Main hoist residual life prediction method and device

Abstract

The present application relates to a kind of main hoist remaining life prediction method and device, wherein, method includes: S1, the vibration time series data and working condition parameter time series data of the main hoist to be measured are acquired, and the measured vibration signal and measured working condition parameter signal of single operation are segmented, detection data group is constructed and time information is generated;S2, using detection data group to identify working condition, input the first self-encoding model corresponding to working condition, obtain the vibration signal after reconstruction and physical constraint residual, and the model is based on physical constraint dynamics equation to build;S3, total deviation is calculated according to reconstructed signal and physical constraint residual, and deviation sequence is obtained according to time sequence;S4, failure curve is generated based on deviation sequence, and then the remaining life is obtained.The present application can effectively separate working condition influence, fusion physical priori knowledge, realize accurate and reliable life prediction under the condition of variable working condition and variable load.

Description

Technical Field

[0001] This invention relates to the field of maintenance technology for main hoists operating under multiple conditions, and in particular to a method and device for predicting the remaining life of a main hoist. Background Technology

[0002] The main hoist is the "throat" of the coal production and transportation system, and its safe and reliable operation is directly related to the overall production efficiency and personnel safety of the mine. Due to its working characteristics, the main hoist is constantly under variable operating conditions with frequent starts and stops and drastic load changes. These drastic changes in operating conditions cause the vibration signals of its key components, such as the gearbox, bearings, and motors, to exhibit strong non-stationarity.

[0003] Traditional predictive maintenance methods primarily rely on spectral analysis techniques such as Fourier transform. For example, Chinese patent application CN115438443A discloses a method for predicting the remaining life of bearings. This method includes: acquiring the time-series data of the bearing to be predicted; extracting the time-domain features of the time-series data; obtaining an estimated value associated with the time-domain features using a time-series processing model (trained from full-cycle data of bearing samples, including health and fault data); determining the residual value based on the estimated value and the time-domain features; and determining the bearing life corresponding to the residual value as the remaining life of the bearing to be predicted, based on the one-to-one correspondence between the residual value and the bearing life. This method provides a more accurate prediction of bearing life.

[0004] The core assumption of this method is that the signal is stationary, meaning its statistical characteristics do not change over time. However, unlike the bearings in the prior art, when the main hoist operates under varying conditions, the amplitude and frequency components of the vibration signal change drastically with changes in load and speed, causing traditional spectrum analysis to fail. Specifically: 1. Operating condition interference: Increased load will naturally lead to an increase in vibration energy such as RMS value, which can easily be misjudged as the onset of a fault, resulting in a large number of false alarms.

[0005] 2. Fault feature overwhelmed: For example, the early and weak fault features of minor impacts caused by bearing pitting are easily overwhelmed by background vibrations caused by strong changes in operating conditions, resulting in missed detection.

[0006] Furthermore, in CN115438443A mentioned above, the residual value is obtained by predicting time-series characteristics. This method of learning patterns entirely from data ignores the known physical and dynamic characteristics of the equipment itself, resulting in insufficient generalization ability and robustness of the model.

[0007] Therefore, how to develop a method that can effectively separate the influence of operating conditions, integrate prior physical knowledge, and accurately and reliably predict life under varying operating conditions and loads is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a method and apparatus for predicting the remaining life of a main hoist in order to overcome the defects of the prior art.

[0009] The objective of this invention can be achieved through the following technical solutions: A method for predicting the remaining life of a main hoist includes: Step S1: Obtain the vibration time series data of the main hoist under test and the working condition parameter time series data reflecting the operating status of the equipment, and divide them into measured vibration signals corresponding to a single operation of the main hoist under test and measured working condition parameter signals reflecting the operating status of the equipment to construct detection data groups, and generate time information corresponding to each detection data group. Step S2: The working conditions are identified using the detection data set and further input into the first autoencoder model corresponding to the trained working conditions to obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set. The working conditions are divided based on the load and running direction of the main hoist. The first autoencoder model is built based on the physical constraint dynamic equation. The input is the measured vibration signal and the measured working condition parameter signal reflecting the operating status of the equipment. The output is the reconstructed vibration signal and the physical constraint residual. Step S3: Calculate the total deviation using the reconstructed vibration signal and physical constraint residual corresponding to the detection data group, and sort all the total deviations according to the time information of the corresponding detection data group to obtain the deviation sequence; Step S4 generates a failure curve based on the deviation sequence, and further obtains the remaining lifetime based on the failure curve.

[0010] The first autoencoder model is obtained using prior data sets and labeled working conditions obtained from the main hoist under healthy conditions; The process of obtaining the prior data set is as follows: Vibration time-series data of key parts of the main hoist under healthy conditions and time-series data of operating parameters reflecting the equipment's operating status are acquired. The measured vibration signal corresponding to a single operation of the main hoist under healthy conditions and the measured operating parameter signal reflecting the equipment's operating status are then segmented as a priori data set, and the operating conditions of the priori data set are marked.

[0011] The types of operating conditions include unloaded upward movement, heavy load upward movement, unloaded downward movement, and heavy load downward movement.

[0012] The operating condition parameter timing data includes current timing data and operating timing speed; The measured operating condition parameter signals include the measured current signal and the operating speed signal; The inputs to the first autoencoder model are measured vibration signals, current signals, and running speed signals.

[0013] The physical constraint dynamic equations are constructed based on the product of equivalent mass and vibration signal, the product of system damping and vibration velocity, the product of system stiffness and vibration displacement, and the equivalent excitation force, wherein the equivalent excitation force is obtained based on operating speed and current.

[0014] The loss function of the first autoencoder model is the total deviation, which includes a data sub-loss and a physical sub-loss. The data sub-loss is obtained based on the difference between the measured vibration signal and the reconstructed vibration signal, and the physical sub-loss is obtained based on the physical constraint residual.

[0015] The process of obtaining the physical sub-loss includes: The acceleration term is obtained by multiplying the reconstructed vibration signal by the equivalent mass. The reconstructed vibration velocity is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration velocity and the system damping is calculated to obtain the velocity term value. The reconstructed vibration displacement is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration displacement and the system stiffness is calculated to obtain the displacement term value. The physical loss is obtained by summing the acceleration, velocity, and displacement values ​​and subtracting the equivalent excitation force obtained based on the measured running speed and measured current.

[0016] Step S2 includes: Step S2-1: Input the detection data set into the working condition classification network to obtain the working condition classification result; Step S2-2: Input the detection data set into the first autoencoder model of the corresponding working condition that has been trained, and obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set.

[0017] The working condition classification network is composed of multiple spatiotemporal graph convolutional blocks stacked together. Each spatiotemporal graph convolutional block contains a graph convolutional layer and a temporal convolutional layer. The graph convolutional layer updates the first node features of the nodes in the sensor topology graph by aggregating the information of neighboring nodes. The temporal convolutional layer applies standard one-dimensional convolution along the time dimension after the graph convolution. In the sensor topology graph, each sensor corresponds to a node in the sensor topology graph. If two sensors have a physical transmission connection relationship, an edge is constructed between the corresponding two nodes.

[0018] Step S4 includes: Step S4-1: Input the bias sequence into the trained time series model to predict the total bias at multiple future time points; Step S4-2: Generate a failure curve based on the total deviation in the deviation sequence and the predicted total deviation, wherein the horizontal axis of the failure curve is time and the vertical axis is the total deviation value; Step S4-3: Find the point in the failure curve whose total deviation value is the first threshold as the failure point; Step S4-4: Take the difference between the time corresponding to the failure point and the current time as the remaining lifetime.

[0019] A device for predicting the remaining life of a main hoist includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. On the one hand, by identifying operating conditions, data separation based on operating conditions is achieved, thereby decoupling the mutual influence of drastic data shifts between different operating conditions when amplitude and frequency components change drastically due to changes in operating conditions. On the other hand, by integrating prior physical knowledge into the first autoencoder model, known equipment physical dynamics knowledge is embedded as a constraint into the model, making the health baseline model not only fit the data but also conform to physical laws. This greatly reduces the dependence on massive amounts of health data and improves the model's generalization ability and robustness.

[0021] 2. The prior data set obtained by collecting data under healthy conditions allows the first autoencoder model to learn accurate physical constraints. While ensuring that the reconstructed vibration signal is highly consistent with the input vibration signal, it also ensures that implicit features and correlations related to current and running speed can be learned.

[0022] 3. Covering unloaded upward travel, heavily loaded upward travel, unloaded downward travel, and heavily loaded downward travel, this approach covers typical operating cycles of the main hoist. Furthermore, the clear boundaries between these four conditions make them easy to identify and separate. These four conditions exhibit significant differences and distinguishability in electrical parameters, operating speed, and load direction. For example, during heavy-load upward travel, the current is high and the speed is relatively stable; during unloaded downward travel, the current is low and the speed may be relatively high. This clear boundary facilitates the training and identification of the subsequent condition classification network, reducing the risk of misclassification and thus improving the accuracy of lifespan prediction. Additionally, the clear physical meaning facilitates the introduction of physical constraints. Finally, it balances engineering practicality with model generalization ability. Although intermediate or transitional conditions may exist in actual operation, classifying them into the four main conditions ensures model simplicity and training efficiency. In practical applications, the condition classification network can reasonably categorize continuously changing conditions, demonstrating good generalization ability. Setting too many subdivided conditions could lead to sparse samples, model overfitting, and reduced prediction robustness.

[0023] 4. It provides a solid physical theoretical foundation and interpretable modeling framework for the entire life prediction method, and links vibration signals with the system's mass, damping, stiffness, and equivalent excitation force determined by current and operating speed. This makes the model no longer a pure black box data fitting, but embeds known equipment physical dynamics knowledge as hard constraints into the neural network. This greatly reduces the dependence on massive health data and significantly improves the model's generalization ability and robustness when facing operating conditions or states not covered by training data.

[0024] 5. A dual supervision mechanism was implemented. The data sub-loss ensures that the model's reconstructed output is numerically close to the actual measured vibration signal, maintaining the model's data fitting ability. At the same time, the physical sub-loss forces the model's output to satisfy the physical dynamic equations. This design allows the model to learn not only the statistical laws in the data during training, but also to be constrained by physical laws. This enables the model to learn a health status representation that is more consistent with the actual physical characteristics of the equipment, fundamentally improving the accuracy and reliability of life prediction.

[0025] 6. It provides a clear, differentiable, and computationally stable optimization objective. By minimizing the mean of the squared Euclidean distance between the measured vibration signal and the model-reconstructed vibration signal, it can effectively drive the encoder and decoder of the autoencoder to learn the core features and patterns of the vibration signal under healthy conditions, providing a high-quality, clean signal reconstruction benchmark for subsequent calculation of the total deviation.

[0026] 7. Directly quantifying the degree of conformity of physical equations into part of the loss function effectively suppresses the model from learning spurious features that violate physical laws, ensuring the physical consistency of the learned healthy baseline, which is the key to improving the generalization of the method.

[0027] 8. This network structure is specifically designed for processing data with spatial and temporal dependencies. For complex systems such as main hoists, their multi-source monitoring signals exhibit complex spatiotemporal correlation patterns under different operating conditions. It can simultaneously capture the graph structure relationship of signals in the sensor dimension and the dynamic evolution law in the time dimension, thereby enabling more refined and accurate identification and classification of complex and continuous operating condition changes, laying a solid foundation for subsequent adaptive analysis of operating conditions. In addition, in the sensor topology graph, each sensor corresponds to a node in the sensor topology graph. If two sensors have a physical transmission connection relationship, an edge is constructed between the corresponding two nodes. This allows the model to focus on the signal correlation between sensors with transmission relationships, improving robustness and accelerating convergence speed.

[0028] 9. First, by predicting deviation trends through a time-series model, a forward-looking prediction of the future state of the equipment is achieved, rather than just an assessment of the current state. Second, the continuous deviation sequence is transformed into a failure curve, making the performance degradation process visible. Finally, a preset, unified first threshold is used to determine the failure point, and the time difference between the current moment and the failure point is defined as the remaining life. This makes the prediction result an intuitive indicator with clear physical meaning, which is convenient for maintenance personnel to use directly to make maintenance decisions. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the main steps of the method of the present invention. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0031] A method for predicting the remaining life of a main hoist, such as Figure 1 As shown, it includes: Step S1: Obtain the vibration time series data of the main hoist under test and the working condition parameter time series data reflecting the operating status of the equipment, and divide them into measured vibration signals corresponding to a single operation of the main hoist under test and measured working condition parameter signals reflecting the operating status of the equipment to construct detection data groups, and generate time information corresponding to each detection data group. Step S2: The working conditions are identified using the detection data set and further input into the first autoencoder model corresponding to the trained working conditions to obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set. The working conditions are divided based on the load and running direction of the main hoist. The first autoencoder model is built based on the physical constraint dynamic equation. The input is the measured vibration signal and the measured working condition parameter signal reflecting the operating status of the equipment. The output is the reconstructed vibration signal and the physical constraint residual. The first autoencoder model is obtained using prior data sets based on the main hoist under healthy conditions and their labeled operating conditions; The process of obtaining the prior data set is as follows: Vibration time-series data of key components of the main hoist under healthy conditions and operating condition parameter time-series data reflecting the equipment's operating status are acquired. These are then segmented to obtain measured vibration signals corresponding to a single operation of the main hoist under healthy conditions, and measured operating condition parameter signals reflecting the equipment's operating status, which are used as prior data sets. The operating conditions of the prior data sets are then labeled. In this embodiment, the operating condition parameter time-series data includes current time-series data and operating speed. Of course, in other embodiments, other data, such as temperature, can also be used, depending on the type of equipment.

[0032] This process mainly involves the synchronous acquisition and preprocessing of multi-source data. The vibration time-series data sources include three triaxial accelerometers, respectively installed on the high-speed shaft, low-speed shaft bearing housing, and motor housing of the hoist reducer. Additionally, the current time-series data consists of the three-phase current of the motor. All multi-source data requires time synchronization to ensure strict temporal synchronization of all data streams. Wavelet threshold denoising is then applied to the vibration signals.

[0033] Then, training samples are constructed using the prior data set, and a first autoencoder model corresponding to the working condition is trained based on the constructed training samples. The first autoencoder model is built based on the physical constraint dynamic equation. The input is the measured vibration signal and the measured working condition parameter signal reflecting the operating status of the equipment. The output is the reconstructed vibration signal and the physical constraint residual. In this embodiment, the measured working condition parameter signal includes the measured current signal and the operating speed signal.

[0034] The physical constraint dynamics equations are constructed based on the product of equivalent mass and vibration signal, the product of system damping and vibration velocity, the product of system stiffness and vibration displacement, and the equivalent excitation force. The equivalent excitation force is derived from the operating speed and current. In this embodiment, the physical constraint dynamics equations are:

[0035]

[0036] in:m For the equivalent quality of the system, for t Vibration displacement at time t, for The second derivative, for t System damping at any moment, for The first derivative, for t The system stiffness at time t, for t Current at any moment for t The running speed at any given moment, For equivalent excitation force, for t The measured vibration signal at a given time.

[0037] It provides a solid physical theoretical foundation and interpretable modeling framework for the entire life prediction method, and links vibration signals with the system's mass, damping, stiffness and the equivalent excitation force determined by current and operating speed. This makes the model no longer a pure black box data fitting, but embeds known equipment physical dynamics knowledge as hard constraints into the neural network. This greatly reduces the dependence on massive health data and significantly improves the model's generalization ability and robustness when facing operating conditions or states not covered by training data.

[0038] The loss function of the first autoencoder model is the total deviation, which includes data sub-loss and physical sub-loss. The data sub-loss is obtained based on the difference between the measured vibration signal and the reconstructed vibration signal, while the physical sub-loss is obtained based on the physical constraint residual.

[0039] The data sub-loss is:

[0040] in: For data sub-loss, N For the sample size, for t i The measured vibration signal at time t, for t i The reconstructed vibration signal at time t, It is a Euclidean distance.

[0041] The process of obtaining the physical sub-loss includes: The acceleration term is obtained by multiplying the reconstructed vibration signal by the equivalent mass. The reconstructed vibration velocity is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration velocity and the system damping is calculated to obtain the velocity term value. The reconstructed vibration displacement is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration displacement and the system stiffness is calculated to obtain the displacement term value. The physical loss is obtained by summing the acceleration, velocity, and displacement values ​​and subtracting the equivalent excitation force obtained based on the measured running speed and measured current.

[0042] In this embodiment, the physical sub-loss is:

[0043]

[0044]

[0045] in: For physical loss, N For the sample size, for t i Reconstructed vibration displacement at time t, for t i Time residuals for The second derivative, For system damping under healthy conditions, for The first derivative, The system stiffness under healthy conditions, for t i Current at any moment for t i The running speed at any given moment, for t i The reconstructed vibration signal at a given moment.

[0046] In this embodiment, the loss function of the first autoencoder model is:

[0047] Wherein, α and β are hyperparameters, used as weights to balance data fitting and physical constraints. In this embodiment, the hyperparameters α and β are set using a calibration strategy based on the alignment of the initial loss magnitude. In step 0 of model training, the unweighted initial data loss L is calculated respectively. data and initial physical loss L physicsTo eliminate the difference in absolute numerical magnitude between the two (for example, the magnitude of the squared error of a vibration signal is usually much smaller than the sum of squares of the physical residuals including the second derivative), the initial weighting ratio β / α is set to L. data / L physics .

[0048] Specifically, in this embodiment, the baseline is set to α=1.0, while the value of β is dynamically calculated based on the initial magnitude ratio of the actual operating condition signal (for example, the ratio is measured to be 3.5×10 under a certain heavy load condition). -5 Therefore, we set β = 3.5 × 10 -5 This ensures that the gradient contributions of the two losses to the network parameters at the training start point are of the same order of magnitude.

[0049] Based on the stage that is manually confirmed as a healthy state, vibration signals, current signals and operating speed signals under specific working conditions are collected, and parameter fitting regression is performed on the aforementioned dynamic equations to calibrate the health benchmark parameters corresponding to the working condition offline, namely the system stiffness and system damping under the healthy state.

[0050] In this embodiment, similarly, the measured operating condition parameter signals include the measured current signal and the operating speed signal.

[0051] Specific step S2 includes: Step S2-1: Input the detection data set into the working condition classification network to obtain the working condition classification result; Step S2-2: Input the detection data set into the first autoencoder model of the corresponding working condition that has been trained, and obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set.

[0052] For the working condition classification network, firstly, it is necessary to construct a sensor topology graph G=(V,E,A), where each sensor is regarded as a graph node. If two sensors are physically connected (such as a motor sensor and a gearbox input shaft sensor), then an edge is established between them, and an adjacency matrix is ​​constructed to describe the connection relationship of the graph.

[0053] The working condition classification network consists of multiple spatiotemporal graph convolutional blocks stacked together. Each spatiotemporal graph convolutional block contains a graph convolutional layer and a temporal convolutional layer.

[0054] At each time step, a spatial domain convolution operation is performed on the first node feature of the graph node to aggregate information from neighboring nodes. After the graph convolution, a standard one-dimensional convolution is applied along the time dimension to capture the dynamic evolution of the features.

[0055] The outputs of multiple spatiotemporal graph convolutional blocks are passed through a global average pooling layer and a fully connected layer. Finally, the Softmax function is used to output the probability distribution of each working condition category. The working condition classification network uses either the cross-entropy loss function for supervised learning or the clustering loss function for unsupervised learning.

[0056] Step S3: Calculate the total deviation using the reconstructed vibration signal and physical constraint residual corresponding to the detection data group, and sort all the total deviations according to the time information of the corresponding detection data group to obtain the deviation sequence; Step S4: Generate a failure curve based on the deviation sequence, and further obtain the remaining lifetime based on the failure curve, including: Step S4-1: Input the bias sequence into the trained time series model to predict the total bias at multiple future time points; Step S4-2: Generate a failure curve based on the total deviation in the deviation sequence and the predicted total deviation, where the horizontal axis of the failure curve is the time and the vertical axis is the total deviation value; Step S4-3: Find the point in the failure curve whose total deviation value is the first threshold as the failure point; Step S4-4: Take the difference between the time corresponding to the failure point and the current time as the remaining lifetime.

[0057] In addition, in some embodiments, the total deviation can be directly compared with a first threshold. If it is greater than the first threshold, an alarm is triggered. Of course, in most cases, since the prediction is not performed when the system is not in a faulty state, this application needs to use a failure curve to find the failure point.

[0058] This application solution can be applied to various fields that require hoisting applications, such as coal production and transportation, the coal mining industry, other solid minerals, ice making, and grain production industries.

[0059] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for predicting the remaining life of a main hoist, characterized in that, include: Step S1: Obtain the vibration time series data of the main hoist under test and the working condition parameter time series data reflecting the operating status of the equipment, and divide them into measured vibration signals corresponding to a single operation of the main hoist under test and measured working condition parameter signals reflecting the operating status of the equipment to construct detection data groups, and generate time information corresponding to each detection data group. Step S2: The working conditions are identified using the detection data set and further input into the first autoencoder model corresponding to the trained working conditions to obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set. The working conditions are divided based on the load and running direction of the main hoist. The first autoencoder model is built based on the physical constraint dynamic equation. The input is the measured vibration signal and the measured working condition parameter signal reflecting the operating status of the equipment. The output is the reconstructed vibration signal and the physical constraint residual. Step S3: Calculate the total deviation using the reconstructed vibration signal and physical constraint residual corresponding to the detection data group, and sort all the total deviations according to the time information of the corresponding detection data group to obtain the deviation sequence; Step S4 generates a failure curve based on the deviation sequence, and further obtains the remaining lifetime based on the failure curve.

2. The method for predicting the remaining life of a main hoist according to claim 1, characterized in that, The first autoencoder model is obtained using prior data sets and labeled working conditions obtained from the main hoist under healthy conditions; The process of obtaining the prior data set is as follows: Vibration time-series data of key parts of the main hoist under healthy conditions and time-series data of operating parameters reflecting the equipment's operating status are acquired. The measured vibration signal corresponding to a single operation of the main hoist under healthy conditions and the measured operating parameter signal reflecting the equipment's operating status are then segmented as a priori data set, and the operating conditions of the priori data set are marked.

3. The method for predicting the remaining life of a main hoist according to claim 1, characterized in that, The types of operating conditions include unloaded upward movement, heavy load upward movement, unloaded downward movement, and heavy load downward movement.

4. The method for predicting the remaining life of a main hoist according to claim 1, characterized in that, The operating condition parameter timing data includes current timing data and operating timing speed; The measured operating condition parameter signals include the measured current signal and the operating speed signal; The inputs to the first autoencoder model are measured vibration signals, current signals, and running speed signals.

5. The method for predicting the remaining life of a main hoist according to claim 4, characterized in that, The physical constraint dynamic equations are constructed based on the product of equivalent mass and vibration signal, the product of system damping and vibration velocity, the product of system stiffness and vibration displacement, and the equivalent excitation force, wherein the equivalent excitation force is obtained based on operating speed and current.

6. The method for predicting the remaining life of a main hoist according to claim 5, characterized in that, The loss function of the first autoencoder model is the total deviation, which includes a data sub-loss and a physical sub-loss. The data sub-loss is obtained based on the difference between the measured vibration signal and the reconstructed vibration signal, and the physical sub-loss is obtained based on the physical constraint residual.

7. The method for predicting the remaining life of a main hoist according to claim 6, characterized in that, The process of obtaining the physical sub-loss includes: The acceleration term is obtained by multiplying the reconstructed vibration signal by the equivalent mass. The reconstructed vibration velocity is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration velocity and the system damping is calculated to obtain the velocity term value. The reconstructed vibration displacement is obtained based on the reconstructed vibration signal, and the product of the reconstructed vibration displacement and the system stiffness is calculated to obtain the displacement term value. The physical loss is obtained by summing the acceleration, velocity, and displacement values ​​and subtracting the equivalent excitation force obtained based on the measured running speed and measured current.

8. The method for predicting the remaining life of a main hoist according to claim 1, characterized in that, Step S2 includes: Step S2-1: Input the detection data set into the working condition classification network to obtain the working condition classification result; Step S2-2: Input the detection data set into the first autoencoder model of the corresponding working condition that has been trained, and obtain the reconstructed vibration signal and physical constraint residual corresponding to the detection data set.

9. The method for predicting the remaining life of a main hoist according to claim 8, characterized in that, The working condition classification network is composed of multiple spatiotemporal graph convolutional blocks stacked together. Each spatiotemporal graph convolutional block contains a graph convolutional layer and a temporal convolutional layer. The graph convolutional layer updates the first node features of the nodes in the sensor topology graph by aggregating the information of neighboring nodes. The temporal convolutional layer applies standard one-dimensional convolution along the time dimension after the graph convolution. In the sensor topology graph, each sensor corresponds to a node in the sensor topology graph. If two sensors have a physical transmission connection relationship, an edge is constructed between the corresponding two nodes.

10. The method for predicting the remaining life of a main hoist according to claim 1, characterized in that, Step S4 includes: Step S4-1: Input the bias sequence into the trained time series model to predict the total bias at multiple future time points; Step S4-2: Generate a failure curve based on the total deviation in the deviation sequence and the predicted total deviation, wherein the horizontal axis of the failure curve is time and the vertical axis is the total deviation value; Step S4-3: Find the point in the failure curve whose total deviation value is the first threshold as the failure point; Step S4-4: Take the difference between the time corresponding to the failure point and the current time as the remaining lifetime.

11. A device for predicting the remaining life of a main hoist, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-10.

Images