Method for predicting residual service life of equipment based on dynamic failure threshold interval adjustment

Abstract

The invention discloses an equipment residual service life prediction method based on dynamic failure threshold interval adjustment, which comprises the following steps of: firstly, constructing a power law attenuation model with residual strength according to physical information of equipment to obtain a dynamic failure threshold changing along with time; then defining an original full-life vibration signal as an upper bound data set; judging the moment when the vibration amplitude exceeds for the first time by using a dynamic threshold value, and carrying out physical truncation on the signal to obtain a lower bound data set; training residual life prediction neural network models of the upper bound data set and the lower bound data set to obtain an upper bound model and a lower bound model respectively; and finally, predicting the real-time life of the equipment through the upper and lower bound models, and restoring a prediction result into absolute physical time by using time step inversion to realize the interval prediction of the RUL of the equipment. According to the method, the safety and stability of an industrial system can be improved, and theoretical support and decision basis with higher robustness are provided for equipment health management under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment failure prediction and health management technology, specifically involving a method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment. Background Technology

[0002] Predicting the remaining useful life (RUL) of industrial machinery such as train bogies and bearings is beneficial for improving the operational safety of industrial systems and reducing the total lifecycle maintenance costs of equipment. However, traditional prediction methods often rely on fixed failure thresholds, ignoring the physical fact of the gradual decline in equipment lifespan caused by cumulative damage, leading to biased prediction results and affecting the safety of industrial systems. Therefore, this invention proposes a method for predicting the remaining useful life of equipment based on dynamic adjustment of the failure threshold range. To accurately predict the remaining useful life of equipment, this invention constructs a power-law decay model with residual strength based on fracture mechanics, cumulative damage theory, and ISO standards, thereby determining the failure threshold that dynamically evolves over time. To further improve the reliability of the prediction results, this invention proposes RUL interval prediction based on dual upper and lower bound datasets to effectively quantify the uncertainty of the prediction results. In addition, to improve the interpretability of the model, this invention defines a physical constraint on equipment degradation and threshold decay, thereby improving the interpretability and engineering verifiability of the remaining useful life prediction results. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment, which helps to improve the safety and stability of industrial systems and provides more robust theoretical support and decision-making basis for equipment health management under complex operating conditions.

[0004] The technical solution adopted in this invention is a method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment, which is implemented according to the following steps:

[0005] Step 1: Construct a power-law decay model with residual strength based on the physical information of the equipment to obtain the dynamic failure threshold that varies with time. ; Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 3: Train the remaining life prediction neural network models for the upper bound dataset and the lower bound dataset to obtain the upper bound model and the lower bound model, respectively; Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0006] The invention is further characterized in that, Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. ,Right now life, The expression is as shown in formula (1): (1) in, This is a lifespan index, representing the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. ; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): (2) in This is the residual strength coefficient, which typically ranges from 0.7 to 0.8. Step 1.4: Based on cumulative damage theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. With failure threshold The linear negative correlation mapping relationship is obtained as follows: (3) (4) Combining the above relationships, the final power-law decay model with residual intensity is obtained, as shown in the following formula: (5) in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the remaining intensity factor of the device at the current moment, ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifespan ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The value range is 3.0 to 4.0.

[0007] Step 2 is implemented in the following steps: Step 2.1: For each bearing sample in the historical monitoring equipment set, perform raw data loading and channel extraction operations, and construct a continuous full-lifecycle sequence. ; Step 2.2: Based on the lifecycle sequence constructed in Step 2.1 Generate the upper bound dataset Dataset_Upper for training the upper bound model; Step 2.3: Utilize the dynamic threshold model constructed in Step 1 For the whole lifetime sequence Perform stepwise scanning to dynamically identify the threshold lower bound cutoff time under threshold decay conditions. ; Step 2.4: Based on the identified The original data from step 2.1 is reconstructed with physical constraints to build a lower bound dataset, Dataset_Lower, for training the lower bound model.

[0008] Step 2.1 is implemented according to the following steps: Step 2.1.1, Data Loading: Read all the stored raw vibration signals, obtain all vibration data files, and sort them in ascending order according to the timestamp index in the file name to ensure that the data sequence strictly follows the physical time evolution order; Step 2.1.2, Channel Selection: Extract signal data of key monitoring channels from each data file, and simultaneously extract vibration signals in the horizontal and vertical directions to construct multi-dimensional feature vectors; Step 2.1.3: Constructing the lifespan sequence: The sorted data files are concatenated along the time dimension to reconstruct a continuous full-life-cycle vibration signal sequence. .

[0009] Step 2.2 is implemented according to the following steps: Step 2.2.1, Data Retention: Set the observation cutoff threshold to the forced shutdown limit of the experimental platform. Because... The data is designed to stop being collected only when the limit is reached, so no truncation is performed, and the complete sequence is preserved. Step 2.2.2, Define the label: Record the total duration of the entire sequence as the physical limit lifetime. This corresponds to the total number of files multiplied by the single-file sampling interval, for any sampling time in the sequence. t Sampling time t The remaining lifetime label is defined as follows: (6) Step 2.2.3, Encapsulation: Assigning the vibration sequence to the corresponding... The labels are encapsulated into a sample dictionary, normalized, and then stored in the upper bound dataset Dataset_Upper.

[0010] Step 2.3 is implemented according to the following steps: Step 2.3.1, Dynamic Threshold Calculation: For the first... k One data file, corresponding to physical time. Substituting into formula (5), the maximum permissible vibration amplitude standard at the current moment is calculated, as shown in formula (7): (7) Step 2.3.2, Amplitude Determination: Calculate the maximum absolute amplitude of the vibration signal. ,like If so, it is determined that the device has reached the functional failure boundary at that moment; Step 2.3.3, Cutoff Point Locking: Record the first moment when the condition in Step 2.3.2 is met. The lower bound of the threshold is the cutoff time. And immediately stop scanning subsequent data.

[0011] Step 2.4 is implemented according to the following steps: Step 2.4.1, Physical Truncation: From the whole lifetime sequence Remove from All data segments after time step 0 are retained, with only valid functional data from time step 0 to Tcut being retained. ; Step 2.4.2, Label Redefinition: Refine the truncation time. Defined as the end of the functional lifetime, for any sampling time in the retained sequence. t Sampling time tThe remaining lifetime label is defined as follows: (8) Step 2.4.3, Encapsulation: Encapsulating the truncated vibration sequence With redefining tags Encapsulate and store the data in the lower bound of the threshold dataset, Dataset_Lower.

[0012] Step 3 is implemented in the following steps: Step 3.1: Construct a CNN-GRU-Attention deep neural network architecture for time series regression prediction, including an input layer, a one-dimensional convolutional layer, a GRU recurrent layer, an attention mechanism layer, and a fully connected output layer connected in sequence, for training the upper bound model and the lower bound model; Step 3.2: Train the upper bound model using the upper bound dataset Dataset_Upper constructed in Step 2.2 and the deep neural network architecture in Step 3.1; Step 3.3: Train the lower bound model using the lower bound dataset Dataset_Lower constructed in Step 2.4 and the deep neural network architecture in Step 3.1.

[0013] Step 3.2 is implemented according to the following steps: Step 3.2.1, Data Input: Input the uncrunted full-lifetime vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.2.2, Optimization Objective: Define a loss function to minimize the difference between the network's predicted value and the upper bound of the lifetime label. Step 3.2.3, Backpropagation: The backpropagation algorithm is used to update the weight parameters of the convolutional layer, GRU layer, attention layer and fully connected layer at the same time, so that the CNN-GRU-Attention deep neural network learns the complete degradation trajectory of the device from a healthy state to the forced shutdown of the experimental platform. The model after training is denoted as the upper bound model Model_Upper. Step 3.3 is implemented in the following steps: Step 3.3.1, Data Input: Input the physically truncated vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.3.2, Independent Training: Use the same hyperparameter settings as in Step 3.2, but perform independent iterations on the truncated dataset; Step 3.3.3, Backpropagation: The backpropagation algorithm is used to simultaneously update the weight parameters of the convolutional layer, GRU layer, attention layer, and fully connected layer, enabling the CNN-GRU-Attention deep neural network to learn the device's transition from a healthy state to triggering a dynamic physical threshold. The degradation characteristics are used to determine the lower bound model, which is denoted as Model_Lower after training.

[0014] Step 4 is implemented in the following steps: The real-time vibration data of the device under test is simultaneously input into the upper bound model Model_Upper and the lower bound model Model_Lower trained in steps 3.2 and 3.3. After time step inversion, the data is restored to absolute physical time, generating a remaining life prediction interval composed of a conservative lower bound and an optimistic upper bound estimate, thereby realizing the interval prediction of the remaining life of the device.

[0015] The beneficial effects of this invention are that the equipment remaining service life prediction method based on dynamic failure threshold range adjustment aims to solve the problem that traditional fixed failure thresholds ignore physical degradation, leading to biased prediction results and reduced reliability. This invention achieves accurate characterization of failure boundaries by constructing a physically-aware dynamic failure threshold model, overcoming the poor interpretability of purely data-driven methods. To further improve prediction reliability, this invention utilizes a data truncation strategy under physical constraints to construct dual datasets within the threshold failure range, realizing the transformation from single-point prediction to range prediction. Furthermore, this invention does not rely on prior assumptions during the inference stage, directly generating the RUL prediction range through the dual model outputs to effectively quantify the uncertainty of the prediction results. This invention provides more robust theoretical support and a quantitative risk assessment basis for equipment maintenance under complex operating conditions. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the equipment remaining service life prediction method based on dynamic failure threshold range adjustment of the present invention; Figure 2 This is an example of the device remaining service life prediction method based on dynamic failure threshold range adjustment of the present invention. The vibration signal diagram of the mechanical bearing when it reaches the upper limit of the failure threshold is shown. Figure 3 This is an example of the device remaining service life prediction method based on dynamic failure threshold range adjustment of the present invention. The vibration signal intersection analysis diagram of the mechanical bearing considering the lower limit dynamic failure threshold is shown. Figure 4 This is an example of the remaining service life prediction method for equipment based on dynamic failure threshold range adjustment of the present invention, shown in the diagram of the remaining service life prediction range for mechanical bearing equipment. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] This invention relates to a method for predicting the remaining service life of equipment based on dynamically adjusted failure threshold ranges. The flowchart is as follows: Figure 1 As shown, please follow these steps: Step 1: Construct a power-law decay model with residual strength based on the physical information of the equipment to obtain the dynamic failure threshold that varies with time. ; Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. ,Right now life, The expression is as shown in formula (1): (1) in, For ball bearings, the life index is taken as... =3, for roller bearings take =10 / 3. This is the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. This is typically the threshold for mandatory shutdown of monitoring equipment; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): (2) in This is the residual strength coefficient, which typically ranges from 0.7 to 0.8. Step 1.4: Based on Cumulative Damage Theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. (Formula (3)) and failure threshold The linear negative correlation mapping relationship (formula (4)) is obtained as follows: (3) (4) Combining the above relationships, the final power-law decay model with residual intensity is obtained, as shown in the following formula: (5) in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the residual intensity factor of the device at the current moment (normalized to 0~1), ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifespan ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The value range is 3.0 to 4.0.

[0019] Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 2 is implemented in the following steps: Step 2.1: For each bearing sample in the historical monitoring equipment set, perform raw data loading and channel extraction operations, and construct a continuous full-lifecycle sequence. ; Step 2.1 is implemented according to the following steps: Step 2.1.1, Data Loading: Read all the stored raw vibration signals, obtain all vibration data files, and sort them in ascending order according to the timestamp index in the file name to ensure that the data sequence strictly follows the physical time evolution order; Step 2.1.2, Channel Selection: Extract signal data of key monitoring channels from each data file, and simultaneously extract vibration signals in the horizontal and vertical directions to construct multi-dimensional feature vectors; Step 2.1.3: Constructing the lifespan sequence: The sorted data files are concatenated along the time dimension to reconstruct a continuous full-life-cycle vibration signal sequence. .

[0020] Step 2.2: Based on the lifecycle sequence constructed in Step 2.1 Generate the upper bound dataset Dataset_Upper for training the upper bound model; Step 2.2 is implemented according to the following steps: Step 2.2.1, Data Retention: Set the observation cutoff threshold to the forced shutdown limit of the experimental platform. Because... The data is designed to stop being collected only when the limit is reached, so no truncation is performed, and the complete sequence is preserved. Step 2.2.2, Define the label: Record the total duration of the entire sequence as the physical limit lifetime. This corresponds to the total number of files multiplied by the single-file sampling interval, for any sampling time in the sequence. t Sampling time t The remaining lifetime label is defined as follows: (6) Step 2.2.3, Encapsulation: Assigning the vibration sequence to the corresponding... The labels are encapsulated into a sample dictionary, normalized, and then stored in the upper bound dataset Dataset_Upper.

[0021] Step 2.3: Utilize the dynamic threshold model constructed in Step 1 For the whole lifetime sequence Perform stepwise scanning to dynamically identify the threshold lower bound cutoff time under threshold decay conditions. ; Step 2.3 is implemented according to the following steps: Step 2.3.1, Dynamic Threshold Calculation: For the first... k One data file, corresponding to physical time. Substituting into formula (5), the maximum permissible vibration amplitude standard at the current moment is calculated, as shown in formula (7): (7) Step 2.3.2, Amplitude Determination: Calculate the maximum absolute amplitude of the vibration signal. ,like If so, it is determined that the device has reached the functional failure boundary at that moment; Step 2.3.3, Cutoff Point Locking: Record the first moment when the condition in Step 2.3.2 is met. The lower bound of the threshold is the cutoff time. And immediately stop scanning subsequent data.

[0022] Step 2.4: Based on the identified The original data from step 2.1 is reconstructed with physical constraints to build a lower bound dataset, Dataset_Lower, for training the lower bound model.

[0023] Step 2.4 is implemented according to the following steps: Step 2.4.1, Physical Truncation: From the whole lifetime sequence Remove from All data segments after time step 0 are retained, with only valid functional data from time step 0 to Tcut being retained. ; Step 2.4.2, Label Redefinition: Refine the truncation time. Defined as the end of the functional lifetime, for any sampling time in the retained sequence. t Sampling time t The remaining lifetime label is defined as follows: (8) Step 2.4.3, Encapsulation: Encapsulating the truncated vibration sequence With redefining tags Encapsulate and store the data in the lower bound of the threshold dataset, Dataset_Lower.

[0024] Step 3: Train the remaining life prediction neural network models for the upper bound dataset and the lower bound dataset to obtain the upper bound model and the lower bound model, respectively; Step 3 is implemented in the following steps: Step 3.1: Construct a CNN-GRU-Attention deep neural network architecture for time series regression prediction, including an input layer, a one-dimensional convolutional layer, a GRU recurrent layer, an attention mechanism layer, and a fully connected output layer connected in sequence, for training the upper bound model and the lower bound model; Step 3.2: Train the upper bound model using the upper bound dataset Dataset_Upper constructed in Step 2.2 and the deep neural network architecture in Step 3.1; Step 3.2 is implemented according to the following steps: Step 3.2.1, Data Input: Input the uncrunted full-lifetime vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.2.2, Optimization Objective: Define a loss function (mean squared error, MSE) to minimize the difference between the network's predicted value and the upper bound of the lifetime label. Step 3.2.3, Backpropagation: The backpropagation algorithm is used to update the weight parameters of the convolutional layer, GRU layer, attention layer and fully connected layer at the same time, so that the CNN-GRU-Attention deep neural network learns the complete degradation trajectory of the device from a healthy state to the forced shutdown of the experimental platform. The model after training is denoted as the upper bound model Model_Upper. Step 3.3: Train the lower bound model using the lower bound dataset Dataset_Lower constructed in Step 2.4 and the deep neural network architecture in Step 3.1.

[0025] Step 3.3 is implemented in the following steps: Step 3.3.1, Data Input: Input the physically truncated vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.3.2, Independent Training: Use the same hyperparameter settings as in Step 3.2, but perform independent iterations on the truncated dataset; Step 3.3.3, Backpropagation: The backpropagation algorithm is used to simultaneously update the weight parameters of the convolutional layer, GRU layer, attention layer, and fully connected layer, enabling the CNN-GRU-Attention deep neural network to learn the device's transition from a healthy state to triggering a dynamic physical threshold. The degradation characteristics are used to determine the lower bound model, which is denoted as Model_Lower after training.

[0026] Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0027] Step 4 is implemented in the following steps: The real-time vibration data of the device under test is simultaneously input into the upper bound model Model_Upper and the lower bound model Model_Lower trained in steps 3.2 and 3.3. After time step inversion, the data is restored to absolute physical time, generating a remaining life prediction interval composed of a conservative lower bound and an optimistic upper bound estimate, thereby realizing the interval prediction of the remaining life of the device.

[0028] Example 1 Due to the complex operating conditions in industrial settings and the physical degradation characteristics of equipment materials, traditional fixed-threshold prediction methods struggle to accurately assess the fuzzy failure process from "functional degradation" to "physical shutdown," resulting in unquantifiable uncertainty in prediction results and the risk of over-maintenance or unexpected downtime. To improve the scientific nature of equipment health management and optimize operation and maintenance strategies, it is necessary to adopt a method for predicting the remaining useful life of equipment based on dynamically adjusted failure threshold ranges to eliminate the limitations of this single assessment. The main process of this prediction method is as follows: First, based on the operating parameters and material properties of the historically observed equipment, the theoretical rated life is calculated using ISO standards and fracture mechanics theory, and a power-law decay model with residual strength is constructed to generate the dynamic failure threshold of the equipment that evolves over time. Then, the original full-life vibration signal set is subjected to dual preprocessing: the forced shutdown limit of the test bench is used as the upper bound of the observation to obtain the upper bound dataset Dataset_Upper; the original data is traversed to identify the moment when the vibration amplitude first exceeds the dynamic failure threshold curve and physical truncation is performed to obtain the lower bound dataset Dataset_Lower. Next, the above two datasets are input into the neural network model for parallel training, independently learning the degradation characteristics of the equipment under different failure definitions. Finally, the real-time vibration data of the equipment under test is simultaneously input into the two trained models, and the absolute physical time is restored through time step inversion to generate a remaining life prediction interval composed of a conservative lower bound estimate and an optimistic upper bound estimate, thereby realizing the interval prediction of the remaining life of the equipment.

[0029] Example 2 This invention relates to a method for predicting the remaining service life of equipment based on dynamically adjusted failure threshold ranges. The flowchart is as follows: Figure 1 As shown, please follow these steps: Step 1: Based on the physical information of historical monitoring equipment (including model parameters, operating condition data, etc.), construct a power-law decay model with residual strength to obtain the dynamic failure threshold that changes over time. ; Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. (Right now life). The expression is as shown in formula (1): (1) in, For ball bearings, the life index is taken as... =3, for roller bearings take =10 / 3. This is the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. This is typically the threshold for mandatory shutdown of monitoring equipment; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): (2) in This is the residual strength coefficient, which typically ranges from 0.7 to 0.8. Step 1.4: Based on Cumulative Damage Theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. (Formula (3)) and failure threshold The linear negative correlation mapping relationship (formula (4)) is obtained as follows: (3) (4) Based on the above relationships, the final power-law attenuation formula with residual strength is obtained as follows: (5) in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the residual intensity factor of the device at the current moment (normalized to 0~1), ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifespan ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The typical value range is 3.0 to 4.0.

[0030] Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 3: Train the remaining lifetime prediction neural network models of Dataset_Upper and Dataset_Lower to obtain the upper bound model and lower bound model, respectively; Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0031] Example 3 This invention relates to a method for predicting the remaining service life of equipment based on dynamically adjusted failure threshold ranges. The flowchart is as follows: Figure 1 As shown, please follow these steps: Step 1: Based on the physical information of historical monitoring equipment (including model parameters, operating condition data, etc.), construct a power-law decay model with residual strength to obtain the dynamic failure threshold that changes over time. ; Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. (Right now life). The expression is as shown in formula (1): (1) in, For ball bearings, the life index is taken as... =3, for roller bearings take =10 / 3. This is the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. This is typically the threshold for mandatory shutdown of monitoring equipment; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): (2) in This is the residual strength coefficient, which typically ranges from 0.7 to 0.8. Step 1.4: Based on Cumulative Damage Theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. (Formula (3)) and failure threshold The linear negative correlation mapping relationship (formula (4)) is obtained as follows: (3) (4) Based on the above relationships, the final power-law attenuation formula with residual strength is obtained as follows: (5) in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the residual intensity factor of the device at the current moment (normalized to 0~1), ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifespan ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The typical value range is 3.0 to 4.0.

[0032] Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 2 is implemented in the following steps: Step 2.1: For each bearing sample in the historical monitoring equipment set, perform raw data loading and channel extraction operations.

[0033] Step 2.1 is as follows: Step 2.1.1, Data Loading: Read the folder storing the original vibration signals, obtain all vibration data files, and sort them in ascending order according to the timestamp index in the file name to ensure that the data sequence strictly follows the physical time evolution order; Step 2.1.2, Channel Selection: Extract signal data of key monitoring channels from each data file, and simultaneously extract vibration signals in the horizontal and vertical directions to construct multi-dimensional feature vectors; Step 2.1.3: Constructing the lifespan sequence: The sorted data files are concatenated along the time dimension to reconstruct a continuous full-life-cycle vibration signal sequence. .

[0034] Step 2.2: Based on the lifecycle sequence constructed in Step 2.1 Generates a dataset, Dataset_Upper, for training the upper bound model.

[0035] Step 2.2 is as follows: Step 2.2.1, Data Retention: Set the observation cutoff threshold to the forced shutdown limit of the experimental platform. Because... The data is designed to stop being collected only when the limit is reached, so no truncation is performed, and the complete sequence is preserved. Step 2.2.2, Define the label: Record the total duration of the entire sequence as the physical limit lifetime. This corresponds to the total number of files multiplied by the single-file sampling interval. For any sampling time in the sequence... t Its remaining lifetime label is defined as follows: (6) Step 2.2.3, Encapsulation: Assigning the vibration sequence to the corresponding... The labels are encapsulated into a sample dictionary, normalized, and then stored in the upper bound dataset Dataset_Upper.

[0036] Step 2.3: Utilize the dynamic threshold model constructed in Step 1 For the whole lifetime sequence Perform stepwise scanning to dynamically identify the threshold lower bound cutoff time under threshold decay conditions. .

[0037] Step 2.3 is as follows: Step 2.3.1, Dynamic Threshold Calculation: For the first... k One data file (corresponding to physical time) Substitute these values ​​into formula (5) to calculate the maximum permissible vibration amplitude standard at the current moment, as shown in formula (7): (7) Step 2.3.2, Amplitude Determination: Calculate the maximum absolute amplitude of the vibration signal. .like If so, it is determined that the device has reached the functional failure boundary at that moment; Step 2.3.3, Cutoff Point Locking: Record the first moment when the condition in Step 2.3.2 is met. The lower bound of the threshold is the cutoff time. And immediately stop scanning subsequent data.

[0038] Step 2.4: Based on the identified The original data is reconstructed with physical constraints to build a dataset, Dataset_Lower, for training the lower bound model.

[0039] Step 2.4 is as follows: Step 2.4.1, Physical Truncation: From the whole lifetime sequence Remove from All data segments after time step 0 are retained, with only valid functional data from time step 0 to Tcut being retained. ; Step 2.4.2, Label Redefinition: Refine the truncation time. Defined as the end of functional lifetime. For any sampling time in the retained sequence. t Its remaining lifetime label is defined as follows: (8) Step 2.4.3, Encapsulation: Encapsulating the truncated vibration sequence With redefining tags Encapsulate and store the data in the lower bound of the threshold dataset, Dataset_Lower.

[0040] Step 3: Train the remaining lifetime prediction neural network models of Dataset_Upper and Dataset_Lower to obtain the upper bound model and lower bound model, respectively; Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0041] Example 4 This invention relates to a method for predicting the remaining service life of equipment based on dynamically adjusted failure threshold ranges. The flowchart is as follows: Figure 1 As shown, please follow these steps: Step 1: Based on the physical information of historical monitoring equipment (including model parameters, operating condition data, etc.), construct a power-law decay model with residual strength to obtain the dynamic failure threshold that changes over time. ; Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. (Right now life). The expression is as shown in formula (1): (1) in, For ball bearings, the life index is taken as... =3, for roller bearings take =10 / 3. This is the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. This is typically the threshold for mandatory shutdown of monitoring equipment; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): (2) in This is the residual strength coefficient, which typically ranges from 0.7 to 0.8. Step 1.4: Based on Cumulative Damage Theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. (Formula (3)) and failure threshold The linear negative correlation mapping relationship (formula (4)) is obtained as follows: (3) (4) Based on the above relationships, the final power-law attenuation formula with residual strength is obtained as follows: (5) in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the residual intensity factor of the device at the current moment (normalized to 0~1), ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifespan ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The typical value range is 3.0 to 4.0.

[0042] Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 2 is implemented in the following steps: Step 2.1: For each bearing sample in the historical monitoring equipment set, perform raw data loading and channel extraction operations.

[0043] Step 2.1 is as follows: Step 2.1.1, Data Loading: Read the folder storing the original vibration signals, obtain all vibration data files, and sort them in ascending order according to the timestamp index in the file name to ensure that the data sequence strictly follows the physical time evolution order; Step 2.1.2, Channel Selection: Extract signal data of key monitoring channels from each data file, and simultaneously extract vibration signals in the horizontal and vertical directions to construct multi-dimensional feature vectors; Step 2.1.3: Constructing the lifespan sequence: The sorted data files are concatenated along the time dimension to reconstruct a continuous full-life-cycle vibration signal sequence. .

[0044] Step 2.2: Based on the lifecycle sequence constructed in Step 2.1 Generates a dataset, Dataset_Upper, for training the upper bound model.

[0045] Step 2.2 is as follows: Step 2.2.1, Data Retention: Set the observation cutoff threshold to the forced shutdown limit of the experimental platform. Because... The data is designed to stop being collected only when the limit is reached, so no truncation is performed, and the complete sequence is preserved. Step 2.2.2, Define the label: Record the total duration of the entire sequence as the physical limit lifetime. This corresponds to the total number of files multiplied by the single-file sampling interval. For any sampling time in the sequence... t Its remaining lifetime label is defined as follows: (6) Step 2.2.3, Encapsulation: Assigning the vibration sequence to the corresponding... The labels are encapsulated into a sample dictionary, normalized, and then stored in the upper bound dataset Dataset_Upper.

[0046] Step 2.3: Utilize the dynamic threshold model constructed in Step 1 For the whole lifetime sequence Perform stepwise scanning to dynamically identify the threshold lower bound cutoff time under threshold decay conditions. .

[0047] Step 2.3 is as follows: Step 2.3.1, Dynamic Threshold Calculation: For the first... k One data file (corresponding to physical time) Substitute these values ​​into formula (5) to calculate the maximum permissible vibration amplitude standard at the current moment, as shown in formula (7): (7) Step 2.3.2, Amplitude Determination: Calculate the maximum absolute amplitude of the vibration signal. .like If so, it is determined that the device has reached the functional failure boundary at that moment; Step 2.3.3, Cutoff Point Locking: Record the first moment when the condition in Step 2.3.2 is met. The lower bound of the threshold is the cutoff time. And immediately stop scanning subsequent data.

[0048] Step 2.4: Based on the identified The original data is reconstructed with physical constraints to build a dataset, Dataset_Lower, for training the lower bound model.

[0049] Step 2.4 is as follows: Step 2.4.1, Physical Truncation: From the whole lifetime sequence Remove from All data segments after time step 0 are retained, with only valid functional data from time step 0 to Tcut being retained. ; Step 2.4.2, Label Redefinition: Refine the truncation time. Defined as the end of functional lifetime. For any sampling time in the retained sequence. t Its remaining lifetime label is defined as follows: (8) Step 2.4.3, Encapsulation: Encapsulating the truncated vibration sequence With redefining tags Encapsulate and store the data in the lower bound of the threshold dataset, Dataset_Lower.

[0050] Step 3: Train the remaining lifetime prediction neural network models of Dataset_Upper and Dataset_Lower to obtain the upper bound model and lower bound model, respectively; Step 3 is implemented in the following steps: Step 3.1: Construct a deep neural network architecture for time series regression prediction, used for training the upper bound model and the lower bound model; Step 3.2: Use the dataset Dataset_Upper constructed in Step 2.2 and the network architecture in Step 3.1 to train the upper bound model.

[0051] Step 3.2 is as follows: Step 3.2.1, Data Input: Input the uncrunted full-lifetime vibration sequence and its corresponding... Label input network; Step 3.2.2, Optimization Objective: Define a loss function (mean squared error, MSE) to minimize the difference between the network's predicted value and the upper bound of the lifetime label. Step 3.2.3, Backpropagation: The network weights are updated using the backpropagation algorithm, enabling the model to learn the complete degradation trajectory of the device from a healthy state to a forced shutdown of the experimental platform. The trained model is denoted as Model_Upper.

[0052] Step 3.3: Use the dataset Dataset_Lower constructed in Step 2.4 and the network architecture in Step 3.1 to train the lower bound model.

[0053] Step 3.3 is as follows: Step 3.3.1, Data Input: Input the physically truncated vibration sequence and its corresponding... Label input network; Step 3.3.2, Independent Training: Use the same hyperparameter settings as in Step 3.2, but perform independent iterations on the truncated dataset; Step 3.3.3, Backpropagation: Update the network weights using the backpropagation algorithm, enabling the model to learn the device's transition from a healthy state to triggering the dynamic physical threshold. The degradation characteristics. The model after training is denoted as Model_Lower.

[0054] Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0055] Example 5 This invention relates to a method for predicting the remaining service life of equipment based on dynamically adjusted failure threshold ranges. The flowchart is as follows: Figure 1 As shown, please follow these steps: Step 1: Construct a power-law decay model with residual strength based on the physical information of the equipment to obtain the dynamic failure threshold that varies with time. ; Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 3: Train the remaining lifetime prediction neural network models of Dataset_Upper and Dataset_Lower to obtain the upper bound model and lower bound model, respectively; Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

[0056] Step 4 is implemented in the following steps: The real-time vibration data of the device under test is simultaneously input into the two models Model_Upper and Model_Lower trained in steps 3.2 and 3.3. The data is then converted back to absolute physical time through time step inversion, generating a remaining life prediction interval composed of a conservative lower bound and an optimistic upper bound estimate, thereby realizing the interval prediction of the remaining life of the device.

[0057] Example 6 like Figure 2 As shown, this method defines equipment failure as the moment when the vibration signal first exceeds the failure threshold. Figure 3As shown, this invention uses a dynamic failure threshold curve to truncate the original vibration signal, which is more consistent with physical facts than a fixed failure threshold and is suitable as a lower bound for the failure threshold. Meanwhile, as... Figure 4 As shown, the RUL prediction range generated by this invention successfully covers the actual RUL curve, effectively providing early warning of potential functional failure risks. When performing tasks such as RUL prediction on rotating machinery, the prediction method based on dynamic failure threshold range adjustment proposed in this invention can effectively quantify the uncertainty of the prediction results and enhance the physical interpretability of traditional data-driven prediction methods.

[0058] To verify the feasibility of this invention, it is further described in conjunction with the embodiments and accompanying drawings. The selected data is the PHM2012 (FEMTO-ST) bearing dataset, which was obtained using the PRONOSTIA experimental platform under accelerated degradation conditions and contains full-life vibration data of 17 rolling bearings under three different operating conditions. This invention selects Bearing1_1 to Bearing1_7 under operating condition 1 (speed 1800 rpm, load 4000 N) as the research object to explain the effectiveness of this invention.

[0059] First, the raw PHM2012 data underwent dual preprocessing. The upper bound of the threshold observation was set according to the forced shutdown standard (20g) of the experimental platform, retaining the full-lifetime data of the Bearing1_x series as the "upper bound dataset." Simultaneously, the rated lifespan was calculated based on the ISO 281 standard. ≈9.26h, combined with fracture mechanics parameters β =3.5, construct a dynamic threshold curve This curve is used to identify the functional failure time of each bearing. ,like Figure 3 As shown. The original data is physically truncated, and data is removed. The subsequent high-amplitude vibration data is used to obtain the "lower bound dataset". Next, the two datasets are divided into sample sequences with a fixed time step of 80. Time-frequency features and long-term dependencies are extracted using a CNN-GRU-Attention network, and mapped to normalized RUL values ​​through a fully connected layer. Then, the trained dual model is used to perform parallel predictions on the bearing under test, Bearing1_3, outputting optimistic upper bound lifetime predictions and conservative lower bound lifetime predictions, respectively. Finally, the normalized prediction results are converted back to absolute physical time (seconds) to generate the RUL prediction interval. The prediction results for Bearing1_3 are visualized, as shown below. Figure 4As shown, the predicted interval represented by the shaded band not only successfully covers the actual RUL curve, but also significantly narrows its width when the equipment is close to failure in the later stage of degradation. This reflects the model's improvement in failure certainty, and the lower bound estimate provided effectively warns of potential functional failure risks. This indicates that the present invention can generate prediction results that conform to the laws of physical degradation, providing an innovative and effective solution for equipment health management and helping to enhance the reliability of intelligent operation and maintenance systems.

Claims

1. A method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment, characterized in that, The specific steps are as follows: Step 1: Construct a power-law decay model with residual strength based on the physical information of the equipment to obtain the dynamic failure threshold that varies with time. ; Step 2, Signal Set Preprocessing: Define the original full-lifetime vibration signal as the upper bound dataset Dataset_Upper; use a dynamic threshold to determine the first time the vibration amplitude exceeds the limit. The moment The signal is physically truncated to obtain the lower bound dataset Dataset_Lower; Step 3: Train the remaining life prediction neural network models for the upper bound dataset and the lower bound dataset to obtain the upper bound model and the lower bound model, respectively; Step 4: Predict the real-time lifespan of the device using upper and lower bound models, and use time step inversion to restore the prediction results to absolute physical time, thereby achieving range prediction of the device's RUL.

2. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 1, characterized in that, Step 1 is implemented in the following steps: Step 1.1: Obtain the model parameters and operating condition data of historical monitoring equipment, including rated dynamic load. Equivalent dynamic load and rotational speed The theoretical life of the equipment is calculated based on the ISO 281 standard formula for rated life of rolling bearings. ,Right now life, The expression is as shown in formula (1): （1） in, This is a lifespan index, representing the theoretical lifespan. This will serve as the baseline parameter for time normalization in subsequent models; Step 1.2: Based on the historical monitoring data of the equipment or the safety protection mechanism of the experimental platform, determine the upper limit of the vibration signal observation cutoff. ; Step 1.3: Introduce the residual strength theory from mechanics of materials to determine the equipment's lifespan at the theoretical end of its service life. The vibration threshold corresponding to the lowest structural strength that is still retained, i.e., the lower bound of the physical threshold. According to the industrial standard proportion method, The calculation expression is as shown in formula (2): （2） in This is the residual strength coefficient, which typically ranges from 0.7 to 0.

8. Step 1.4: Based on cumulative damage theory and boundary condition constraints, derive the dynamic threshold model and construct a power-law decay model with residual strength. This model is based on the cumulative damage degree of the equipment. With failure threshold The linear negative correlation mapping relationship is obtained as follows: （3） （4） Combining the above relationships, the final power-law decay model with residual intensity is obtained, as shown in the following formula: （5） in, This represents the maximum threshold decay range of the device throughout its entire lifespan, i.e., the intensity loss space from the healthy state to the lower threshold state. It is a non-linear shape factor. This represents the remaining intensity factor of the device at the current moment, ensuring that the threshold curve changes from the initial moment. Smooth decay to the end of the theoretical lifetime ; Step 1.5: Based on Paris's crack propagation law and Weibull distribution shape parameters in fracture mechanics. Determine nonlinear shape parameters The value of : fatigue crack propagation rate of metallic materials With stress intensity factor amplitude The relationship exhibits a power-law relationship. Furthermore, based on the accelerated crack propagation characteristics under fatigue loading, The value range is 3.0 to 4.

0.

3. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 2, characterized in that, Step 2 is implemented in the following steps: Step 2.1: For each bearing sample in the historical monitoring equipment set, perform raw data loading and channel extraction operations, and construct a continuous full-lifecycle sequence. ; Step 2.2: Based on the lifecycle sequence constructed in Step 2.1 Generate the upper bound dataset Dataset_Upper for training the upper bound model; Step 2.3: Utilize the dynamic threshold model constructed in Step 1 For the whole lifetime sequence Perform stepwise scanning to dynamically identify the threshold lower bound cutoff time under threshold decay conditions. ; Step 2.4: Based on the identified The original data from step 2.1 is reconstructed with physical constraints to build a lower bound dataset, Dataset_Lower, for training the lower bound model.

4. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 3, characterized in that, Step 2.1 is implemented in the following steps: Step 2.1.1, Data Loading: Read all the stored raw vibration signals, obtain all vibration data files, and sort them in ascending order according to the timestamp index in the file name to ensure that the data sequence strictly follows the physical time evolution order; Step 2.1.2, Channel Selection: Extract signal data of key monitoring channels from each data file, and simultaneously extract vibration signals in the horizontal and vertical directions to construct multi-dimensional feature vectors; Step 2.1.3: Constructing the lifespan sequence: The sorted data files are concatenated along the time dimension to reconstruct a continuous full-life-cycle vibration signal sequence. .

5. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 4, characterized in that, Step 2.2 is implemented in the following steps: Step 2.2.1, Data Retention: Set the observation cutoff threshold to the forced shutdown limit of the experimental platform. Because... The data is designed to stop being collected only when the limit is reached, so no truncation is performed, and the complete sequence is preserved. Step 2.2.2, Define the label: Record the total duration of the entire sequence as the physical limit lifetime. This corresponds to the total number of files multiplied by the single-file sampling interval, for any sampling time in the sequence. t Sampling time t The remaining lifetime label is defined as follows: （6） Step 2.2.3, Encapsulation: Assigning the vibration sequence to the corresponding... The labels are encapsulated into a sample dictionary, normalized, and then stored in the upper bound dataset Dataset_Upper.

6. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 5, characterized in that, Step 2.3 is implemented in the following steps: Step 2.3.1, Dynamic Threshold Calculation: For the first... k One data file, corresponding to physical time Substituting into formula (5), the maximum permissible vibration amplitude standard at the current moment is calculated, as shown in formula (7): （7） Step 2.3.2, Amplitude Determination: Calculate the maximum absolute amplitude of the vibration signal. ,like If so, it is determined that the device has reached the functional failure boundary at that moment; Step 2.3.3, Cutoff Point Locking: Record the first moment when the condition in Step 2.3.2 is met. The lower bound of the threshold is the cutoff time. And immediately stop scanning subsequent data.

7. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 6, characterized in that, Step 2.4 is implemented in the following steps: Step 2.4.1, Physical Truncation: From the whole lifetime sequence Remove from All data segments after time step 0 are retained, with only valid functional data from time step 0 to Tcut being retained. ; Step 2.4.2, Label Redefinition: Refine the truncation time. Defined as the end of the functional lifetime, for any sampling time in the retained sequence. t Sampling time t The remaining lifetime label is defined as follows: （8） Step 2.4.3, Encapsulation: Encapsulating the truncated vibration sequence With redefining tags Encapsulate and store the data in the lower bound of the threshold dataset, Dataset_Lower.

8. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 7, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Construct a CNN-GRU-Attention deep neural network architecture for time series regression prediction, including an input layer, a one-dimensional convolutional layer, a GRU recurrent layer, an attention mechanism layer, and a fully connected output layer connected in sequence, for training the upper bound model and the lower bound model; Step 3.2: Train the upper bound model using the upper bound dataset Dataset_Upper constructed in Step 2.2 and the deep neural network architecture in Step 3.1; Step 3.3: Train the lower bound model using the lower bound dataset Dataset_Lower constructed in Step 2.4 and the deep neural network architecture in Step 3.

1.

9. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 8, characterized in that, Step 3.2 is implemented in the following steps: Step 3.2.1, Data Input: Input the uncrunted full-lifetime vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.2.2, Optimization Objective: Define a loss function to minimize the difference between the network's predicted value and the upper bound of the lifetime label. Step 3.2.3, Backpropagation: The backpropagation algorithm is used to update the weight parameters of the convolutional layer, GRU layer, attention layer and fully connected layer at the same time, so that the CNN-GRU-Attention deep neural network learns the complete degradation trajectory of the device from a healthy state to the forced shutdown of the experimental platform. The model after training is denoted as the upper bound model Model_Upper. Step 3.3 is implemented in the following steps: Step 3.3.1, Data Input: Input the physically truncated vibration sequence and its corresponding... The input layer of the deep neural network in step 3.1 is used for label input; Step 3.3.2, Independent Training: Use the same hyperparameter settings as in Step 3.2, but perform independent iterations on the truncated dataset; Step 3.3.3, Backpropagation: The backpropagation algorithm is used to simultaneously update the weight parameters of the convolutional layer, GRU layer, attention layer, and fully connected layer, enabling the CNN-GRU-Attention deep neural network to learn the device's transition from a healthy state to triggering a dynamic physical threshold. The degradation characteristics are used to determine the lower bound model, which is denoted as Model_Lower after training.

10. The method for predicting the remaining service life of equipment based on dynamic failure threshold range adjustment according to claim 9, characterized in that, Step 4 is implemented in the following steps: The real-time vibration data of the device under test is simultaneously input into the upper bound model Model_Upper and the lower bound model Model_Lower trained in steps 3.2 and 3.

3. After time step inversion, the data is restored to absolute physical time, generating a remaining life prediction interval composed of a conservative lower bound and an optimistic upper bound estimate, thereby realizing the interval prediction of the remaining life of the device.

Images