Shield machine fault detection method and system based on edge computing

Abstract

The present invention provides a shield machine fault detection method and system based on edge computing, which relates to the field of industrial Internet technology, including deploying edge computing nodes on the shield machine, aligning data collected by different sensors in time and space and fusing them at the feature level to obtain unified feature data; inputting the fused feature data into an online learning model pre-deployed in the edge computing node to obtain a fault detection result; the edge computing node uploads the fault detection result and fault-related feature data to a cloud server, generates a fault analysis report and feeds it back to the edge node and the user terminal; the edge node regularly pulls the updated online learning model and knowledge base from the cloud to realize cloud-edge collaborative model update and knowledge evolution.

Description

Technical Field

[0001] The present invention relates to industrial Internet technology, and in particular to a shield machine fault detection method and system based on edge computing. Background Art

[0002] The shield machine has a complex structure and a harsh working environment. Various mechanical failures and anomalies often occur during the tunneling process, which seriously affects the tunneling efficiency and construction progress. Traditional shield machine fault detection mainly relies on manual inspections and regular maintenance, which has problems such as untimely detection and low accuracy. Edge computing, as an emerging distributed computing paradigm, has received widespread attention.

[0003] However, the existing edge computing still has the following shortcomings when applied to shield machine fault detection:

[0004] There is a lack of consideration for the special working conditions of the shield machine, and no targeted extraction and fusion of multi-source heterogeneous features related to faults;

[0005] The machine learning model using static offline training lacks the ability to update and optimize the model online according to the changes in the shield machine working conditions;

[0006] There is no knowledge base framework for cloud-edge collaboration, which fails to effectively integrate the incremental knowledge generated by local diagnosis;

[0007] Model and knowledge base updates rely on massive historical data, and their robustness and generalization need to be improved. Summary of the invention

[0008] The embodiments of the present invention provide a shield machine fault detection method and system based on edge computing, which can solve the problems in the prior art.

[0009] According to a first aspect of the embodiments of the present invention,

[0010] Provides a shield machine fault detection method based on edge computing, including:

[0011] An edge computing node is deployed on the shield machine. The edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication. The shield machine operation status data collected by the multi-source heterogeneous sensors is received in real time through the data acquisition module. The shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine. The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through the multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level to obtain unified feature data.

[0012] The fused feature data is input into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal;

[0013] The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. At the same time, it uses the knowledge evolution mechanism to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge node regularly pulls the updated online learning model and knowledge base from the cloud to achieve cloud-edge collaborative model update and knowledge evolution.

[0014] In an optional embodiment,

[0015] The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level. The steps to obtain unified feature data include:

[0016] The edge computing node sets a threshold and uses the Raida criterion to identify and remove outliers and outliers in the raw data collected from multi-source sensors, performs preliminary data cleaning, and then uses the isolation forest or local anomaly factor algorithm to detect anomalies in the cleaned data to improve data quality and reliability.

[0017] The edge computing node uses signal processing technology to adaptively reduce noise and smooth the cleaned data, dynamically adjusts filter parameters according to the statistical characteristics of the data, and suppresses high-frequency noise and random disturbances in the data. The signal processing technology includes Kalman filtering, moving average filtering or wavelet transform.

[0018] The edge computing node maps the data collected by different sensors to a unified time scale through linear interpolation to achieve time alignment of the data. It also uses the spatial mapping method, combined with the optimization algorithm and the error minimization criterion, to convert the sensor data with spatial position differences into a unified spatial framework to achieve spatial alignment of the data.

[0019] The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to a shared latent space, learn the intrinsic associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes a combination of multi-core learning and deep neural networks.

[0020] The edge computing node performs maximum-minimum normalization on the fused low-dimensional feature data to eliminate the dimensional differences of different sensor data, improve the comparability and robustness of the low-dimensional feature data, and use maximum norm regularization to constrain the sparsity and overfitting risks of the low-dimensional feature data, ultimately obtaining unified feature data.

[0021] In an optional embodiment,

[0022] The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to the shared latent space, learn the inherent associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes the steps of combining multi-core learning with deep neural networks:

[0023] The extracted temporal features and spatial features are respectively input into two independent convolutional neural networks for feature extraction. The feature learning ability of deep neural networks is used to extract high-level semantic representations through convolution and pooling operations.

[0024] A multi-core mapping layer containing multiple kernel functions is introduced to combine multi-core learning with deep neural networks. Multi-core learning is used to learn semantic alignment between modalities from different scales and angles. By optimizing the multi-core maximum average difference target, high-level semantic representations of different modalities are adaptively fused to improve the flexibility and adaptability of feature fusion.

[0025] A modality discriminator is introduced after the multi-core mapping layer. The idea of ​​deep adversarial network is used to explicitly constrain the distribution consistency of fusion features in different modalities through adversarial learning, making the fusion features more balanced and generalized, and improving the quality of feature fusion.

[0026] The fused features obtained by multi-core mapping are input into the fully connected network for feature reconstruction. The nonlinear fitting ability of the deep neural network is used to minimize the reconstruction error to ensure that the fused features contain the information of the original multimodal features, while achieving feature dimensionality reduction and optimization.

[0027] Through an end-to-end network structure, multi-core mapping, adversarial discrimination and feature reconstruction are jointly optimized, the advantages of multi-core learning and deep neural networks are brought into play, the semantic associations of different modal features are adaptively learned, and consistent, compact and robust multimodal fusion feature representations are extracted.

[0028] In an optional embodiment,

[0029] The fused feature data is input into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner. The steps of obtaining the fault detection result include:

[0030] In the adaptive feature extraction layer, the gating unit and feature transformation unit are introduced to calculate the gating signal according to the current feature and the hidden state at the previous moment, and the input feature is adaptively transformed to adapt to the changes in feature distribution under different working conditions;

[0031] In the attention fusion layer, the self-attention mechanism is used to generate the attention weight matrix by calculating the correlation between features, and multiply it with the value matrix to obtain the weighted fused feature matrix to highlight the importance of different features;

[0032] In the bidirectional LSTM network layer, the fused feature sequence is encoded through the forward and backward LSTM networks to capture the long-term and short-term dependencies of the feature sequence.

[0033] In the fully connected output layer, the output of the bidirectional long short-term memory network layer is mapped to the probability space of the fault mode through the fully connected layer to obtain the final diagnosis result;

[0034] The online learning model adopts an active learning strategy, dynamically selects samples for annotation and learning according to the uncertainty of prediction, measures the uncertainty of the model for the current sample by calculating the prediction entropy, selects the sample with the highest entropy value to add to the training set, and updates the model parameters of the online learning model;

[0035] The online learning model also adopts an adaptive architecture evolution strategy to dynamically adjust the number of layers and neurons of the model according to the complexity of the diagnostic task, so as to adaptively balance the diagnostic performance and computational efficiency and achieve long-term online diagnosis.

[0036] In an optional embodiment,

[0037] The edge computing node uploads the fault detection results and fault-related feature data to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and user terminal. The steps include:

[0038] According to the fault detection results uploaded by the edge nodes, the entities and relationships related to the current fault mode are retrieved in the knowledge graph to form a local fault subgraph;

[0039] Performing a hybrid search on the local fault subgraph, wherein the hybrid search adopts an adaptive search strategy and dynamically selects a depth-first search or a breadth-first search according to the characteristics of the current node and historical information;

[0040] When performing a hybrid search, a heuristic function is defined to evaluate the importance and priority of the current node. When the value of the heuristic function is higher than a preset threshold, a depth-first search is used; when the value of the heuristic function is lower than the preset threshold, a breadth-first search is used.

[0041] When performing a hybrid search, set a depth threshold. When the search depth exceeds the depth threshold, switch from depth-first search to breadth-first search to avoid falling into a search branch that is too deep.

[0042] The knowledge reasoning engine is used to perform semantic understanding and association analysis on the local fault subgraph obtained by hybrid search, and the cause, severity and predictive maintenance measures of the fault are inferred by matching predefined reasoning rules.

[0043] A fault analysis report is generated based on the inference result, wherein the fault analysis report presents the inferred fault cause, severity, predictive maintenance measures and impact assessment information in a structured manner; and the fault analysis report is fed back to the edge node and the user terminal.

[0044] In an optional embodiment,

[0045] The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. The steps include:

[0046] The cloud server receives the fault detection results and fault-related feature data uploaded by the edge node, and pre-processes the received feature data, including denoising, anomaly detection, and normalization;

[0047] The cloud server extracts a batch of new fault samples from the fault database as the incremental training data set, and uses the time decay strategy to assign a smaller weight coefficient to the earlier fault samples to balance the ratio of new and old samples;

[0048] The cloud server sends the incremental training data set to the edge node. After receiving the data set, the edge node starts to perform incremental learning on the online learning model, and uses the new data to fine-tune and update the online learning model, adapting to the new fault mode and feature distribution while retaining the original online learning model knowledge.

[0049] When performing incremental learning, the edge node loads the structure, parameters, and hyperparameters of the original online learning model, freezes the underlying feature extraction layer of the original online learning model, and only fine-tunes the top classification layer or regression layer. The incremental training dataset is used to train the unfrozen layers and update their parameters.

[0050] During incremental learning, edge nodes control the learning rate and regularization, use the gradient descent algorithm, balance the learning and retention of new and old knowledge, evaluate the performance of the updated online learning model on the incremental training data set and the original test data set, and save and replace the original online learning model after meeting the preset performance indicators.

[0051] In an optional embodiment,

[0052] At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph to expand and improve the knowledge base. The edge nodes regularly pull the updated online learning model and knowledge base from the cloud. The steps to achieve cloud-edge collaborative model update and knowledge evolution include:

[0053] The edge node uses the online learning model updated by incremental learning to analyze and identify real-time data, obtain fault detection results and feature data, and store the new fault modes, new feature indicators and new diagnostic rules learned during the diagnosis process in a structured form in the local knowledge base;

[0054] The edge node regularly uploads new knowledge in the local knowledge base to the cloud server. The cloud server updates and expands the cloud knowledge graph by adding fault type nodes, feature attribute nodes, and rule nodes to the knowledge graph, establishing association edges between nodes, and modifying or supplementing the attribute values ​​of existing nodes and the weights and direction parameters of edges according to the type and content of the new knowledge received.

[0055] The cloud server uses the updated cloud knowledge graph to build a teacher model, learn the low-dimensional embedding representation of nodes and edges in the knowledge graph, and then train the student model based on the knowledge representation of the teacher model, so that the student model simulates the prediction behavior of the teacher model and compresses the dimension and complexity of the knowledge representation. The difference in prediction probability or regression value between the student model and the teacher model is measured by the distillation loss function, and the student model parameters are optimized by back propagation to obtain a knowledge representation model that is a compressed version of the knowledge graph.

[0056] The cloud server will regularly send the updated and optimized online learning model and knowledge base to the edge nodes. After receiving the new version of the model and knowledge base, the edge nodes will replace the old local version and selectively deploy the compressed version of the knowledge representation model or the full version of the knowledge base according to the local computing resources and performance requirements.

[0057] The cloud server will regularly send the updated and optimized online learning model and knowledge representation model to the edge nodes. The edge nodes will replace the local old version and selectively deploy the knowledge representation model or the complete knowledge base according to the local computing resources and performance requirements.

[0058] The edge nodes use the updated online learning model and the deployed knowledge representation model or the complete knowledge base to perform real-time fault diagnosis and analysis. The online learning model uses the structured knowledge contained in the knowledge representation model as input features or constraints, combined with real-time data for fault diagnosis and prediction. At the same time, the new diagnostic results and knowledge are fed back to the cloud server. The cloud server aggregates and analyzes the feedback data of multiple edge nodes to further optimize the knowledge graph, knowledge representation model and online learning model, forming a closed-loop learning process that is collaborative and continuously evolving between the cloud and the edge.

[0059] According to a second aspect of the embodiments of the present invention,

[0060] Provides a shield machine fault detection system based on edge computing, including:

[0061] The first unit is used to deploy an edge computing node on the shield machine. The edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication. The shield machine operation status data collected by the multi-source heterogeneous sensors is received in real time through the data acquisition module. The shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine. The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through the multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level to obtain unified feature data.

[0062] The second unit is used to input the fused feature data into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the cause of the fault, the severity and the predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal;

[0063] The third unit is used for cloud servers to use fault detection results and fault-related feature data to update and optimize the online learning model of edge nodes through incremental learning algorithms. At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned in the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge nodes regularly pull the updated online learning models and knowledge base from the cloud to achieve cloud-edge collaborative model updates and knowledge evolution.

[0064] According to a third aspect of the embodiments of the present invention,

[0065] An electronic device is provided, comprising:

[0066] processor;

[0067] a memory for storing processor-executable instructions;

[0068] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0069] According to a fourth aspect of the embodiments of the present invention,

[0070] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0071] In summary, the online fault diagnosis method in this paper adopts strategies such as adaptive feature extraction, attention fusion, bidirectional long short-term memory network, active learning and adaptive architecture evolution to build an efficient, flexible and intelligent online learning model, which can adaptively extract and fuse features according to the working conditions of the shield machine, capture the dynamic characteristics of the fault mode, and continuously optimize and upgrade the model through active learning and architecture evolution to achieve long-term, real-time and accurate fault diagnosis. This method provides a new idea and paradigm for the intelligent operation and maintenance of complex engineering systems and has broad application prospects.

[0072] The main technical details of the fault diagnosis method for shield machines based on knowledge graph and reasoning proposed in this paper. This method combines the experience and knowledge of experts in the shield field with the data analysis capabilities of machine learning. It can quickly discover and locate faults in massive equipment monitoring data, explore the deep-seated causes of faults, and provide practical maintenance countermeasures, providing new ideas and methods for realizing intelligent operation and maintenance of shield machines.

[0073] This paper uses cloud servers to optimize the updated model and knowledge base, and sends it to edge nodes again, over and over again, forming a continuous learning and evolution process of the model-knowledge-data closed loop. Through this cloud-edge collaborative and data-driven learning paradigm, the fault diagnosis system can continuously accumulate and optimize industrial knowledge, continuously improve the real-time, accuracy and intelligence of fault diagnosis, and provide strong support for the health management of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of a method for detecting faults of a shield machine based on edge computing according to an embodiment of the present invention;

[0075] Figure 2 It is a structural schematic diagram of a shield machine fault detection system based on edge computing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0078] Figure 1 FIG. 1 is a flow chart of a method for detecting faults of a shield machine based on edge computing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0079] S1. Deploy an edge computing node on the shield machine, wherein the edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication, and receives the shield machine operation status data collected by the multi-source heterogeneous sensors in real time through the data acquisition module, wherein the shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine; the edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data, and performs spatiotemporal alignment and feature-level fusion of the data collected by different sensors through the multi-source heterogeneous sensor data fusion technology to obtain unified feature data;

[0080] S2. Input the fused feature data into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal;

[0081] S3. The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. At the same time, it uses the knowledge evolution mechanism to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge node regularly pulls the updated online learning model and knowledge base from the cloud to achieve cloud-edge collaborative model update and knowledge evolution.

[0082] In an optional embodiment,

[0083] The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level. The steps to obtain unified feature data include:

[0084] The edge computing node sets a threshold and uses the Raida criterion to identify and remove outliers and outliers in the raw data collected from multi-source sensors, performs preliminary data cleaning, and then uses the isolation forest or local anomaly factor algorithm to detect anomalies in the cleaned data to improve data quality and reliability.

[0085] The edge computing node uses signal processing technology to adaptively reduce noise and smooth the cleaned data, dynamically adjusts filter parameters according to the statistical characteristics of the data, and suppresses high-frequency noise and random disturbances in the data. The signal processing technology includes Kalman filtering, moving average filtering or wavelet transform.

[0086] The edge computing node maps the data collected by different sensors to a unified time scale through linear interpolation to achieve time alignment of the data. It also uses the spatial mapping method, combined with the optimization algorithm and the error minimization criterion, to convert the sensor data with spatial position differences into a unified spatial framework to achieve spatial alignment of the data.

[0087] The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to a shared latent space, learn the intrinsic associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes a combination of multi-core learning and deep neural networks.

[0088] The edge computing node performs maximum-minimum normalization on the fused low-dimensional feature data to eliminate the dimensional differences of different sensor data, improve the comparability and robustness of the low-dimensional feature data, and use maximum norm regularization to constrain the sparsity and overfitting risks of the low-dimensional feature data, ultimately obtaining unified feature data.

[0089] Exemplarily, the edge computing node first performs preliminary cleaning on the raw data collected from the shield machine's multi-source sensors, identifies and removes outliers and outliers. By setting a reasonable threshold, the Raida criterion (3-σ criterion) is used for outlier detection. Specifically, assuming that the data collected by a sensor at time t is x_t, the mean of the sensor's historical data is μ, and the standard deviation is σ, then when |x_t - μ| > 3σ, x_t can be determined to be an outlier. For example, if the mean of the historical data of the shield machine's cutter head speed sensor is 2.5 r / min and the standard deviation is 0.4 r / min, when the speed data collected at a certain moment is 4.1 r / min, since the absolute value of the difference from the mean, 1.6 r / min, is greater than 3 times the standard deviation (i.e., 1.2 r / min), the data can be determined to be an outlier and removed.

[0090] After completing the initial cleaning, the edge computing node uses the Isolation Forest or Local Outlier Factor algorithm to perform further anomaly detection on the cleaned data to improve data quality and reliability. The Isolation Forest algorithm randomly selects features and recursively builds a tree structure to divide the data. Abnormal samples are usually isolated separately in the earlier division process, so they have a shorter average path length. Assume that a batch of data after initial cleaning contains 100 samples, each sample contains 10 feature dimensions, and 50 isolation trees are constructed with a maximum tree depth of 8. If the average path length of a sample on these 50 trees is 3.2, which is significantly shorter than other samples (such as the average path length is 6.5), then the sample is very likely to be an abnormal sample.

[0091] For the cleaned sensor data, the edge computing node uses signal processing technology to perform adaptive noise reduction and smoothing. According to the statistical characteristics of the data, such as mean, variance, kurtosis, etc., the filter parameters are dynamically adjusted to suppress high-frequency noise and random disturbances in the data. Common signal processing techniques include Kalman filtering, moving average filtering or wavelet transform. Taking Kalman filtering as an example, assuming that the data collected by the shield machine propulsion speed sensor contains Gaussian white noise, the mean of the noise is 0, and the variance is 0.01 m 2 / min 2 By establishing the state space model of propulsion speed, setting the state transfer matrix and observation matrix, and setting the process noise covariance matrix and measurement noise covariance matrix according to the statistical characteristics of noise, iteratively updating the state prediction and state estimation, finally obtaining the smoothed propulsion speed data. If the mean value of the original data in a certain period is 1.2 m / min and the variance is 0.15 m / min, the propulsion speed data can be obtained. 2 / min 2 After Kalman filtering, the mean propulsion speed in this period becomes 1.18 m / min, and the variance is reduced to 0.05 m 2 / min 2 , effectively suppressing noise and random disturbances.

[0092] Due to the differences in sampling frequency and spatial position of different sensors, edge computing nodes need to perform spatiotemporal alignment of multi-source sensor data. In terms of time alignment, the data collected by different sensors are mapped to a unified time scale through linear interpolation. For example, the shield machine cutter head speed sensor samples once every 0.1 seconds, and the propulsion speed sensor samples once every 1 second. By performing linear interpolation on the propulsion speed data, it can be mapped to the same 0.1 second sampling interval as the cutter head speed data. If the cutter head speed is 2.4 r / min at a certain moment, and the propulsion speeds collected in two adjacent 1-second intervals are 1.2 m / min and 1.3 m / min respectively, the propulsion speed corresponding to that moment can be obtained by linear interpolation as 1.24 m / min.

[0093] In terms of spatial alignment, the edge computing node uses the spatial mapping method, combined with the optimization algorithm and the error minimization criterion, to convert the sensor data with spatial position differences into a unified spatial framework. For example, the shield machine has four propulsion cylinder displacement sensors, which are located at the top, bottom, left and right of the tunneling head. By establishing the geometric structure model of the shield machine, taking the center line of the shield machine as the reference, the coordinate transformation matrix is ​​solved by the least squares method, and the displacement data of the four sensors are uniformly mapped to the longitudinal propulsion direction of the shield machine. If the displacement readings of the four sensors at a certain moment are (10.2, 9.8, 10.5, 10.1) mm, the longitudinal displacement of the shield machine can be obtained by multiplying the coordinate transformation matrix [0.707, -0.707, 0, 0; 0.707, 0.707,0, 0; 0, 0, 1, 0; 0, 0, 0, 1] with the displacement reading vector, and the longitudinal displacement of the shield machine is 14.2 mm.

[0094] For aligned time series sensor data, the edge computing node uses sliding window technology to extract statistical features in the time domain, frequency domain, and time-frequency domain. For example, for the shield machine cutter head speed data, the sliding window length is set to 10 seconds and the step length is 5 seconds. The mean, standard deviation, peak-to-peak value, root mean square and other time domain features are extracted in each window. The frequency domain features such as spectral entropy and center frequency are extracted through Fourier transform. The time-frequency domain features such as energy distribution are extracted by short-time Fourier transform or wavelet transform. If the mean of the cutter head speed data in a sliding window is 2.6 r / min, the standard deviation is 0.3 r / min, the peak-to-peak value is 1.5 r / min, and the center frequency is 0.2 Hz, then these features can be used as the feature representation of the cutter head speed data in the time window.

[0095] For spatially distributed sensor data, edge computing nodes extract geometric features and topological features. For example, for the displacement data of shield machine propulsion cylinders, the mean and variance of the cylinder displacement can be calculated to characterize the overall propulsion state, the maximum and minimum values ​​of the displacement difference between cylinders can be extracted to characterize the degree of propulsion deviation, and the adjacency matrix of the relative position of the cylinders can be constructed to characterize the cylinder distribution topology. If the mean of the displacement data of the four cylinders mapped to the longitudinal direction at a certain moment is 15 mm, the variance is 4 mm^2, and the maximum displacement difference between two adjacent cylinders is 1.2 mm and the minimum is 0.4 mm, then these features can be used as the feature representation of the cylinder displacement data at that moment.

[0096] After extracting the multimodal features of different sensor data, the edge computing node uses multimodal feature fusion technology to map these features to a shared latent space, learn the intrinsic correlation between features, and obtain unified low-dimensional feature data. Commonly used multimodal feature fusion technologies include the combination of multi-core learning and deep neural networks. Taking multi-core learning as an example, assuming that the extracted cutter head speed feature dimension is 20, the propulsion speed feature dimension is 15, and the propulsion cylinder displacement feature dimension is 25, by constructing multiple kernel functions (such as Gaussian kernel, polynomial kernel) to calculate the similarity matrix between different features, and then use the weighted summation method to obtain the fused similarity matrix. Finally, a dimensionality reduction method such as kernel principal component analysis is used to map the fused features to a shared latent space of 10 dimensions to achieve feature fusion and dimensionality reduction.

[0097] In order to eliminate the dimensional differences of different sensor data features and improve the comparability and robustness of the fused features, the edge computing node performs maximum-minimum normalization on the low-dimensional feature data. For the i-th dimension feature x_i, its maximum-minimum normalization formula is: (x_i - min(x_i)) / (max(x_i) - min(x_i)), where max(x_i) and min(x_i) are the maximum and minimum values ​​of the feature of this dimension, respectively. The normalized feature value range is [0, 1]. For example, if the maximum value of a feature dimension after fusion is 10.5 and the minimum value is 2.4, then the feature value of 5.7 becomes (5.7-2.4) / (10.5-2.4)=0.41 after maximum-minimum normalization.

[0098] In addition, in order to constrain the sparsity and overfitting risk of low-dimensional feature data, the edge computing node uses maximum norm regularization to constrain the feature data. Maximum norm regularization tends to learn sparse feature representations and reduce the risk of overfitting. If the fused low-dimensional feature data is represented as a matrix W, the maximum norm regularization term can be expressed as λ||W||_∞, where λ is the regularization coefficient and ||W||_∞ represents the maximum norm of the matrix W, that is, the maximum absolute value of the elements in the matrix. When optimizing the objective function, the maximum norm regularization term is added, and the elements of the matrix W are updated and constrained by methods such as gradient descent, so that it maintains sparsity while meeting the optimization goal.

[0099] Through the above steps, the edge computing node can preprocess, extract features and fuse the multi-source heterogeneous data collected from the shield machine, and finally obtain unified, low-dimensional, robust and interpretable feature data, providing high-quality input for subsequent fault diagnosis and prediction.

[0100] In an optional embodiment,

[0101] The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to the shared latent space, learn the inherent associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes the steps of combining multi-core learning with deep neural networks:

[0102] The extracted temporal features and spatial features are respectively input into two independent convolutional neural networks for feature extraction. The feature learning ability of deep neural networks is used to extract high-level semantic representations through convolution and pooling operations.

[0103] A multi-core mapping layer containing multiple kernel functions is introduced to combine multi-core learning with deep neural networks. Multi-core learning is used to learn semantic alignment between modalities from different scales and angles. By optimizing the multi-core maximum average difference target, high-level semantic representations of different modalities are adaptively fused to improve the flexibility and adaptability of feature fusion.

[0104] A modality discriminator is introduced after the multi-core mapping layer. The idea of ​​deep adversarial network is used to explicitly constrain the distribution consistency of fusion features in different modalities through adversarial learning, making the fusion features more balanced and generalized, and improving the quality of feature fusion.

[0105] The fused features obtained by multi-core mapping are input into the fully connected network for feature reconstruction. The nonlinear fitting ability of the deep neural network is used to minimize the reconstruction error to ensure that the fused features contain the information of the original multimodal features, while achieving feature dimensionality reduction and optimization.

[0106] Through an end-to-end network structure, multi-core mapping, adversarial discrimination and feature reconstruction are jointly optimized, the advantages of multi-core learning and deep neural networks are brought into play, the semantic associations of different modal features are adaptively learned, and consistent, compact and robust multimodal fusion feature representations are extracted.

[0107] For example, the temporal features and spatial features extracted from the shield machine sensor data are input into two independent convolutional neural networks respectively, and the feature learning ability of the deep neural network is used to extract high-level semantic representations through convolution and pooling operations. For the temporal features, a one-dimensional convolutional neural network (1D-CNN) is used for processing. Assuming that the dimension of the extracted temporal features is 20 and the sequence length is 100, it is reshaped into a 20×100 two-dimensional matrix and input into the 1D-CNN. The network structure includes 3 one-dimensional convolutional layers and 2 maximum pooling layers, with convolution kernel sizes of 3, 4, and 5, respectively, and the number of convolution kernels is 32, 64, and 128, respectively. The pooling window size is 2, and the activation function uses the rectified linear unit (ReLU). After convolution and pooling operations, the temporal features are encoded into a 128-dimensional high-level semantic representation.

[0108] For spatial features, a two-dimensional convolutional neural network (2D-CNN) is used for processing. Assuming that the extracted spatial features are 25×25 two-dimensional matrices, they are directly input into the 2D-CNN. The network structure includes 2 two-dimensional convolutional layers and 2 maximum pooling layers, with convolution kernel sizes of 3×3 and 4×4, respectively, the number of convolution kernels of 16 and 32, respectively, the pooling window size of 2×2, and the activation function of ReLU. After convolution and pooling operations, the spatial features are encoded into a 32-dimensional high-level semantic representation.

[0109] After the independent convolutional neural network, a multi-core mapping layer containing multiple kernel functions is introduced to combine multi-core learning with deep neural networks, and multi-core learning is used to learn semantic alignment between modalities from different scales and angles. The multi-core mapping layer contains three kernel functions, namely linear kernel, Gaussian kernel and polynomial kernel. For the linear kernel, its form is k(x,y)=x^Ty, which directly calculates the inner product of two feature vectors to measure their linear similarity, where k represents the kernel function, x and y represent two input feature vectors, and T represents the transpose operation of the matrix or vector. For the Gaussian kernel, its form is k(x,y)=exp(-||xy||^2 / (2σ^2)), where σ is the Gaussian kernel width parameter, and it captures the nonlinear similarity between feature vectors by mapping to infinite dimensional space. For the polynomial kernel, its form is k(x,y)=(x^T y+c)^d, where c is the bias term and d is the polynomial degree, and it learns the high-order interaction information between feature vectors by mapping to high-dimensional space.

[0110] The multi-core mapping layer concatenates the 128-dimensional representation of temporal features and the 32-dimensional representation of spatial features into a 160-dimensional multimodal feature vector, and then maps them to high-dimensional space through three kernel functions to obtain three similarity matrices of different scales. Let the similarity matrix after the i-th kernel function is mapped as K_i, then the output of the multi-core mapping layer is the weighted sum Σ(wi×K_i), where w_i is the weight coefficient of the i-th kernel function, which is adaptively learned through back propagation. In order to make the similarity matrices of different kernel functions consistent in scale, the kernel normalization operation is introduced to divide each similarity matrix by its nuclear norm to ensure that the similarity values ​​of different kernel functions are in the same order of magnitude. Through weighted summation, the semantic alignment information between modalities learned at different scales and angles is adaptively fused to improve the flexibility and adaptability of feature fusion.

[0111] The modality discriminator is introduced after the multi-core mapping layer. The idea of ​​deep adversarial network is used to explicitly constrain the distribution consistency of fusion features in different modalities through adversarial learning. The modality discriminator consists of 3 layers of fully connected networks. The first layer takes the fusion features output by the multi-core mapping layer as input, and the dimension is the output dimension of the multi-core mapping layer. The second layer contains 64 neurons, and the third layer contains 1 neuron. The Sigmoid activation function is used to output the probability of the modality label. The goal of the modality discriminator is to minimize the modality classification loss of the fusion features, even if the fusion features are difficult to distinguish in different modalities.

[0112] During the training process, adversarial learning is achieved through gradient reversal. First, the parameters of the multi-core mapping layer are fixed, and the modal discriminator is trained to minimize the classification loss so that it can accurately distinguish which modality the fused features come from. Then the parameters of the modal discriminator are fixed, and the multi-core mapping layer is trained to maximize the classification loss, making the fused features difficult to distinguish in different modalities. Through alternating training, the multi-core mapping layer and the modal discriminator are continuously optimized in the adversary, and finally the fused features with consistent distribution in different modalities are obtained. For example, when the dimension of the fused feature is 128, the classification accuracy of the modal discriminator drops from more than 90% at the beginning to nearly 50%, indicating that the distribution of the fused features in different modalities gradually tends to be consistent, and the quality of feature fusion is improved.

[0113] The fusion features obtained by multi-core mapping are input into the fully connected network for feature reconstruction. The nonlinear fitting ability of the deep neural network is used to minimize the reconstruction error to ensure that the fusion features contain the information of the original multimodal features, while achieving feature dimensionality reduction and optimization. The fully connected network contains 3 hidden layers and 1 output layer. The number of neurons in each layer is 128, 64, 32, and 10 respectively. The activation function uses ReLU, and the activation function of the output layer is a linear function. The input of the network is the 128-dimensional fusion feature output by the multi-core mapping layer, and the output is the reconstructed 10-dimensional feature vector.

[0114] During the training process, the fully connected network optimizes the network parameters by minimizing the mean square error loss function between the fused features and the reconstructed features. For example, if the fused features are [0.85, 1.24, ..., 0.36, -0.47] and the reconstructed features are [0.92, 1.15, ..., 0.29, -0.52], the mean square error loss is ((0.85-0.92)^2 + (1.24-1.15)^2+ ... + (0.36-0.29)^2 + (-0.47+0.52)^2) / 128. The loss function is minimized by back propagation and gradient descent, and the weights and biases of the fully connected network are updated to make the reconstructed features as close to the fused features as possible. At the same time, the dimensionality reduction effect of the fully connected network makes the final 10-dimensional feature vector contain the main information of the original multimodal features, achieving feature optimization.

[0115] By constructing an end-to-end network structure including independent convolutional neural networks, multi-core mapping layers, modal discriminators and fully connected networks, jointly optimizing multi-core mapping, adversarial discrimination and feature reconstruction, and leveraging the advantages of multi-core learning and deep neural networks, the semantic associations of different modal features are adaptively learned, and a consistent, compact and robust multimodal fusion feature representation is extracted.

[0116] In the end-to-end training process, the adaptive moment estimation algorithm is used, the learning rate is set to 0.001, the batch size is set to 64, and the number of training rounds is set to 200. In each round of training, the output of the independent convolutional neural network, the multi-core mapping layer, the modal discriminator and the fully connected network is first calculated by forward propagation, and then the weighted sum of the kernel function of the multi-core mapping layer, the classification loss of the modal discriminator, and the reconstruction loss of the fully connected network are calculated. Then the gradient is back-propagated to update the parameters of each network module. By alternately optimizing the multi-core mapping layer and the modal discriminator, adversarial learning is achieved and the distribution consistency of the fusion features is improved. After multiple rounds of iterations, the network converges and the quality of the fusion features is evaluated on the test set. For example, if the fusion features are used for the shield machine fault diagnosis task, if the diagnostic accuracy is increased from 85% of the single modal features to 92% of the fusion features, it means that the fusion features can explore the complementarity of different modal information and have better discrimination ability.

[0117] In summary, the multimodal feature fusion method adaptively learns the semantic association of temporal features and spatial features through the collaborative optimization of independent convolutional neural networks, multi-core mapping layers, modal discriminators and fully connected networks, constrains the distribution consistency of fused features, and extracts compact and robust fused feature representations through end-to-end training, providing high-quality feature input for subsequent shield machine fault diagnosis tasks.

[0118] In an optional embodiment,

[0119] The fused feature data is input into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner. The steps of obtaining the fault detection result include:

[0120] In the adaptive feature extraction layer, the gating unit and feature transformation unit are introduced to calculate the gating signal according to the current feature and the hidden state at the previous moment, and the input feature is adaptively transformed to adapt to the changes in feature distribution under different working conditions;

[0121] In the attention fusion layer, the self-attention mechanism is used to generate the attention weight matrix by calculating the correlation between features, and multiply it with the value matrix to obtain the weighted fused feature matrix to highlight the importance of different features;

[0122] In the bidirectional LSTM network layer, the fused feature sequence is encoded through the forward and backward LSTM networks to capture the long-term and short-term dependencies of the feature sequence.

[0123] In the fully connected output layer, the output of the bidirectional long short-term memory network layer is mapped to the probability space of the fault mode through the fully connected layer to obtain the final diagnosis result;

[0124] The online learning model adopts an active learning strategy, dynamically selects samples for annotation and learning according to the uncertainty of the prediction, measures the uncertainty of the model for the current sample by calculating the prediction entropy, selects the sample with the highest entropy value to add to the training set, and updates the model parameters of the online learning model;

[0125] The online learning model also adopts an adaptive architecture evolution strategy to dynamically adjust the number of layers and neurons of the model according to the complexity of the diagnostic task, so as to adaptively balance the diagnostic performance and computational efficiency and achieve long-term online diagnosis.

[0126] Exemplarily, in the adaptive feature extraction layer, a gating unit and a feature transformation unit are introduced to adaptively adjust the feature representation according to the current feature and the hidden state at the previous moment to adapt to the feature distribution changes under different working conditions of the shield machine. The gating unit includes an update gate and a reset gate. The update gate controls the influence of the hidden state at the previous moment on the current feature, and the reset gate controls the degree of reset of the hidden state by the current feature. The gating signal is calculated by the concatenation vector of the current feature and the hidden state at the previous moment through the fully connected layer and the Sigmoid activation function, and the value range is [0, 1].

[0127] The feature transformation unit performs affine transformation on the input features, maps the original features to the new feature space through linear transformation and translation operations, and enhances the representation ability of the features. The transformed features are point-multiplied with the gated signal to obtain the adaptively adjusted feature representation.

[0128] For example, assume that the feature vector at the current moment is x_t, with a dimension of 50, and the hidden state vector at the previous moment is h_(t-1), with a dimension of 64. First, concatenate x_t and h_(t-1) into a 114-dimensional vector, and then calculate the 128-dimensional gate signal vector g_t through a 114×128 fully connected layer and the Sigmoid function. Next, linearly transform x_t through a 50×64 fully connected layer, and add a 64-dimensional bias vector to obtain the transformed feature vector x_t'. Finally, perform a dot multiplication of the gate signal g_t and the transformed feature x_t' to obtain the adaptively adjusted feature vector x_t''.

[0129] Through the gating mechanism and feature transformation, the adaptive feature extraction layer can dynamically adjust the feature representation according to the changes in the working conditions of the shield machine, highlight the features related to the current working conditions, suppress noise and redundant information, and improve the discriminability and robustness of the features.

[0130] In the attention fusion layer, the self-attention mechanism is used to generate the attention weight matrix by calculating the correlation between features, and multiply it with the value matrix to obtain the weighted fusion feature matrix, highlighting the importance of different features. The self-attention mechanism includes three matrices: query, key, and value, which are obtained by linear transformation of the input feature matrix. The product of the query matrix and the key matrix is ​​scaled and normalized by the normalized exponential function (Softmax) to obtain the attention weight matrix, which represents the correlation between different features. The attention weight matrix is ​​multiplied by the value matrix to obtain the weighted fusion feature matrix.

[0131] For example, assume that the dimension of the input feature matrix X is n×d, where n is the number of features and d is the dimension of the feature. First, three d×d parameter matrices W_q, W_k, and W_v are used to linearly transform X to obtain the query matrix Q, key matrix K, and value matrix V. Then, the product of the transpose of Q and K is calculated, and then divided by the square root of d for scaling to obtain the n×n correlation matrix S. Then, the Softmax function is applied to S for normalization to obtain the attention weight matrix A. Finally, A is multiplied by V to obtain the weighted fusion feature matrix X'.

[0132] Assuming n=10, d=32, the dimensions of Q, K, and V are all 10×32. The dimension of the correlation matrix S is 10×10, which represents the correlation between the 10 features. After Softmax normalization, each row of the attention weight matrix A represents the correlation weight distribution between the corresponding feature and other features, and the sum of the weights is 1. Multiply A and V to obtain the weighted fusion feature matrix X' with a dimension of 10×32. Each feature vector is a weighted combination of the original features, and the weight is adaptively calculated by the attention mechanism.

[0133] Through the self-attention mechanism, the attention fusion layer can adaptively calculate the correlation between features, highlight important features, suppress redundant features, and improve the effect of feature fusion. At the same time, the self-attention mechanism has the advantage of being parallelizable and can efficiently process large-scale feature data.

[0134] In the bidirectional LSTM network layer, the fused feature sequence is encoded through the forward and backward LSTM networks to capture the long-term and short-term dependencies of the feature sequence. The LSTM network consists of four parts: input gate, forget gate, output gate, and cell state. Through the gating mechanism and memory unit, it can effectively solve the gradient vanishing and gradient exploding problems and model long-distance dependencies.

[0135] In the forward LSTM network, the feature sequence is input in chronological order, and the network updates the hidden state and cell state according to the input at the current moment and the hidden state at the previous moment. In the backward LSTM network, the feature sequence is input in reverse chronological order, and the network updates the hidden state and cell state according to the input at the current moment and the hidden state at the next moment. Through bidirectional processing, the forward and backward dependencies of the feature sequence can be captured simultaneously, enhancing the context perception ability of feature representation.

[0136] For example, assume that the length of the fused feature sequence is T and the feature dimension is d. For the forward long short-term memory network, assume that the input at the tth moment is x_t, the hidden state is h_t, and the cell state is c_t, then the forward propagation process of the network is as follows:

[0137] Input gate: i_t = Sigmoid(W_i[x_t, h_(t-1)] + b_i);

[0138] Forget gate: f_t = Sigmoid(W_f[x_t, h_(t-1)] + b_f);

[0139] Cell state: c_t = f_t * c_(t-1) + i_t * tanh(W_c[x_t, h_(t-1)] + b_c);

[0140] Output gate: o_t = Sigmoid(W_o[x_t, h_(t-1)] + b_o);

[0141] Hidden state: h_t = o_t * tanh(c_t);

[0142] Where Wi, Wf, Wc, Wo are the weight matrices of the network, and bi, bf, bc, bo are the bias vectors.

[0143] For the backward LSTM network, let the input at the tth moment be x_t, the hidden state be h'_t, and the cell state be c'_t. The network propagation process is similar to the forward process, except that the subscripts of the hidden state and the cell state are changed to t+1.

[0144] Finally, the hidden state sequence h_1, h_2, ..., h_T of the forward network and the hidden state sequence h'_1, h'_2, ..., h'_T of the backward network are concatenated to obtain the bidirectionally encoded feature sequence. Assuming that the hidden state dimension is h, the dimension of the concatenated feature sequence is T×2h, and the feature vector at each moment contains the forward and backward context information.

[0145] Through the bidirectional long short-term memory network layer, the temporal dependency of feature sequences can be fully mined, the dynamic evolution of shield machine failure modes can be captured, and a richer and more comprehensive feature representation can be provided for fault diagnosis.

[0146] In the fully connected output layer, the output of the bidirectional long short-term memory network layer is mapped to the probability space of the fault mode through the fully connected layer to obtain the final diagnosis result. The fully connected layer converts the feature representation into the posterior probability of the fault mode through linear transformation and nonlinear activation function. By setting a suitable decision threshold, the fault state of the shield machine can be judged according to the probability value.

[0147] For example, assume that the output dimension of the bidirectional long short-term memory network layer is 2h, and the number of fault mode categories is C. First, a 2h×C fully connected layer is used to map the features to the C-dimensional output space, and then the posterior probability of each fault mode is calculated by the Softmax function. Let the feature vector of the i-th sample be x_i, the weight matrix of the fully connected layer be W, and the bias vector be b, then the calculation process of the output probability vector y_i is as follows:

[0148] y_i = Softmax(W * x_i + b);

[0149] Where y_i is a C-dimensional probability vector, indicating the probability that the sample belongs to each fault mode. The Softmax function normalizes the output value to a probability value between [0, 1] and ensures that the sum of the probabilities is 1.

[0150] Assuming that the shield machine has four states: normal, cutterhead wear, support shoe damage, and segment disengagement, then C=4. If the true label of the training sample x_i is [0, 1, 0, 0], indicating that it belongs to the cutterhead wear fault, then the ideal output probability vector should be [0, 1, 0, 0]. By minimizing the cross entropy loss function, the parameters of the fully connected layer can be trained to make the predicted probability as close to the true probability distribution as possible.

[0151] In the inference stage, a probability threshold can be set, such as 0.5. When the predicted probability of a certain fault mode exceeds the threshold, the shield machine is judged to be in the fault state, otherwise it is considered to be in a normal state. The threshold can also be adjusted according to actual needs to balance the accuracy of diagnosis and the false alarm rate.

[0152] Through the fully connected output layer, the high-level features extracted by the bidirectional long short-term memory network layer are mapped to the probability space of the fault mode to achieve end-to-end fault diagnosis. The fully connected layer has a strong nonlinear fitting ability and can adaptively learn the complex mapping relationship between features and fault modes, thus improving the accuracy and reliability of diagnosis.

[0153] The online learning model adopts an active learning strategy, dynamically selecting samples for annotation and learning according to the uncertainty of prediction, so as to improve the generalization ability and learning efficiency of the model. The active learning strategy measures the uncertainty of the model for the current sample by calculating the prediction entropy, selects the sample with the highest entropy value to add to the training set, and updates the parameters of the online learning model. The prediction entropy indicates the uncertainty of the model for the sample category. The higher the entropy value, the more uncertain the model's prediction of the sample, the greater the amount of information contained, and the more obvious the optimization effect on the model.

[0154] Through active learning strategies, the online learning model can adaptively select the most valuable samples for learning, avoid labeling a large number of redundant or invalid samples, and improve learning efficiency. At the same time, by continuously learning new samples, the online learning model can continuously optimize and expand knowledge, adapt to changes in shield machine working conditions, and achieve long-term online diagnosis.

[0155] For example, suppose there are 100 samples in the initial training set, and 10 samples are selected for training each round. In the first round of active learning, 1,000 unlabeled samples are predicted, the prediction entropy is calculated, and the 10 samples with the highest entropy values ​​are selected and labeled before being added to the training set, resulting in a new training set of 110 samples. The model parameters are updated using the new training set, and the next round of active learning is performed. This is repeated for 5 rounds, and finally a training set of 150 samples is obtained. The diagnostic accuracy of the model is increased from the initial 85% to 93%, and the generalization ability and robustness are significantly improved.

[0156] The online learning model also uses an adaptive architecture evolution strategy to dynamically adjust the number of layers and neurons of the model according to the complexity of the diagnostic task, so as to adaptively balance the diagnostic performance and computational efficiency and achieve long-term online diagnosis. The adaptive architecture evolution strategy sets an initial model architecture, such as 3 hidden layers with 64 neurons in each layer, and then dynamically increases or decreases the number of hidden layers and neurons according to the complexity of the diagnostic task and the performance of the model, optimizing the model architecture to adapt to the characteristics of the data and the requirements of the diagnosis.

[0157] During the model training process, regularly evaluate the performance of the model on the validation set, such as accuracy, recall, and other indicators. If the performance improvement of the model slows down or a bottleneck occurs, consider increasing the capacity of the model, such as adding a hidden layer or the number of neurons in each layer. If the performance of the model meets the requirements, but the inference speed is slow or the computing resources are more occupied, consider reducing the capacity of the model, such as reducing a hidden layer or the number of neurons in each layer. At the same time, it is also necessary to consider the impact of increasing or decreasing capacity on model performance to avoid underfitting or overfitting.

[0158] For example, suppose the initial model architecture is an input layer, three hidden layers, and an output layer, and each hidden layer contains 64 neurons. After 5 rounds of training, the accuracy of the model on the validation set is 90%, but the inference time is 50ms, which exceeds the real-time requirement. Therefore, we try to reduce the model capacity, reduce the number of neurons in the third hidden layer from 64 to 32, and retrain the model. The accuracy is slightly reduced to 88%, but the inference time is shortened to 30ms, meeting the real-time requirement.

[0159] Then, we continued to train the model and found that the accuracy rate increased slowly, reaching a maximum of about 91%. In order to further improve the performance, we tried to increase the model capacity by adding a new hidden layer after the second hidden layer, containing 128 neurons, and appropriately increasing the number of neurons in other layers to form a new model architecture. After training, the accuracy of the new model increased to 94%, and the inference time was 45ms, which was within an acceptable range.

[0160] Through the adaptive architecture evolution strategy, the online learning model can dynamically adjust its own structure according to the needs of the diagnostic task, and seek the best balance between diagnostic performance and computational efficiency. At the same time, the model can adapt to changes in data distribution, continuously optimize and upgrade, and maintain long-term stable and reliable diagnostic capabilities.

[0161] In summary, the online fault diagnosis method in this paper adopts strategies such as adaptive feature extraction, attention fusion, bidirectional long short-term memory network, active learning and adaptive architecture evolution to build an efficient, flexible and intelligent online learning model, which can adaptively extract and fuse features according to the working conditions of the shield machine, capture the dynamic characteristics of the fault mode, and continuously optimize and upgrade the model through active learning and architecture evolution to achieve long-term, real-time and accurate fault diagnosis. This method provides a new idea and paradigm for the intelligent operation and maintenance of complex engineering systems and has broad application prospects.

[0162] In an optional embodiment,

[0163] The edge computing node uploads the fault detection results and fault-related feature data to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and user terminal. The steps include:

[0164] According to the fault detection results uploaded by the edge nodes, the entities and relationships related to the current fault mode are retrieved in the knowledge graph to form a local fault subgraph;

[0165] Performing a hybrid search on the local fault subgraph, wherein the hybrid search adopts an adaptive search strategy and dynamically selects a depth-first search or a breadth-first search according to the characteristics of the current node and historical information;

[0166] When performing a hybrid search, a heuristic function is defined to evaluate the importance and priority of the current node. When the value of the heuristic function is higher than a preset threshold, a depth-first search is used; when the value of the heuristic function is lower than the preset threshold, a breadth-first search is used.

[0167] When performing a hybrid search, set a depth threshold. When the search depth exceeds the depth threshold, switch from depth-first search to breadth-first search to avoid falling into a search branch that is too deep.

[0168] The knowledge reasoning engine is used to perform semantic understanding and association analysis on the local fault subgraph obtained by hybrid search, and the cause, severity and predictive maintenance measures of the fault are inferred by matching predefined reasoning rules.

[0169] A fault analysis report is generated based on the inference result, wherein the fault analysis report presents the inferred fault cause, severity, predictive maintenance measures and impact assessment information in a structured manner; and the fault analysis report is fed back to the edge node and the user terminal.

[0170] For example, when the edge node of the shield machine detects a fault and uploads relevant data to the cloud, the cloud server first needs to retrieve the entities and relationships related to the current fault mode in the knowledge graph to form a local fault subgraph. The shield machine knowledge graph contains a large amount of information related to shield construction, such as concepts, components, fault modes, and treatment methods, as well as the associations between them.

[0171] For example, suppose the edge node detects that the screw conveyor of the shield machine has a "motor current is too high" fault and uploads the fault code "F013" and related sensor data. After receiving the data, the cloud uses "F013" as the search entry in the knowledge graph to find the entities directly connected to it, including the fault name "screw conveyor motor current is too high", possible causes "motor bearing damage", "screw blade blockage", etc., affected components "screw conveyor", "screw motor", "reducer", etc., as well as related historical cases, maintenance manuals, etc. These entities and their relationships constitute a local subgraph around "F013".

[0172] After obtaining the local fault subgraph, it is necessary to further search and expand the relevant knowledge in order to mine more information that is helpful for shield machine fault diagnosis. This paper proposes an adaptive hybrid search strategy that dynamically selects depth-first search (DFS) or breadth-first search (BFS) according to the characteristics of the current node and historical information.

[0173] DFS is suitable for drilling down into a specific fault concept or relationship chain. For example, by drilling down along the "screw conveyor motor bearing damage", you can find more specific fault modes such as "bearing ball damage" and "bearing cage deformation".

[0174] BFS is suitable for divergent thinking and covers more fault-related fields. For example, starting from "screw conveyor current is too high", we can not only find the relevant causes, but also find the impact of the fault on tunneling efficiency and related preventive maintenance measures.

[0175] In order to dynamically choose between DFS and BFS, this paper defines a heuristic function f to evaluate the importance and priority of the current node:

[0176] f = w1 × r + w2 × n + w3 × h;

[0177] Among them, r represents the correlation between the current node and the root node (i.e., the initial fault code), n represents the in-degree of the current node, and h represents the level of the current node. w1, w2, and w3 are three weight coefficients.

[0178] When the value of f exceeds a preset threshold (such as 0.7), DFS is preferred; when the value of f is lower, BFS is selected. In addition, a depth threshold (such as 4) is set. When the depth of DFS exceeds this threshold, it automatically switches to BFS.

[0179] Taking "screw conveyor motor bearing damage" as an example, assuming that the shortest path length between it and the root node "F013" is 2, the in-degree is 5, the level is 3, and the weight coefficients are 0.6, 0.3, and 0.1 respectively, then its f value is:

[0180] f = 0.6 × 2 + 0.3 × 5 + 0.1 × 3 = 3.0;

[0181] The threshold of 0.7 is exceeded, so DFS is selected to continue searching downward for nodes such as "Bearing Ball Damage" until the search depth exceeds 4 layers, then switch to BFS.

[0182] Through the hybrid search strategy, a more comprehensive local fault subgraph can be obtained. Next, the knowledge reasoning engine is used to perform semantic understanding and association analysis on the subgraph to infer the cause, severity and predictive maintenance measures of the fault.

[0183] This paper adopts a rule-based reasoning method and predefines a series of reasoning rules for common faults of shield machines, such as:

[0184] Rule1: IF fault code = F013 AND screw conveyor motor winding temperature > 130℃ THEN fault cause = "screw conveyor motor bearing damage" AND fault level = 3 AND maintenance measures = "replace screw conveyor motor bearing";

[0185] Rule2: IF fault code = F013 AND spiral blade clearance < 10mm THEN fault cause = "spiral blade obstruction" AND fault level = 2 AND maintenance measures = "clean up foreign matter between spiral blades and check the wear of spiral blades";

[0186] The inference engine will check whether the premise of each rule is satisfied in the fault subgraph one by one, and if so, trigger the corresponding conclusion. When multiple rules are triggered at the same time, the priority is determined according to the fault level and other attributes, and the final inference result is selected.

[0187] The diagnostic results such as the cause of the fault, severity, and maintenance measures output by the inference engine need to be fed back to the edge nodes and operators of the shield machine in a way that is easy to understand and execute. This paper proposes to generate a structured fault analysis report, which mainly includes:

[0188] Basic fault information: TBM number, fault time, location, fault code, fault description, etc.

[0189] Fault cause analysis: List the possible causes diagnosed by the inference engine and their credibility, such as:

[0190] The screw conveyor motor bearing is damaged (credibility 0.9);

[0191] The spiral blade is blocked (confidence level 0.7);

[0192] Fault severity assessment: According to the reasoning results and the pre-established shield machine fault classification standards, the severity is given, such as:

[0193] Severity: Level 3 (more serious, requiring shutdown for maintenance);

[0194] Predictive maintenance measures: For each possible cause of failure, corresponding treatment measures and maintenance suggestions are given, such as:

[0195] Replacement of screw conveyor motor bearings:

[0196] Lift the motor and remove the coupling;

[0197] Remove the damaged bearing and check the bearing seat;

[0198] Install new bearings and adjust the motor position;

[0199] Cleaning the spiral blades:

[0200] Cut off the power supply of the screw conveyor;

[0201] Open the inspection door and clean the foreign matter between the spiral blades;

[0202] Check the blade wear and adjust the gap if necessary;

[0203] Required tools: high pressure water gun, jack, measuring tool;

[0204] Impact assessment: Analyze and predict the impact of faults on shield tunneling parameters, efficiency, project progress, and the costs of downtime and maintenance, such as:

[0205] The excavation speed is reduced by about 30%;

[0206] Daily footage loss: about 5m / day;

[0207] Maintenance time: about 8 hours;

[0208] Economic loss: about 30,000 yuan;

[0209] After the report is generated, it is sent to the edge node of the shield machine in HTML or PDF format, and pushed to relevant technicians and managers through the Web or App. They can quickly and accurately carry out targeted inspection and maintenance work based on the information provided in the report to minimize the adverse effects of the failure.

[0210] The main technical details of the fault diagnosis method for shield machines based on knowledge graph and reasoning proposed in this paper. This method combines the experience and knowledge of experts in the shield field with the data analysis capabilities of machine learning. It can quickly discover and locate faults in massive equipment monitoring data, explore the deep-seated causes of faults, and provide practical maintenance countermeasures, providing new ideas and methods for realizing intelligent operation and maintenance of shield machines.

[0211] In an optional embodiment,

[0212] The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. The steps include:

[0213] The cloud server receives the fault detection results and fault-related feature data uploaded by the edge node, and pre-processes the received feature data, including denoising, anomaly detection, and normalization;

[0214] The cloud server extracts a batch of new fault samples from the fault database as the incremental training data set, and uses the time decay strategy to assign a smaller weight coefficient to the earlier fault samples to balance the ratio of new and old samples;

[0215] The cloud server sends the incremental training data set to the edge node. After receiving the data set, the edge node starts to perform incremental learning on the online learning model, and uses the new data to fine-tune and update the online learning model, adapting to the new fault mode and feature distribution while retaining the original online learning model knowledge.

[0216] When performing incremental learning, the edge node loads the structure, parameters, and hyperparameters of the original online learning model, freezes the underlying feature extraction layer of the original online learning model, and only fine-tunes the top classification layer or regression layer. The incremental training dataset is used to train the unfrozen layers and update their parameters.

[0217] During incremental learning, edge nodes control the learning rate and regularization, use the gradient descent algorithm, balance the learning and retention of new and old knowledge, evaluate the performance of the updated online learning model on the incremental training data set and the original test data set, and save and replace the original online learning model after meeting the preset performance indicators.

[0218] Exemplarily, the cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm, which mainly includes the following steps:

[0219] After the cloud server receives the fault detection results and related feature data uploaded by the edge node, it first needs to preprocess these data to improve data quality and learning efficiency. Preprocessing mainly includes:

[0220] Denoising: Due to factors such as sensor accuracy and electromagnetic interference, the uploaded data may contain noise. Use median filtering, wavelet transform and other methods to remove random noise and impulse noise. For example, for vibration data collected by the sensor, if the vibration value at a certain moment is 10m / s², and the previous and next moments are around 1m / s², then this value is likely to be impulse noise and needs to be removed or corrected.

[0221] Anomaly detection: There may be outliers in the fault data, that is, data points that are significantly different from the data distribution under normal conditions. The 3-σ criterion, local outlier factor (LOF) and other methods are used to identify abnormal data. Taking the bearing temperature data as an example, the temperature under normal conditions is usually between 40℃ and 60℃. If a temperature value above 100℃ appears, it is considered to be abnormal data.

[0222] Normalization: The data collected by different sensors may have different dimensions and numerical ranges, so they need to be normalized to map the data to the interval of [0, 1] or [-1, 1] to eliminate the dimension effect. Common normalization methods include maximum and minimum value normalization and zero mean normalization. For example, the bearing vibration data is normalized from 0m / s²~10m / s² to the interval of [0,1], and the speed data is normalized from 0r / min~3000r / min to the interval of [-1, 1].

[0223] In order to make the online learning model of edge nodes adapt to new fault modes and data distribution, it is necessary to select a batch of new fault samples from the cloud fault database to build an incremental training data set. The following issues need to be paid attention to during the construction process:

[0224] Sample selection: Select fault samples that are relatively recent and relevant to the current edge node environment from the fault database. Taking the fault diagnosis of a wind turbine as an example, the fault data of the same model wind turbine or adjacent wind farm within the past year can be selected as incremental samples.

[0225] Time decay: Since the working conditions and failure modes of equipment may change over time, a smaller weight is assigned to earlier failure samples and a larger weight is assigned to the latest samples. The weight can be represented by an exponential decay function, such as w = e^(-λt), where w is the weight, t is the time interval between samples, and λ is the decay coefficient, which is usually between 0.01 and 0.1. For example, when λ=0.05, the weight of the sample one year ago is e^(-0.05×365)≈0.16, while the weight of the sample one month ago is e^(-0.05×30)≈0.78.

[0226] Balance new and old samples: The incremental training data set should contain a certain proportion of new and old samples to balance the learning of new and old knowledge. The proportion of new samples can be dynamically adjusted according to factors such as model performance and fault frequency, and is generally 20% to 50%. Taking bearing fault diagnosis as an example, if there are 10,000 samples in the original training set and 1,000 new samples are collected, an incremental training set containing 2,000 new samples and 8,000 old samples can be constructed.

[0227] After receiving the incremental training data set sent by the cloud, the edge node starts to perform incremental learning and optimization on the original online learning model. The specific steps are as follows:

[0228] Load the original model: Load the structure, parameters and hyperparameters of the saved online learning model. Taking the CNN model as an example, load the network structure of the convolution layer, pooling layer, fully connected layer, and the weight, bias, learning rate, activation function and other parameters of each layer.

[0229] Freeze feature extraction layer: In the incremental learning phase, freeze the underlying feature extraction layer of the original model (such as the convolution layer and pooling layer of CNN), and only fine-tune the parameters of the top classification layer or regression layer. This can adapt to new failure modes while retaining the original feature extraction capabilities. If all parameters are completely retrained, the original knowledge may be forgotten.

[0230] Fine-tune the classification layer: Use the incremental training data set to update the weights and biases of the classification layer or regression layer through forward propagation and back propagation. Take the Softmax classification layer as an example. Assuming that the original model can identify 10 fault modes, and 2 new fault modes appear in the incremental data set, the number of neurons in the Softmax layer is increased from 10 to 12, and the weights of the newly added neurons are randomly initialized, and training is performed using the incremental data set.

[0231] Optimize hyperparameters: When performing incremental learning, you need to set hyperparameters such as learning rate, regularization coefficient, batch size, etc. The learning rate is usually set to a small value, such as 0.001~0.01, to avoid causing too much damage to the original knowledge. Regularization helps prevent overfitting. L1 or L2 regularization can be used, with coefficients between 0.01~0.1. The batch size is selected based on the number of samples and memory limitations, and is generally 32~256.

[0232] Evaluate model performance: Evaluate the performance of the updated model on the incremental training data set and the original test set, and calculate indicators such as accuracy, recall, and F1 value. Taking fault diagnosis as an example, if after incremental learning, the accuracy of the model on the new fault category is improved to more than 90%, and the accuracy on the original fault category does not drop significantly (such as the drop is within 5%), then the incremental learning is considered effective.

[0233] Save the updated model: If the updated model meets the preset performance indicators, it will be saved and replaced with the original model for subsequent online fault diagnosis tasks. If the performance improvement is not obvious or there is a significant decline, it is necessary to further adjust the incremental training data, optimize the algorithm, etc. until a satisfactory model is obtained.

[0234] The above are the main technical details of the cloud server optimizing the edge node online learning model through incremental learning. This method can continuously learn new fault knowledge without interrupting the edge node diagnosis service, while retaining the original diagnosis ability, improving the accuracy, adaptability and real-time performance of fault diagnosis.

[0235] In an optional embodiment,

[0236] At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph to expand and improve the knowledge base. The edge nodes regularly pull the updated online learning model and knowledge base from the cloud. The steps to achieve cloud-edge collaborative model update and knowledge evolution include:

[0237] The edge node uses the online learning model updated by incremental learning to analyze and identify real-time data, obtain fault detection results and feature data, and store the new fault modes, new feature indicators and new diagnostic rules learned during the diagnosis process in a structured form in the local knowledge base;

[0238] The edge node regularly uploads new knowledge in the local knowledge base to the cloud server. The cloud server updates and expands the cloud knowledge graph by adding fault type nodes, feature attribute nodes, and rule nodes to the knowledge graph, establishing association edges between nodes, and modifying or supplementing the attribute values ​​of existing nodes and the weights and direction parameters of edges according to the type and content of the new knowledge received.

[0239] The cloud server uses the updated cloud knowledge graph to build a teacher model, learn the low-dimensional embedding representation of nodes and edges in the knowledge graph, and then train the student model based on the knowledge representation of the teacher model, so that the student model simulates the prediction behavior of the teacher model and compresses the dimension and complexity of the knowledge representation. The difference in prediction probability or regression value between the student model and the teacher model is measured by the distillation loss function, and the student model parameters are optimized by back propagation to obtain a knowledge representation model that is a compressed version of the knowledge graph.

[0240] The cloud server will regularly send the updated and optimized online learning model and knowledge base to the edge nodes. After receiving the new version of the model and knowledge base, the edge nodes will replace the old local version and selectively deploy the compressed version of the knowledge representation model or the full version of the knowledge base according to the local computing resources and performance requirements.

[0241] The cloud server will regularly send the updated and optimized online learning model and knowledge representation model to the edge nodes. The edge nodes will replace the local old version and selectively deploy the knowledge representation model or the complete knowledge base according to the local computing resources and performance requirements.

[0242] The edge nodes use the updated online learning model and the deployed knowledge representation model or the complete knowledge base to perform real-time fault diagnosis and analysis. The online learning model uses the structured knowledge contained in the knowledge representation model as input features or constraints, combined with real-time data for fault diagnosis and prediction. At the same time, the new diagnostic results and knowledge are fed back to the cloud server. The cloud server aggregates and analyzes the feedback data of multiple edge nodes to further optimize the knowledge graph, knowledge representation model and online learning model, forming a closed-loop learning process that is collaborative and continuously evolving between the cloud and the edge.

[0243] For example, the cloud-edge collaborative fault diagnosis system dynamically updates the new knowledge learned by the edge node to the cloud knowledge graph through the knowledge evolution mechanism, realizes the continuous expansion and optimization of the knowledge base, and synchronizes the updated online learning model and knowledge base to the edge node, forming a closed-loop knowledge evolution and model update process. The specific technical solution is as follows:

[0244] The edge node uses the online learning model after incremental learning to perform fault diagnosis on real-time data and obtain fault detection results and related feature data. During the diagnosis process, the model may discover some new fault modes, feature indicators or diagnostic rules. These new knowledge are stored in the local knowledge base of the edge node in a structured form. The structured representation of knowledge can be in JSON, XML and other formats, for example:

[0245] {

[0246] "fault_type": "Gear wear",

[0247] "features": [

[0248] {

[0249] "name": "Gearbox vibration peak value",

[0250] "value": 5.2,

[0251] "unit": "m / s²",

[0252] "threshold": 4.0

[0253] },

[0254] {

[0255] "name": "Gearbox Temperature",

[0256] "value": 78.5,

[0257] "unit": "℃",

[0258] "threshold": 70.0

[0259] }

[0260] ],

[0261] "rules": [

[0262] {

[0263] "antecedent": "Gearbox vibration peak value > 4.0 m / s² and gearbox temperature > 70.0℃",

[0264] "consequent": "Gear wear",

[0265] "confidence": 0.85

[0266] } ]

[0268] }

[0269] The above JSON data indicates that the edge node has diagnosed a new fault type "gear wear". The characteristics of this fault include the gearbox vibration peak exceeding 4.0 m / s² and the gearbox temperature exceeding 70.0 ℃. The confidence of the diagnosis rule is 0.85.

[0270] The edge node uploads the new knowledge in the local knowledge base to the cloud server regularly (such as every day or every week). The upload process adopts the incremental synchronization method, that is, only the modified or newly added knowledge is transmitted to reduce the amount of data transmission.

[0271] After receiving the new knowledge uploaded by the edge node, the cloud server integrates it into the cloud knowledge graph. The knowledge graph is stored in a graph database (such as Neo4j). The nodes represent the fault type, feature attributes, diagnosis rules, etc., and the edges represent the relationship between nodes, such as what features the fault type contains, the logical relationship between features and rules, etc.

[0272] Adding nodes: If new fault types, characteristic attributes or rules appear in the knowledge, corresponding nodes are created in the knowledge graph. For example, for the above fault knowledge of "gear wear", a new fault type node of "gear wear" is added to the graph.

[0273] Establishing associative edges: Establishing associative edges between related nodes based on the logical relationships in the new knowledge. For example, establishing inclusion edges between the "gear wear" node and the "gearbox vibration peak" and "gearbox temperature" feature nodes, and establishing triggering edges between the "gear wear" node and the diagnosis rule node.

[0274] Update attribute values: If the new knowledge contains changes in the attribute values ​​of existing nodes, update the attributes of the corresponding nodes. For example, if the "gearbox temperature" feature node already exists in the knowledge graph, but its threshold is updated from 65.0 ℃ to 70.0 ℃ in the new knowledge, modify the threshold attribute of the node.

[0275] Adjust edge weights: Adjust the weight values ​​of associated edges based on the importance of features or rules reflected by the new knowledge. For example, if the knowledge uploaded by multiple edge nodes indicates that "gearbox vibration peak" is an important feature of "gear wear", increase the weight of the associated edge between the two.

[0276] By continuously integrating new knowledge, the cloud-based knowledge graph can be continuously expanded and optimized to promptly reflect changes in equipment failure characteristics and diagnostic knowledge.

[0277] In order to facilitate the online learning model to utilize structured knowledge, the cloud server embeds the knowledge graph into a low-dimensional continuous space and learns the distributed representation of nodes and edges. This process is implemented through the knowledge representation learning algorithm and is mainly divided into the following steps:

[0278] Constructing the teacher model: The teacher model learns the low-dimensional embedding vector of each node and edge in the knowledge graph through random walks, node attribute encoding, etc. Taking the TransE algorithm as an example, it regards fault types, features, rules, etc. as entities and relations in the vector space, and learns the embedding vectors of entities and relations by minimizing the loss function of h + r ≈ t, where h, r, t represent the head entity, relation, and tail entity, respectively.

[0279] Distillation of student model: The knowledge representation dimensionality learned by the teacher model may be high, which is not conducive to the deployment and real-time reasoning of edge nodes. Therefore, a student model is further trained, and the knowledge representation of the teacher model is used as a supervisory signal to compress the dimension of the knowledge representation through distillation and other techniques. Specifically, the prediction probability or similarity score of the teacher model on the knowledge graph is used as a soft label to train the student model to imitate the behavior of the teacher model. The student model adopts a lighter network structure, such as an embedded lookup table. By minimizing the cross entropy loss or square loss of the prediction results of the student model and the teacher model, the student model is made to approach the performance of the teacher model, while compressing the number of parameters and the amount of calculation.

[0280] Knowledge base compression: To further reduce the storage and communication overhead of the knowledge base, the entity and relationship embedding vectors learned by the student model are compressed. Common compression methods include quantization, pruning, low-rank decomposition, etc. For example, using 8-bit quantization to convert floating-point embedding vectors to integers can reduce the storage space of each embedding vector by 4 times. For another example, using singular value decomposition to approximate the embedding matrix as the product of several low-rank matrices can greatly reduce the storage space of the embedding matrix while ensuring a certain reconstruction error.

[0281] After knowledge representation learning, the cloud knowledge graph is compressed into a lightweight knowledge representation model that not only contains the structural information of the graph but is also easy to deploy and reason about.

[0282] The cloud server regularly sends the updated and optimized online learning model, compressed knowledge base, and knowledge representation model to the edge node. The sending frequency can be dynamically adjusted according to factors such as the frequency of equipment failures and network bandwidth, usually once a day or once a week.

[0283] After receiving the new version of the model and knowledge base, the edge node replaces the old local version. According to the node's own computing resources and performance requirements, the full version of the knowledge base or the compressed version of the knowledge representation model can be selectively deployed. Nodes with sufficient resources can load the complete knowledge base and use the original knowledge graph for reasoning; nodes with limited resources can only deploy the knowledge representation model and use the embedded vector for fast matching and query.

[0284] The edge node uses the updated online learning model and knowledge base to perform real-time monitoring and fault diagnosis of the equipment. During the diagnosis process, the online learning model matches the real-time data collected by the equipment sensor with the fault mode and feature threshold in the knowledge base to determine the health status and fault type of the equipment.

[0285] Specifically, the online learning model calculates the similarity between the real-time data and the feature embedding vector learned by the knowledge representation model. If the feature vector of the data is highly similar to the feature vector of a certain fault mode, it can be determined that the equipment has experienced this type of fault. At the same time, the model can also use the diagnostic rules in the knowledge base to confirm and segment the fault. For example, if the online learning model determines that the equipment has a "gear wear" fault, further using the rules in the knowledge base that vibration_peak > 4.0 and temperature > 70.0, it can be confirmed that the fault is of a serious level and requires immediate shutdown for maintenance.

[0286] Online learning models can also use knowledge representation as input features and combine real-time data to train end-to-end fault diagnosis networks. Taking convolutional neural networks as an example, the multi-sensor data of the equipment is constructed into time series or images and sent to the input layer of CNN; at the same time, the fault types, features, rules, etc. learned by the knowledge representation model are embedded in vectors and mapped to another set of input features through a lookup table, which are fused with the sensor data features in the fully connected layer, and finally the fault category is output through the Softmax layer. This end-to-end learning paradigm can simultaneously model data features and structured knowledge, enhancing the generalization and robustness of online learning models.

[0287] After a period of time (such as one week or one month) after deploying the updated model and knowledge base, the edge node will feed back the diagnosis results and newly discovered fault knowledge to the cloud server. The feedback content includes but is not limited to:

[0288] Statistics of performance indicators such as fault detection rate, false alarm rate, and missed alarm rate of the model;

[0289] Evaluation results such as accuracy and recall of knowledge base matching;

[0290] Structured knowledge of newly discovered fault types, abnormal operating conditions, etc.;

[0291] Suggestions for revising the threshold and confidence of existing rules in the knowledge base;

[0292] Performance evaluation and improvement suggestions for the compressed knowledge representation model in resource-constrained nodes.

[0293] The cloud server aggregates the diagnostic feedback data from multiple edge nodes and further optimizes and updates the online learning model, knowledge graph, and knowledge representation model. The optimization measures include:

[0294] For cases of false positives and missed positives, the network structure and loss function of the online learning model are adjusted to improve the accuracy of fault detection;

[0295] For incorrect cases of knowledge matching, correct the rule threshold, confidence and other parameters in the knowledge graph to improve the reliability of knowledge reasoning;

[0296] Integrate the knowledge of newly reported fault types and features from different nodes to expand the scale and diversity of the cloud knowledge graph;

[0297] Based on the usage feedback of resource-constrained nodes, the number of parameters and calculation amount of the knowledge representation model are further compressed, reducing the model complexity while meeting performance requirements.

[0298] This paper uses cloud servers to optimize the updated model and knowledge base, and sends it to edge nodes again, over and over again, forming a continuous learning and evolution process of the model-knowledge-data closed loop. Through this cloud-edge collaborative and data-driven learning paradigm, the fault diagnosis system can continuously accumulate and optimize industrial knowledge, continuously improve the real-time, accuracy and intelligence of fault diagnosis, and provide strong support for the health management of industrial equipment.

[0299] Figure 2 FIG. 1 is a schematic diagram of the structure of a shield machine fault detection system based on edge computing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0300] The first unit is used to deploy an edge computing node on the shield machine. The edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication. The shield machine operation status data collected by the multi-source heterogeneous sensors is received in real time through the data acquisition module. The shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine. The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through the multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level to obtain unified feature data.

[0301] The second unit is used to input the fused feature data into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the cause of the fault, the severity and the predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal;

[0302] The third unit is used for cloud servers to use fault detection results and fault-related feature data to update and optimize the online learning model of edge nodes through incremental learning algorithms. At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned in the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge nodes regularly pull the updated online learning models and knowledge base from the cloud to achieve cloud-edge collaborative model updates and knowledge evolution.

[0303] According to a third aspect of the embodiments of the present invention,

[0304] An electronic device is provided, comprising:

[0305] processor;

[0306] a memory for storing processor-executable instructions;

[0307] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0308] According to a fourth aspect of the embodiments of the present invention,

[0309] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0310] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0311] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A shield machine fault detection method based on edge computing, characterized in that: include: An edge computing node is deployed on the shield machine. The edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication. The shield machine operation status data collected by the multi-source heterogeneous sensors is received in real time through the data acquisition module. The shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine. The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through the multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level to obtain unified feature data. The fused feature data is input into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal; The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. At the same time, it uses the knowledge evolution mechanism to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge node regularly pulls the updated online learning model and knowledge base from the cloud to achieve cloud-edge collaborative model update and knowledge evolution.

2. The method according to claim 1, characterized in that: The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level. The steps to obtain unified feature data include: The edge computing node sets a threshold and uses the Raida criterion to identify and remove outliers and outliers in the raw data collected from multi-source sensors, performs preliminary data cleaning, and then uses the isolation forest or local anomaly factor algorithm to detect anomalies in the cleaned data to improve data quality and reliability. The edge computing node uses signal processing technology to adaptively reduce noise and smooth the cleaned data, dynamically adjusts filter parameters according to the statistical characteristics of the data, and suppresses high-frequency noise and random disturbances in the data. The signal processing technology includes Kalman filtering, moving average filtering or wavelet transform. The edge computing node maps the data collected by different sensors to a unified time scale through linear interpolation to achieve time alignment of the data. It also uses the spatial mapping method, combined with the optimization algorithm and the error minimization criterion, to convert the sensor data with spatial position differences into a unified spatial framework to achieve spatial alignment of the data. The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to a shared latent space, learn the intrinsic associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes a combination of multi-core learning and deep neural networks. The edge computing node performs maximum-minimum normalization on the fused low-dimensional feature data to eliminate the dimensional differences of different sensor data, improve the comparability and robustness of the low-dimensional feature data, and use maximum norm regularization to constrain the sparsity and overfitting risks of the low-dimensional feature data, ultimately obtaining unified feature data.

3. The method according to claim 2, characterized in that The edge computing node extracts statistical features in the time domain, frequency domain, and time-frequency domain from the aligned time series sensor data using the sliding window technology, extracts geometric features and topological features from the spatial distribution sensor data, and then uses the multimodal feature fusion technology to map the extracted different modal features to the shared latent space, learn the inherent associations between the features, and obtain unified low-dimensional feature data. The multimodal feature fusion technology includes the steps of combining multi-core learning with deep neural networks: The extracted temporal features and spatial features are respectively input into two independent convolutional neural networks for feature extraction. The feature learning ability of deep neural networks is used to extract high-level semantic representations through convolution and pooling operations. A multi-core mapping layer containing multiple kernel functions is introduced to combine multi-core learning with deep neural networks. Multi-core learning is used to learn semantic alignment between modalities from different scales and angles. By optimizing the multi-core maximum average difference target, high-level semantic representations of different modalities are adaptively fused to improve the flexibility and adaptability of feature fusion. A modality discriminator is introduced after the multi-core mapping layer. The idea of ​​deep adversarial network is used to explicitly constrain the distribution consistency of fusion features in different modalities through adversarial learning, making the fusion features more balanced and generalized, and improving the quality of feature fusion. The fused features obtained by multi-core mapping are input into the fully connected network for feature reconstruction. The nonlinear fitting ability of the deep neural network is used to minimize the reconstruction error to ensure that the fused features contain the information of the original multimodal features, while achieving feature dimensionality reduction and optimization. Through an end-to-end network structure, multi-core mapping, adversarial discrimination and feature reconstruction are jointly optimized, the advantages of multi-core learning and deep neural networks are brought into play, the semantic associations of different modal features are adaptively learned, and consistent, compact and robust multimodal fusion feature representations are extracted.

4. The method according to claim 1, characterized in that: The fused feature data is input into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner. The steps of obtaining the fault detection result include: In the adaptive feature extraction layer, the gating unit and feature transformation unit are introduced to calculate the gating signal according to the current feature and the hidden state at the previous moment, and the input feature is adaptively transformed to adapt to the changes in feature distribution under different working conditions; In the attention fusion layer, the self-attention mechanism is used to generate the attention weight matrix by calculating the correlation between features, and multiply it with the value matrix to obtain the weighted fused feature matrix to highlight the importance of different features; In the bidirectional LSTM network layer, the fused feature sequence is encoded through the forward and backward LSTM networks to capture the long-term and short-term dependencies of the feature sequence. In the fully connected output layer, the output of the bidirectional long short-term memory network layer is mapped to the probability space of the fault mode through the fully connected layer to obtain the final diagnosis result; The online learning model adopts an active learning strategy, dynamically selects samples for annotation and learning according to the uncertainty of prediction, measures the uncertainty of the model for the current sample by calculating the prediction entropy, selects the sample with the highest entropy value to add to the training set, and updates the model parameters of the online learning model; The online learning model also adopts an adaptive architecture evolution strategy to dynamically adjust the number of layers and neurons of the model according to the complexity of the diagnostic task, so as to adaptively balance the diagnostic performance and computational efficiency and achieve long-term online diagnosis.

5. The method according to claim 1, characterized in that The edge computing node uploads the fault detection results and fault-related feature data to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on knowledge graph technology, outputs the fault cause, severity and predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and user terminal. The steps include: According to the fault detection results uploaded by the edge nodes, the entities and relationships related to the current fault mode are retrieved in the knowledge graph to form a local fault subgraph; Performing a hybrid search on the local fault subgraph, wherein the hybrid search adopts an adaptive search strategy and dynamically selects a depth-first search or a breadth-first search according to the characteristics of the current node and historical information; When performing a hybrid search, a heuristic function is defined to evaluate the importance and priority of the current node. When the value of the heuristic function is higher than a preset threshold, a depth-first search is used; when the value of the heuristic function is lower than the preset threshold, a breadth-first search is used. When performing a hybrid search, set a depth threshold. When the search depth exceeds the depth threshold, switch from depth-first search to breadth-first search to avoid falling into a search branch that is too deep. The knowledge reasoning engine is used to perform semantic understanding and association analysis on the local fault subgraph obtained by hybrid search, and the cause, severity and predictive maintenance measures of the fault are inferred by matching predefined reasoning rules. A fault analysis report is generated based on the inference result, wherein the fault analysis report presents the inferred fault cause, severity, predictive maintenance measures and impact assessment information in a structured manner; and the fault analysis report is fed back to the edge node and the user terminal.

6. The method according to claim 1, characterized in that The cloud server uses the fault detection results and fault-related feature data to update and optimize the online learning model of the edge node through an incremental learning algorithm. The steps include: The cloud server receives the fault detection results and fault-related feature data uploaded by the edge node, and pre-processes the received feature data, including denoising, anomaly detection, and normalization; The cloud server extracts a batch of new fault samples from the fault database as the incremental training data set, and uses the time decay strategy to assign a smaller weight coefficient to the earlier fault samples to balance the ratio of new and old samples; The cloud server sends the incremental training data set to the edge node. After receiving the data set, the edge node starts to perform incremental learning on the online learning model, and uses the new data to fine-tune and update the online learning model, adapting to the new fault mode and feature distribution while retaining the original online learning model knowledge. When performing incremental learning, the edge node loads the structure, parameters, and hyperparameters of the original online learning model, freezes the underlying feature extraction layer of the original online learning model, and only fine-tunes the top classification layer or regression layer. The incremental training dataset is used to train the unfrozen layers and update their parameters. During incremental learning, the edge node controls the learning rate and regularization, uses the gradient descent algorithm, balances the learning and retention of new and old knowledge, evaluates the performance of the updated online learning model on the incremental training data set and the original test data set, and saves and replaces the original online learning model after meeting the preset performance indicators.

7. The method according to claim 1, characterized in that At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned during the fault diagnosis process to the knowledge graph to expand and improve the knowledge base. The edge nodes regularly pull the updated online learning model and knowledge base from the cloud. The steps to achieve cloud-edge collaborative model update and knowledge evolution include: The edge node uses the online learning model updated by incremental learning to analyze and identify real-time data, obtain fault detection results and feature data, and store the new fault modes, new feature indicators and new diagnostic rules learned during the diagnosis process in a structured form in the local knowledge base; The edge node regularly uploads new knowledge in the local knowledge base to the cloud server. The cloud server updates and expands the cloud knowledge graph by adding fault type nodes, feature attribute nodes, and rule nodes to the knowledge graph, establishing association edges between nodes, and modifying or supplementing the attribute values ​​of existing nodes and the weights and direction parameters of edges according to the type and content of the new knowledge received. The cloud server uses the updated cloud knowledge graph to build a teacher model, learn the low-dimensional embedding representation of nodes and edges in the knowledge graph, and then train the student model based on the knowledge representation of the teacher model, so that the student model simulates the prediction behavior of the teacher model and compresses the dimension and complexity of the knowledge representation. The difference in prediction probability or regression value between the student model and the teacher model is measured by the distillation loss function, and the student model parameters are optimized by back propagation to obtain a knowledge representation model that is a compressed version of the knowledge graph. The cloud server will regularly send the updated and optimized online learning model and knowledge base to the edge nodes. After receiving the new version of the model and knowledge base, the edge nodes will replace the old local version and selectively deploy the compressed version of the knowledge representation model or the full version of the knowledge base according to the local computing resources and performance requirements. The cloud server will regularly send the updated and optimized online learning model and knowledge representation model to the edge nodes. The edge nodes will replace the local old version and selectively deploy the knowledge representation model or the complete knowledge base according to the local computing resources and performance requirements. The edge nodes use the updated online learning model and the deployed knowledge representation model or the complete knowledge base to perform real-time fault diagnosis and analysis. The online learning model uses the structured knowledge contained in the knowledge representation model as input features or constraints, combined with real-time data for fault diagnosis and prediction. At the same time, the new diagnostic results and knowledge are fed back to the cloud server. The cloud server aggregates and analyzes the feedback data of multiple edge nodes to further optimize the knowledge graph, knowledge representation model and online learning model, forming a closed-loop learning process that is collaborative and continuously evolving between the cloud and the edge.

8. A shield machine fault detection system based on edge computing, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy an edge computing node on the shield machine. The edge computing node is connected to the multi-source heterogeneous sensors on the shield machine for communication. The shield machine operation status data collected by the multi-source heterogeneous sensors is received in real time through the data acquisition module. The shield machine operation status data includes the propulsion speed, propulsion pressure, cutter head speed, cutter head torque, slag bin pressure, slag discharge volume, grouting pressure and grouting volume of the shield machine. The edge computing node pre-processes the received shield machine operation status data to eliminate noise and outliers in the data. Through the multi-source heterogeneous sensor data fusion technology, the data collected by different sensors are aligned in time and space and fused at the feature level to obtain unified feature data. The second unit is used to input the fused feature data into the online learning model pre-deployed in the edge computing node. The online learning model adopts an active learning strategy to dynamically adjust the parameters and structure of the online learning model according to the working condition changes of the shield machine. The online learning model includes an adaptive feature extraction layer, an attention fusion layer, a bidirectional long short-term memory network layer and a fully connected output layer. The fault mode of the shield machine is detected and identified in real time in an end-to-end manner to obtain the fault detection result; the edge computing node uploads the fault detection result and the feature data related to the fault to the cloud server. The cloud server performs semantic understanding and association analysis on the received data based on the knowledge graph technology, outputs the cause of the fault, the severity and the predictive maintenance measures through the knowledge reasoning engine, and generates a fault analysis report to feed back to the edge node and the user terminal; The third unit is used for cloud servers to use fault detection results and fault-related feature data to update and optimize the online learning model of edge nodes through incremental learning algorithms. At the same time, the knowledge evolution mechanism is used to dynamically update the new knowledge learned in the fault diagnosis process to the knowledge graph, expand and improve the knowledge base, and the edge nodes regularly pull the updated online learning models and knowledge base from the cloud to achieve cloud-edge collaborative model updates and knowledge evolution.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images