Crane transmission system life prediction method and device, electronic equipment and storage medium

Abstract

This invention provides a method, device, electronic device, and storage medium for predicting the lifespan of a crane transmission system, belonging to the field of crane maintenance technology. The method includes: acquiring a time series of health indicators reflecting the overall degradation trend of the crane transmission system, the time series being constructed based on vibration signals collected from the transmission system under variable load and variable speed conditions; inputting the health indicator time series into a pre-trained residual nested long short-term memory attention prediction network; the residual nested long short-term memory attention prediction network being composed of stacked residual blocks; and outputting predicted remaining service life values ​​of key components of the transmission system through sequence learning and attention weighting processing of the residual nested long short-term memory attention prediction network. This invention improves the accuracy of crane transmission system lifespan prediction.

Description

Technical Field

[0001] This invention relates to the field of crane maintenance technology, specifically to a method, device, electronic equipment, and storage medium for predicting the lifespan of a crane transmission system. Background Technology

[0002] Large cranes are key equipment for heavy material handling and the installation of large components. Accurately predicting the Remaining Useful Life (RUL) of critical components in the crane's transmission system is a core technology for shifting from traditional periodic maintenance or reactive repair to state-based predictive intelligent operation and maintenance. Currently, RUL prediction methods for complex mechanical equipment such as crane transmission systems can be mainly divided into two categories: physical model-based and data-driven methods. Physical model-based methods rely on precise mathematical modeling of component failure mechanisms, while traditional data-driven methods are mainly implemented through techniques such as support vector machines and shallow neural networks.

[0003] In the process of realizing this invention, the inventors discovered that existing methods have at least the following problems: the crane transmission system has a complex structure and variable operating conditions, and its degradation process involves multi-physics coupling. It is extremely difficult to establish a general physical model using existing physical model-based methods, and the model parameters are difficult to obtain and update online, resulting in poor generalization ability and insufficient predictive leadership in practical engineering applications. While existing data-driven methods reduce the dependence on accurate physical models, their feature extraction relies heavily on human experience, making it difficult to effectively capture long-term temporal dependencies in the performance degradation process, and also unable to fully model the complex nonlinear correlation between early weak degradation symptoms and final failure, resulting in insufficient prediction accuracy and robustness.

[0004] In summary, existing technologies have not yet effectively solved the key technical challenge of how to achieve high-precision and high-robustness prediction of remaining service life under strong noise and variable operating conditions. Summary of the Invention

[0005] In view of this, it is necessary to provide a method, device, electronic device and storage medium for predicting the life of a crane transmission system, in order to solve the technical problems of inaccurate prediction and low stability caused by the existing method of using a data-driven model based on historical monitoring data to learn the degradation law under strong noise and variable working conditions.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for predicting the lifespan of a crane transmission system, comprising:

[0007] Obtain a time series of health indicators that reflect the overall degradation trend of the crane transmission system. The time series of health indicators is constructed based on vibration signals collected from the transmission system under variable load and variable speed conditions. The time series of the health indicators are input into a pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, each of which uses nested long short-term memory network units as core feature transformers and a one-dimensional convolutional network as identity shortcut connections, and an attention layer is connected at the end of the network. Through sequence learning and attention-weighted processing of the residual nested long short-term memory attention prediction network, the remaining service life prediction value of the key components of the transmission system is output.

[0008] In one possible implementation, obtaining the time series of health indicators reflecting the overall degradation trend of the crane drive system includes: Acquire monitoring data of multi-source asynchronous heterogeneous crane drive systems; Extract fault features related to system degradation from the monitoring data; Based on the fault characteristics and monitoring data, the time series of the health indicators reflecting the overall degradation trend of the system is calculated.

[0009] In one possible implementation, extracting fault features related to system degradation from the monitoring data includes: Vibration signals of the transmission system during operation are collected by sensors placed on key components of the transmission system. A non-stationary signal processing method is used to extract fault features characterizing the health status of components from the vibration signal.

[0010] In one possible implementation, the residual block is a network structure with nested long short-term memory network units as the core operation layer and a one-dimensional convolutional network as the parallel fast path. The residual nested long short-term memory network is formed by sequentially stacking multiple residual blocks; An attention layer is connected before the final output of the residual nested long short-term memory network. The attention layer is used to learn and assign importance weights to the hidden state features at different time steps for the current lifespan prediction.

[0011] In one possible implementation, the nested long short-term memory network unit includes an outer long short-term memory hierarchy and an inner long short-term memory hierarchy; The input of the internal long short-term memory level is jointly regulated by the input gate output and the forget gate output of the external long short-term memory level; The output of the internal long short-term memory level, after being transformed, is used to update the memory cell state of the external long short-term memory level.

[0012] In one possible implementation, before inputting the health indicator time series into the pre-trained residual nested long short-term memory attention prediction network, the crane transmission system life prediction method includes a pre-training step of the residual nested long short-term memory attention prediction network, which includes: Obtain a historical dataset containing time series of health indicators and their corresponding labels of actual remaining lifespan; The historical dataset is divided into a training set and a validation set; Using the time series of health indicators in the training set as input and the corresponding real remaining lifespan labels as supervision signals, the residual nested long short-term memory attention prediction network is trained, and the training process is monitored and stopped early using the validation set until the network converges.

[0013] In one possible implementation, obtaining the historical dataset containing time series of health indicators and their corresponding true remaining lifespan labels includes: The original monitoring signal is subjected to Hilbert-Huang transform to obtain the marginal spectrum; The marginal spectrum is normalized and regularized to obtain normalized data. The regularized data is input into a stacked denoising autoencoder for deep feature extraction and fusion to obtain fused features, which are then used as the health indicator time series for network training.

[0014] On the other hand, the present invention also provides a crane transmission system life prediction device, comprising: The acquisition module is used to acquire time series of health indicators that reflect the overall degradation trend of the crane's transmission system; The input module is used to input the health indicator time series into a pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, wherein each residual block uses nested long short-term memory network units as core feature transformers, and uses a one-dimensional convolutional network as identity shortcut connections, and has an attention layer connected at the end of the network. The prediction module is used to output the remaining service life prediction value of the key components of the transmission system through sequence learning and attention weighting processing of the residual nested long short-term memory attention prediction network.

[0015] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the crane transmission system life prediction method described in any of the above implementations.

[0016] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instructions, which, when executed by a processor, can implement the steps in the crane transmission system life prediction method described in any of the above implementations.

[0017] The beneficial effects of this invention are: the crane transmission system life prediction method provided by this invention, through... Obtaining time series of health indicators reflecting the overall degradation trend of the system provides a quantitative and direct input basis for life prediction. By using a residual nested long short-term memory attention prediction network composed of stacked residual blocks, the residual structure is used to alleviate the gradient degradation problem in deep network training. The nested long short-term memory units enhance the ability to capture long-term complex temporal dependencies in performance degradation, and the attention mechanism is used to adaptively focus on historical states that are more critical to the current prediction. This allows for more accurate modeling of the system degradation process under complex operating conditions, ultimately achieving the prediction of remaining service life and improving the accuracy and reliability of the prediction of the remaining service life of crane transmission systems. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of an embodiment of the crane transmission system life prediction method provided by the present invention; Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of S101; Figure 3 For the present invention Figure 2 A schematic diagram of an embodiment of S202; Figure 4 A schematic flowchart of another embodiment of the crane transmission system life prediction method provided by the present invention; Figure 5 For the present invention Figure 4 A schematic diagram of an embodiment of S401; Figure 6A schematic diagram of an embodiment of the crane transmission system life prediction device provided by the present invention; Figure 7 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0022] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This invention provides a method, apparatus, electronic device, and storage medium for predicting the lifespan of a crane transmission system. The technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] Figure 1 A schematic flowchart of an embodiment of the crane transmission system life prediction method provided by the present invention is shown below. Figure 1 As shown, the life prediction method for crane transmission systems includes: S101. Obtain the time series of health indicators that reflect the overall degradation trend of the crane transmission system. The time series of health indicators is constructed based on the vibration signals collected by the transmission system under variable load and variable speed conditions. S102. Input the health indicator time series into the pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, where each residual block uses nested long short-term memory network units as core feature transformers and a one-dimensional convolutional network as identity shortcut connections, and an attention layer is connected at the end of the network. S103. Through sequence learning and attention-weighted processing of the residual nested long short-term memory attention prediction network, the remaining service life prediction value of the key components of the transmission system is output.

[0026] This embodiment provides a specific implementation of a method for predicting the lifespan of a crane transmission system.

[0027] First, obtain a time series of health indicators that reflect the overall degradation trend of the crane's transmission system.

[0028] The foundation for constructing the health indicator time series comes from vibration signals collected from the transmission system under real-world variable load and speed conditions. These vibration signals are acquired through accelerometers placed on key components such as the high-speed shaft, low-speed shaft, or bearing housings of the gearbox, effectively capturing the equipment's operating status. By performing time-domain, frequency-domain, and time-frequency-domain feature extraction and fusion calculations on these raw vibration signals, a quantitative sequence characterizing the evolution of the system's health status can be formed, thus providing direct and reliable input for subsequent intelligent prediction.

[0029] Next, the time series of this health indicator is fed into a pre-trained dedicated prediction network, namely the Residual Nested Long Short-Term Memory Attention Prediction Network (Prediction Network). This prediction network is a deep sequence model specifically built for modeling the complex degradation process of mechanical systems. Its core architecture consists of multiple stacked residual blocks.

[0030] Each residual block employs nested long short-term memory network units as core feature transformers to deeply mine the long-term temporal dependencies and complex patterns contained in the performance degradation data. At the same time, each residual block is connected in parallel with a one-dimensional convolutional network as an identity shortcut connection. This ensures that gradient information is effectively transmitted during deep network training, alleviates the gradient vanishing or exploding problem, and guarantees the stable learning ability of the prediction network.

[0031] At the end of the network, there is an attention layer that can autonomously learn and evaluate the differences in the importance of different state features in historical time steps for the current lifetime prediction, thereby enabling focus on key degradation stages.

[0032] Finally, through the aforementioned prediction network, which performs progressive feature learning and context modeling on the input health indicator time series, the network is able to capture the complete evolutionary pattern from early slight degradation to near failure. After weighted integration of important information via an attention mechanism, the network output generates predicted values ​​for the remaining service life of key components of the transmission system (such as high-speed gears or support bearings in the main hoisting gearbox).

[0033] This embodiment uses a time series of health indicators constructed from vibration signals as input and employs a dedicated deep neural network combining residual structures, nested long short-term memory units, and attention mechanisms for time series modeling and prediction. This effectively addresses the problem of insufficient modeling capability of traditional methods for the complex nonlinear degradation process of crane transmission systems. This approach enhances the capture of long-period time series dependencies and highlights the impact of key degradation stages, thereby significantly improving the accuracy and reliability of remaining service life prediction.

[0034] In some embodiments of the present invention, such as Figure 2 As shown, step S101, obtaining the time series of health indicators reflecting the overall degradation trend of the crane transmission system, includes: S201. Obtain monitoring data of the multi-source asynchronous heterogeneous transmission system of the crane; S202. Extract fault characteristics related to system degradation from monitoring data; S203. Based on fault characteristics and monitoring data, calculate the time series of health indicators that reflect the overall degradation trend of the system.

[0035] Specifically, this embodiment provides a specific implementation process for obtaining the health indicator time series reflecting the overall degradation trend of the crane transmission system in step S101. This process aims to transform the original and diverse equipment condition monitoring data into a quantitative sequence that can clearly and stably characterize the overall performance degradation path of the system, providing high-quality input for subsequent intelligent prediction models.

[0036] First, execute step S201 to obtain monitoring data of the multi-source asynchronous heterogeneous crane transmission system.

[0037] Among them, multi-source asynchronous heterogeneous monitoring data refers to a set of physical quantity data that are asynchronous in time and different in format and physical meaning, collected from different components of the transmission system through different types of sensors.

[0038] Specifically, this dataset may include, but is not limited to: vibration signals from the gearbox or bearing housing acquired by an accelerometer, bearing or lubricating oil temperature signals acquired by a temperature sensor, stress-strain signals from key structural components measured by strain gauges, and motor current and voltage signals obtained from the electrical cabinet. These data differ in sampling frequency, data dimensions, and the physical processes they reflect, collectively forming a three-dimensional information network describing the operating status of the transmission system.

[0039] Next, step S202 is performed to extract fault features related to system degradation from the above monitoring data.

[0040] Among them, fault characteristics related to system degradation refer to state parameters that can sensitively reflect the occurrence and development of damage such as wear, fatigue, and cracks in mechanical parts.

[0041] This step extracts core information about the system's health status from high-dimensional raw data. For example, time-domain features (such as root mean square value, peak value, and kurtosis), frequency-domain features (such as amplitude at specific fault frequencies and sideband energy), and time-frequency-domain features (such as wavelet packet node energy) can be extracted from vibration signals; trend features such as temperature rise rate and steady-state temperature value can be extracted from temperature signals. These features together constitute a multivariate feature vector describing the system's health status.

[0042] Then, step S203 is executed, and based on the extracted fault characteristics and monitoring data, a time series of health indicators reflecting the overall degradation trend of the system is calculated.

[0043] Among them, the health index is a scalar indicator designed to fuse or map multiple feature vectors into a one-dimensional, monotonic trend quantity, used to intuitively quantify the overall health of the system. Its purpose is to eliminate the interference of single-feature fluctuations and highlight the main trend of degradation. This step can be specifically implemented using statistical measurement methods, that is, calculating the intrinsic relationship between each feature sequence and the equipment degradation process. Commonly used quantitative indicators and their calculation formulas include:

[0044]

[0045]

[0046] In the formula, , and These are monotonicity, trend, and correlation. The total lifetime time series length. For the first The time series of the column, for Sequence data, belong , The total number of sensors, For time step, This is the average value. express The differential.

[0047] Will , and By comprehensively considering these three indicators, a reasonable linear feature optimization index is constructed for the overall evaluation of sensor data. The formula is

[0048] In the formula, and These are all weighting coefficients. Since the sensor data sequence directly affects the lifespan prediction results and degradation trends, the correlation index and monotonicity index have independent weighting coefficients. The trend index may be affected by local trends due to data fluctuations during the calculation process, and does not have a significant contribution to the overall feature selection index.

[0049] By comprehensively calculating one or more of the above indicators and weighting and fusing or selecting each feature, a health indicator time series with a value range typically between [0,1] and monotonically approaching the failure threshold (e.g., 1) as the equipment deteriorates can be generated.

[0050] This embodiment overcomes the limitations of a single signal source by integrating multi-source heterogeneous data, providing a data foundation for a comprehensive assessment of system status. It then selectively extracts fault features directly related to mechanical degradation, achieving an effective transformation from raw data to status information, which helps capture early, subtle signs of degradation. Furthermore, mathematical modeling fuses these multi-dimensional features into a single health indicator, which filters out interference such as operating condition fluctuations and clearly characterizes the gradual trend of performance degradation. This transforms complex, high-dimensional raw monitoring data into a stable, interpretable time-series signal with strong correlation to lifespan, providing accurate and reliable input for subsequent deep learning prediction models and fundamentally improving the accuracy and predictive power of lifespan prediction.

[0051] In some embodiments of the present invention, such as Figure 3 As shown, step S202 extracts fault features related to system degradation from the monitoring data, including: S301. Vibration signals of the transmission system during operation are collected by sensors placed on key components of the transmission system. S302. Employ non-stationary signal processing methods to extract fault features characterizing the health status of components from vibration signals.

[0052] This embodiment uses signal processing technology to accurately extract key information characterizing the early damage and degradation state of components from vibration monitoring signals.

[0053] First, step S301 is executed, in which vibration signals of the transmission system during operation are collected by sensors placed on key components of the transmission system.

[0054] Specifically, vibration acceleration sensors are typically placed at key locations along the drivetrain, such as the input shaft bearing housing, output shaft bearing housing, and drum bearing housing of the main hoisting gearbox—areas where vibration energy is concentrated and fault-sensitive. These sensors can acquire, in real time, broadband vibration response signals generated by gear meshing, bearing rolling, and structural resonance during hoisting, luffing, and slewing operations. This signal contains the richest and most direct information about the equipment's health status, making it one of the most important data sources for fault diagnosis and lifespan prediction.

[0055] Next, the core step S302 is executed, employing a non-stationary signal processing method to extract fault features characterizing the health status of components from the aforementioned vibration signals. The purpose of this step is to address the analytical challenges arising from the non-stationary characteristics of crane vibration signals under varying load and speed conditions.

[0056] In this embodiment, non-stationary signal processing methods specifically refer to a class of analytical techniques designed to process signals whose statistical characteristics (such as mean, variance, and frequency components) change over time. Because the crane's operating speed and load constantly change, the amplitude and frequency components of its vibration signal also change over time, making traditional spectrum analysis methods based on the stationary assumption ineffective. In this embodiment, the following targeted methods can be specifically employed: A direct, fast iterative filtering method for extracting weak fault features: This method is suitable for separating low-energy periodic impact components caused by early local damage (such as initial spalling of bearings or micro-cracks in gears) from strong background noise. Through adaptive filtering, it can effectively enhance fault features with extremely low signal-to-noise ratios in the signal. An improved multiple synchronous compression transform method for extracting time-varying instantaneous frequency features under variable speed conditions: This method is particularly suitable for tracking the continuous variation trajectory of fault characteristic frequencies (such as gear meshing frequency and its sidebands, bearing passage frequency) in the time-frequency domain under fluctuating speed conditions. Its discretized expression describes the energy redistribution process of the signal in the time-frequency plane, and can be represented as an integral and summation form over a specific operator to achieve higher time-frequency convergence. A dual-dictionary sparse representation method for extracting features of combined gear and bearing faults: This method is applicable to complex situations where different morphologies of impact components in the vibration signal are coupled together when both gears and bearings are faulty. By constructing overcomplete dictionaries matching the gear fault waveform and the bearing fault waveform respectively, and using an optimization algorithm to solve for the sparsest representation of the signal under these dictionaries, effective separation and reconstruction of various source features in the combined fault can be achieved. For example, when applying the improved multiple synchronous compression transformation method, the acquired bearing housing vibration signal is first transformed to obtain a high-resolution time-frequency distribution map. If there is damage to the outer ring of the bearing, a characteristic ridge line that is proportional to the shaft rotation frequency and whose energy changes with time (corresponding to the rotation speed) can be observed in the time-frequency map. By extracting this ridge line, the fault type can be diagnosed and its intensity can be quantified. This quantified intensity can be used as an effective fault feature.

[0057] This embodiment directly collects vibration signals to ensure that the raw data contains rich state information. Then, it adopts a non-stationary signal processing method to effectively extract weak early fault signs, fault frequencies that vary with rotational speed, and mutually coupled composite fault features from complex time-varying signals. This provides reliable feature data for constructing high-quality health indicators, significantly improving the accuracy and effectiveness of the input data for subsequent life prediction models, so that the final prediction results can better reflect the true remaining life of the equipment.

[0058] In some embodiments of the present invention, the residual block is a network structure with nested long short-term memory network units as the core operation layer and a one-dimensional convolutional network as the parallel fast path. Nested residual long short-term memory networks are formed by sequentially stacking multiple residual blocks; An attention layer is connected before the final output of the residual nested long short-term memory network. The attention layer is used to learn and assign importance weights to the hidden state features at different time steps for the current lifespan prediction.

[0059] In some embodiments of the present invention, the nested long short-term memory network unit includes an outer long short-term memory hierarchy and an inner long short-term memory hierarchy; The input of the internal long short-term memory level is jointly regulated by the input gate output and the forget gate output of the external long short-term memory level; The output of the internal long short-term memory level is transformed and used to update the memory cell state of the external long short-term memory level.

[0060] The core architecture of the prediction network defined in this embodiment aims to build a dedicated deep learning model that can deeply mine long-term temporal dependencies in health indicator sequences and perform stable training.

[0061] In this embodiment, the residual block adopts a specific fusion structure: nested long short-term memory network units are used as the core operation layer, which are responsible for performing the main feature transformation and memory functions of the sequence data; at the same time, a one-dimensional convolutional network is used as a parallel shortcut path.

[0062] In this system, the one-dimensional convolutional network performs a linear projection or simple feature extraction on the input sequence, and its output is added to the output of the nested long short-term memory network units along the channel dimension. This constitutes a residual connection.

[0063] Specifically, if the input of the residual block is denoted as x The output after transformation by nested long short-term memory network units is denoted as F(x) The output after transformation by a one-dimensional convolutional network is denoted as G(x) Then the final output of the residual block H(x) It can be represented as H(x) = F(x) + G(x) This structure allows gradients to pass through a fast path during backpropagation training. G(x) Direct transmission effectively alleviates the gradient vanishing or exploding problems common in deep networks, allowing for the construction of deeper networks to enhance model capacity.

[0064] Multiple residual blocks as defined above are stacked sequentially to form a deep residual nested long short-term memory network. Each residual block receives the output sequence of the previous block as input for higher-level feature abstraction.

[0065] At the end of the stacked network, before the final prediction output, an attention layer is connected. This layer works by receiving the hidden state sequence from all time steps of the residual network's output and calculating a weight score for the state at each time step. This weight score is learned by the network itself and represents the importance of the state information at that historical moment for making the current remaining lifetime prediction decision.

[0066] Finally, the hidden states of all time steps are weighted and summed to form a comprehensive context vector, which is then fed into subsequent fully connected layers for regression prediction. This allows the model to autonomously focus on more decisive stages in the degradation process (such as the period of accelerated failure deterioration), rather than treating all historical information equally.

[0067] Furthermore, nested long short-term memory network units are a structural refinement of standard long short-term memory network units, containing a two-level LSTM structure: an outer long short-term memory level and an inner long short-term memory level.

[0068] The outer layer functions similarly to a standard LSTM, processing the main time series and generating the main memory stream. Its unique characteristic lies in the fact that the input to the inner layer does not come directly from the original input or the previous time-series internal state, but is dynamically controlled by the immediate decision signals from the outer layer.

[0069] Specifically, in the computation of a time step, the input gate of the outer layer... And the Gate of Oblivion The outputs of these gates are used together to control the computation of the internal layers. The input gates of the outer layers output... After a transformation (such as multiplying with the candidate memory cell state and processing with an activation function), the input vector for the inner hierarchy is generated. Meanwhile, the forget gate output at the outer level... external memory cell state at the previous time step Multiply to generate the hidden state of the previous time step in the inner hierarchy. Internal hierarchy is based on ( , Using the input pair as its own, it independently performs a round of memory update including its own gating operation to obtain the internal memory state. Subsequently, the output of the inner level undergoes a transformation (e.g., through the output gate of the outer level). Regulation and Multiplication) is used to update the state of memory cells in the outer short-term memory level at the current time step. This nesting mechanism enables more refined and deeper temporal dependency modeling within the unit, and introduces secondary thinking and information filtering capabilities to the updates of the external main memory stream, thereby enhancing the model's ability to capture complex, long-term degradation patterns.

[0070] This embodiment addresses the gradient transfer challenge in building deep time-series prediction models through residual connections, resulting in a more stable training process and enabling the extraction of more abstract features from deeper networks. A two-layer processing mechanism with nested long short-term memory units significantly enhances the model's depth and granularity in depicting the slow, long-term, and complex nonlinear time-series process of crane drive system performance degradation. By introducing an attention mechanism, the model gains the ability to dynamically and differentially assess the importance of historical information, allowing its predictive decisions to rely more on key turning points or significant characteristic periods in the degradation process, rather than being diluted by a large amount of data from normal or gradual degradation phases. Through the synergistic effect of these multiple network structures, the accuracy, robustness, and interpretability of the final remaining useful life prediction are improved.

[0071] In some embodiments of the present invention, before inputting the time series of health indicators into a pre-trained residual nested long short-term memory attention prediction network, such as Figure 4As shown, the crane transmission system life prediction method includes a pre-training step of a residual nested long short-term memory attention prediction network. This pre-training step includes: S401. Obtain a historical dataset containing time series of health indicators and their corresponding real remaining lifespan labels; S402. Divide the historical dataset into a training set and a validation set; S403. Using the time series of health indicators in the training set as input and the corresponding real remaining lifespan labels as supervision signals, train the residual nested long short-term memory attention prediction network, and use the validation set to monitor the training process and stop it early until the network converges.

[0072] Specifically, to enable the residual nested long short-term memory attention prediction network to have accurate lifespan prediction capabilities, it needs to be pre-trained. This pre-training step is a process of optimizing the network's internal parameters using supervised learning based on historical data. Its purpose is to enable the network to learn to map the correct remaining lifespan value from the input health indicator sequence.

[0073] The pre-training process begins with step S401, which involves acquiring a historical dataset containing time series of health indicators and their corresponding labels for the actual remaining lifespan.

[0074] In this context, the historical dataset specifically refers to the collection of data that provides paradigms for model learning, and its construction forms the basis for training. Each sample in this dataset typically contains data from a complete device lifecycle: from initial use to failure or severe performance degradation.

[0075] The health indicator time series is a numerical sequence constructed according to the aforementioned embodiments, characterizing the evolution of the device's health status throughout its entire lifespan. The true remaining useful life label corresponds to the actual value at each point in time in this sequence, typically expressed in units of cycle count or runtime.

[0076] For example, for a bearing that fails after 1000 hours of operation, the actual remaining service life label corresponding to the health indicator sequence collected at the 800th hour would be 200 hours. Obtaining a sufficiently large dataset that covers different failure modes and operating conditions is a prerequisite for the model to achieve strong generalization ability.

[0077] Next, step S402 is executed to divide the historical dataset into a training set and a validation set.

[0078] The training set is a subset of data directly used to adjust the network weights and parameters. The model minimizes prediction error by repeatedly learning from the sample patterns within it. The validation set, on the other hand, is not directly involved in parameter adjustment. It is used to monitor the model's performance on unseen data, prevent overfitting, select the best model, and determine when to stop training.

[0079] A common method of partitioning is to randomly divide the data according to a preset ratio (such as 7:3 or 8:2) to ensure that the data distribution in the two sets is basically consistent.

[0080] Model optimization is completed in step S403. This step uses the time series of health indicators in the training set as input features and their corresponding true remaining lifespan labels as supervision signals (i.e., the target values ​​that the model learns to fit) to iteratively train the residual nested long short-term memory attention prediction network.

[0081] Specifically, the training process is implemented using the backpropagation algorithm and an optimizer (such as the Adam optimizer). In each iteration, the network makes a prediction based on the current parameters and calculates the difference between the predicted value and the true label (i.e., the loss) using a defined loss function. The optimizer calculates the gradient based on the loss value and updates all the network's weight parameters to make the loss value change in the direction of decreasing.

[0082] The entire training process is conducted under the monitoring of the validation set. After a certain number of training rounds (called an epoch), the model's performance is evaluated on the validation set (e.g., the loss value is calculated).

[0083] Early stopping, also known as the early stopping method, is a commonly used regularization technique. Specifically, it works by automatically terminating training when the loss on the validation set stops decreasing or even starts to increase after several consecutive epochs (e.g., 10 epochs), indicating that the model's performance has reached its limit. The network then rolls back to the set of network parameters corresponding to the lowest loss on the validation set. This mechanism effectively prevents the model from overfitting the training data, ensuring that it learns universal degradation patterns rather than noise from the training data, thus guaranteeing that the trained network has good generalization ability.

[0084] This embodiment establishes a reliable model performance evaluation mechanism by constructing a labeled historical dataset and scientifically dividing the dataset. Then, by combining supervised learning, backpropagation, optimizers, and early stopping techniques, it can automatically mine the complex nonlinear mapping relationship between health indicators and remaining lifespan from historical data and encode this relationship into the weights of the network.

[0085] In some embodiments of the present invention, such as Figure 5As shown, step S401, obtaining a historical dataset containing time series of health indicators and their corresponding true remaining lifespan labels, includes: S501. Perform Hilbert-Huang transform on the original monitoring signal to obtain the marginal spectrum; S502. Perform data normalization and regularization on the marginal spectrum to obtain normalized data; S503. Input the regularized data into the stacked denoising autoencoder for deep feature extraction and fusion to obtain fused features, and use the fused features as a health indicator time series for network training.

[0086] In step S501, the original monitoring signal is subjected to Hilbert-Huang transform to obtain the marginal spectrum, which aims to transform the time-domain signal into a frequency-domain representation that can better reveal the energy distribution of the fault frequency components.

[0087] The Hilbert-Huang transform is a method suitable for analyzing nonlinear and non-stationary signals. Its process mainly consists of two steps: the first step is empirical mode decomposition, which transforms the original signal... x(t) Adaptively decomposed into a series of eigenmode function components from high frequency to low frequency. c j (t) and a residual term r n (t) ,Right now The second step is to analyze each IMF component. c j (t) Perform a Hilbert transform to obtain its analytic signal, and then calculate the instantaneous frequency. ω j (t) and instantaneous amplitude a j (t) .

[0088] Among them, marginal spectrum h(ω) It is obtained by integrating the Hilbert amplitude spectrum over the entire time range, and its calculation formula is as follows: h(ω) = ,in H(ω,t) This is the Hilbert spectrum. The marginal spectrum describes the total energy accumulation of the signal at each frequency point, and can clearly and stably show the energy peak at the fault characteristic frequency. It is more suitable for non-stationary vibration signals than the traditional Fourier spectrum.

[0089] In step S502, the obtained marginal spectrum is normalized and data regularized to obtain regularized data. Since the signal energy levels of different devices and different acquisitions may vary, data regularization and normalization are key steps to ensure data scale consistency and facilitate subsequent deep network processing.

[0090] Specific operations may include: unifying marginal spectra of different lengths to the same number of frequency points using interpolation methods; subsequently, normalizing the energy value at each frequency point, for example, by using max-min normalization to map the value to the [0,1] interval. After this step, the input data is standardized, which helps improve the stability and convergence speed of model training.

[0091] Finally, step S503 is executed, in which the normalized marginal spectrum data is input into the stacked denoising autoencoder for deep feature extraction and fusion, and the output fused features are used as a health indicator time series for network training.

[0092] Stacked denoising autoencoders (DAEs) are deep neural networks composed of multiple cascaded denoising autoencoder (DAE) layers, used for unsupervised feature learning. A basic denoising autoencoder consists of an encoder and a decoder. The encoder maps input data with random noise to a low-dimensional representation in the hidden layers. The decoder then attempts to reconstruct the clean original input data from the noisy low-dimensional representation. Its training objective is to minimize the reconstruction error, such as mean squared error.

[0093] In this way, the network is forced to learn essential features in the data that are resistant to noise interference. By stacking multiple such DAEs, with the hidden layer output of the previous DAE serving as the input of the next DAE, abstract features from low to high levels can be extracted layer by layer. Finally, the output of an intermediate layer in the stack (usually the output of the last encoder) is taken as the fused feature. This feature integrates multi-level information from the original marginal spectrum, possessing higher representational power and stronger indicativeness of degenerate states.

[0094] This embodiment obtains the marginal spectrum through the Hilbert-Huang transform, converting the non-stationary time-domain vibration signal into frequency-domain features that can stably characterize the energy of the fault frequency, providing high-quality initial input for deep learning. Furthermore, through normalization and regularization, dimensional and scale differences are eliminated, creating conditions for stable model learning. Then, using a stacked denoising autoencoder, highly abstract, noise-free, and closely related to the intrinsic mechanism of equipment degradation deep fusion features are automatically learned from a large amount of marginal spectrum data in an unsupervised manner. These features contain richer predictive information, thus providing more discriminative and predictive training samples for the subsequent residual nested long short-term memory attention prediction network, improving the accuracy and generalization ability of the final lifetime prediction model from the data source.

[0095] To better implement the crane transmission system life prediction method in this invention embodiment, based on the crane transmission system life prediction method, correspondingly, as follows: Figure 6As shown, this embodiment of the invention also provides a crane transmission system life prediction device, the crane transmission system life prediction device 600 comprising: Module 601 is used to acquire time series of health indicators that reflect the overall degradation trend of the crane transmission system. The input module 602 is used to input the health indicator time series into a pre-trained residual nested long short-term memory attention prediction network; the residual nested long short-term memory attention prediction network is composed of stacked residual blocks, wherein each residual block uses nested long short-term memory network units as core feature transformers, and uses a one-dimensional convolutional network as identity shortcut connections, and has an attention layer connected at the end of the network. The prediction module 603 is used to output the remaining service life prediction value of the key components of the transmission system through the sequence learning and attention weighting processing of the residual nested long short-term memory attention prediction network.

[0096] The crane transmission system life prediction device 600 provided in the above embodiments can realize the technical solutions described in the above embodiments of the crane transmission system life prediction method. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the crane transmission system life prediction method, which will not be repeated here.

[0097] like Figure 7 As shown, the present invention also provides an electronic device 700. The electronic device 700 includes a processor 701, a memory 702, and a display 703. Figure 7 Only some components of the electronic device 700 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0098] In some embodiments, processor 701 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 702 or process data, such as the crane drive system life prediction method of the present invention.

[0099] In some embodiments, processor 701 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 701 may be local or remote. In some embodiments, processor 701 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, intranet, multi-cloud, etc., or any combination thereof.

[0100] In some embodiments, memory 702 may be an internal storage unit of electronic device 700, such as a hard disk or memory of electronic device 700. In other embodiments, memory 702 may also be an external storage device of electronic device 700, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 700.

[0101] Furthermore, the memory 702 may include both internal storage units of the electronic device 700 and external storage devices. The memory 702 is used to store application software and various types of data installed on the electronic device 700.

[0102] In some embodiments, display 703 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 703 is used to display information from electronic device 700 and to display a visual user interface. Components 701-703 of electronic device 700 communicate with each other via a system bus.

[0103] In one embodiment, when processor 701 executes the crane drive system life prediction program in memory 702, the following steps can be implemented: Obtain time series of health indicators reflecting the overall degradation trend of crane transmission systems; The time series of the health indicators are input into a pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, each of which uses nested long short-term memory network units as core feature transformers and a one-dimensional convolutional network as identity shortcut connections, and an attention layer is connected at the end of the network. Through sequence learning and attention-weighted processing of the residual nested long short-term memory attention prediction network, the remaining service life prediction value of the key components of the transmission system is output.

[0104] It should be understood that when the processor 701 executes the crane transmission system life prediction program in the memory 702, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0105] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 700 mentioned. Electronic device 700 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, electronic device 700 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0106] Accordingly, this application also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the programs or instructions are executed by a processor, they can implement the steps or functions of the crane transmission system life prediction method provided in the above-described method embodiments.

[0107] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0108] The crane transmission system life prediction method, device, electronic device, and storage medium provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting the lifespan of a crane transmission system, characterized in that, include: Obtain a time series of health indicators that reflect the overall degradation trend of the crane transmission system. The time series of health indicators is constructed based on vibration signals collected from the transmission system under variable load and variable speed conditions. The time series of the health indicators are input into a pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, each of which uses nested long short-term memory network units as core feature transformers and a one-dimensional convolutional network as identity shortcut connections, and an attention layer is connected at the end of the network. Through sequence learning and attention-weighted processing of the residual nested long short-term memory attention prediction network, the remaining service life prediction value of the key components of the transmission system is output.

2. The method for predicting the lifespan of a crane transmission system according to claim 1, characterized in that, The acquisition of time series health indicators reflecting the overall degradation trend of the crane transmission system includes: Acquire monitoring data of multi-source asynchronous heterogeneous crane drive systems; Extract fault features related to system degradation from the monitoring data; Based on the fault characteristics and monitoring data, the time series of the health indicators reflecting the overall degradation trend of the system is calculated.

3. The method for predicting the lifespan of a crane transmission system according to claim 2, characterized in that, The extraction of fault features related to system degradation from the monitoring data includes: Vibration signals of the transmission system during operation are collected by sensors placed on key components of the transmission system. A non-stationary signal processing method is used to extract fault features characterizing the health status of components from the vibration signal.

4. The method for predicting the lifespan of a crane transmission system according to claim 1, characterized in that, The residual block is a network structure with nested long short-term memory network units as the core operation layer and a one-dimensional convolutional network as the parallel fast path. The residual nested long short-term memory network is formed by sequentially stacking multiple residual blocks; An attention layer is connected before the final output of the residual nested long short-term memory network. The attention layer is used to learn and assign importance weights to the hidden state features at different time steps for the current lifespan prediction.

5. The method for predicting the lifespan of a crane transmission system according to claim 4, characterized in that, The nested long short-term memory network unit includes an outer long short-term memory hierarchy and an inner long short-term memory hierarchy; The input of the internal long short-term memory level is jointly regulated by the input gate output and the forget gate output of the external long short-term memory level; The output of the internal long short-term memory level, after being transformed, is used to update the memory cell state of the external long short-term memory level.

6. The method for predicting the lifespan of a crane transmission system according to claim 1, characterized in that, Before inputting the health indicator time series into the pre-trained residual nested long short-term memory attention prediction network, the crane transmission system life prediction method includes a pre-training step of the residual nested long short-term memory attention prediction network, which includes: Obtain a historical dataset containing time series of health indicators and their corresponding labels of actual remaining lifespan; The historical dataset is divided into a training set and a validation set; Using the time series of health indicators in the training set as input and the corresponding real remaining lifespan labels as supervision signals, the residual nested long short-term memory attention prediction network is trained, and the training process is monitored and stopped early using the validation set until the network converges.

7. The method for predicting the lifespan of a crane transmission system according to claim 6, characterized in that, The process of obtaining a historical dataset containing time series of health indicators and their corresponding true remaining lifespan labels includes: The original monitoring signal is subjected to Hilbert-Huang transform to obtain the marginal spectrum; The marginal spectrum is normalized and regularized to obtain normalized data. The regularized data is input into a stacked denoising autoencoder for deep feature extraction and fusion to obtain fused features, which are then used as the health indicator time series for network training.

8. A life prediction device for a crane transmission system, characterized in that, include: The acquisition module is used to acquire a time series of health indicators that reflect the overall degradation trend of the crane transmission system. The time series of health indicators is constructed based on vibration signals collected by the transmission system under variable load and variable speed conditions. The input module is used to input the health indicator time series into a pre-trained residual nested long short-term memory attention prediction network. The residual nested long short-term memory attention prediction network is composed of stacked residual blocks, wherein each residual block uses nested long short-term memory network units as core feature transformers, and uses a one-dimensional convolutional network as identity shortcut connections, and has an attention layer connected at the end of the network. The prediction module is used to output the remaining service life prediction value of the key components of the transmission system through sequence learning and attention weighting processing of the residual nested long short-term memory attention prediction network.

9. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the crane transmission system life prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the crane transmission system life prediction method according to any one of claims 1 to 7.

Images