Shield tunneling machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning

Abstract

The invention discloses a shield tunneling machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning, and belongs to the technical field of engineering machinery and artificial intelligence. According to the method, pressure, flow and displacement signals are processed according to channels, local time domain features are extracted by using a one-dimensional feature fusion convolutional neural network based on an attention mechanism, weights are dynamically distributed in combination with a channel attention mechanism, key fault features are strengthened, and noise is suppressed; the feature distribution of the source domain and the target domain is aligned through the maximum mean value difference, and the field offset problem under target domain data scarcity is solved in combination with a classifier adaptive optimization module. According to the method, high-precision fault classification can be realized in a small sample scene, the dependence on labeled data is reduced, the real-time performance and robustness of fault diagnosis of the shield tunneling machine hydraulic system are improved, and the shutdown maintenance cost is reduced.

Description

Technical Field

[0001] The present application relates to the fields of engineering machinery and artificial intelligence technology, and specifically to a fault diagnosis method for a shield machine hydraulic system based on deep convolution and shallow transfer learning. Background Art

[0002] Shield machines are large-scale engineering equipment used in tunnel construction. Their hydraulic systems are one of the core systems that maintain their normal operation. Due to the complex components and the harsh construction environment, hydraulic system failures are common, posing risks to construction safety.

[0003] Based on the aforementioned factors, due to the high reliability of hydraulic system components and the precision involved in shield machine development, shield machine failures present a "few samples, many types" characteristic, which increases the complexity of fault diagnosis. Traditional fault diagnosis relies heavily on fault samples, leading to insufficient understanding of fault diagnosis and identification among engineers. Summary of the Invention

[0004] This application provides a shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning, which solves the problem that the shield machine hydraulic system fault diagnosis method cannot be separated from a large number of fault samples.

[0005] The technical solution of this application is as follows: A fault diagnosis method for a shield machine hydraulic system based on deep convolution and shallow transfer learning includes the following methods: S1. Obtain pressure signal, flow signal and displacement signal of the shield machine hydraulic system; S2. Input the pressure signal, flow signal, and displacement signal as independent channel signals into the convolution layer and pooling layer of a one-dimensional feature fusion convolutional neural network based on an attention mechanism to extract local time domain features of each channel; S3. The channel attention mechanism of the one-dimensional feature fusion convolutional neural network based on the attention mechanism calculates the weights of the extracted local time domain feature signals of each channel and distributes the weights according to the global statistical information of the channel signals; S4, weighting the extracted local time domain feature signals of each channel and fusing them into a unified feature vector to obtain a depth feature vector; S5. Use the maximum mean difference to characterize the distribution difference of deep features between the source domain and the target domain, and map the data of the source domain and the target domain to the same feature space by minimizing the maximum mean difference loss function; S6. Add a dynamic adjustment mechanism to the fully connected layer of the one-dimensional feature fusion convolutional neural network. Based on the distribution characteristics of the target domain data in the feature space, the dynamic adjustment mechanism is used to optimize the weight parameters of the fully connected layer to correct the classification boundary. S7. Determine the fault of the shield machine hydraulic system based on the classification result of the target domain data.

[0006] Furthermore, in step S1, the pressure signal includes a hydraulic pump output pressure signal, a relief valve output pressure signal, a speed control valve output pressure signal, a reversing valve output pressure signal, and a hydraulic cylinder output pressure signal, and in step S2, each pressure signal serves as an independent channel.

[0007] Furthermore, the flow signal in step S1 is a flow signal output by the speed control valve, and the displacement signal is a displacement signal of the hydraulic cylinder piston.

[0008] Furthermore, in step S3, the global statistical information includes mean and variance.

[0009] Furthermore, the dynamic adjustment mechanism in step S6 is an adversarial training mechanism or a parameter adaptive fine-tuning mechanism.

[0010] Furthermore, in step S7, the classification result is used as a query input of the knowledge graph, and the fault of the shield machine hydraulic system is determined based on the knowledge graph.

[0011] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows: 1. This application improves feature extraction accuracy by processing multi-source signals in separate channels. Specifically, this application addresses the problem of feature interference caused by dimensional differences in pressure, flow, and displacement signals. It extracts local time-domain features through independent channels, avoiding signal superposition conflicts and reducing the risk of misjudgment.

[0012] 2. This application dynamically calculates and assigns weights to each channel based on global correlation. Specifically, when abnormal fault characteristics appear, the channel weights are significantly strengthened and normal signals are suppressed. This can significantly increase the contribution of key features, significantly reduce noise interference, and improve the accuracy of fault identification.

[0013] 3. This application addresses the issue of a limited number of fault samples through transfer learning. Specifically, in step S5, the distribution difference between the source and target domains is measured using MMD, and the MMD loss function is minimized to map the features of the two domains into the same space. This allows the model to leverage source domain knowledge to achieve high-accuracy fault identification even when the target domain has fewer annotations, significantly reducing its reliance on field fault samples.

[0014] 4. This application uses classifier adaptive optimization to optimize the decision boundary. Specifically, this application introduces a parameter fine-tuning mechanism in step S6. This dynamic adjustment mechanism uses adaptive parameter fine-tuning to dynamically adjust the weights or biases of the classification layer based on the distribution of the target domain data (such as feature mean and variance), thereby optimizing the location of the classification boundary. The parameter fine-tuning mechanism in this application can dynamically adjust the classification boundary based on the target domain distribution.

[0015] 5. This application realizes a direct association between fault classification and diagnosis. Specifically, this application directly associates the diagnosis results through fault type classification. Each type of fault corresponds to a preset solution, realizing a closed-loop solution of diagnosis-classification-decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0017] Figure 1 This is the architecture diagram of the one-dimensional feature fusion convolutional neural network based on the attention mechanism in this application; Figure 2 Schematic diagram of the principle of one-dimensional feature fusion convolutional neural network based on attention mechanism in this application. DETAILED DESCRIPTION

[0018] Based on the background technology, see the attached Figure 1 and attached Figure 2 , this application provides a shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning, including the following methods: S1. Obtain the pressure signal, flow signal and displacement signal of the shield machine hydraulic system.

[0019] In step S1, the pressure signals include the hydraulic pump output pressure signal, the overflow valve output pressure signal, the speed control valve output pressure signal, the reversing valve output pressure signal, and the hydraulic cylinder output pressure signal. In step S2, each pressure signal is treated as an independent channel. The flow signal is the speed control valve output flow signal, and the displacement signal is the hydraulic cylinder piston displacement signal. In specific implementation, the seven independent signals of the shield machine hydraulic system are received through a multi-channel input layer.

[0020] S2. Input the pressure signal, flow signal, and displacement signal as independent channel signals into the convolution layer and pooling layer of the one-dimensional feature fusion convolutional neural network based on the attention mechanism to extract the local time domain features of each channel.

[0021] The one-dimensional feature fusion convolutional neural network based on the attention mechanism is a multi-channel convolution processing. Each channel performs convolution operation independently, which is to distinguish the local time domain features of pressure, flow, and displacement.

[0022] A convolutional neural network uses an attention mechanism to extract features by fusing one-dimensional features. The convolutional layer uses convolution kernels to extract local features of the data. In the multi-channel convolution processing of a convolutional neural network, each channel undergoes an independent convolution operation, and the convolution results of each channel are then accumulated at the corresponding position to form the final output. The convolution kernel is three-dimensional in this process, with the size of its third dimension matching the number of channels in the input data. Specifically, for each channel of the input data, a corresponding two-dimensional convolution kernel is used for convolution. The purpose of the pooling layer is to reduce the dimensionality of the data and extract effective features. Max pooling extracts features by selecting the maximum value from the covered input region, while average pooling achieves this by calculating the average of all values ​​within the region. Both pooling techniques reduce the computational burden of subsequent network layers while still capturing important information in the input data. The one-dimensional convolution module performs an independent convolution operation on each channel to extract local temporal features. The pooling module downsamples the convolution features (max pooling), compressing redundant information and enhancing feature robustness.

[0023] S3, S3, the channel attention mechanism of the one-dimensional feature fusion convolutional neural network based on the attention mechanism calculates the weights of the extracted local time domain feature signals of each channel, and distributes the weights according to the global statistical information of the channel signal. Global statistical information includes mean and variance. After extracting local features by channel, the attention mechanism (such as the channel attention module) can dynamically allocate weights based on the global statistical information (mean, variance) of the signal. For example: if the flow signal fluctuates abnormally, the attention module will enhance its channel weight; if the pressure signal is within the normal range, its channel weight is automatically suppressed. The channel attention module can calculate global statistics → mean and variance for each channel feature. The fault type of the shield machine hydraulic system is usually strongly correlated with specific signals: Abnormal pressure → pump valve failure or overflow valve blockage; Abnormal flow → Pipeline leakage or speed regulating valve failure; Abnormal displacement → Hydraulic cylinder stuck or sensor drifting.

[0024] S4. Weighting the extracted local time domain feature signals of each channel and fusing them into a unified feature vector to obtain a depth feature vector. In specific implementation, the weighted multi-channel features are fused into a unified depth feature vector through a weighted fusion module.

[0025] S5. Use the maximum mean difference to characterize the difference in deep feature distribution between the source domain and the target domain, and map the data of the source domain and the target domain to the same feature space by minimizing the maximum mean difference loss function.

[0026] This step calculates the maximum mean difference loss using the MMD domain alignment module. The network parameters are updated via gradient backpropagation to align the feature distributions of the two domains. The target domain data consists of sensor signals collected in real time by the current shield machine (which may be unlabeled or have a small number of labels). These signals are also extracted as deep features through steps S1-S4, but their distribution may differ from that of the source domain, such as feature drift caused by equipment aging. The source domain consists of labeled sensor signals from a historical database, such as hydraulic system fault records from a tunnel project. These signals have been extracted as deep features through steps S1-S4, such as normal operating condition features and fault category A / B / C features.

[0027] S6. A dynamic adjustment mechanism is added to the fully connected layer of the one-dimensional feature fusion convolutional neural network to optimize the classification boundary based on the distribution characteristics of the target domain data in the feature space. The distribution characteristics of the target domain data mainly refer to domain shift. The joint distribution of input features and labels may differ (for example, due to aging of field equipment, the characteristic representation of the same failure mode may differ from that in the source domain). The dynamic fully connected layer can achieve adaptive optimization of the classification boundary.

[0028] The dynamic adjustment mechanism in this step is an adversarial training mechanism or a parameter adaptive fine-tuning mechanism.

[0029] S7. Determine the fault of the shield machine hydraulic system based on the classification results of the target domain data. In step S7, the classification results are used as query input for the knowledge graph, and the shield machine hydraulic system fault is determined based on the knowledge graph. A neural network model is used to extract features and perform transfer learning on sensor signals, classifying the target domain data into predefined fault categories (e.g., Category A "hydraulic pump leakage" and Category B "stuck reversing valve"). The classification results trigger domain knowledge in the knowledge graph, combining the fault category, signal characteristics, and context to infer the specific fault cause and solution. The same fault category (e.g., Category A) may correspond to multiple solutions, requiring dynamic matching through the knowledge graph. The specific classification results output by the model are then used for further screening within the knowledge graph. Solution priority is adjusted based on the equipment's operating status (e.g., construction phase, load conditions).

[0030] The knowledge graph is built based on domain knowledge of shield machine hydraulic systems, historical failure cases, equipment maintenance manuals, and expert experience. The fault-cause-solution triples in historical maintenance records are integrated with expert-annotated fault feature rules to define the hierarchical structure and causal relationships of faults. When target domain data is scarce, predefined expert rule reasoning (such as "sudden drop in flow signal + sudden rise in pressure signal → directional valve stuck") is used to directly associate fault features with causes. The semantic similarity between entities in the graph (such as the strong association between "overflow valve failure" and "pressure runaway") is utilized to generalize the diagnostic capabilities of a small number of samples. The construction of knowledge graphs is an existing technology in this field, and technicians can complete it through professional knowledge.

[0031] Anything not described in this application can be achieved by adopting or drawing on existing technologies.

[0032] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A fault diagnosis method for shield machine hydraulic system based on deep convolution and shallow transfer learning, characterized in that: This includes the following methods: S1. Obtain pressure signal, flow signal and displacement signal of the shield machine hydraulic system; S2. Input the pressure signal, flow signal, and displacement signal as independent channel signals into the convolution layer and pooling layer of a one-dimensional feature fusion convolutional neural network based on an attention mechanism to extract local time domain features of each channel; S3. The channel attention mechanism of the one-dimensional feature fusion convolutional neural network based on the attention mechanism calculates the weights of the extracted local time domain feature signals of each channel and distributes the weights according to the global statistical information of the channel signals; S4, weighting the extracted local time domain feature signals of each channel and fusing them into a unified feature vector to obtain a depth feature vector; S5. Use the maximum mean difference to characterize the distribution difference of deep features between the source domain and the target domain, and map the data of the source domain and the target domain to the same feature space by minimizing the maximum mean difference loss function; S6. Add a dynamic adjustment mechanism to the fully connected layer of the one-dimensional feature fusion convolutional neural network. Based on the distribution characteristics of the target domain data in the feature space, the dynamic adjustment mechanism is used to optimize the weight parameters of the fully connected layer to correct the classification boundary. S7. Determine the fault of the shield machine hydraulic system based on the classification result of the target domain data.

2. A shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning according to claim 1, characterized in that: In step S1, the pressure signals include the hydraulic pump output pressure signal, the overflow valve output pressure signal, the speed control valve output pressure signal, the reversing valve output pressure signal, and the hydraulic cylinder output pressure signal, and in step S2, each pressure signal is used as an independent channel.

3. A shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning according to claim 2, characterized in that: The flow signal in step S1 is the flow signal output by the speed control valve, and the displacement signal is the displacement signal of the hydraulic cylinder piston.

4. A shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning according to claim 3, characterized in that: In step S3, the global statistical information includes mean and variance.

5. A shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning according to claim 4, characterized in that: The dynamic adjustment mechanism in step S6 is an adversarial training mechanism or a parameter adaptive fine-tuning mechanism.

6. A shield machine hydraulic system fault diagnosis method based on deep convolution and shallow transfer learning according to claim 5, characterized in that: In step S7, the classification result is used as a query input of the knowledge graph, and the fault of the shield machine hydraulic system is determined based on the knowledge graph.

Images