) at frequency point k. Value to obtain center frequency Center frequency The calculation formula is: ; Specifically, in the calculation process, for each frequency point k, the calculation is first performed. Then the molecule is calculated as k and The sum of the products, with the denominator calculated as follows: The sum of these terms is used to calculate the center frequency by dividing the numerator by the denominator. Thus, due to the introduction of the vibration transmission coefficient, the energy weight at the gear meshing point is amplified because of the transmission coefficient. A value greater than 1 makes the calculated center frequency more reflective of the energy distribution changes caused by damage at the gear meshing parts, and more closely matches the actual damage energy distribution. Spectral entropy To reflect the degree of disorder in the spectrum, the number of frequency points M is first fixed at 512, corresponding to the 0-5kHz frequency band, matching the sensor's acquisition bandwidth, and then the obstacle characteristic frequency neighborhood is analyzed. The energy percentage is weighted and enhanced. For each frequency point k, calculate its corresponding spectral amplitude. Then calculate the energy at that frequency point. And obtain the total energy of all 512 frequency points. ; Let the fault characteristic frequency be For each frequency point k, determine whether it is in the neighborhood of the fault characteristic frequency. Internal, judgment If the condition is met, a weighted amplification is applied to the energy proportion at that frequency point, with an amplification coefficient of 1.2. Otherwise, the energy proportion is calculated normally. The calculation formula is: ; Based on energy percentage Calculate the spectral entropy : ; Specifically, by using a base-2 logarithm, the spectral entropy unit is bits. At the same time, by weighting and enhancing the energy proportion of the fault characteristic frequency neighborhood, these fault-related frequency components contribute more to the calculation of spectral entropy, thereby strengthening the entropy contribution of the damaged frequency and better reflecting the degree of disorder in the spectrum when the mechanical structure fails. The final time-frequency domain features are expressed as follows ; The process of obtaining a high-dimensional degenerative feature set representing damage evolution is as follows: Time-frequency domain features from which spectral peak, center frequency, and spectral entropy features were extracted. The original signal is subjected to wavelet packet decomposition. After wavelet packet decomposition, energy features are extracted from each sub-band signal obtained by decomposition. Let the signal of the i-th sub-band in the j-th layer be... and obtain time-frequency domain features. discrete-time signal energy The calculation formula is: ; Specifically, by calculating time-frequency domain features Discrete-time signal energy of each subband of the original signal A set of high-dimensional energy features based on wavelet packet decomposition was obtained. These high-dimensional energy characteristics This reflects the energy distribution of the signal in different frequency subbands. The damage evolution of mechanical structural components will cause the energy of the signal in a specific frequency subband to change. For example, early damage may increase the energy of the high-frequency subband, and as the damage develops, the energy will transfer to the low-frequency subband. Extracted time-frequency domain features (Including spectral peak, center frequency, spectral entropy, etc.) and high-dimensional energy features obtained from wavelet packet decomposition To integrate the features, suppose there are m time-frequency domain features, and wavelet packet decomposition yields n sub-band energy features. Then the integrated high-dimensional degenerate feature set... Represented as: in, For transpose; Specifically, the high-dimensional degradation feature set obtained through integration includes time-frequency domain statistical characteristics and wavelet packet multi-resolution energy distribution, which can reflect information on the damage evolution of mechanical structural components from multiple perspectives. The spectral entropy, center frequency, and spectral entropy in the time-frequency domain features are good at capturing the macroscopic trend of damage and the overall law from normal to failure. The high-dimensional energy features of wavelet packet energy are good at capturing the microscopic details of damage and the local signal changes of early small defects.
[0020] S3: Adaptive dimensionality reduction optimization of high-dimensional feature sets is performed using the pre-trained physical memory neural network fusion model CP-NMN to obtain the same-dimensional health index across working conditions. The Physical Memory Neural Network Fusion Model (CP-NMN) combines the physical knowledge of mechanical structural components with neural networks to adaptively reduce the dimensionality of high-dimensional degenerate feature sets in order to obtain same-dimensional health indicators across working conditions. The mechanical properties of materials, the dynamic characteristics of structures, and the physical laws of damage evolution are embedded as constraints into the structure or training process of neural networks; Set the high-dimensional degradation features as As input to the CP-NMN model, the features are transformed by combining physical constraints with the various layers of the neural network, achieving dimensionality reduction and constraint fusion. The structure of the physical memory neural network fusion model CP-NMN is represented as follows: ; in, Indicates the first The output of the layer, and These are the weight matrix and bias vector of the l-th layer, respectively, and σ is the activation function. This refers to the physical memory constraint term of the l-th layer; The physical memory constraint term is embedded into the structure (forward propagation formula) and training process (loss function optimization) of the neural network in the form of weight bias initialization and loss function regularization term. This ensures that the CP-NMN model always follows the physical laws of mechanical components when reducing dimensionality and fusing high-dimensional features. The specific implementation process is as follows: By minimizing the model's physical loss function To optimize the network's weight matrix. and bias vector During training, the CP-NMN model adaptively reduces the dimensionality of the high-dimensional degenerate features F, removing redundant feature information, while fusing complementary features to reduce the high-dimensional features. After transformation through a multi-layer network, a low-dimensional, cross-condition, same-dimensional health index H is output. Furthermore, this index H conforms to the actual damage evolution law of mechanical structural components, enabling it to uniformly characterize the health status and physical losses of mechanical structural components under different operating conditions (such as different loads and speeds). The implementation process is as follows: Physical loss function Based on the fatigue Miner's rule for fatigue damage accumulation, let the physical quantity output by the model be... The physical quantity output according to the theoretical relationship of Miner's rule is The physical loss is then expressed as: m is the number of physical constraints. The physical quantity output through the physical memory constraint term; Furthermore, physical memory constraints Including the constraints of the physical laws governing damage evolution and structural dynamics constraints : The physical laws governing damage evolution are constrained by the regularization term of the loss function and the physical memory constraint term. Combining embedding, based on the fatigue Miner's law, the cumulative fatigue damage satisfies ,in, This represents the actual number of loops. For fatigue life, the physical laws governing damage evolution are constrained. satisfy: ; in, This indicates the damage characteristics corresponding to the current loop count; Specifically, by introducing constraints based on the physical laws of damage evolution. When training the CP-NMN model, ensure that the damage features processed by the model conform to the damage accumulation law of the fatigue Miner rule, so that the model can more accurately characterize the damage evolution process of mechanical structural components based on physical laws, and provide a more reliable foundation for subsequent tasks such as remaining life prediction. The structural dynamics constraint is to ensure that the frequency characteristics learned by the network do not produce resonance errors with the natural frequency of the mechanical structure, and conform to the dynamic law that the vibration frequency is far away from the natural frequency during normal operation; For the l-th layer of the CP-NMN model, the high-dimensional degenerate feature set containing vibrational frequency features... During processing, structural dynamic characteristics are constrained to ensure that the frequency characteristics learned by the network conform to this physical law; Determine the natural frequency of mechanical structural components And calculate the natural angular frequency. The formula is: ; For the l-th layer of the CP-NMN model, to ensure that the frequency-related features obtained after processing this layer do not produce resonant errors with the natural frequency, structural dynamics constraints are introduced. Requirements for structural dynamic characteristics satisfy: ; Specifically, The vector constructed based on the network structure and input features of layer l, and the inherent angular frequency The dot product is 0, which means that the structural dynamics are constrained. The CP-NMN model is constrained in the frequency dimension corresponding to the natural angular frequency, so that the features related to the vibration frequency will avoid the range corresponding to the natural frequency during the learning process. This ensures that the learned frequency features conform to the dynamic law that the vibration frequency of the mechanical structure is far away from the natural frequency when it is running normally, and avoids feature learning bias caused by resonance risk. After training for a preset number of rounds (50), the CP-NMN model can adaptively process high-dimensional degenerate features, transforming multi-dimensional high-dimensional degenerate feature sets. The table is transformed into a low-dimensional same-dimensional health index H. The same-dimensional health index H can characterize the health status of mechanical structural components under different operating conditions, such as different loads, speeds, and ambient temperatures. The closer the H value is to 1, the closer the mechanical structural component is to a healthy state. The closer the value is to 0, the more severe the damage to the mechanical structural component is, and the closer it is to a failure state. This enables a unified assessment of the health status of mechanical structural components across operating conditions.
[0021] S4: Based on the same-dimensional health indicators, the remaining lifespan is dynamically predicted using a long short-term memory network, and an engineering early warning is achieved by setting a lifespan threshold. The same-dimensional health indicators H obtained through the CP-NMN model are arranged in chronological order to form a time series. , where t is time, the time series of the same-dimensional health indicator H is used as the input data of the LSTM network to learn the evolution pattern of the health indicator over time. Using a trained LSTM network, based on current and historical time series of health indicators in the same dimension... To predict future health indicators, a predicted health indicator sequence is obtained. ; Then, based on a preset lifespan threshold, failure is determined when the H value drops to 0.1, and the moment when the lifespan threshold is first reached is found in the predicted health indicator sequence. Then the predicted remaining lifespan is ; Specifically, due to the uncertainty of the operating conditions and damage evolution of mechanical structural components, LSTM networks can continuously receive new health indicator data, constantly update predictions, and achieve dynamic prediction of remaining lifespan. That is, as time goes by and new data is acquired, the prediction results of remaining lifespan are adjusted in real time. A preset lifetime threshold is set when the predicted remaining lifetime... When the value is less than or equal to the preset warning threshold of 0.1, an engineering warning is triggered and communicated to relevant engineering personnel or control systems through audible and visual alarms and system prompts, so as to avoid more serious consequences caused by the failure of mechanical structural components and ensure the normal operation of the project.
[0022] Example 2: Figure 2As shown, a system for predicting the remaining life of an engineered mechanical structural component includes: Data acquisition module: Collects multi-dimensional raw datasets of mechanical structural components through multi-source sensors; High-dimensional acquisition module: Based on the matching of multi-dimensional original datasets, time-frequency domain features are extracted by combining inherent frequency and fault feature frequency, and high-dimensional degradation feature set is obtained by wavelet packet analysis and combined with characterization of damage evolution; Same-dimensional decision module: The pre-trained physical memory neural network fusion model CP-NMN is used to adaptively reduce the dimensionality of the high-dimensional feature set to obtain same-dimensional health indicators across working conditions. Lifespan prediction module: Based on the same-dimensional health indicators, it uses a long short-term memory network to dynamically predict the remaining lifespan and presets a lifespan threshold to achieve engineering early warning.
[0023] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.
[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0025] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting the remaining life of engineered mechanical structural components, characterized in that, include: S1: Collect multi-dimensional raw datasets of mechanical structural components using multi-source sensors; S2: Time-frequency domain feature extraction is performed by matching the multi-dimensional original dataset with the inherent frequency and the fault characteristic frequency, and a high-dimensional degradation feature set is obtained by wavelet packet analysis and combining it with the characterization of damage evolution. S3: Adaptive dimensionality reduction optimization of high-dimensional feature sets is performed using the pre-trained physical memory neural network fusion model CP-NMN to obtain the same-dimensional health index across working conditions. Set the high-dimensional degradation features as As input to the CP-NMN model, the features are transformed by combining physical constraints with the various layers of the neural network, achieving dimensionality reduction and constraint fusion. The structure of the physical memory neural network fusion model CP-NMN is represented as follows: ； in, Indicates the first The output of the layer, and These are the weight matrix and bias vector of the l-th layer, respectively, and σ is the activation function. This refers to the physical memory constraint term of the l-th layer; S4: Based on the same-dimensional health indicators, a long short-term memory network is used to dynamically predict the remaining lifespan, and a preset lifespan threshold is used to achieve engineering early warning.
2. The method for predicting the remaining life of an engineered mechanical structural component according to claim 1, characterized in that, The time-frequency domain features include spectral peak value, center frequency, and spectral entropy.
3. The method for predicting the remaining life of an engineered mechanical structural component according to claim 2, characterized in that, The process for obtaining the spectrum peak value is as follows: For the original multidimensional raw dataset Perform a Fourier transform and set the frequency resolution to 10Hz. Based on each frequency point k after the discrete Fourier transform, calculate the spectral amplitude obtained from the discrete Fourier transform. Its corresponding frequency value is Iterate through the values of k, when When the frequency falls within the range of 100-500Hz or 2000-5000Hz, record the spectral amplitude. The maximum value is taken as the peak value of the spectrum. .
4. The method for predicting the remaining life of an engineered mechanical structural component according to claim 2, characterized in that, The process for obtaining the center frequency is as follows: Determine the set of frequency points to be used in the calculation Where M is the number of frequency points, for each frequency point Corresponding spectral amplitude Introducing the vibration transmission coefficient at the damaged site For gear meshing, the transmission coefficient For the bearing raceway, the transmission coefficient ; Assign a corresponding transmission coefficient based on the damage location corresponding to frequency point k. Value to obtain center frequency Center frequency The calculation formula is: .
5. The method for predicting the remaining life of an engineered mechanical structural component according to claim 2, characterized in that, The process of obtaining the spectral entropy is as follows: For each frequency point k, calculate its corresponding spectral amplitude. Then calculate the energy at that frequency point. And obtain the total energy of all 512 frequency points. ; Let the fault characteristic frequency be For each frequency point k, determine Whether the condition is met or not, if it is met, the energy proportion at that frequency point is weighted and enhanced with an enhancement coefficient of 1.2; otherwise, the energy proportion is calculated normally. The calculation formula is: ; Based on energy percentage Calculate the spectral entropy : , where i is the index.
6. The method for predicting the remaining life of an engineered mechanical structural component according to claim 1, characterized in that, The process of obtaining a high-dimensional degenerative feature set representing damage evolution is as follows: Time-frequency domain features The original signal is decomposed into wavelet packets, and energy features are extracted from each sub-band signal obtained by decomposition. Let the signal of the i-th sub-band in the j-th layer be... and obtain time-frequency domain features. discrete-time signal energy The calculation formula is: The extracted time-frequency domain features High-dimensional energy features obtained from wavelet packet decomposition To integrate the features, suppose there are m time-frequency domain features, and wavelet packet decomposition yields n sub-band energy features. Then the integrated high-dimensional degenerate feature set... Represented as: ;in, This is a transpose.
7. The method for predicting the remaining life of an engineered mechanical structural component according to claim 1, characterized in that, The training process of the physical memory neural network fusion model is as follows: Physical memory constraint Including the constraints of the physical laws governing damage evolution and structural dynamics constraints Physical memory constraint terms It is embedded into the structure and training process of the neural network in the form of weight bias initialization and loss function regularization, by minimizing the physical loss function of the model. Optimize the network's weight matrix and bias vector Output a low-dimensional, cross-condition, same-dimensional health indicator H; Theoretical loss function Based on the fatigue Miner's rule for fatigue damage accumulation, let the physical quantity output by the model be... According to the theoretical relationship of Miner's rule, the output physical quantity is: The physical loss is then expressed as: m is the number of physical constraints. The physical quantity output by the physical memory constraint term model; After a predetermined number of training rounds, the CP-NMN model can adaptively process high-dimensional degenerate features, transforming multi-dimensional high-dimensional degenerate feature sets. The table is transformed into a low-dimensional, same-dimensional health indicator H.
8. The method for predicting the remaining life of an engineered mechanical structural component according to claim 7, characterized in that, The physical laws governing damage evolution constrain The implementation process is as follows: Based on Miner's law of fatigue, the cumulative fatigue damage satisfies... ,in, This represents the actual number of loops. For fatigue life, the physical laws governing damage evolution are constrained. satisfy: ; in, This indicates the damage characteristics corresponding to the current loop count.
9. The method for predicting the remaining life of an engineered mechanical structural component according to claim 7, characterized in that, The structural dynamics constraints The implementation process is as follows: Determine the natural frequency of mechanical structural components And calculate the natural angular frequency. The formula is: ; For the l-th layer of the CP-NMN model, to ensure that the frequency-related features obtained after processing this layer do not produce resonant errors with the natural frequency, structural dynamics constraints are introduced. Requirements for structural dynamic characteristics satisfy: .
10. A system for predicting the remaining life of engineered mechanical structural components, using the method described in any one of claims 1-9, characterized in that, include: Data acquisition module: Collects multi-dimensional raw datasets of mechanical structural components through multi-source sensors; High-dimensional acquisition module: Based on the matching of multi-dimensional original datasets, time-frequency domain features are extracted by combining inherent frequency and fault feature frequency, and high-dimensional degradation feature set is obtained by wavelet packet analysis and combined with characterization of damage evolution; Same-dimensional decision module: The pre-trained physical memory neural network fusion model CP-NMN is used to adaptively reduce the dimensionality of the high-dimensional feature set to obtain same-dimensional health indicators across working conditions. Lifespan prediction module: Based on the same-dimensional health indicators, it uses a long short-term memory network to dynamically predict the remaining lifespan and presets a lifespan threshold to achieve engineering early warning.