Lifting motor fault diagnosis model, method and system based on multi-scale star aggregation network

Abstract

The invention discloses a hoisting motor fault diagnosis model, method and system based on a star aggregation network, and belongs to the technical field of hoisting machinery fault diagnosis. The method comprises the following steps: acquiring vibration signal data of a lifting motor, and performing multi-working-condition fault diagnosis data set construction and preliminary data screening; performing time-frequency feature extraction on the data of the data set by using synchronous compression wavelet transform; constructing a multi-scale star aggregation network fault diagnosis model; inputting the time-frequency image data training set into a fault diagnosis model for model training, and obtaining optimal model parameters of the fault diagnosis model based on a loss function; and constructing a multi-working-condition fault diagnosis model of the lifting motor according to the determined model parameters, and performing fault diagnosis. According to the method, the problems of difficulty in fault feature extraction and low fault diagnosis precision under variable working conditions of the lifting motor of the hoisting machinery can be solved, the limitation of single feature extraction of a traditional algorithm is overcome, and the accuracy, the adaptability and the generalization capability of fault diagnosis of the lifting motor are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of crane machinery fault diagnosis technology, specifically relating to a hoisting motor fault diagnosis model, method and system based on a multi-scale star-shaped aggregation network. Background Technology

[0002] Large-scale lifting machinery is widely used in many key fields such as port loading and unloading, construction, and industrial manufacturing, occupying an indispensable position in modern industrial production and infrastructure construction. As the core drive device of the lifting mechanism in large-scale lifting machinery such as bridge erecting machines, the stability and reliability of the hoisting motor plays a decisive role in the production efficiency and safe operation of the entire lifting operation. Hoisting motors operate in complex environments, needing to withstand harsh working conditions and frequent start-stop heavy loads, making them prone to faults such as stator inter-turn short circuits and rotor bar breakage. These faults can range from affecting performance to causing safety accidents, making efficient fault diagnosis crucial. The current mainstream fault diagnosis methods and their shortcomings are as follows: (1) Time-frequency analysis method.

[0003] This method employs techniques such as Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), Variable Mode Decomposition (VMD), and Synchronous Compressed Wavelet Transform (SSWT) to process motor vibration signals and extract features. However, traditional time-frequency analysis methods have several drawbacks: STFT suffers from a trade-off between time and frequency resolution, making it difficult to capture high-frequency transient components of early faults; CWT requires specific selection of wavelet basis functions, resulting in poor adaptability; and VMD requires manual parameter setting, leading to insufficient multi-channel data collaborative analysis and mode aliasing under noise, thus affecting data preprocessing and feature extraction performance.

[0004] (2) Core methods of fault diagnosis: including expert knowledge-based diagnosis methods, analytical model-based diagnosis methods, and data-driven diagnosis methods.

[0005] Diagnostic methods based on expert knowledge rely on domain experts to formulate rules for reasoning and analysis, and are suitable for scenarios with large-scale knowledge bases. Their limitations include: a lack of mathematical logic support, resulting in poor applicability in computer vision and high-level pattern recognition tasks involving big data; high cost of acquiring expert knowledge, difficulty in continuous learning and evolution, and limited generalization ability as the knowledge system struggles to cover all faults in complex equipment failure scenarios.

[0006] Diagnostic methods based on analytical models establish accurate mathematical models of equipment and diagnose faults by comparing the model output with actual measurements. However, these methods have several drawbacks: high modeling costs for complex systems; parameter correction or model reconstruction required when system dynamics change; exponentially increasing maintenance costs in large equipment; and a significant conflict between computational complexity and modeling accuracy when dealing with complex systems coupled with multi-physics fields.

[0007] Data-driven diagnostic methods employ techniques such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), combined with signal processing and AI algorithms, to uncover latent patterns in data and achieve fault identification, overcoming some of the shortcomings of the first two types of methods. However, these methods suffer from several drawbacks: traditional models often employ channel-independent strategies, ignoring channel correlations and leading to performance degradation; while attention mechanisms alleviate some issues, in multi-sensor scenarios, they suffer from high computational complexity due to multi-channel interaction, susceptibility to weight distribution drift, and insufficient multi-scale feature fusion, making them unsuitable for the real-time diagnostic needs of hoisting motors under complex operating conditions. Summary of the Invention

[0008] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0009] Therefore, the purpose of this invention is to provide a hoisting motor fault diagnosis model, method and system based on a multi-scale star-shaped aggregation network, which can solve the problems of difficulty in fault diagnosis of hoisting motors under different working conditions and the single feature extraction method of existing fault diagnosis algorithms, and improve the accuracy and robustness of fault diagnosis.

[0010] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a fault diagnosis model for hoisting motors based on a star-shaped aggregation network. The model is a multi-scale star-shaped aggregation network fault diagnosis model based on time-frequency domain data analysis. The model integrates residual neural networks, multilayer perceptrons, and star-shaped aggregation modules to form a network architecture that features feature extraction, cross-channel interaction, and multi-scale fusion.

[0011] In addition, the hoisting motor fault diagnosis model based on a star-shaped aggregation network according to the present invention may also have the following additional technical features: In some implementations, the residual neural network mitigates gradient degradation through a skip connection mechanism to extract multi-level features from time-frequency images; the multilayer perceptron is used to realize dimensional transformation and nonlinear mapping of sequence representation; the star-shaped aggregation module aggregates multi-channel information through a centralized strategy, constructs a global core representation, and copies and distributes the global core representation to each channel. After being concatenated with local features, it is mapped back to the original dimension by the multilayer perceptron to achieve cross-channel multi-scale feature fusion.

[0012] In some implementations, the multilayer perceptron maps the input feature information of all channels to a higher-dimensional space, and the star-shaped aggregation module obtains the global core representation through random pooling, which encodes the global information of all channels. The global core representation is copied and distributed to each channel, and then concatenated with the original sequence representation of each channel. The concatenated representation is mapped back to the original dimension through a multilayer perceptron, and then added to the original representation to achieve the fusion of the global core representation and the representation of each channel.

[0013] This invention provides a method for diagnosing hoisting motor faults based on a star-shaped aggregation network, the method comprising: S1: Construction of hoist motor fault diagnosis dataset. By acquiring hoist motor vibration signal data, a multi-condition fault diagnosis dataset is constructed and preliminary data screening is performed. S2: Fault data feature extraction and data processing, using synchronous compressed wavelet transform to extract time-frequency features from the dataset; S3: Construction of a fault diagnosis model for hoisting motors. By integrating residual neural networks, multilayer perceptrons, and star aggregation modules, a multi-scale star aggregation network fault diagnosis model is constructed. S4: Data-driven fault diagnosis model training. The time-frequency image data training set is input into the fault diagnosis model for model training, and the optimal model parameters of the fault diagnosis model are obtained based on the loss function.

[0014] In addition, the hoisting motor fault diagnosis method based on a star-shaped aggregation network according to the present invention may also have the following additional technical features: In some implementations, step S1 uses a six-element collaborative approach to construct a multi-condition fault diagnosis dataset.

[0015] In some of these implementations, the dataset is constructed using the following formula: , In the formula, DS c For the dataset, M To diagnose the operational status of the object, L To diagnose the stress conditions on the object, F For operating speed or frequency, N For the number of data collections, S For sampling frequency and data type, A This describes the data filtering process for the dataset.

[0016] In some implementations, the data filtering in step S1 includes: analyzing the time-domain waveform using Matlab, using the sliding window method to extract several segments of steady-state signal, and removing transient interference data during start-stop operations.

[0017] In some implementations, step S2 includes: processing the pre-screened vibration signal using time-frequency analysis to construct a time-frequency distribution spectrum that can simultaneously characterize the signal's time-domain characteristics and frequency-domain resolution, thereby depicting in detail the variation of the signal's frequency components over time. The time-frequency analysis method specifically adopts the synchronous compressed wavelet transform method. Based on the wavelet transform principle, this method can extract multi-scale features of a signal by capturing its instantaneous frequency, abrupt change points, and local details. The rich feature information facilitates the fusion and classification of information at different scales by the neural network. The synchronous compressed wavelet transform effectively suppresses spectral energy diffusion by introducing the concept of instantaneous frequency, causing energy to accumulate near the wavelet center frequency, thereby enhancing the time-frequency representation capability and achieving the technical effect of accelerating the training process of fault diagnosis models and reducing computational resource consumption.

[0018] In some implementations, the fault diagnosis process in step S5 includes: The four-dimensional time-frequency image dataset, after feature extraction by synchronous compressed wavelet transform, is input into the trained multi-scale star-shaped aggregation network fault diagnosis model; the four dimensions of the four-dimensional time-frequency image dataset correspond to the input batch, the number of monitoring channels, time, and frequency, respectively. Feature extraction is performed by three residual modules connected in series in the model to reduce the dimensionality of the time-frequency graph. Multi-scale global core representations are obtained through multi-layer perceptron and random pooling layer in the model. Then, the global core representations are spliced ​​and fused with the information of each channel through multi-layer perceptron and fully connected layer in the model to obtain the final hoisting motor fault diagnosis result.

[0019] This invention also provides a multi-condition fault diagnosis system for hoisting motors based on star-shaped aggregation networks, used to implement the multi-condition fault diagnosis method for hoisting motors based on multi-scale star-shaped aggregation networks as described above; the system includes: The dataset construction and filtering module is configured to acquire hoist motor vibration signal data, and to construct multi-condition fault diagnosis datasets and perform preliminary data filtering. The feature extraction module is configured to extract time-frequency features from the dataset using synchronous compressed wavelet transform. The model building module is configured to build a multi-scale star-shaped aggregation network fault diagnosis model architecture. The model parameter determination module is configured to input the time-frequency image data training set into the fault diagnosis model for model training, and obtain the optimal model parameters of the fault diagnosis model based on the loss function. The fault diagnosis module is configured to construct a multi-public fault diagnosis model for the hoisting motor based on the multi-scale star-shaped aggregation network fault diagnosis model architecture and the optimal model parameters, and then perform fault diagnosis.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the hoisting motor fault diagnosis method based on multi-scale star aggregation network provides a specific hoisting motor fault diagnosis dataset, extracts fault features using synchronous compressed wavelet transform, and combines a multi-scale star aggregation network algorithm to achieve fault diagnosis. Finally, it constructs a fault diagnosis system architecture based on digital twin, which solves the problems of difficulty in fault diagnosis of hoisting motors under varying operating conditions and the single feature extraction method of existing fault diagnosis algorithms, thereby improving the accuracy and robustness of fault diagnosis. In this embodiment of the invention, the hoisting motor fault diagnosis method based on multi-scale star aggregation network provides a multi-condition dataset construction method, proposes a six-element collaborative dataset construction formula (Equation 1), and the sampling elements cover key operating condition parameters such as equipment status, load, and frequency, and perform data filtering; In this embodiment of the invention, the hoisting motor fault diagnosis method based on a multi-scale star-shaped aggregation network provides an improved feature extraction technique: synchronous compressed wavelet transform (SSWT) is used to process non-stationary vibration signals, solving the modal aliasing and resolution limitation problems of traditional time-frequency analysis methods (such as STFT / CWT); In this embodiment of the invention, the hoisting motor fault diagnosis method based on a multi-scale star-shaped aggregation network provides a multi-scale star-shaped aggregation network architecture: innovatively integrating residual networks, multilayer perceptrons (MLPs), and star-shaped aggregation modules; replacing the traditional distributed attention mechanism with a centralized information interaction strategy (aggregation → distribution fusion), significantly reducing computational complexity (linear complexity); achieving efficient fusion of multi-scale features across channels, and improving the generalization ability across working conditions.

[0021] The hoisting motor fault diagnosis system based on a multi-scale star-shaped aggregation network of the present invention can realize the hoisting motor fault diagnosis method based on a multi-scale star-shaped aggregation network, and therefore has at least all the features and advantages of the hoisting motor fault diagnosis method based on a multi-scale star-shaped aggregation network, which will not be repeated here. Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of a multi-condition fault diagnosis process for a hoisting motor based on a multi-scale star-shaped aggregation network, as disclosed in one embodiment of the present invention. Figure 2This is a flowchart of the construction process for a fault diagnosis dataset for a hoisting motor, as disclosed in one embodiment of the present invention. Figure 3 This is a time-domain waveform diagram of the vibration signal of a hoisting motor with an inter-turn short circuit fault, as disclosed in an embodiment of the present invention. Figure 4 This is a data filtering diagram of a hoisting motor fault diagnosis dataset disclosed in one embodiment of the present invention; Figure 5 This is a schematic diagram of a centralized information interaction method and a distributed interaction method of the attention mechanism in a star-shaped aggregation module, as disclosed in an embodiment of the present invention. Figure 6 This is a diagram of a star-shaped aggregation network structure for multi-scale core fusion disclosed in one embodiment of the present invention; Figure 7 This is an architecture diagram of a bridge erecting machine hoisting motor fault diagnosis system disclosed in one embodiment of the present invention; Figure 8 This is a functional block diagram of a bridge erecting machine lifting motor digital twin platform system disclosed in one embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0025] This invention addresses the difficulties in fault diagnosis of hoisting motors under varying operating conditions and the limited feature extraction methods of existing fault diagnosis algorithms. Based on typical fault conditions of hoisting motors, it studies the construction of fault diagnosis datasets and data preprocessing methods, enabling effective acquisition and feature extraction of vibration signals under these conditions. Further research into multi-condition fault diagnosis technology for hoisting motors is conducted. Based on algorithms such as multilayer perceptrons and deep residual networks, combined with a star-shaped aggregation module, a multi-scale star-shaped aggregation network fault diagnosis model is developed. This enhances the information mining capabilities of vibration signal data feature representation, improves the robustness of fault diagnosis algorithms, and ultimately increases the accuracy of fault diagnosis systems for hoisting motors under varying loads and frequencies. The vibration signal of a motor contains rich information about its operating status. Due to the complexity of its operating status and the continuous generation of multi-part coupled vibration signals that change over time, the construction of a two-dimensional joint time-frequency distribution using time-frequency analysis can characterize the variation of signal frequency components over time in detail. Constructing a time-frequency distribution map that simultaneously represents the time-domain characteristics and frequency-domain resolution of the signal can comprehensively extract the time-varying features of different frequency components, effectively solving the problem of difficulty in extracting the complex vibration signal features of hoisting motors. This is an effective means of preprocessing motor vibration signal data.

[0026] Please see Figure 1 As shown, in some embodiments of the present invention, a multi-condition fault diagnosis method for hoisting motors based on a multi-scale star-shaped aggregation network is provided, the steps of which include: Step 1: Obtain vibration signal data of the hoisting motor, construct a multi-condition fault diagnosis dataset and perform preliminary data screening; Step 2: Extract time-frequency features from the dataset using synchronous compressed wavelet transform; Step 3: Construct a multi-scale star-shaped aggregation network fault diagnosis model based on multilayer perceptron, residual neural network and star aggregation module; Step 4: Input the time-frequency image data training set into the fault diagnosis model for model training, and obtain the optimal model parameters of the fault diagnosis model based on the loss function; Step 5: Develop a multi-fault diagnosis system for hoisting motors based on the determined model parameters and obtain the diagnosis results.

[0027] The following section explains the construction and selection of the hoist motor fault diagnosis dataset.

[0028] To diagnose faults in the hoisting motor of a bridge erecting machine, it is necessary to collect vibration signal data under various operating conditions and construct a fault diagnosis dataset for the hoisting motor. However, existing open-source fault diagnosis datasets are generally too general, mostly built for common scenarios, and their construction does not consider the impact of actual equipment application scenarios and changing operating conditions on data accuracy.

[0029] This invention proposes a method for constructing a multi-condition fault diagnosis dataset for key components of large-scale engineering equipment, as shown in Equation 1: (1) In the formula, The operating status of important components or diagnostic objects usually covers their typical fault status and normal operating status. Typical fault statuses include motor status with stator inter-turn short circuit fault, motor status with rotor bar breakage fault, and motor status with bearing fault. This indicates the stress conditions of the components of the diagnostic object, such as load and torque. This represents its operating speed or frequency; This represents the number of times data needs to be collected; Indicates the sampling frequency and data type. Data filtering for the dataset; the above six elements work together to form a method for constructing a multi-condition fault diagnosis dataset for important components of large-scale engineering equipment.

[0030] Furthermore, regarding the fault diagnosis dataset of the bridge erecting machine's hoisting motor... This represents the types of fault conditions and normal operating conditions of the bridge erecting machine's hoisting motor. To obtain various load data by changing the number of load plates during the operation of the hoisting motor, the load conditions during dataset construction can be determined. Corresponding to the motor operating frequency; each operating condition was repeated 10 times to avoid data randomness, i.e., the corresponding number of data collections. The value is 10; the sampling frequency is 10000Hz; the data type is vibration signal and it is saved as a .CSV file, which corresponds to the dataset construction method. .

[0031] In some embodiments of the present invention, the typical working condition fault diagnosis dataset of the bridge erecting machine hoisting motor, compared with the existing general open source dataset, can better simulate the real state and special working conditions of the bridge erecting machine hoisting motor in actual operation while ensuring the convenience of the test, thereby enhancing the diversity and authenticity of the data and further improving the accuracy of subsequent fault diagnosis results. Figure 2 This document demonstrates the process of building a dataset for fault diagnosis of hoisting motors.

[0032] In some embodiments of the present invention, considering the working conditions of vibration signal acquisition from the hoisting motor of the bridge erecting machine, the present invention requires preliminary data screening of the constructed original dataset, that is, the elements in the dataset construction method corresponding to formula (1). The hoisting motor test uses the "lifting operation cycle" as the basic measurement unit. Unlike traditional time-series-based data acquisition methods, this invention considers the period from load lifting off the ground to the triggering of the limit device as a complete working condition test cycle. During the test, the hoisting motor experiences severe vibrations due to mechanical impact during startup and shutdown, resulting in non-stationary characteristics in the time domain signal. The data exhibits strong fluctuations and cannot accurately reflect the hoisting motor's operating status for most of the time.

[0033] In some embodiments of the present invention, such as Figure 3 As shown, the vibration signals of the hoisting motor under stator inter-turn short-circuit fault conditions were imported into Matlab and analyzed in the time domain. It was found that a significant transient impact response exists during the motor start-up and shutdown phase, accounting for approximately 20% of the single measurement duration. The vibration acceleration amplitude is 3-5 times that of the steady-state operation phase. This characteristic is fundamentally different from the steady-state operation of the motor, and if directly used for subsequent fault diagnosis model training, it will interfere with the classification accuracy. To eliminate the interference of transient processes during hoisting motor start-up and shutdown and to ensure the effectiveness and representativeness of the dataset, data screening was conducted on the collected vibration signals. After importing the raw vibration data in .CSV format into Matlab, the steady-state operation region in the operating cycle was identified based on the time domain waveform characteristics. A sliding window method (window length 2048) was used to extract 10 segments of vibration signal sequences from the intermediate steady-state operation phase as valid samples, such as... Figure 4 As shown, this screening method effectively eliminates interfering data during the motor start-up and shutdown phases, laying a data foundation for subsequent fault feature extraction and intelligent diagnostic model construction.

[0034] In some embodiments of the present invention, the constructed dataset comprehensively covers the typical operating conditions of the hoisting motor through multi-dimensional parameter combinations. It includes four types of typical operating conditions of the hoisting motor, three types of load conditions, four types of operating frequencies, 10 hoisting operation cycles, and 10 segments of steady-state operation data samples. Finally, it forms a standardized data sample with a sample size of 4×4×3×10×10=4800 sets, which completely records the vibration signal data generated by the hoisting motor during the hoisting operation cycle and retains the key feature information of the motor during operation. This lays a solid data foundation for subsequent research on fault feature extraction and fault diagnosis algorithms for bridge erecting machine hoisting motors.

[0035] The following section explains the fault feature extraction based on synchronous compressed wavelet transform.

[0036] For highly integrated mechanical components such as the hoisting motor of a bridge erecting machine, the complex operating conditions and the continuous generation of multi-part coupled vibration signals that change over time make standalone time-domain or frequency-domain analysis ineffective. To address these limitations, a two-dimensional joint time-frequency distribution is constructed using time-frequency analysis methods. This method details the variation of signal frequency components over time, creating a time-frequency distribution map that simultaneously characterizes the signal's time-domain properties and frequency-domain resolution. This effectively solves the problem of extracting complex vibration signal characteristics from the hoisting motor. Common time-frequency analysis techniques include Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), and Synchronous Compressed Wavelet Transform (SSWT).

[0037] In some embodiments of the present invention, the Synchronous Compressed Wavelet Transform (SSWT) introduces the concept of instantaneous frequency to compress the energy of each frequency band after wavelet transform to a certain threshold range, thereby concentrating energy near the wavelet center frequency and enhancing the time-frequency representation capability. Its expression is as follows: (2) In the formula, The scaling factor. The translation factor is... The function is the conjugate function of the wavelet basis. In Matlab, synchronous compressed wavelet transform is used for time-frequency analysis of vibration signals under three motor fault and normal operating conditions, employing Amor wavelets as the basis functions. Synchronous compressed wavelet transform significantly improves the clustering and clarity of the hoisting motor vibration signal time-frequency map by performing secondary processing on the fuzzy time-frequency map after continuous wavelet transform, thus enhancing the recognizability of time-frequency information. Furthermore, synchronous compressed wavelet transform fully utilizes the sparsity characteristics of the signal in the time-frequency domain, reducing redundant information while achieving efficient data compression, thereby improving the efficiency of vibration signal feature extraction and data storage.

[0038] In some embodiments of the present invention, a comparison of the above three time-frequency analysis methods reveals that: (1) The short-time Fourier transform achieves local spectrum analysis of the signal through a sliding window function, but its time-frequency resolution is limited by a fixed window function.

[0039] (2) Compared with the short-time Fourier transform, the continuous wavelet transform has higher time-frequency resolution by means of multi-resolution analysis capability, and exhibits stronger adaptability when processing non-stationary signals.

[0040] (3) Synchronous compressed wavelet transform captures the instantaneous frequency, abrupt change points, and local details of the signal to extract multi-scale features. The rich feature information facilitates the fusion and classification of information at different scales by the neural network. At the same time, it makes the results more accurate by suppressing spectral energy diffusion, reduces the amount of data input to the neural network, accelerates the training process, and reduces the consumption of computing resources. Therefore, synchronous compressed wavelet transform is selected for multi-condition hoisting motor fault feature extraction based on the characteristics of the variable operating conditions of the bridge erecting machine's hoisting motor.

[0041] The following section describes the fault diagnosis algorithm for hoisting motors based on multi-scale star-shaped aggregation networks.

[0042] In some embodiments of the present invention, the multi-scale star-shaped aggregation network integrates a residual neural network, a multilayer perceptron, and a star-shaped aggregation module, forming a multi-scale star-shaped aggregation network architecture of "feature extraction—cross-channel interaction—multi-scale fusion". Specifically, the residual neural network alleviates gradient degradation by adding a skip connection mechanism, ensuring the feature extraction capability of deep networks; the multilayer perceptron performs dimensional transformation and nonlinear mapping of sequence representation during star-shaped aggregation; and the star-shaped aggregation module, as the core component, aggregates multi-channel information through a centralized strategy, constructs a global core representation, and redistributes and concatenates local features, providing an efficient cross-channel interaction foundation for subsequent multi-scale feature fusion.

[0043] In some embodiments of the present invention, the star-shaped aggregation module first aggregates information from all channels to obtain a global core representation through a centralized information interaction strategy, and then redistributes the core representation and fuses it with the feature input of each channel, such as... Figure 5 As shown, the design of the star-shaped aggregation module avoids, to some extent, the secondary complexity problem caused by the distributed interaction methods used in deep learning algorithms such as attention mechanisms, GNNs, and mixers. This makes the computational complexity of the star-shaped aggregation module linearly related to both the number of channels and the window length. Therefore, the star-shaped aggregation module can adaptively extract effective information from multiple channels. This interaction mode not only reduces the complexity of the interaction but also reduces the dependence on channel quality, making it suitable for efficiently capturing channel correlations in multivariate time series.

[0044] In the above embodiments, the working principle of star-shaped polymerization mainly includes the following two steps: (1) Aggregation: The star aggregation module first maps the input feature information of all channels to a higher-dimensional space through MLP, and uses stochastic pooling to obtain the global core representation. This core representation encodes the global information of all channels.

[0045] Given a C channel Multivariate sequences, core representationO It is composed of any function f The generated vector has the following form: (3) The core represents encoding information from all channels, calculated as follows: (4) In the formula, Representing sequences from the hidden layer dimension d Mapping to core dimensions d' ; Stoch_Pool This is called random pooling, which combines the advantages of mean pooling and max pooling. It randomly selects elements from the feature map based solely on their probability values; through aggregation... C The core representation is obtained by using a sequence of channels. O .

[0046] (2) Distribution and integration: The core representation O The representation is copied and distributed to each channel, and then concatenated with the original sequence representation of each channel. After processing by another MLP module, the concatenated representation is mapped back to the original dimension and added to the original representation, thus achieving the fusion of the core representation and the representation of each channel.

[0047] The core representation is fused with the representations of all sequences as follows: (5) (6) In the formula, Repeat_Concat The operation connects the core representation to the feature information of each channel, resulting in... Then use another MLP ( : This is used to aggregate the representation of the connections and map it back to the original hidden layer dimensions. d .

[0048] In some embodiments of the present invention, the fault diagnosis process is as follows: First, such as Figure 6 As shown, the network takes the four-dimensional time-frequency data image after feature extraction via SSWT as input (corresponding to input batch, monitoring channel, time, and frequency, respectively). Three concatenated residual modules further extract features to reduce the dimensionality of the SSWT time-frequency image. Using residual modules helps deep models reduce the occurrence of gradient vanishing or exploding. Subsequently, a multi-scale global core representation is obtained through MLP and random pooling layers to capture channel correlations in multivariate time series. The multi-scale core is sequentially copied and concatenated to each channel. Finally, a multilayer perceptron and a fully connected layer fuse the multi-scale core with the information from each channel to obtain the final hoist motor fault diagnosis result.

[0049] The multi-scale star-shaped aggregation network proposed in this invention is constructed using the Python programming language and PyCharm software within the PyTorch framework. Training data is input into the network model, and the objective function value is obtained through iterative calculation. The network weights are then iteratively updated through the training module to complete the network training. Finally, test data is input into the network to identify motor fault states. Each hoisting mechanism state contains 1200 samples, and the specific settings of the training and test sets are shown in Table 1.

[0050] Table 1 Training and Test Set Information

[0051] The model training and testing experiments were conducted using the pyTorch framework on a computer with 256GB of RAM, an RTX 4090 graphics card, and a Windows 10 system. During model training, each batch contained 200 samples, the Adam optimization algorithm was used, the learning rate was set to 0.01, and the cross-entropy loss function was employed.

[0052] Table 2 Experimental Parameter Settings

[0054] In the above embodiments, the fault diagnosis test results are as follows: To verify the diagnostic performance of the model, data from three operating frequencies (20, 30, and 50 Hz) in the fault diagnosis dataset constructed in this invention were used as the fault diagnosis training set, and the remaining data from one frequency (40 Hz) was used as the test set. That is, the number of samples in the training set was 3600, and the number of samples in the test set was 1200, which effectively ensured that the test data was independent of the training data.

[0055] Furthermore, in the rotating hoist motor fault diagnosis test, four cross-speed diagnostic tasks (A, B, C, and D) were generated, as shown in Table 3. The multi-scale star-shaped aggregation network constructed in this invention was then used for fault diagnosis. The test results are shown in Table 3, with an average diagnostic accuracy of 96.56%, further demonstrating that the algorithm proposed in this invention still has excellent diagnostic performance under different operating frequencies of the hoist motor. Moreover, the test frequencies used in each task are outside the range of the training frequencies, which also proves to some extent that the constructed fault diagnosis model has a certain degree of extrapolation capability.

[0056] Table 3. Test accuracy under cross-speed diagnostic tasks

[0057] Similarly, the proposed fault model was tested and diagnosed under the lifting motor across lifting load tasks. That is, the vibration signal data of the lifting motor under two different load conditions was used as the training set, and the motor data under the remaining load condition was used as the test set. The test set data was rotated to generate three cross-load diagnosis tasks E, F and G as shown in the table below. The test results are shown in Table 4. The average diagnosis accuracy rate is 98.81%, which further proves that the fault diagnosis algorithm proposed in this invention has excellent diagnostic performance under the condition of lifting motor across load.

[0058] Table 4 Test accuracy under cross-load diagnostic tasks

[0059] The fault diagnosis system is described below.

[0060] The goal of the bridge erecting machine hoisting motor fault diagnosis system based on digital twins is to achieve real-time status monitoring and intelligent fault diagnosis of the hoisting motor. Leveraging the advantages of digital twins, an intelligent fault diagnosis system architecture for the bridge erecting machine hoisting motor is proposed, mainly composed of four parts: physical layer, data layer, model layer, and application layer. Figure 7 As shown.

[0061] The physical entity layer is the bottom layer of the entire fault diagnosis system, including the physical test platform of the bridge erecting machine's hoisting mechanism, PLC equipment, and various sensors. The operation of the physical equipment continuously generates data, which is collected in real time by the sensors and PLC equipment.

[0062] The data layer is mainly responsible for collecting, storing, processing and transmitting data from the physical layer, and providing data sources for the model layer and application layer. It is the key to the normal operation of the digital twin intelligent fault diagnosis system. Its main functions are to store, manage and transmit the parameter attribute data (rated power, load, etc.), operating status data (speed, load, etc.) of the hoisting motor, and multi-condition vibration signal data of the hoisting motor for fault diagnosis. Model Layer: The digital twin layer is an intermediate layer located between the user layer and the physical entity layer, serving as a bridge for data transmission between the two layers. Within the digital twin layer, not only is the physical entity of the bridge erecting machine's hoisting mechanism synchronously mapped, achieving simulation of the entire lifecycle of the hoisting mechanism's physical entity, but a fault diagnosis model is also encapsulated. By collecting real-time data to perform fault analysis on the hoisting motor, a comprehensive understanding of the hoisting motor's operating status can be achieved.

[0063] Application Layer: This is the front-end web page, primarily providing service entry points for staff and responsible for the visual display of the page. In this layer, staff can view the hoist motor's operating information in real time, including both real-time and historical data, facilitating the assessment of subsequent work. The fault diagnosis page allows staff to view fault diagnosis results at any time, facilitating intervention and maintenance.

[0064] The functional module design of the bridge erecting machine hoisting motor fault diagnosis system mainly revolves around core objectives such as real-time equipment monitoring, data preprocessing and feature extraction, and fault diagnosis and early warning. A system functional analysis is conducted as follows: Figure 8 As shown: The fault diagnosis system is designed and built based on a B / S architecture. The system front-end provides a human-machine interface via a web interface, supporting cross-platform access and multi-terminal adaptation. The main interface is centered on a 3D virtual scene, adopting a wraparound functional navigation layout, integrating a real-time equipment status panel, historical data trend charts, and a fault diagnosis result visualization module, allowing for intuitive access to comprehensive information on the hoist motor's operating status. The back-end service layer relies on the Advantech WISE-IoTSuite industrial IoT platform, integrating a MySQL database and OPC UA communication protocol to complete data acquisition, storage, and fusion analysis. The system embeds fault diagnosis algorithms into the virtual model through a RESTful API interface, supporting online model updates and diagnostic result display, forming a closed-loop diagnostic process of "perception-analysis-decision".

[0065] This invention uses a MySQL database for data storage and acquires and transmits data via the OPC UA data transmission framework and TCP / IP protocol. A digital twin of the bridge erecting machine's lifting mechanism is constructed using the WISE-IoTSuite / SaaS Composer digital twin visualization platform and modeling tools such as Solidworks and 3DS Max. The system's fault diagnosis and prediction functions are developed using PyCharm as the IDE for Python. Backend data processing uses the .NET Framework as the system framework, and the host computer interface and backend code are developed using C# and Visual Studio Code.

[0066] The feasibility of this invention has been verified through experiments, and the specific results are as follows: Dataset construction: 4800 standardized samples were formed, covering 4 types of fault states, 3 types of loads, and 4 frequencies, providing a reliable data foundation for model training.

[0067] Model performance: The multi-scale star-shaped aggregation network achieved a diagnostic accuracy of 99.92% on a specific test set, an average diagnostic accuracy of 96.56% across rotational speeds, and an average diagnostic accuracy of 98.81% across loads, with a training loss as low as 5.41e-6, demonstrating the model's high accuracy and robustness under varying operating conditions.

[0068] System Implementation: A fault diagnosis system was built based on a B / S architecture, integrating a MySQL database, OPC UA protocol, and a digital twin visualization platform to achieve real-time monitoring, fault diagnosis, and visualization, thus verifying the practical application value of the system.

[0069] All parts of this invention not described in detail herein can be referred to in the prior art or are known to those skilled in the art. This embodiment does not limit these aspects and will not describe them in detail here.

[0070] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A fault diagnosis model for hoisting motors based on star-shaped aggregation networks, characterized in that, The model is a multi-scale star-shaped aggregation network fault diagnosis model based on time-frequency domain data analysis. The model integrates residual neural networks, multilayer perceptrons, and star-shaped aggregation modules to form a network architecture that features feature extraction, cross-channel interaction, and multi-scale fusion.

2. The hoisting motor fault diagnosis model based on a star-shaped aggregation network according to claim 1, characterized in that, The residual neural network alleviates gradient degradation through a skip connection mechanism and is used to extract multi-level features from time-frequency images; the multilayer perceptron is used to realize dimensional transformation and nonlinear mapping of sequence representation; the star-shaped aggregation module aggregates multi-channel information through a centralized strategy, constructs a global core representation, and copies and distributes the global core representation to each channel. After being concatenated with local features, it is mapped back to the original dimension by the multilayer perceptron, realizing cross-channel multi-scale feature fusion.

3. The hoisting motor fault diagnosis model based on a star-shaped aggregation network according to claim 2, characterized in that, The multilayer perceptron maps the input feature information of all channels to a higher-dimensional space, and the star-shaped aggregation module obtains the global core representation through random pooling technology. The global core representation encodes the global information of all channels. The global core representation is copied and distributed to each channel, and then concatenated with the original sequence representation of each channel. The concatenated representation is mapped back to the original dimension through a multilayer perceptron, and then added to the original representation to achieve the fusion of the global core representation and the representation of each channel.

4. A method for fault diagnosis of hoisting motors based on star-shaped aggregation networks, characterized in that, The method includes the following steps: S1: Construction of hoist motor fault diagnosis dataset. By acquiring hoist motor vibration signal data, a multi-condition fault diagnosis dataset is constructed and preliminary data screening is performed. S2: Fault data feature extraction and data processing, using synchronous compressed wavelet transform to extract time-frequency features from the dataset; S3: Construction of a fault diagnosis model for hoisting motors. By integrating residual neural networks, multilayer perceptrons, and star aggregation modules, a multi-scale star aggregation network fault diagnosis model is constructed. S4: Data-driven fault diagnosis model training. The time-frequency image data training set is input into the fault diagnosis model for model training, and the optimal model parameters of the fault diagnosis model are obtained based on the loss function.

5. The hoisting motor fault diagnosis method based on a star-shaped aggregation network according to claim 4, characterized in that, In step S1, a six-element collaborative approach is used to construct a multi-condition fault diagnosis dataset.

6. The hoisting motor fault diagnosis method based on a star-shaped aggregation network according to claim 5, characterized in that, The formula for constructing the dataset is: ， In the formula, DS c For the dataset, M To diagnose the operational status of the object, L To diagnose the stress conditions on the object, F For operating speed or frequency, N For the number of data collections, S For sampling frequency and data type, A This describes the data filtering process for the dataset.

7. The hoisting motor fault diagnosis method based on a star-shaped aggregation network according to claim 4, characterized in that, The data filtering in step S1 includes: analyzing the time-domain waveform using Matlab, using the sliding window method to extract several segments of steady-state signal, and removing transient interference data during start-stop operations.

8. The hoisting motor fault diagnosis method based on a star-shaped aggregation network according to claim 4, characterized in that, Step S2 includes: processing the vibration signal after preliminary screening using time-frequency analysis method, constructing a time-frequency distribution spectrum that can simultaneously characterize the signal's time-domain characteristics and frequency-domain resolution, and then describing in detail the variation law of the signal's frequency components over time; The time-frequency analysis method specifically adopts the synchronous compressed wavelet transform method. Based on the wavelet transform principle, this method can extract multi-scale features of a signal by capturing its instantaneous frequency, abrupt change points, and local details. The rich feature information facilitates the fusion and classification of information at different scales by the neural network. The synchronous compressed wavelet transform effectively suppresses spectral energy diffusion by introducing the concept of instantaneous frequency, causing energy to accumulate near the wavelet center frequency, thereby enhancing the time-frequency representation capability and achieving the technical effect of accelerating the training process of fault diagnosis models and reducing computational resource consumption.

9. The hoisting motor fault diagnosis method based on a multi-scale star-shaped aggregation network according to claim 4, characterized in that, The fault diagnosis process in step S5 includes: The four-dimensional time-frequency image dataset, after feature extraction by synchronous compressed wavelet transform, is input into the trained multi-scale star-shaped aggregation network fault diagnosis model; the four dimensions of the four-dimensional time-frequency image dataset correspond to the input batch, the number of monitoring channels, time, and frequency, respectively. Feature extraction is performed by three residual modules connected in series in the model to reduce the dimensionality of the time-frequency graph. Multi-scale global core representations are obtained through multi-layer perceptron and random pooling layer in the model. Then, the global core representations are spliced ​​and fused with the information of each channel through multi-layer perceptron and fully connected layer in the model to obtain the final hoisting motor fault diagnosis result.

10. A hoisting motor fault diagnosis system based on a star-shaped aggregation network, characterized in that, The system is used to implement the hoisting motor fault diagnosis method based on a star-shaped aggregation network as described in any one of claims 4 to 9; the system comprises: The dataset construction and filtering module is configured to acquire hoist motor vibration signal data, and to construct multi-condition fault diagnosis datasets and perform preliminary data filtering. The feature extraction module is configured to extract time-frequency features from the dataset using synchronous compressed wavelet transform. The model building module is configured to build a multi-scale star-shaped aggregation network fault diagnosis model architecture. The model parameter determination module is configured to input the time-frequency image data training set into the fault diagnosis model for model training, and obtain the optimal model parameters of the fault diagnosis model based on the loss function. The fault diagnosis module is configured to construct a multi-public fault diagnosis model for the hoisting motor based on the multi-scale star-shaped aggregation network fault diagnosis model architecture and the optimal model parameters, and then perform fault diagnosis.

Images