Methods, devices, equipment and storage media for diagnosing faults in rotating components of cranes

Abstract

This invention provides a method, apparatus, device, and storage medium for fault diagnosis of rotating components of a crane, relating to the field of mechanical fault diagnosis technology. The method includes: processing source and target domain data of the crane's rotating components in a unified format to output a normalized vibration signal; aligning feature distributions through a shared encoder and adversarial training to output domain-invariant features; constructing conditional embedding vectors based on the domain-invariant features and generating target domain fault data using a diffusion model; combining the generated fault data with real data to form a diagnostic dataset, extracting multimodal features, fusing them, and outputting the fault diagnosis result through a classifier. This invention effectively solves the problem of scarce crane fault data through a diffusion model and, combined with multimodal feature fusion technology, improves the accuracy, robustness, and cross-domain generalization ability of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis, and in particular to a crane rotating component fault diagnosis method, device, equipment and storage medium. BACKGROUND

[0002] The rotating components of heavy machinery such as cranes are prone to failure during long-term high-load operation, and accurate diagnosis is crucial to ensuring equipment safety and operational efficiency.

[0003] Currently, the machine learning method based on vibration signal analysis is mainly used in the field of fault diagnosis, which has obvious limitations. First, the deep learning method relies heavily on a large number of labeled fault data, but the fault samples in the actual operation of the crane are scarce and costly to obtain, making it difficult to train the model. Second, traditional methods mostly use single modal features, which are difficult to fully capture the spatiotemporal characteristics of complex faults and have poor robustness in noisy environments. In addition, when migrating diagnostic knowledge between different equipment and working conditions, the model's generalization ability is generally insufficient due to significant differences in data distribution.

[0004] Therefore, there is an urgent need for a fault diagnosis scheme that can overcome the limitations of data scarcity, effectively integrate multi-source information, and have strong generalization ability to improve the practicality and reliability of crane rotating component fault diagnosis. SUMMARY

[0005] The present application provides a crane rotating component fault diagnosis method, device, equipment and storage medium to solve the defects of existing crane rotating component fault diagnosis methods such as fault data scarcity, incomplete feature extraction, and poor cross-domain generalization ability, and to realize multi-modal fusion of data to improve diagnostic performance.

[0006] The present application provides a crane rotating component fault diagnosis method, which includes the following steps: first, uniformly format processing the source domain data and target domain data of the crane rotating component, and outputting the normalized vibration signal; then inputting the normalized vibration signal into a shared encoder for feature mapping, realizing feature distribution alignment through adversarial training, and outputting domain-invariant features; then constructing a conditional embedding vector based on the domain-invariant features, generating target domain fault data using a diffusion model, and outputting synthetic fault data; finally, combining the synthetic fault data with the real data into a diagnostic data set, extracting multi-modal features from the data set, and outputting the fault diagnosis result after feature fusion through a classifier.

[0007] According to the crane rotating component fault diagnosis method provided by the present application, the specific implementation of the data uniform format processing includes: aligning the sampling frequencies of the source domain data and the target domain data, outputting the resampled data; amplitude normalizing the resampled data, outputting the standardized signal; dividing the standardized signal into fixed-length sequences to form the normalized vibration signal.

[0008] According to the crane rotating part fault diagnosis method provided by the application, the specific implementation of the feature distribution alignment includes: using a shared encoder to map a normalized vibration signal to an initial feature representation; inputting the initial feature representation into a domain discriminator for domain classification, and calculating a domain discrimination loss; jointly optimizing the shared encoder based on the domain discrimination loss and a distribution distance loss, and outputting domain-invariant features.

[0009] According to the crane rotating part fault diagnosis method provided by the application, the specific implementation of generating target domain fault data includes: extracting target domain background features representing the background of the target domain equipment from the domain-invariant features; extracting source domain fault mode features representing the fault types of the source domain from the domain-invariant features; combining the target domain background features and the source domain fault mode features into a conditional embedding vector; and using a diffusion model to generate synthetic fault data guided by the conditional embedding vector.

[0010] According to the crane rotating part fault diagnosis method provided by the application, the specific implementation of the multi-modal feature extraction includes: converting the vibration signal in the diagnostic data set into a Gabor angle field image through Gabor angle field coding; extracting image modal features from the Gabor angle field image; and directly extracting time series modal features from the vibration signal in the diagnostic data set.

[0011] According to the crane rotating part fault diagnosis method provided by the application, the further optimization scheme of feature extraction includes: using a convolutional neural network to perform spatial feature extraction on the Gabor angle field image, and outputting the image modal features; and simultaneously using a time series modeling network to model the time dependence relationship of the vibration signal, and outputting the time series modal features.

[0012] According to the crane rotating part fault diagnosis method provided by the application, the specific implementation of the feature fusion and classification includes: fusing the image modal features and the time series modal features to obtain a joint feature representation; converting the joint feature representation into a classification feature vector; processing the classification feature vector through a classifier to output a fault diagnosis result.

[0013] The application further provides a crane rotating part fault diagnosis device, comprising the following modules: a processing module, configured to perform unified format processing on source domain data and target domain data of the crane rotating part, output normalized vibration signals, the source domain data is from fault vibration signals of other rotating machines, and the target domain data is from vibration signals collected on site of the crane; an alignment module, configured to input the normalized vibration signals into a shared encoder for feature mapping, and realize feature distribution alignment through adversarial training, and output domain invariant features; a fusion module, configured to construct a conditional embedding vector based on the domain invariant features, generate target domain fault data by using a diffusion model, and output synthetic fault data; and a diagnosis module, configured to combine the synthetic fault data and real data into a diagnosis data set, perform multi-modal feature extraction on the diagnosis data set, and output a fault diagnosis result through feature fusion by a classifier.

[0014] The application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the crane rotating part fault diagnosis method according to any one of the above when executing the computer program.

[0015] The application further provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the crane rotating part fault diagnosis method according to any one of the above.

[0016] The application further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the crane rotating part fault diagnosis method according to any one of the above.

[0017] The crane rotating part fault diagnosis method, device, equipment and storage medium provided by the application can generate conditional data by introducing a diffusion model, can construct high-quality target domain fault samples by using limited source domain fault data and target domain normal data, and solve the problems of insufficient model training and overfitting caused by data scarcity; a multi-modal feature fusion framework is designed, vibration signals are converted into image modalities and time modalities at the same time, spatial and time features are extracted by using convolutional neural networks and time modeling networks respectively, full global information of faults is fully captured, and the representation ability of complex fault modes is improved; domain invariant features are learned by a feature alignment module, data distribution differences between different devices and working conditions are overcome, and the cross-domain adaptability of the model is improved; meanwhile, the application of multi-modal fusion and attention mechanism enables the model to focus on key fault features under strong noise interference, so that stable and reliable diagnosis in a high-noise industrial environment is realized. BRIEF DESCRIPTION OF DRAWINGS

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

[0019] Figure 1 This is a system framework diagram of a fault diagnosis method for rotating components of a crane provided by the present invention.

[0020] Figure 2 This is a flowchart illustrating a method for diagnosing faults in rotating components of a crane, provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the structure of a RepLKNet provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the internal structure of a GRU neural network provided by the present invention.

[0023] Figure 5 This is a schematic diagram of a bidirectional GRU network structure provided by the present invention.

[0024] Figure 6 This is a schematic diagram of the structure of a global attention mechanism provided by the present invention.

[0025] Figure 7 This is a structural schematic diagram of a fault diagnosis device for rotating components of a crane provided by the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions 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.

[0028] Figure 1 This is a system framework diagram of a fault diagnosis method for rotating components of a crane provided by the present invention.

[0029] like Figure 1 As shown, the system framework includes a data generation module 101, a multimodal data module 102, a multimodal feature extraction module 103, and a feature fusion and classification module 104.

[0030] For example, the data generation module 101 can use a diffusion model to generate target domain fault data and synthesize vibration signals that conform to the characteristics of the target domain through a forward noise injection and reverse denoising process.

[0031] For example, the multimodal data module 102 can process the image and time-series modes of the vibration signal.

[0032] Specifically, the image mode converts the one-dimensional signal into a two-dimensional image through Gram angle field coding, while the temporal mode preserves the original vibration sequence.

[0033] For example, the multimodal feature extraction module 103 can extract features from images and time-series data respectively.

[0034] Specifically, image feature extraction can employ a reparameterized large kernel network (RepLKNet), while temporal feature extraction can employ a bidirectionally gated recurrent unit (BiGRU) and a global attention mechanism.

[0035] For example, the feature fusion and classification module 104 can fuse multimodal features and output diagnostic results through a fully connected layer and a Softmax classifier.

[0036] Thus, the system framework forms a complete diagnostic closed loop, ensuring the efficiency and reliability of the method, and is suitable for intelligent operation and maintenance and condition monitoring of rotating parts of cranes.

[0037] The following is combined Figures 2 to 6 This invention describes a fault diagnosis method for rotating crane components. Addressing the problems of data scarcity, incomplete feature extraction, and poor model generalization ability in fault diagnosis of rotating crane components, this invention proposes a method based on a diffusion model and multimodal deep learning. This method achieves high-precision and highly robust fault diagnosis through steps such as data generation, feature alignment, multimodal feature extraction, and fusion. The invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Figure 2 This is a flowchart illustrating a method for diagnosing faults in rotating components of a crane provided by the present invention, as shown below. Figure 2 As shown, the method includes the following:

[0039] S201. Perform unified format processing on the source domain data and target domain data of the rotating parts of the crane, and output normalized vibration signal.

[0040] In this embodiment, the source domain data comes from the fault vibration signals of other rotating machinery, which is obtained by introducing artificial defects on the test bench.

[0041] For example, publicly available industrial fault datasets can be selected as source domain data. Source domain data can contain various typical fault types, such as broken teeth, pitting, and spalling.

[0042] Specifically, source domain data can be acquired from a high-precision test bench with a sampling frequency set to a high value (e.g., 12kHz) to capture detailed fault characteristics. Vibration signals cover a variety of operating conditions, including different speeds (e.g., 500rpm to 2000rpm) and loads (e.g., no load, half load, full load), to provide rich fault mode samples for model training.

[0043] In actual data acquisition, accelerometers can be installed at key locations on the experimental platform to ensure signal quality.

[0044] In addition, the dataset can also include environmental noise and interference signals to simulate real industrial environments and enhance the robustness of the model.

[0045] In this embodiment, the target domain data comes from normal vibration signals collected on-site by the crane.

[0046] For example, an acceleration sensor can be installed on the rotating parts of the crane (such as the gearbox of a reducer) to collect data. The sensor is connected to the data recording system via wired or wireless means to ensure the stability and real-time performance of the signal transmission.

[0047] Specifically, the target domain data can be collected from the actual operating environment of the crane, and the sampling frequency is set to an appropriate value (e.g., 6.8kHz) according to the characteristics of the equipment. Vibration signals under different workloads (e.g., no load, half load, full load) and operating speeds are recorded to reflect the normal operating characteristics of the crane under real working conditions.

[0048] In the actual data collection process, it is necessary to consider the complexity of the on-site environment, such as temperature changes and mechanical vibration interference, in order to ensure the representativeness of the data.

[0049] Optionally, normal categories can be labeled for the target domain data to provide accurate label information for subsequent model training.

[0050] In some embodiments, the sampling frequencies of the source domain data and the target domain data can be aligned first, and then resampled data can be output.

[0051] For example, the sampling frequency of the source domain data can be adjusted to match that of the target domain data to eliminate the impact of sampling rate differences.

[0052] Specifically, sampling frequency matching can be achieved using anti-aliasing filtering and linear interpolation algorithms. Anti-aliasing filtering removes high-frequency components to prevent signal aliasing; linear interpolation algorithms generate new sampling points by calculating the values ​​between adjacent sampling points, ensuring the continuity of the signal on the time axis.

[0053] Furthermore, the amplitude of the frequency-aligned resampled data is normalized to output a standardized signal.

[0054] For example, the amplitude of the vibration signal can be adjusted using the Z-score normalization method to eliminate dimensional differences caused by different sensors and acquisition environments.

[0055] Specifically, the mean and standard deviation of each signal segment can be calculated first. By subtracting the mean and dividing by the standard deviation, the waveform can be normalized to a mean of zero and a standard deviation of zero, thereby reducing the impact of outliers and enhancing the reliability of the data.

[0056] Furthermore, the normalized signal is divided into fixed-length sequences to form a normalized vibration signal.

[0057] For example, continuous vibration signals can be segmented into time-series samples of fixed length to increase the number of samples and facilitate model processing.

[0058] Specifically, each time series sample can be divided into 1024 sampling points, with 512 sampling points overlapping between adjacent samples. Multiple samples can be generated by using a sliding window method to increase the number of training samples and capture the temporal continuity features in the signal.

[0059] Optionally, source domain samples can be labeled with fault categories (such as normal, broken teeth, pitting, peeling), and target domain samples can be uniformly labeled as normal to ensure the accuracy and consistency of label management.

[0060] Thus, this invention, through a systematic data preprocessing workflow including sampling frequency alignment, amplitude normalization, and segmented windowing, resolves the differences between source and target domain data in terms of sampling rate, amplitude scale, and noise characteristics, laying a solid foundation for subsequent feature alignment and model training, and ensuring data quality and consistency.

[0061] S202. Input the normalized vibration signal into the shared encoder for feature mapping, and achieve feature distribution alignment through adversarial training, so as to output domain invariant features.

[0062] In some embodiments, a shared encoder may be used to first map the normalized vibration signal to an initial feature representation.

[0063] For example, a shared encoder can be built on a temporal convolutional network (TCN) to extract multi-scale temporal features.

[0064] Specifically, the encoder structure includes multiple convolutional layers: the first layer uses a large convolutional kernel (e.g., 64) to capture low-frequency trends; the second layer uses a smaller convolutional kernel (e.g., 32) and dilated convolution (dilation=2) to capture multi-scale patterns; the third layer further expands the receptive field by using a convolutional kernel size of 16 and dilated convolution (dilation=4); and finally, a low-dimensional feature vector (e.g., 128-dimensional) is output through a global average pooling layer.

[0065] Alternatively, batch normalization layers and activation functions (such as ReLU) can be used during encoder training to accelerate convergence and improve model stability.

[0066] Furthermore, the initial feature representations are input into the domain discriminator for domain classification, and the domain discrimination loss is calculated.

[0067] For example, the domain discriminator can be a multi-layer fully connected network used to distinguish the source of features: the source domain or the target domain.

[0068] Specifically, the domain discriminator may include an input layer (128-dimensional), a hidden layer (64-dimensional, using the LeakyReLU activation function with a negative slope of 0.2), and an output layer (2-dimensional, using the Softmax function to output probabilities).

[0069] In actual training, the domain discrimination loss can adopt the binary classification cross-entropy loss function to drive feature distribution alignment. The domain discriminator improves its ability to distinguish between source and target domain features by continuously optimizing the weight parameters, thereby prompting the encoder to generate consistent feature representations.

[0070] Furthermore, the shared encoder is jointly optimized based on domain discrimination loss and distribution distance loss, resulting in domain-invariant features at output.

[0071] For example, the maximum mean discrepancy (MMD) loss can be introduced to force the feature distributions of normal data in the source and target domains to be close in a high-dimensional space.

[0072] Specifically, MMD loss optimizes the encoder by comparing the distribution differences of source and target domain features in a high-dimensional space and combining it with adversarial training.

[0073] In this embodiment, the encoder's total loss function is a weighted combination of the task loss, domain discrimination loss, and MMD loss. For example, the domain discrimination loss weight can be set to 0.1, and the MMD loss weight can be set to 1.0.

[0074] In actual training, the encoder parameters can be updated using the backpropagation algorithm to generate features whose source cannot be distinguished by the domain discriminator.

[0075] Thus, by jointly optimizing adversarial training and distribution distance loss, this invention achieves effective alignment of source domain fault features and target domain normal features in the feature space, generating domain-invariant features, providing reliable input for subsequent cross-domain data generation, and ensuring the consistency between generated data and target domain characteristics.

[0076] S203. Construct conditional embedding vectors based on domain-invariant features, generate target domain fault data using a diffusion model, and output synthetic fault data.

[0077] In some embodiments, target domain background features characterizing the target domain device background can be extracted from domain-invariant features.

[0078] For example, the background features of the target domain reflect the inherent characteristics of the crane equipment, such as its physical structure and operating conditions.

[0079] Specifically, a low-dimensional vector can be extracted from domain-invariant features using a feature separation network. This low-dimensional vector encodes properties of the target domain device, such as the modal response of a gearbox and the influence of sensor mounting positions.

[0080] It should be noted that feature separation networks typically consist of fully connected layers and attention mechanisms, enabling them to automatically select feature dimensions relevant to the device context.

[0081] In some embodiments, source domain fault mode features characterizing the source domain fault type can also be extracted from the domain-invariant features.

[0082] For example, the source domain fault mode features include vibration characteristics of specific faults, such as tooth breakage, pitting, and spalling.

[0083] Specifically, high-dimensional vectors can be extracted from domain-invariant features. These high-dimensional vectors encode the dynamic patterns of source domain faults, such as variations in fault feature frequencies and amplitudes.

[0084] In the actual feature extraction process, clustering algorithms or deep learning models can be used to encode fault modes to ensure that the extracted features can accurately reflect the unique attributes of different fault types.

[0085] Furthermore, the extracted target domain background features and source domain fault mode features are combined into a conditional embedding vector.

[0086] For example, target domain background features and source domain fault mode features can be fused into a conditional embedding vector through a concatenation operation.

[0087] Specifically, the target domain background feature vector (e.g., 128-dimensional) and the source domain fault mode feature vector (e.g., 128-dimensional) are concatenated in terms of feature dimensions to form a 256-dimensional conditional embedding vector. This vector is then used as guidance information for the diffusion model to ensure that the generated data retains both the target domain background and injects fault modes.

[0088] Optionally, the conditional embedding vector can also be optimized through weighted summation or attention mechanisms to balance the contributions of different features.

[0089] Furthermore, synthetic fault data is generated by using a diffusion model guided by conditional embedding vectors.

[0090] For example, the diffusion model includes a forward noise injection and a reverse denoising process to generate high-quality fault data.

[0091] Specifically, the forward process adds Gaussian noise to the data step by step through a Markov chain, with the noise level increasing with the time step; the reverse process reconstructs the target data from the noisy data using a trained denoising model.

[0092] In this embodiment of the application, the generator network can adopt a time series Transformer structure to make full use of temporal dependencies and multi-scale information.

[0093] In the actual generation phase, a reverse diffusion process (e.g., 1000 steps) is run starting from pure noise to iteratively generate fault vibration data that conforms to the characteristics of the target domain. After generation, filtering and amplitude adjustment are performed to ensure physical availability.

[0094] For example, use a Butterworth low-pass filter to filter out unreasonable high-frequency components and adjust the amplitude range to match the sensor range.

[0095] Thus, by using a conditional diffusion model, this invention injects mode variations of source domain faults while preserving the time-frequency characteristics of normal crane signals, generating high-quality and diverse fault data. This solves the problem of scarce fault data in industrial scenarios and provides rich samples for model training.

[0096] S204. Combine the synthetic fault data with the real data to form a diagnostic dataset, and extract multimodal features from the diagnostic dataset.

[0097] In one alternative implementation, vibration signals in the diagnostic dataset can be converted into Gram angle field images via Gram angle field coding, and then image modal features can be extracted from the Gram angle field images.

[0098] For example, gramian angular field (GAF) coding can convert one-dimensional vibration signals into two-dimensional images to preserve the correlation of time series.

[0099] Specifically, the GAF encoding steps include: rescaling the normalized signal to the [-1,1] interval; converting to polar coordinates, calculating the Gramian matrix, and then uniformly adjusting the size of the generated GAF ​​image to 224×224 pixels to match the input requirements of the image network.

[0100] In the polar coordinate system, the angle represents the value of the time series, and the radius represents the timestamp; each element in the Gramian matrix is ​​a trigonometric function value of the angle (such as the cosine value).

[0101] In some embodiments, a convolutional neural network can be used to extract spatial features from a Gram angle field image and output image modal features.

[0102] For example, RepLKNet, a convolutional neural network that uses large convolutional kernels, can be used for image feature extraction.

[0103] Figure 3 This is a schematic diagram of the structure of a RepLKNet provided by the present invention.

[0104] like Figure 3 As shown, RepLKNet consists of a start layer (stem) and four stages (stages 1-4). The start layer consists of two 3×3 convolutional layers, used to capture more detailed information at the beginning. Stages 1 to 4 each contain multiple RepLK blocks, with stage 1 containing 4 RepLK blocks and 128 output channels; stage 2 containing 6 RepLK blocks and 256 output channels; stage 3 containing 8 RepLK blocks and 512 output channels; and stage 4 containing 10 RepLK blocks and 1024 output channels.

[0105] In the embodiments of this application, RepLK blocks can use large convolutional kernels (such as 3×3) to capture a wider range of local features and effectively extract large-scale contextual information (such as long-term frequency change trends).

[0106] For example, a Convolutional Feedforward Network (ConvFFN) block (similar to a convolutional neural network block) can be placed after each RepLK block, inspired by feedforward networks (FFNs), to increase nonlinear and cross-channel information communication.

[0107] For example, the processing flow of a RepLK block is: batch normalization, 1×1 convolutional layer, 3×3 depthwise separable convolutional layer, and 1×1 convolutional layer.

[0108] For example, the processing flow of the convolutional feedforward network block is as follows: batch normalization, 1×1 convolutional layer, Gaussian error linear units (GELU), 1×1 convolutional layer.

[0109] For example, the processing flow of the starting layer is: a 3×3 convolutional layer with stride 2, a 3×3 depth-separable convolutional layer, a 1×1 convolutional layer, and a 3×3 depth-separable convolutional layer with stride 2.

[0110] For example, the processing flow of the transition layer is: a 1×1 convolutional layer and a 3×3 depth-separable convolutional layer with a stride of 2.

[0111] In the embodiments of this application, RepLKNet can maintain computational efficiency and enhance generalization ability during the training phase through reparameterization techniques.

[0112] Alternatively, a pre-trained RepLKNet-B mini-model can be used, with fine-tuning of the last three layers to output 1024-dimensional image features from the global average pooling layer.

[0113] During training, image augmentation techniques such as random rotation and cropping can be used to improve the robustness of the model.

[0114] Thus, this invention converts one-dimensional vibration signals into two-dimensional images and extracts spatial features through GAF encoding and RepLKNet feature extraction, fully capturing the global structural information of fault signals, making up for the shortcomings of traditional methods in spatial feature extraction, and improving the accuracy of fault diagnosis.

[0115] In one alternative implementation, temporal modal features can be extracted directly from vibration signals in a diagnostic dataset.

[0116] For example, a time-series modeling network can be used to model the time dependence of vibration signals in order to extract dynamic evolution features.

[0117] Specifically, BiGRU combined with a global attention mechanism can be used.

[0118] Figure 4 This is a schematic diagram of the internal structure of a GRU neural network provided by the present invention.

[0119] like Figure 4 As shown, the gated recurrent unit (GRU) is a simplified version of the long short-term memory (LSTM) network, including a reset gate and an update gate.

[0120] For example, resetting the gate controls the degree of forgetting the previous state information, updating the gate controls the degree of retention of the previous state information, and managing the flow of information through the gating mechanism reduces network complexity and improves computational efficiency.

[0121] Among them, h t-1 The hidden state of the previous time step; x t This is the input for the current time step; r t This represents a "reset gate" (as in a GRU) or some intermediate variable; σ is the Sigmoid activation function, with an output range of (0,1), used for gating signals; z t Indicates "update gate" or another gated signal; tanh is the hyperbolic tangent activation function, with an output range of (-1, 1), used to generate candidate values; + is vector addition; × is element-wise multiplication; h t The hidden state at the current time step; It is based on the input x at the current time step. t and after resetting the door r t The hidden state h of the previous time step being filtered t-1 The calculated alternative hidden states.

[0122] Figure 5 This is a schematic diagram of a bidirectional GRU network structure provided by the present invention.

[0123] like Figure 5 As shown, BiGRU consists of two GRU networks, one for forward propagation and one for backward propagation, with no connection between the forward hidden layer and the backward hidden layer.

[0124] In this embodiment, BiGRU is configured as a multi-layer network with an input dimension of 1024, a hidden layer dimension of 256, and 2 layers. A dropout rate (e.g., 0.2) is used to prevent overfitting.

[0125] Among them, h t-1 The hidden state from the previous time step is used to calculate h. t h t The hidden state at the current time step is used to calculate h. t+1 h t+1 This is the hidden state for the next time step.

[0126] Figure 6 This is a schematic diagram of the structure of a global attention mechanism provided by the present invention.

[0127] like Figure 6 As shown, the global attention model is a mechanism used to enhance the model's attention to different parts of the input signal sequence.

[0128] For example, at each time step, attention weights can be calculated to represent the degree of attention the model pays to each part of the input sequence. These weights are then used to weight the feature representation, enabling the model to focus more on important temporal features. By dynamically allocating importance weights for different time steps, the model can pay more attention to key moment information while ignoring noise or unimportant time periods, ultimately outputting 256-dimensional temporal features.

[0129] Where x1, x2, x3, ..., x k The elements of the input sequence (input vector), such as hidden states or word embeddings at the encoder end; h1, h2, h3, ..., h k The hidden states of the encoder are processed by GRU / LSTM, etc., x1, x2, x3, ..., x k α ki h represents the attention weight, indicating the degree of attention given to the i-th input when generating the k′-th output; k′ Let be the hidden state of the decoder at time step k′; β represents the context vector, i.e., the weighted sum.

[0130] In actual training, gradient clipping and learning rate decay strategies can also be used to optimize model parameters.

[0131] Thus, this invention effectively extracts the temporal evolution information of fault signals by bidirectionally modeling temporal dependencies using BiGRU and combining it with a global attention mechanism to enhance the features of key time segments, thereby improving the accuracy and robustness of temporal feature extraction and laying the foundation for multimodal fusion.

[0132] S105. Perform feature fusion and output the fault diagnosis results through the classifier.

[0133] In some embodiments, image modal features and temporal modal features can be fused first to obtain a joint feature representation, and then the joint feature representation can be converted into a classification feature vector.

[0134] For example, two strategies, feature concatenation and weighted fusion, can be used for multimodal feature fusion.

[0135] Specifically, feature concatenation directly combines 1024-dimensional image features and 256-dimensional temporal features to form a 1280-dimensional joint feature. Subsequently, the joint feature representation is input into a fully connected layer for dimensionality reduction, outputting a 512-dimensional classification feature vector. The fully connected layer can use the ReLU activation function and a dropout rate (e.g., 0.5) to prevent overfitting and ensure the simplicity and effectiveness of the feature representation.

[0136] Specifically, the weighted fusion uses learnable weights to sum the features of the two modalities. The weights are automatically learned through an attention mechanism to balance the contributions of different modalities.

[0137] Furthermore, the classification feature vectors can be processed by a classifier to output fault diagnosis results.

[0138] For example, the classifier can employ a combination of fully connected layers and a Softmax classifier for multi-fault type classification.

[0139] Specifically, the fully connected layer consists of a hidden layer (512-dimensional, using the ReLU activation function and a dropout rate of 0.5) and an output layer (the output dimension is the number of fault categories). The Softmax classifier outputs the probability distribution of each fault category.

[0140] In this embodiment, the loss function can be cross-entropy loss, and a phased strategy is adopted during the training process: first train the feature alignment module, then train the diffusion model, and finally train the multimodal diagnostic network.

[0141] The batch size can be set to 32; the optimizer uses Adam with an initial learning rate of 0.001, which decays by 0.1 every 50 rounds; and the training rounds are 200.

[0142] Thus, this invention achieves high-precision and robust diagnosis of faults in rotating components of cranes through multimodal feature fusion and classifier optimization. This method effectively solves the problems of data scarcity, incomplete feature extraction, and poor model generalization ability, providing reliable technical support for the intelligent operation and maintenance of cranes.

[0143] Figure 7 This is a structural diagram of a crane rotating component fault diagnosis device provided in an embodiment of this application. The crane rotating component fault diagnosis device includes: a processing module 701, an alignment module 702, a fusion module 703, and a diagnosis module 704.

[0144] The system comprises: a processing module 701, which processes the source and target domain data of the crane's rotating components in a unified format and outputs a normalized vibration signal; the source domain data comes from fault vibration signals from other rotating machinery, and the target domain data comes from vibration signals collected on-site from the crane; an alignment module 702, which inputs the normalized vibration signal into a shared encoder for feature mapping and achieves feature distribution alignment through adversarial training, outputting domain-invariant features; a fusion module 703, which constructs conditional embedding vectors based on domain-invariant features, generates target domain fault data using a diffusion model, and outputs synthetic fault data; and a diagnosis module 704, which combines the synthetic fault data with real data to form a diagnostic dataset, extracts multimodal features from the diagnostic dataset, performs feature fusion, and outputs fault diagnosis results through a classifier.

[0145] In some embodiments, the processing module 701 is specifically used to: align the sampling frequencies of the source domain data and the target domain data, and output resampled data; normalize the amplitude of the resampled data, and output a standardized signal; and divide the standardized signal into a fixed-length sequence to form a normalized vibration signal.

[0146] In some embodiments, the alignment module 702 is specifically used to: map the normalized vibration signal to an initial feature representation using a shared encoder; input the initial feature representation into a domain discriminator for domain classification and calculate the domain discrimination loss; jointly optimize the shared encoder based on the domain discrimination loss and the distribution distance loss, and output domain-invariant features.

[0147] In some embodiments, the fusion module 703 is specifically used to: extract target domain background features that characterize the background of the target domain device and source domain fault mode features that characterize the source domain fault type from the domain-invariant features; combine the target domain background features and source domain fault mode features into a conditional embedding vector; and generate synthetic fault data using a diffusion model guided by the conditional embedding vector.

[0148] In some embodiments, the diagnostic module 704 is specifically used to: convert the vibration signals in the diagnostic dataset into Gram angle field images through Gram angle field encoding; extract image modal features from the Gram angle field images; and directly extract temporal modal features from the vibration signals in the diagnostic dataset.

[0149] In some embodiments, the diagnostic module 704 is specifically used to: extract spatial features from the Gram angle field image using a convolutional neural network and output image modal features; and model the time dependence of the vibration signal using a temporal modeling network and output temporal modal features.

[0150] In some embodiments, the diagnostic module 704 is specifically used to: fuse image modal features and temporal modal features to obtain a joint feature representation; convert the joint feature representation into a classification feature vector; process the classification feature vector through a classifier and output the fault diagnosis result.

[0151] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 880. The processor 810, communication interface 820, and memory 830 communicate with each other via the communication bus 880. The processor 810 can call logical instructions in the memory 830 to execute a fault diagnosis method for a crane rotating component. This method includes: first, processing the source domain data and target domain data of the crane rotating component in a unified format to output a normalized vibration signal; then, inputting the normalized vibration signal into a shared encoder for feature mapping, aligning the feature distribution through adversarial training, and outputting domain-invariant features; next, constructing a conditional embedding vector based on the domain-invariant features, generating target domain fault data using a diffusion model, and outputting synthetic fault data; finally, combining the synthetic fault data with real data to form a diagnostic dataset, extracting multimodal features from this dataset, fusing features, and outputting the fault diagnosis result through a classifier.

[0152] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the crane rotating component fault diagnosis method provided by the above methods. The method includes: firstly, processing the source domain data and target domain data of the crane rotating component in a unified format to output a normalized vibration signal; then, inputting the normalized vibration signal into a shared encoder for feature mapping, aligning the feature distribution through adversarial training, and outputting domain-invariant features; next, constructing a conditional embedding vector based on the domain-invariant features, generating target domain fault data using a diffusion model, and outputting synthetic fault data; finally, combining the synthetic fault data with real data to form a diagnostic dataset, extracting multimodal features from the dataset, fusing features, and outputting fault diagnosis results through a classifier.

[0154] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the crane rotating component fault diagnosis method provided by the above methods. The method includes: firstly, processing the source domain data and target domain data of the crane rotating component in a unified format to output a normalized vibration signal; then, inputting the normalized vibration signal into a shared encoder for feature mapping, aligning the feature distribution through adversarial training, and outputting domain-invariant features; next, constructing a conditional embedding vector based on the domain-invariant features, generating target domain fault data using a diffusion model, and outputting synthetic fault data; finally, combining the synthetic fault data with real data to form a diagnostic dataset, extracting multimodal features from the dataset, performing feature fusion, and outputting the fault diagnosis result through a classifier.

[0155] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of diagnosing a failure of a rotating component of a crane, characterized by, The method comprises: The source domain data and the target domain data of the crane rotating part are uniformly formatted, and normalized vibration signals are output, the source domain data is from the fault vibration signals of other rotating machines, and the target domain data is from the normal vibration signals collected on site of the crane; The normalized vibration signals are input into a shared encoder for feature mapping, and feature distribution alignment is realized through adversarial training, and domain invariant features are output, the adversarial training comprises jointly optimizing the shared encoder using domain discriminant loss and maximum mean difference loss; wherein, the maximum mean difference loss forces domain alignment by comparing the feature distribution difference between the source domain and the target domain, and is combined with the domain discriminant loss with a preset weight as a total loss to optimize the shared encoder, and the preset weight is configured as a domain discriminant loss weight of 0.1 and a maximum mean difference loss weight of 1.0; The feature separation network is used to separate the target domain background features representing the target domain equipment background and the source domain fault mode features representing the source domain fault type from the domain invariant features; and the target domain background features are spliced with the source domain fault mode features to form a conditional embedding vector; A diffusion model is used to generate target domain fault data through a reverse denoising process using the conditional embedding vector as a guide, and synthetic fault data are output; wherein, the generation network of the diffusion model adopts a time series Transformer structure; The synthetic fault data and the real data are combined into a diagnostic data set, multi-modal feature extraction is performed on the diagnostic data set, and feature fusion is performed to output a fault diagnosis result through a classifier.

2. The crane rotating component failure diagnosis method according to claim 1, characterized by, The source domain data and the target domain data of the crane rotating part are uniformly formatted, and normalized vibration signals are output, the source domain data is from the fault vibration signals of other rotating machines, and the target domain data is from the normal vibration signals collected on site of the crane; The source domain data and the target domain data are sampled to align the sampling frequencies, and resampled data are output; The resampled data are amplitude normalized, and standardized signals are output; The standardized signals are segmented into fixed-length sequences to form the normalized vibration signals.

3. The crane rotating component failure diagnosis method according to claim 1, characterized by, The normalized vibration signals are input into a shared encoder for feature mapping, and feature distribution alignment is realized through adversarial training, and domain invariant features are output, which comprises: The normalized vibration signals are mapped into initial feature representations using the shared encoder; The initial feature representations are input into a domain discriminator for domain classification, and domain discriminant loss is calculated; The shared encoder is jointly optimized based on the domain discriminant loss and the distribution distance loss, and the domain invariant features are output.

4. The crane rotating part failure diagnostic method according to Claim 1, characterized by, The diagnostic data set is subjected to multi-modal feature extraction, which comprises: The vibration signals in the diagnostic data set are converted into Gram angle field images through Gram angle field coding; Image modal features are extracted from the Gram angle field images; Time series modal features are directly extracted from the vibration signals in the diagnostic data set.

5. The crane rotating component failure diagnostic method according to claim 4, characterized by, The image modal features are extracted from the Gram angle field images, which comprises: Convolutional neural networks are used to extract spatial features from the Gram angle field images, and the image modal features are output; The time series modal features are extracted from the vibration signals in the diagnostic data set, which comprises: The time-dependent relationship of the vibration signal is modeled using a time series modeling network, and a time series modal feature is output.

6. The method of claim 4 or 5, characterized in that The feature fusion outputs a fault diagnosis result through a classifier, including: The image modal feature and the time series modal feature are fused to obtain a joint feature representation. The joint feature representation is converted into a classification feature vector. The classification feature vector is processed by a classifier to output the fault diagnosis result.

7. A crane rotating part failure diagnosing device characterized by comprising: The device includes: A processing module for performing uniform format processing on source domain data and target domain data of a crane rotating component, and outputting a normalized vibration signal, the source domain data being from fault vibration signals of other rotating machines, and the target domain data being from vibration signals collected on site of the crane; An alignment module for inputting the normalized vibration signal into a shared encoder for feature mapping, and realizing feature distribution alignment through adversarial training, and outputting domain-invariant features, the adversarial training including jointly optimizing the shared encoder using a domain discriminant loss and a maximum mean difference loss; wherein the maximum mean difference loss enforces domain alignment by comparing the feature distribution difference between the source domain and the target domain, and is combined with the domain discriminant loss with a preset weight as a total loss to optimize the shared encoder, the preset weight being configured as a domain discriminant loss weight of 0.1 and a maximum mean difference loss weight of 1.0; A fusion module for: using a feature separation network to separate, from the domain-invariant features, a target domain background feature representing a background of the target domain device and a source domain fault pattern feature representing a fault type of the source domain; and concatenating the target domain background feature and the source domain fault pattern feature to combine into a conditional embedding vector; using a diffusion model to generate target domain fault data through a reverse denoising process guided by the conditional embedding vector, and output synthetic fault data; wherein a generation network of the diffusion model adopts a time series Transformer structure; A diagnosis module for combining the synthetic fault data and real data into a diagnosis data set, performing multi-modal feature extraction on the diagnosis data set, and outputting a fault diagnosis result through feature fusion by a classifier.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to implement the crane rotating component fault diagnosis method of any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the crane rotating component fault diagnosis method of any one of claims 1 to 6.

Images