Large machinery dumping trend monitoring and fault diagnosis method

Abstract

The invention relates to the field of large-scale machinery comprehensive guarantee, in particular to a large-scale machinery dumping trend monitoring and fault diagnosis method, which comprises the following steps: acquiring mechanical signals to construct a source domain data set and a target domain data set, and extracting to obtain source domain time-frequency features and target domain time-frequency features; extracting shared features between the source domain time-frequency features and the target domain time-frequency features by taking the KAN-embedded deformable rednet as a feature extractor; calculating the difference of related sub-domain distribution between data set vectors of a source domain and a target domain in a specific layer of the deep neural network by using LMMD; and based on the extracted shared features and the aligned data, carrying out large machine toppling trend monitoring, and determining whether the large machine has a fault and the fault type. A KAN module is embedded into a deep residual network and is combined with deformable convolution, a sub-domain adaptation strategy based on LMMD is adopted, a dynamic weight adjustment factor is introduced, precise modeling of complex nonlinear dynamic characteristics is achieved, and then the dumping trend of large machinery is efficiently and accurately monitored and faults are diagnosed.

Description

Technical Field

[0001] The present application relates to the technical field of comprehensive protection of large-scale machinery, and more specifically, to a method for monitoring the tipping trend and diagnosing faults of large-scale machinery. Background Art

[0002] In the modern construction machinery sector, the overturning of large machinery and equipment has long been a major safety hazard plaguing the industry. While various preventative measures have been implemented to prevent overturning accidents, most focus on immediate intervention before tipping occurs, such as reinforcing the machinery, installing emergency devices, and strengthening monitoring and early warning systems. However, these methods have significant limitations, relying primarily on the operator's experience and immediate response capabilities, and lacking automated and precise monitoring methods.

[0003] With the increasing complexity of construction machinery and the increasing diversity of operating environments, existing monitoring technology systems are no longer able to meet the high demands for real-time safety monitoring of modern construction machinery. Therefore, developing an accurate, efficient, and automated method for monitoring and diagnosing tipping trends has become a critical issue that needs to be addressed. Summary of the Invention

[0004] In view of this, the present application provides a fault diagnosis method for monitoring the tipping tendency of large machines. By monitoring the operating status of the machine in real time and quickly issuing an early warning when the machine is at risk of tipping, the response speed to potential dangers can be significantly improved, the safety of operators can be ensured, and the risk of equipment damage can be reduced.

[0005] The technical solutions provided in this application are as follows:

[0006] A method for monitoring and diagnosing the tipping tendency of large machinery, comprising:

[0007] Collect mechanical signals to construct labeled source domain datasets and unlabeled target domain datasets, and extract source domain time-frequency features and target domain time-frequency features;

[0008] A deformable deep neural network embedded with a specific nonlinear feature enhancement module KAN is used as a feature extractor to extract shared features between the source domain time-frequency features and the target domain time-frequency features;

[0009] The local maximum mean difference (LMMD) is used to calculate the difference in the distribution of related subdomains between the source domain and target domain dataset vectors in a specific layer of the deep neural network, which is used to align the distribution of the source domain and the target domain;

[0010] Based on the extracted shared features and aligned data, the overturning trend of large machinery is monitored to determine whether there is a fault in the large machinery and the type of fault.

[0011] In one possible approach, the source domain dataset contains n samples, each of which contains data features and the data characteristics The corresponding true label The source domain dataset is represented as:

[0012]

[0013] The target domain dataset contains n samples, each of which contains data features X i t , the target domain dataset is expressed as:

[0014] In one possible manner, extracting source domain time-frequency features and target domain time-frequency features includes:

[0015] The mechanical signal is transformed into a wavelet time-frequency image of a specific size through wavelet transform, which is used to characterize the characteristics of the source domain data and the target domain data at different times and frequencies, thereby obtaining the source domain time-frequency characteristics and the target domain time-frequency characteristics.

[0016] In one possible approach, the deformable deep neural network embedded with the specific nonlinear feature enhancement module KAN includes a 7*7 convolutional layer, a pooling layer, and 16 3*3 deformable convolutional layers constituting a residual block, which is used to extract features from the wavelet time-frequency map layer by layer;

[0017] KAN is embedded into the main path of each residual block and combined with the deformable convolution output;

[0018] The output of the residual block is expressed as:

[0019] H(X)=DCN(X)+KAN(X)

[0020] Where H(X) represents the output of the residual block, DCN(X) is the output of the deformable convolutional residual module, and KAN(X) is the nonlinear feature map generated by KAN.

[0021] In one possible approach, after extracting features layer by layer from the wavelet time-frequency graph using a deformable deep neural network embedded with a KAN as a feature extractor, the method further includes:

[0022] The high-dimensional features obtained by the feature extractor are sent to the fully connected layer for integration and transformation to obtain a 1*256 vector of the source domain dataset. and the 1*256 vector of the target domain dataset

[0023] In one possible approach, the method of calculating the difference in relevant subdomain distributions between source domain and target domain dataset vectors in a specific layer of a deep neural network using a local maximum mean difference (LMMD) may further include:

[0024] The activation L layer is calculated by LMMD and The difference in the distribution of related subdomains between them is used as the loss of different domains. The loss is used as the optimization goal to align the distribution of the source domain and the target domain, removing the working condition fluctuations and the distribution differences between domains.

[0025] In one possible approach, based on the extracted shared features and aligned data, large machinery tipping trend monitoring is performed to determine whether the large machinery has a fault and the type of fault, including:

[0026] Measuring the actual fault conditions in the source domain dataset through the cross entropy loss function Predicting fault conditions with target domain dataset The difference between them is used to obtain the classification loss, which is expressed as:

[0027]

[0028] Where c is the number of bearing fault categories, Represents the true label of the source domain dataset sample; predicted label It is calculated by the softmax function in the domain-sharing classifier and is expressed as:

[0029]

[0030] Where, Represents the output of the source domain dataset in the fully connected layer;

[0031] Determine the LMMD domain adaptation loss function ε l , expressed as:

[0032]

[0033]

[0034] Where n is the number of samples in the target domain dataset, λ is the trade-off parameter, e is the number of training iterations of the current model, and es is the total number; α is the dynamic weight adjustment factor, and λ is the trade-off parameter;

[0035] Based on the classification loss ε d and LMMD domain adaptation loss function ε l Determine the overall loss function of the network model, expressed as:

[0036] ε total =εd +ε l

[0037] Taking minimizing the overall loss function as the overall optimization goal, a fault diagnosis model is obtained through iterative training to determine whether a large-scale machinery has a fault and the type of fault.

[0038] Compared with the existing technology, the technical solution provided by this application has the following beneficial effects:

[0039] This application improves the traditional residual network ResNet and adopts the method of embedding Kolmogorov-Arnold Networks to enhance the network's feature representation ability. KAN optimizes the information transmission and expression capabilities of deep networks through its unique network structure and feature mapping method, so that the network can more efficiently extract transferable features between the source domain and the target domain. ResNet embedded with KAN can adaptively adjust the network structure, making its feature migration between different working conditions (such as different load conditions) more flexible and accurate, and improving the monitoring and diagnosis capabilities of potential dumping trends.

[0040] Traditional methods based on adversarial neural networks (DANNs) and global maximum mean difference (MMD) typically rely on adversarial training mechanisms to narrow the distribution differences between the source and target domains. However, this method has a complex training process and high computational overhead. This invention simplifies the model structure and improves diagnostic efficiency by removing the adversarial training mechanism and using local maximum mean difference (LMMD) to align the distributions of the same fault type under different operating conditions. LMMD can accurately align the distributions of the source and target domains within a smaller local area, thereby improving the accuracy and robustness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a method for monitoring and diagnosing the tipping trend of large machinery provided in Example 1 of the present application.

[0042] Figure 2 This is a flow chart of a method for monitoring the tipping trend and diagnosing faults of large machinery provided in Example 2 of the present application.

[0043] Figure 3 This is a framework diagram of the fault diagnosis model provided in Example 2 of this application.

[0044] Figure 4 This is a flowchart of the method for monitoring tipping trends and diagnosing faults based on a trained model provided in Example 3 of the present application. DETAILED DESCRIPTION

[0045] The following will combine the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0046] To solve the problems raised in the background technology, this application proposes a method for monitoring the tipping tendency and fault diagnosis of large machinery using pressure sensors based on a subdomain adaptive network embedded in a deep residual KAN. This method combines the advantages of deep transfer learning and Kolmogorov-Arnold Networks (KAN), and can effectively monitor the tipping tendency of large machinery under different working conditions and improve its fault diagnosis capabilities. This model fully utilizes the advantages of unsupervised learning. Without the need for a large amount of labeled data, it can effectively extract fault features from sensor data and accurately determine whether the machinery has a tipping tendency.

[0047] Below, the technical solution provided by this application is described in detail through specific embodiments.

[0048] Example 1

[0049] See also Figure 1 , is a flow chart of a method for monitoring and diagnosing the tipping trend of large machinery provided in Example 1 of this application. Figure 1 As shown in , the specific implementation steps of the above method include:

[0050] Step 101: Acquire mechanical signals to construct a source domain dataset and a target domain dataset, and extract source domain time-frequency features and target domain time-frequency features.

[0051] Step 102: Using a deformable deep neural network embedded with a specific nonlinear feature enhancement module KAN as a feature extractor, the shared features between the source domain time-frequency features and the target domain time-frequency features are extracted.

[0052] Step 103: Use the local maximum mean difference (LMMD) to calculate the difference in the distribution of relevant subdomains between the source domain and target domain dataset vectors in a specific layer of the deep neural network, and use it as the loss of different domains to align the distribution of the source domain and the target domain, eliminating working condition fluctuations and inter-domain distribution differences.

[0053] Step 104: Based on the extracted shared features and aligned data, monitor the tendency of large machinery toppling to determine whether there is a fault in the large machinery and the type of fault.

[0054] Compared with the prior art, the technical solution provided in Example 1 of the present application has the following beneficial effects:

[0055] This application improves the traditional residual network ResNet and adopts the method of embedding Kolmogorov-Arnold Networks to enhance the network's feature representation ability. KAN optimizes the information transmission and expression capabilities of deep networks through its unique network structure and feature mapping method, so that the network can more efficiently extract transferable features between the source domain and the target domain. ResNet embedded with KAN can adaptively adjust the network structure, making its feature migration between different working conditions (such as different load conditions) more flexible and accurate, and improving the monitoring and diagnosis capabilities of potential dumping trends.

[0056] Traditional methods based on adversarial neural networks (DANNs) and global maximum mean difference (MMD) typically rely on adversarial training mechanisms to narrow the distribution differences between the source and target domains. However, this method has a complex training process and high computational overhead. This invention simplifies the model structure and improves diagnostic efficiency by removing the adversarial training mechanism and using local maximum mean difference (LMMD) to align the distributions of the same fault type under different operating conditions. LMMD can accurately align the distributions of the source and target domains within a smaller local area, thereby improving the accuracy and robustness of fault diagnosis.

[0057] Example 2

[0058] See also Figure 2 , is a flow chart of a method for monitoring and diagnosing the overturning trend of large machinery provided in Example 2 of this application. Figure 2 As shown in , the specific implementation steps of the above method include:

[0059] Step 201: Collect mechanical signals of the monitored large-scale machinery to construct a source domain dataset and a target domain dataset.

[0060] Mechanical signals of large machinery reflect changes in mechanical properties during operation. These signals include operational data under different operating conditions and the dynamics of mechanical parameters such as force, pressure, torque, and vibration experienced by various components. For example, when a large crane is lifting a heavy object, its boom experiences tension, as well as vibration caused by the swaying of the object.

[0061] As a feasible method, pressure sensors pre-installed on large machinery can be used to obtain the above-mentioned mechanical signals. Specifically, pressure sensors are installed on key load-bearing components of large machinery. When the large machinery is in operation, the pressure changes on these components will cause the pressure sensors to produce corresponding electrical signal changes. These electrical signals carry relevant information about the mechanical signals. These electrical signals are transmitted to a specific data acquisition system via cables or wirelessly, and then input into a pre-built fault diagnosis model to process and analyze the above signals and conduct real-time monitoring to determine whether the large machinery is at risk of overturning.

[0062] In this embodiment, operating data under normal operating conditions is used as the source domain data, and operating data under abnormal operating conditions is used as the target domain data, forming the source and target domain datasets, respectively. The different operating conditions of large machinery include normal operating conditions and abnormal operating conditions with potential for tipping or component failure. The source domain data, each of which is assigned corresponding label information, constitutes the source domain dataset.

[0063] The target domain data is constructed using operational data collected during tipping trends or fault conditions as target domain data. These tipping trends and fault conditions are diverse and complex. When acquiring data, the data is unlabeled because it's difficult to accurately determine the type and severity of the fault beforehand. For example, when an excavator malfunctions, it could be caused by a variety of different reasons. During the data collection phase, it's impossible to immediately classify and label this data in detail. Instead, only the raw data is collected, leaving the specific labels to be determined later during further analysis and processing.

[0064] Step 202: transform the collected mechanical signals of the large-scale machinery into wavelet time-frequency images of a specific size through wavelet transform to extract source domain time-frequency features and target domain time-frequency features.

[0065] Specifically, wavelet transforms are used to convert time-series data into time-frequency images, revealing the data's characteristics at different times and frequencies. In this embodiment, the mechanical signal is transformed into a 224*224 wavelet time-frequency image through wavelet transform, enabling more effective identification of the machine's operating status, particularly detecting abnormal tipping trends or fault conditions.

[0066] Step 203: Use the deformable deep residual network Resnet as a feature extractor to capture high-dimensional features of the source domain dataset and the target domain dataset respectively.

[0067] Due to different working conditions, there are differences in the data distribution of the source domain and the target domain, but there are also some common features between them. For example, regardless of whether the machine is in normal working conditions or fault conditions, certain basic mechanical principles and signal characteristics remain unchanged, such as the basic frequency range of mechanical vibration, the approximate fluctuation law of the pressure signal, etc. Learning shared features is to let the network find these representative features that exist in both the source domain and the target domain, so that the model can use these common features to perform fault diagnosis and improve the generalization ability of the model in the target domain. This application uses the subdomain adaptive network of the improved deep residual embedding KAN to extract features from the data and learn the shared features between the source domain data and the target domain data.

[0068] Specifically, the wavelet time-frequency map obtained after the above-mentioned wavelet transform and other preprocessing is input into the deformable deep residual network ResNet network to extract the features in the wavelet time-frequency map layer by layer.

[0069] As a classic deep learning architecture, the deep residual network ResNet is composed of multiple residual blocks stacked together and has a very deep network structure. Depending on the depth, the deep residual network ResNet can have many variants, including but not limited to ResNet-18, ResNet-34, etc. The embodiment of the present application uses ResNet-18 as a feature extractor. By introducing the residual learning unit, the gradient disappearance problem in the deep network is effectively alleviated. The deformable convolution further enhances the network's adaptability to spatial features, and dynamically adjusts the sampling position of the convolution kernel by learning the offset to adapt to complex target deformations.

[0070] Specifically, if Figure 3 As shown in Figure 2, the ResNet-18 architecture consists of one 7x7 convolutional layer, one pooling layer, and 16 3x3 deformable convolutional layers forming a residual block. The 7x7 convolutional layer is used for initial downsampling and feature extraction, while the 3x3 deformable convolutional network (DCN) layer adaptively adjusts the position of the convolution kernel to more accurately capture complex features. The number and dimensionality of feature maps change with each convolution layer. The pooling layer, located after the 7x7 convolutional layer, reduces the data dimensionality and computational effort through downsampling while preserving important features.

[0071] In addition, to further enhance the ability to express nonlinear features, this application also embeds the Kolmogorov-Arnold Networks (KAN) module. The KAN module is embedded in the main path of each residual block and combined with the deformable convolution output. The output of the residual block is expressed as:

[0072] H(X)=DCN(X)+KAN(X)

[0073] Where H(X) represents the output of the residual block, DCN(X) is the output of the deformable convolutional residual module, and KAN(X) is the nonlinear feature map generated by the KAN module.

[0074] In the embodiment of the present application, after multi-layer feature extraction by deformable ResNet-18, the high-dimensional features of the output are representative 1*512 vector features, and the above vector features are used as shared features between the source domain and the target domain.

[0075] Step 204: Send the high-dimensional features obtained by the feature extractor to the fully connected layer to integrate and transform the high-dimensional features.

[0076] like Figure 3 As shown in , the fully connected layer also includes a Flatten layer before the fully connected layer, which is used to convert the multi-dimensional feature map output by the previous convolutional layer into a one-dimensional vector, which is convenient for subsequent fully connected layer processing. Each neuron in the fully connected layer is connected to all the neurons of the one-dimensional vector output by the previous layer, namely the Flatten layer, to integrate and transform the input features, and further extract and process the features. In the embodiment of the present application, the deformable ResNet-18 is used as a feature extractor to extract high-dimensional features from the source domain and target domain data sets, which are represented by 1*512 vector features. The deformable Resnet-18 feature extractor is connected to the fully connected layer to integrate and transform the high-dimensional features extracted previously to obtain a 1*256 vector of the source domain data set. and the 1*256 vector of the target domain dataset That is, after the operation of the fully connected layer, the feature vectors of the source domain dataset and the target domain dataset change from 1*512 dimensions to 1*256 dimensions.

[0077] The Softmax layer follows the fully connected layer, receiving its output and converting it into a probability distribution. The Softmax layer calculates the probability value for each category and uses this probability to determine whether the input data represents a normal operating condition or a specific abnormal condition, thereby achieving classification.

[0078] Step 205: Align the distributions of the source domain and the target domain through the local maximum difference (LMMD) to remove the working condition fluctuations and the distribution differences between domains.

[0079] Local Maximum Mean Difference (LMMD) is a method for measuring the difference between two probability distributions. The embodiment of the present application adopts a subdomain adaptation strategy based on the local maximum mean difference (LMMD) to calculate the distribution difference between the source domain and target domain features in different subdomains. The difference in data distribution between the source domain and the target domain is reduced based on subdomain adaptation. In contrast, traditional global domain adaptation methods have difficulty in accurately aligning feature distributions of different subdomains, while subdomain adaptation can reduce distribution differences more finely by decomposing the feature space and independently optimizing the corresponding subdomains.

[0080] As an achievable approach, after completing the probability distribution calculation of the Softmax layer, in a specific layer of the model, such as the activation L layer, the difference in the distribution of the relevant subdomains between the source domain dataset vector and the target domain dataset vector is calculated and regarded as the loss of different domains. In this embodiment of the present application, the activation value of the L layer is selected. and This is used to calculate the LMMD, which is considered the loss between different domains. This loss can be used as an optimization target during machine learning optimization. During model training, minimizing this loss helps ensure that the feature distributions of the source and target domains at layer L are as similar as possible, thereby reducing inter-domain differences and improving model performance on the target domain.

[0081] Under different operating conditions, the operating state of machinery changes, resulting in fluctuations in the collected data. Furthermore, inherent distribution differences exist between the source and target domains. Using LMMD to align distributions effectively reduces the impact of these operating condition fluctuations and inter-domain distribution differences on model performance. For example, even if the distribution of mechanical signals in the target domain differs significantly from that in the source domain due to a mechanical fault, LMMD adjustments can still enable the model to more accurately identify the fault type, improving the accuracy and reliability of fault diagnosis.

[0082] After the distribution alignment between the source domain and the target domain is completed, the fault classification and diagnosis of large-scale machinery tipping trends can be carried out based on the extracted features and the aligned data, and real-time monitoring and early warning can be issued. The specific steps include the following:

[0083] Step 206: Measure the actual fault conditions in the source domain dataset using the cross entropy loss function Predicting fault conditions with target domain dataset The difference between them is used to obtain the classification loss.

[0084] Among them, the above classification loss is expressed as:

[0085]

[0086] Where c is the number of bearing fault categories, Represents the true label of the source domain dataset sample. Predicted label It is calculated by the softmax function in the domain-sharing classifier, which is defined as:

[0087]

[0088] in, Represents the output of the source domain dataset in the fully connected layer.

[0089] Step 207: Calculate the LMMD domain adaptive loss function ε l .

[0090] Another optimization objective ε l , is to reduce the distribution difference of related subdomains of the same class in the source domain dataset and the target domain dataset. The calculation formula is:

[0091]

[0092] Where n is the number of samples in the target domain dataset, λ is the trade-off parameter, e is the number of training iterations for the current model, and es is the total number. α is the dynamic weight adjustment factor, and λ is the trade-off parameter.

[0093] Step 208: Based on the above classification loss ε and LMMD domain adaptive loss function ε l Determine the overall loss function of the network model.

[0094] Specifically, the above overall loss function is defined as:

[0095] ε total =ε d +ε l

[0096] In summary, there are two loss functions in the model proposed in the embodiment of the present application, namely the LMMD domain adaptive loss function ε l And the classification loss function ε d Minimize ε l It can reduce the relevant local distribution difference between each pair of source and target domains. Similarly, minimizing the classification loss ε d The source domain samples can be predicted more accurately.

[0097] Step 209 : Taking minimizing the above-mentioned overall loss function as the overall optimization goal, a fault diagnosis model is obtained through iterative training.

[0098] In the embodiment of the present application, the overall optimization target is gradually reduced by continuously adjusting the parameters of the model. During the training process, the model continuously learns the characteristic patterns of the source domain and target domain data based on the input features and labels, thereby improving the accuracy of fault classification.

[0099] After model training is complete, real-time operational data from large machinery is fed into the trained model. The fault diagnosis model then classifies the input data based on the extracted features and learned classification patterns, determining whether the large machinery is prone to tipping or experiencing other faults.

[0100] Compared with the prior art, the technical solution provided in Example 2 of the present application has the following beneficial effects:

[0101] The technical solution of this application processes the signal of the pressure sensor and inputs it into the deep learning model for fault diagnosis. The signal is first denoised and feature extracted through the preprocessing module to filter out the noise signal and enhance the effective features. Then, feature learning and fault pattern recognition are performed through the subdomain adaptive network of the deep residual embedding KAN. The network can dynamically adjust the weight of the features according to different working conditions, and align the distribution of the source domain and the target domain through the local maximum mean difference LMMD, thereby effectively overcoming the impact of the working condition difference and ensuring the accuracy of fault diagnosis.

[0102] Finally, the trained model compares the feature distributions of the source and target domains, outputs a classification result of potential tipping or other faults, and provides real-time feedback to operators through the early warning system. This process ensures that when a large machine is about to tip, a warning signal is issued in a timely manner, preventing equipment damage and casualties.

[0103] Example 3

[0104] See also Figure 4 , is a flow chart of the method for monitoring the tipping trend and diagnosing the fault based on the trained model provided in the third embodiment of the present application. Figure 4 As shown in , the specific implementation steps of the above method include:

[0105] Step 301: Obtain a labeled source domain dataset and unlabeled target domain datasets

[0106] Each sample in the source domain dataset contains data features X i s and its corresponding true label Y i s Each sample in the target domain dataset only contains data features X i t , representing data that may have a dumping trend or an unknown fault state.

[0107] Step 302: Randomly extract samples from the source domain dataset and the target domain dataset, and feed the samples into the KAN embedded deformable ResNet-18 feature extractor to extract high-dimensional features. and

[0108] Step 303: Use the fault classifier to classify the high-dimensional features corresponding to the source domain data. Perform classification and obtain predicted labels

[0109] Step 304: Calculate the loss function ε a and ε l , determine the total loss function ε total .

[0110] Step 305: According to the total loss ε total Update the feature extractor and classifier.

[0111] Step 306: After completing model training and updating, input the target domain data into the model and output the fault diagnosis result, i.e. the predicted target domain data set label. Determine whether there is a fault in large machinery and the type of fault.

[0112] The technical solution of this application embeds Kolmogorov-Arnold Networks into a deep residual network for the first time and combines it with deformable convolution, enabling the model to accurately capture complex nonlinear dynamic features, greatly improving the ability to extract features of the operating status of large machinery under different working conditions, and thus more accurately diagnosing faults. The adoption of a subdomain adaptation strategy based on LMMD effectively improves the alignment accuracy of cross-domain feature distributions, overcomes the limitations of traditional global adaptation methods, and enables the model to more accurately identify faults and reduce the misjudgment rate when processing data under different working conditions. The dynamic weight adjustment factor introduced in the domain adaptation process achieves loss balance in different training stages, optimizes the model training process, improves training efficiency, enables the model to converge to a better solution faster, and enhances the stability and generalization ability of the model.

[0113] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring and diagnosing the tipping trend of large machinery, characterized in that: include: Collect mechanical signals to construct labeled source domain datasets and unlabeled target domain datasets, and extract source domain time-frequency features and target domain time-frequency features; A deformable deep neural network embedded with a specific nonlinear feature enhancement module KAN is used as a feature extractor to extract shared features between the source domain time-frequency features and the target domain time-frequency features; The local maximum mean difference (LMMD) is used to calculate the difference in the distribution of related subdomains between the source domain and target domain dataset vectors in a specific layer of the deep neural network, which is used to align the distribution of the source domain and the target domain; Based on the extracted shared features and aligned data, the overturning trend of large machinery is monitored to determine whether there is a fault in the large machinery and the type of fault.

2. A method for monitoring and diagnosing the overturning tendency of large machinery according to claim 1, characterized in that: The source domain dataset contains n samples, each of which contains data features. and the data characteristics The corresponding true label The source domain dataset is represented as: The target domain dataset contains n samples, each of which contains data features The target domain dataset is represented as:

3. A method for monitoring and diagnosing the overturning tendency of large machinery according to claim 1, characterized in that: The extracting of source domain time-frequency features and target domain time-frequency features includes: The mechanical signal is transformed into a wavelet time-frequency image of a specific size through wavelet transform, which is used to characterize the characteristics of the source domain data and the target domain data at different times and frequencies, thereby obtaining the source domain time-frequency characteristics and the target domain time-frequency characteristics.

4. A method for monitoring and diagnosing the tipping tendency of large machinery according to claim 1, characterized in that: The deformable deep neural network embedded with the specific nonlinear feature enhancement module KAN includes a 7*7 convolutional layer, a pooling layer, and 16 3*3 deformable convolutional layers constituting a residual block, which is used to extract features from the wavelet time-frequency graph layer by layer; KAN is embedded into the main path of each residual block and combined with the deformable convolution output; The output of the residual block is expressed as: H(X)=DCN(X)+KAN(X) Where H(X) represents the output of the residual block, DCN(X) is the output of the deformable convolutional residual module, and KAN(X) is the nonlinear feature map generated by KAN.

5. A method for monitoring and diagnosing the overturning tendency of large machinery according to claim 1, characterized in that: After extracting features layer by layer from the wavelet time-frequency graph using the deformable deep neural network embedded with the KAN as a feature extractor, the method further includes: The high-dimensional features obtained by the feature extractor are sent to the fully connected layer for integration and transformation to obtain a 1*256 vector of the source domain dataset. and the 1*256 vector of the target domain dataset 6. A method for monitoring and diagnosing the overturning tendency of large machinery according to claim 5, characterized in that: The method of calculating the difference in the distribution of related subdomains between the source domain and target domain dataset vectors in a specific layer of the deep neural network using the local maximum mean difference (LMMD) further includes: The activation L layer is calculated by LMMD and The difference in the distribution of related subdomains between them is used as the loss of different domains. The loss is used as the optimization goal to align the distribution of the source domain and the target domain, removing the working condition fluctuations and the distribution differences between domains.

7. A method for monitoring and diagnosing the overturning tendency of large machinery according to claim 1, characterized in that: Based on the extracted shared features and aligned data, large machinery tipping trend monitoring is performed to determine whether there is a fault in the large machinery and the type of fault, including: Measuring the actual fault conditions in the source domain dataset through the cross entropy loss function Predicting fault conditions with target domain dataset The difference between them is used to obtain the classification loss, which is expressed as: Where c is the number of bearing fault categories, Represents the true label of the source domain dataset sample; predicted label It is calculated by the softmax function in the domain-sharing classifier and is expressed as: Where, Represents the output of the source domain dataset in the fully connected layer; Determine the LMMD domain adaptation loss function ε l , expressed as: Where n is the number of samples in the target domain dataset, λ is the trade-off parameter, e is the number of training iterations of the current model, and es is the total number; α is the dynamic weight adjustment factor, and λ is the trade-off parameter; Based on the classification loss ε d and LMMD domain adaptation loss function ε l Determine the overall loss function of the network model, expressed as: e total =e d +e l Taking minimizing the overall loss function as the overall optimization goal, a fault diagnosis model is obtained through iterative training to determine whether a large-scale machinery has a fault and the type of fault.

Images