A machine learning based engineering machinery failure prediction and prevention method

Abstract

The application discloses an engineering machinery fault prediction and prevention method based on machine learning, comprising the following steps: collecting multi-source operation data of equipment, and constructing a structured multi-dimensional time sequence; constructing a heterogeneous graph structure based on the connection relationship, control path and energy transmission path of equipment components; extracting node time sequence dynamic characteristics by using an improved expansion time convolution network, and generating node embedding representation by weighting through a channel attention mechanism; inputting the embedding representation into an edge attention enhanced heterogeneous graph neural network, and calculating a fault risk score by fusing edge type and time bias information; combining historical fault cases to construct a retrieval index mechanism, generating a candidate set according to similarity matching, and outputting a structured prevention suggestion. The application can realize higher-accuracy fault risk prediction under complex structural conditions, significantly improve early warning response speed and processing suggestion generation efficiency, and effectively enhance the operation reliability and maintenance intelligent level of the engineering machinery system.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction and maintenance technology for construction machinery, and in particular to a method for predicting and preventing construction machinery faults based on machine learning. Background Technology

[0002] As core equipment in infrastructure construction and resource extraction, the continuous and stable operation of construction machinery has a significant impact on construction efficiency and operational safety. With increasing operating time, key components such as hydraulic systems, power systems, and control systems are susceptible to the combined effects of impact loads, thermal stress, vibration fatigue, and other factors, gradually leading to performance degradation and potential failure risks. Failure to promptly identify early signs of failure may result in sudden shutdowns, equipment damage, or even safety accidents, causing severe economic losses and safety hazards.

[0003] Currently, fault prediction for construction machinery equipment in the industry largely relies on traditional methods such as empirical rules, manual judgment, or single-variable anomaly alarms. These methods suffer from slow response, high false alarm rates, and dependence on human experience, making it difficult to accurately capture latent trends before faults occur. Furthermore, although some equipment is equipped with sensors to collect operating parameter data, the relevant analysis systems are mostly limited to threshold detection based on single-channel data or shallow modeling. They struggle to integrate the dynamic correlation characteristics of multi-dimensional data such as pressure, temperature, vibration, and current, and are unable to characterize the evolution of equipment status under complex operating conditions.

[0004] On the other hand, current commonly used methods generally ignore the impact of topological connections and signal transmission paths between different components in the equipment structure on fault propagation, and lack the ability to model the correlation between sensor data and the physical structure of the equipment. This results in poor performance of prediction models in multi-source heterogeneous data scenarios, making it difficult to achieve high-precision equipment condition assessment and early warning inference. Fault detection results and historical experience knowledge are also not effectively linked, which cannot help to formulate targeted prevention strategies, thus limiting the feasibility and engineering application value of the prediction results.

[0005] Therefore, how to provide a machine learning-based method for predicting and preventing engineering machinery failures is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a machine learning-based method for predicting and preventing faults in engineering machinery. This invention fully integrates an improved dilated temporal convolutional network and an edge-attention-enhanced heterogeneous graph neural network to construct a comprehensive intelligent prediction system encompassing multi-source operational data acquisition and preprocessing, topology-driven heterogeneous graph construction, temporal embedding representation learning, edge-attention aggregation and temporal bias gating propagation, historical case vector matching, and strategy generation. This achieves accurate identification of potential fault risks in key components of engineering machinery under complex operating conditions and generates preventative control suggestions. This invention possesses advantages such as strong multi-source data modeling capabilities, accurate structural topology representation, multi-scale temporal feature perception, high depth of fault mode mining, and highly practical strategy output. It can significantly improve the foresight and accuracy of equipment fault prediction, promoting a shift from fault response to predictive prevention, thereby effectively addressing the problems of lagging identification, coarse structural modeling, and lack of preventative mechanisms in existing methods.

[0007] A method for predicting and preventing engineering machinery faults based on machine learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source operation data of engineering machinery and equipment to generate structured multi-dimensional time series data;

[0009] S2. Based on structured multidimensional time series data and the component connection relationships, control signal paths and energy transmission paths of the equipment, construct a heterogeneous graph;

[0010] S3. For the time series information attached to each node in the heterogeneous graph, an improved dilated temporal convolutional network is used to extract dynamic features, and the feature channels are weighted through a channel attention mechanism to output the time series embedding representation of the node.

[0011] S4. Input the time series embedding representation of the node into the edge attention-enhanced heterogeneous graph neural network, introduce the time series embedding representation of the node and the edge type embedding, use the time bias gating aggregation method when fusing neighbor information, and output the fault risk score of each key node.

[0012] S5. Structure and encode existing fault events, build a historical case library and retrieval vector and matching index mechanism, and perform similarity matching between fault risk features and case features in the historical case library based on fault risk scores to generate a candidate set.

[0013] S6. Calculate the weighted average of the matching score and effect score of historical cases with the current failure risk characteristics, select the top three highest-scoring processing strategies, and generate structured prevention suggestions.

[0014] Optionally, S1 specifically includes: collecting various raw sensor data of engineering machinery and equipment during operation, aligning the collected data according to a unified timestamp, removing missing values ​​and outliers, and using an interpolation algorithm to complete the missing segments, unifying the dimensions of the collected data before normalization according to physical unit standards, performing standardization and normalization operations, and obtaining structured multidimensional time series data.

[0015] Optionally, S2 specifically includes:

[0016] S21. Read the structured multidimensional time series data, associate each piece of structured multidimensional time series data with the corresponding sensor number, physical installation location, and connection component information, and establish a correspondence table between sensor data and equipment components;

[0017] S22. Obtain the component structure diagram and control system configuration file of the engineering machinery equipment, analyze the physical connection method, signal control path and energy transmission channel between each functional component, and extract the three types of connection relationships that form physical connection relationship, control path relationship and energy path relationship;

[0018] S23. Set each sensor in the structured multidimensional time series data as a sensor node, set the execution component connected to the sensor as a functional component node, and set the controller, driver, logic circuit or communication interface connected to the functional component as a control node.

[0019] S24. Based on the physical connection relationship, add edges between sensor nodes and functional component nodes and mark the edge type as physical connection edge; based on the control path relationship, add edges between functional component nodes and control nodes and mark them as control signal edges; based on the energy path relationship, add edges between functional component nodes and mark them as energy transmission edges.

[0020] S25. For each node, read the corresponding structured multidimensional time series data and embed it into the graph structure as the node's attribute information. Combine all the constructed sensor nodes, functional component nodes and control nodes, physical connection edges, control signal edges and energy transmission edges into a heterogeneous graph.

[0021] Optionally, S3 specifically includes:

[0022] S31. Read the structured multidimensional time series data attached to each node in the heterogeneous graph in sequence, align the various raw sensor data in time order and stack them in the channel direction to generate a two-dimensional matrix with the dimension of the number of channels and the time step, and combine the inputs of multiple nodes to form a three-dimensional input tensor.

[0023] S32. Input the input tensor into the improved dilated temporal convolutional network, extract the short-term and long-term change features in the time series layer by layer to form an intermediate representation. During the convolution process, the time series features of each channel are weighted and fused through the channel attention mechanism.

[0024] S33. Output the fused feature representation as a time series embedding representation into the graph neural network structure.

[0025] Optionally, the improved dilated temporal convolutional network specifically includes:

[0026] In the structured multidimensional time series input corresponding to each node, the time interval sequence between two adjacent time points is calculated based on the sensor sampling timestamp. The maximum, minimum, average and standard deviation of the time interval for each channel are statistically analyzed to obtain the sampling stability index of the channel. The change value between two adjacent points is calculated for the numerical sequence of each channel. The average increase or decrease of the change value sequence and the standard deviation of the rate of change are statistically analyzed to form a set of parameters reflecting the dynamic fluctuation intensity and frequency characteristics of the channel. The time interval statistics and rate of change statistics of all channels are arranged in order of channel number, and a time feature matrix is ​​constructed with channels as rows and features as columns.

[0027] In each convolutional layer, multiple dilation factors with different values ​​are preset to represent the receptive field length ranges in the short, medium, and long term, respectively. Then, for time series inputs with the same channel dimension, three sets of one-dimensional dilation convolution operations are performed using the dilation factors. Each set of convolutions achieves different receptive field spans by inserting equally spaced holes in the convolution kernel, thereby expanding the time-aware range while keeping the number of convolution kernel parameters unchanged. All convolution operations are completed in parallel, outputting multiple sets of feature representations corresponding to different receptive field scales, and constructing receptive field attention vectors.

[0028] Read the weight coefficients and temporal feature matrices corresponding to each scale in the receptive field attention vector, perform channel-level alignment on the same time step for all convolution outputs with different dilation factors in the previous layer, multiply the outputs at the corresponding positions of each scale by the corresponding weights and add them element by element, summarize and fuse the results according to the channel direction, and generate the dynamic weighted output of the current layer.

[0029] Before each convolutional layer input, the mean, maximum, standard deviation and sign trend change of the difference between adjacent time slices at the current time slice are calculated for all channels of the input tensor. These are combined to generate channel-level statistical feature vectors. The input is a weight calculation network containing a two-layer fully connected structure. The first layer performs linear transformation and ReLU activation, and the second layer outputs a scalar as the attention score of the channel. Softmax normalization is performed on all channel scores to generate channel attention weight vectors.

[0030] Each dimension of the channel attention weight vector is multiplied by the value of all pixels in the corresponding channel intermediate feature tensor. Channel-wise numerical scaling is performed. The mean of all dimensions of the channel attention weight vector is calculated. Channels with a value lower than the mean are considered to have lower weights. The corresponding intermediate feature tensor is multiplied by a scaling factor of less than 1. Channels with a value higher than the mean are considered to have higher weights. The intermediate feature tensor retains its original value to preserve the response amplitude. Finally, the channel-weighted input tensor is generated.

[0031] Repeatedly perform dynamic expansion factor selection and channel weighting operations to extract embedding vectors containing short-term variation features and long-term dependency structures layer by layer, complete multi-scale dynamic modeling, and output node time series embedding representations with consistent dimensions.

[0032] Optionally, S4 specifically includes:

[0033] S41. Read the time series embedding representation of each node and the structural information of the adjacent edges of the node, find the edge type number of each edge, extract the corresponding edge type embedding, combine the time series embedding representation of the current node with the time series embedding representation of each neighboring node, and construct the neighboring nodes.

[0034] S42. Embed the time series of the current node into the input edge attention-enhanced heterogeneous graph neural network, combine the edge type embedding of each neighboring edge, calculate the corresponding edge attention weights and perform weighted summation to form the fused neighbor features;

[0035] S43. Input the neighbor features into a gating function with a time bias. Together with the time series context information of the current node, the gating node features are used as the output of the current propagation layer and passed into the next layer of the graph neural network structure. Repeat the processing flow and finally output the fault risk score of each key node.

[0036] Optionally, the edge-attention-enhanced heterogeneous graph neural network specifically includes:

[0037] For each adjacent edge of the current node, the time series embedding representation of the current node, the time series embedding representation of the neighboring nodes, and the edge type embedding vector of the edge are concatenated in sequence to form a triplet feature vector of a single edge, which is used as the attention input sample of the edge. The summaries constitute the attention input batch of all adjacent edges of the current node.

[0038] The triplet feature input vector of each edge is input into the edge attention score calculation network. An affine transformation layer is used followed by a LeakyReLU activation function to generate an intermediate score vector. The original attention score is output through a projection layer. Softmax normalization is performed on the scores of all adjacent edges of the current node to generate edge attention weights.

[0039] For all adjacent edges of the current node, multiply the time series embedding representation of each neighboring node by the corresponding edge attention weight coefficient, perform a weighted operation and sum them to generate the neighbor feature aggregation vector of the current node;

[0040] Extract the latest time segments from the original time series corresponding to the current node, and encode the trend of statistical value changes and mean fluctuation characteristics into a fixed-length time state vector as the local time context representation of the current node.

[0041] The fused neighbor embedding vector and its time state vector of the current node are input into two sets of fully connected layers respectively. After generating representations with the same dimension, vector addition is performed and the result is fed into the Sigmoid activation function to obtain the gating coefficient vector. The gating coefficient vector is multiplied element by element with the neighbor embedding vector to suppress and amplify each feature dimension. The gating adjusted node update representation is output and used as the state of the current layer of the current node.

[0042] The message passing and update process is repeated to advance the state evolution of all nodes layer by layer. Finally, in the last output stage of the edge attention-enhanced heterogeneous graph neural network, the fully connected regression layer is input to perform weighted summation and bias superposition, and the final failure risk score of each key node is output.

[0043] Optionally, S5 specifically includes:

[0044] S51. Read the archived historical fault data records, extract the original text or table content containing the fault occurrence time, equipment number, target component name, sensor abnormal parameters and corresponding processing actions, perform field mapping and type conversion according to the preset field format, and organize the extracted information into a structured coded representation.

[0045] S52. For each structured fault record, extract the component number from the target component field and associate it with the equipment topology node. Extract the abnormal channel and numerical features from the abnormal parameters to generate a fault feature vector. Map the corresponding handling measures and effect labels to preset label categories and construct a multi-field combination vector.

[0046] S53. Input the structured multi-field combination vector into the feature encoder, perform numerical standardization and dimension alignment, generate a fixed-length historical case vector, and establish a historical case library containing case vectors, label identifiers and field indexes;

[0047] S54. Using a vector similarity-based retrieval method, calculate the Euclidean distance and cosine similarity between the new fault risk feature vector and the historical case vector, construct a retrieval vector set that supports multiple indexing methods, and establish a multi-level indexing mechanism to accelerate the matching process, and select cases with similarity scores higher than a preset threshold as candidate sets.

[0048] Optionally, S6 specifically includes: for each historical case in the candidate set, reading the processing measure field and the corresponding effect label field, standardizing and rewriting the processing measure text, converting the effect label into a numerical representation, calculating the weighted average of the matching score and effect score of each historical case with the current fault risk characteristics, sorting the weighted results from high to low, selecting the top three processing strategies with the highest scores, and splicing them into structured suggestion content to output to the system platform as the final prevention suggestion.

[0049] The beneficial effects of this invention are:

[0050] This invention proposes a machine learning-based method for predicting and preventing faults in engineering machinery. Addressing the problems of delayed fault warnings, coarse feature recognition, lack of topological correlation, and separation of prevention strategies in existing equipment operation status monitoring, this invention constructs a heterogeneous graph structure with physical connections, control paths, and energy flow relationships. It integrates an improved dilated temporal convolutional network to extract multi-scale dynamic features from sensors, enhancing the perception of key fluctuation patterns and long-range lag behaviors. Furthermore, it introduces an edge-attention-enhanced heterogeneous graph neural network, combined with edge-type embedding and temporal bias gating mechanisms, to achieve risk focusing and dynamic propagation modeling of key impact paths in the equipment's structural topology. Based on this, it matches historical fault handling cases to generate multi-dimensional prevention suggestions with labels, structure, and effect feedback, enabling early detection and proactive intervention of high-risk faults. This invention possesses advantages such as refined temporal feature modeling, in-depth structural relationship perception, strong risk scoring interpretability, and automatic generation of prevention suggestions, effectively solving the problems of reliance on experience in engineering machinery fault handling, untimely identification and response, and disconnect between strategies. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a machine learning-based method for predicting and preventing engineering machinery faults, as proposed in this invention.

[0053] Figure 2 This is a flowchart of the multi-scale dynamic feature extraction process of the improved dilated temporal convolutional network proposed in this invention.

[0054] Figure 3 This is a flowchart of the fault propagation modeling and risk scoring process for the edge attention-enhanced heterogeneous graph neural network proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figures 1-3 A machine learning-based method for predicting and preventing engineering machinery faults includes the following steps:

[0057] S1. Collect multi-source operating data of engineering machinery and equipment, including key parameters such as pressure, temperature, vibration, current, and speed, and perform preprocessing operations such as time synchronization, missing data compensation, anomaly removal, and normalization to generate structured multidimensional time series data.

[0058] S2. Based on structured multidimensional time series data and the component connection relationship, control signal path and energy transmission path of the equipment, sensor nodes, functional component nodes, control nodes, etc. are constructed as multi-type nodes in the heterogeneous graph, and different types of edges are set to form a heterogeneous graph that can reflect the topology of the engineering machinery system.

[0059] S3. For the time series information attached to each node in the heterogeneous graph, an improved dilated temporal convolutional network is used to extract dynamic features, and the feature channels are weighted through a channel attention mechanism to output the time series embedding representation of the node.

[0060] S4. Input the time series embedding representation of the node into the edge attention-enhanced heterogeneous graph neural network, introduce the time series embedding representation of the node and the edge type embedding, use the time bias gating aggregation method when fusing neighbor information, and output the fault risk score of each key node.

[0061] S5. Structure and encode existing fault events, build a historical case library containing target components, fault characteristics, handling measures and effect tags, and a retrieval vector and matching index mechanism. Based on the fault risk score, perform similarity matching between the fault risk characteristics and the case characteristics in the historical case library to generate a candidate set.

[0062] S6. Calculate the weighted average of the matching score and effect score of historical cases with the current failure risk characteristics, select the top three highest-scoring processing strategies, extract the validated processing strategies, and generate structured prevention recommendations.

[0063] This implementation achieves high-precision prediction and intelligent prevention of faults in engineering machinery by fusing an improved dilated temporal convolutional network with an edge-attention-enhanced heterogeneous graph neural network. A structured multidimensional time series is constructed from multi-source operational data. By introducing a channel-level temporal feature matrix, the dilation factor of the convolutional structure is dynamically adjusted, and weighted modeling is performed on the input channels, enabling the improved dilated temporal convolutional network to have adaptive receptive field scheduling and non-uniform feature modeling capabilities. A heterogeneous graph structure is constructed based on the equipment topology, introducing edge type embedding and node temporal embedding. An edge attention mechanism is used to model the importance of edges, and a time-biased gating aggregation function is used to enhance the expressive power of cross-temporal states, outputting fault risk scores for key nodes. At the prevention level, vector-level similarity matching is performed between the risk scores and processing schemes in the historical case library to generate structured prevention strategies, achieving closed-loop linkage between prediction and control. This method exhibits higher robustness and response speed in anomaly evolution capture and strategy recommendation.

[0064] In this embodiment, S1 specifically includes: collecting various raw sensor data of pressure, temperature, vibration, current and speed of engineering machinery equipment during operation; aligning the collected data according to a unified timestamp; removing missing values ​​and outliers; using an interpolation algorithm to complete missing segments; unifying the dimensions of the collected data before normalization according to physical unit standards; performing standardization and normalization operations to obtain structured multidimensional time series data.

[0065] In this embodiment, S2 specifically includes:

[0066] S21. Read the structured multidimensional time series data, associate each piece of structured multidimensional time series data with the corresponding sensor number, physical installation location, and connection component information, and establish a correspondence table between sensor data and equipment components;

[0067] S22. Obtain the component structure diagram and control system configuration file of the engineering machinery equipment, analyze the physical connection method, signal control path and energy transmission channel between each functional component, and extract the three types of connection relationships that form physical connection relationship, control path relationship and energy path relationship;

[0068] S23. Set each sensor in the structured multidimensional time series data as a sensor node, set the execution component connected to the sensor as a functional component node, and set the controller, driver, logic circuit or communication interface connected to the functional component as a control node.

[0069] S24. Based on the physical connection relationship, add edges between sensor nodes and functional component nodes and mark the edge type as physical connection edge; based on the control path relationship, add edges between functional component nodes and control nodes and mark them as control signal edges; based on the energy path relationship, add edges between functional component nodes and mark them as energy transmission edges.

[0070] S25. For each node, read the corresponding structured multidimensional time series data and embed it into the graph structure as the node's attribute information to ensure that each node carries real-time or historical operation data. Combine all the constructed sensor nodes, functional component nodes and control nodes, physical connection edges, control signal edges and energy transmission edges into a heterogeneous graph with a complete topological structure and data mapping relationship, which serves as the basic graph structure for subsequent graph neural network input.

[0071] This implementation constructs a heterogeneous graph with the ability to recognize the structure of engineering machinery, achieving precise fusion of multi-source operational data and equipment topology. Structured multi-dimensional time-series data and sensor numbers, installation locations, and connection component information are extracted to establish a mapping relationship between equipment data and physical structure. Based on the component structure diagram and control system configuration, the physical connections, signal paths, and energy flow relationships between functional components are analyzed, generating three types of edge types. Sensors, functional components, and controllers are further constructed as multi-type nodes, and edges are added according to the three types of connection relationships to construct a heterogeneous graph containing physical connection edges, control signal edges, and energy transmission edges. Each node carries corresponding time-series data, embedded as attribute information, achieving deep fusion of graph structure and operational data. The generated heterogeneous graph can fully reflect the spatial structure and dynamic behavior of the equipment system, providing a structurally complete and data-accurate foundation for subsequent fault reasoning in graph neural networks, significantly improving the adaptability and modeling accuracy of the prediction model for complex structural equipment.

[0072] In this embodiment, S3 specifically includes:

[0073] S31. Read the structured multidimensional time series data attached to each node in the heterogeneous graph in sequence. Align the various raw sensor data of temperature, pressure, vibration, current and speed in time order and stack them in the channel direction to generate a two-dimensional matrix with the dimension of the number of channels and time steps. Combine the inputs of multiple nodes to form a three-dimensional input tensor.

[0074] S32. Input the input tensor into the improved dilated temporal convolutional network, extract the short-term and long-term change features in the time series layer by layer to form an intermediate representation with temporal context awareness. During the convolution process, the time series features of each channel are weighted and fused through the channel attention mechanism to enhance the responsiveness of key feature channels.

[0075] S33. The fused feature representation is output as a time series embedding representation into the graph neural network structure as the input feature vector of each node.

[0076] This implementation employs an improved dilated temporal convolutional network to dynamically model structured multidimensional time-series data, fully exploring the key features and temporal dependencies between sensor channels. By constructing a channel temporal feature matrix that includes statistical features of time intervals and rate of change indicators, the sampling stability and dynamic fluctuation characteristics of each channel are effectively characterized. Dilated convolutions with different dilation factors are executed in parallel within each network layer to extract short-term local abrupt changes, periodic oscillations, and long-term trend responses, respectively. Multi-scale features are dynamically fused based on the receptive field attention vector, enhancing the modeling ability for non-equidistant and long-distance dependencies. Simultaneously, a channel attention mechanism is introduced to automatically calculate the importance score of each channel feature, enhancing key channels and compressing redundant channel responses, significantly optimizing feature representation efficiency and discriminability. Through the above multi-scale dynamic perception and channel weighted fusion, the generated time-series embedding vector can accurately capture the operational status changes of key components in engineering machinery, providing rich and discriminative input features for subsequent graph neural network processing.

[0077] In this embodiment, the improved dilated temporal convolutional network specifically includes:

[0078] In the structured multidimensional time series input corresponding to each node, the time interval sequence between two adjacent time points is calculated based on the sensor sampling timestamp. The maximum, minimum, average and standard deviation of the time interval for each channel are statistically analyzed to obtain the sampling stability index of the channel. The change value between two adjacent points is calculated for the numerical sequence of each channel. The average increase or decrease of the change value sequence and the standard deviation of the change rate are statistically analyzed to form a parameter set reflecting the dynamic fluctuation intensity and frequency characteristics of the channel. The time interval statistics and change rate statistics of all channels are arranged in order of channel number. A time feature matrix is ​​constructed with channels as rows and features as columns, which serves as the basis for the subsequent dynamic convolution strategy.

[0079] In each convolutional layer, multiple dilation factors with different values ​​are preset to represent the receptive field length ranges for short, medium, and long term, respectively. Then, for time series inputs with the same channel dimension, three sets of one-dimensional dilation convolution operations are performed using the dilation factors. Each set of convolutions achieves different receptive field spans by inserting equally spaced holes in the convolution kernel, expanding the time perception range while keeping the number of convolution kernel parameters constant. Each set of dilation convolutions will cover the change features at different time scales. The short receptive field mainly extracts local drastic changes and fine-grained fluctuations, the medium receptive field is suitable for periodic oscillations and adjacent interference, and the long receptive field helps to capture long-distance lag responses and cross-period trends. All convolution operations are completed in parallel, outputting multiple sets of feature representations corresponding to different receptive field scales. A receptive field attention vector is constructed as the basis for dynamic dilation factor selection.

[0080] Read the weight coefficients and temporal feature matrix corresponding to each scale in the receptive field attention vector, perform channel-level alignment on the same time step for all convolution outputs with different dilation factors in the previous layer, multiply the outputs at the corresponding positions of each scale by the corresponding weights and add them element by element, summarize and fuse the results according to the channel direction, and generate the dynamic weighted output of the current layer to adapt to the dependency range of different time series features.

[0081] Before each convolutional layer input, the mean, maximum, standard deviation, and sign trend change of the difference between adjacent time slices at the current time slice are calculated for all channels of the input tensor. These are combined to generate channel-level statistical feature vectors. The input is a weight calculation network containing two fully connected layers. The first layer performs linear transformation and ReLU activation, and the second layer outputs a scalar as the attention score of the channel. Softmax normalization is performed on all channel scores to generate channel attention weight vectors, which are used as weight coefficients for subsequent weighted feature fusion.

[0082] Each dimension of the channel attention weight vector is multiplied by the value of all pixels in the corresponding channel intermediate feature tensor. Channel-wise numerical scaling is performed. The mean of all dimensions of the channel attention weight vector is calculated. Channels with a value lower than the mean are considered to have lower weights. The corresponding intermediate feature tensor is multiplied by a scaling factor less than 1 to compress the response intensity. Channels with a value higher than the mean are considered to have higher weights. The intermediate feature tensor retains its original value to preserve the response amplitude. Finally, a channel-weighted input tensor is generated as the input for subsequent dilated convolution operations.

[0083] Repeatedly perform dynamic expansion factor selection and channel weighting operations to extract embedding vectors containing short-term variation features and long-term dependency structures layer by layer, complete multi-scale dynamic modeling, and output node time series embedding representations with consistent dimensions.

[0084] In this embodiment, S4 specifically includes:

[0085] S41. Read the time series embedding representation of each node and the structural information of the adjacent edges of the node, find the edge type number of each edge, extract the corresponding edge type embedding, combine the time series embedding representation of the current node with the time series embedding representation of each neighboring node, and construct the neighboring nodes.

[0086] S42. Embed the time series of the current node into the input edge attention-enhanced heterogeneous graph neural network, combine the edge type embedding of each neighboring edge, calculate the corresponding edge attention weights and perform weighted summation to form the fused neighbor features;

[0087] S43. Input the neighbor features into a gating function with a time bias. Together with the time series context information of the current node, the gating node features are used as the output of the current propagation layer and passed into the next layer of the graph neural network structure. Repeat the processing flow and finally output the fault risk score of each key node.

[0088] This implementation employs an edge-attention-enhanced heterogeneous graph neural network to perform deep modeling of the structural diagram of engineering machinery equipment, fully integrating heterogeneous relationships and temporal features between nodes. By constructing triplet features containing time-series embeddings of the current node, neighboring nodes, and edge type embeddings, the semantics of structural connections are accurately expressed. An edge-attention mechanism is introduced to learn the contribution strength of different types of edges, improving the selectivity of information aggregation. During each layer of propagation, a time bias control is introduced by calculating a gating coefficient vector. Combined with the local temporal context state of the current node, the influence of neighboring features on the evolution of the current node's state is automatically adjusted, effectively adapting to rhythmic changes and delayed response phenomena during equipment operation. Compared to traditional graph neural network structures, this method significantly enhances the modeling capability for relationships between multiple types of components, strengthens the interactive expression between key nodes, and the output fault risk score is more consistent with time-varying correlations and actual topological logic, supporting early identification and causal reasoning of potential fault propagation paths and key anomaly sources.

[0089] In this embodiment, the edge-attention enhanced heterogeneous graph neural network specifically includes:

[0090] For each adjacent edge of the current node, the time series embedding representation of the current node, the time series embedding representation of the neighboring nodes, and the edge type embedding vector of the edge are concatenated in sequence to form a triplet feature vector of a single edge, which is used as the attention input sample of the edge. The summaries constitute the attention input batch of all adjacent edges of the current node.

[0091] The triplet feature input vector of each edge is input into the edge attention score calculation network. An affine transformation layer is used followed by a LeakyReLU activation function to generate an intermediate score vector. The original attention score is output through a projection layer. Softmax normalization is performed on the scores of all adjacent edges of the current node to generate edge attention weights.

[0092] For all adjacent edges of the current node, multiply the time series embedding representation of each neighboring node by the corresponding edge attention weight coefficient, perform a weighted operation and sum them to generate the neighbor feature aggregation vector of the current node as the fused neighbor embedding representation;

[0093] Extract the latest time segments from the original time series corresponding to the current node, and encode the statistical value change trend and mean fluctuation characteristics into a fixed-length time state vector as the local time context representation of the current node, which is used to provide a time bias adjustment signal in gated aggregation.

[0094] The fused neighbor embedding vector and its time state vector of the current node are input into two sets of fully connected layers respectively. After generating representations with the same dimension, vector addition is performed and the result is fed into the Sigmoid activation function to obtain the gating coefficient vector. The gating coefficient vector is multiplied element by element with the neighbor embedding vector to suppress and amplify each feature dimension. The gating adjusted node update representation is output and used as the state of the current layer of the current node.

[0095] The message passing and update process is repeated to advance the state evolution of all nodes layer by layer. Finally, in the last output stage of the edge attention-enhanced heterogeneous graph neural network, the fully connected regression layer is input to perform weighted summation and bias superposition, and the final failure risk score of each key node is output.

[0096] In this embodiment, S5 specifically includes:

[0097] S51. Read the archived historical fault data records, extract the original text or table content containing the fault occurrence time, equipment number, target component name, sensor abnormal parameters and corresponding processing actions, perform field mapping and type conversion according to the preset field format, and organize the extracted information into a structured coded representation.

[0098] S52. For each structured fault record, extract the component number from the target component field and associate it with the equipment topology node. Extract the abnormal channel and numerical features from the abnormal parameters to generate a fault feature vector. Map the corresponding handling measures and effect labels to preset label categories and construct a multi-field combination vector.

[0099] S53. Input the structured multi-field combination vector into the feature encoder, perform numerical standardization and dimension alignment, generate a fixed-length historical case vector, and establish a historical case library containing case vectors, label identifiers and field indexes;

[0100] S54. Using a vector similarity-based retrieval method, calculate the Euclidean distance and cosine similarity between the new fault risk feature vector and the historical case vector, construct a retrieval vector set that supports multiple indexing methods, and establish a multi-level indexing mechanism to accelerate the matching process, and select cases with similarity scores higher than a preset threshold as candidate sets.

[0101] This implementation method automatically extracts key fields such as equipment number, fault location, abnormal channel, and handling measures from archived fault data records, and converts them into a unified structured coding format to effectively avoid information loss due to inconsistent descriptions. Based on the association between the target component number and the equipment topology, a fault feature vector is constructed by combining abnormal sensor parameters, and the handling measures and effect labels are standardized and coded to establish a multi-field combination vector with complete fields. A feature encoder is used to perform numerical normalization and dimensionality unification to generate a set of historical cases that can be used for vectorized retrieval. During the retrieval stage, a similarity matching mechanism is established by combining Euclidean distance and cosine similarity. A multi-level index structure is used to improve search efficiency, achieving high-precision and high-efficiency matching between new fault risk features and verified historical cases, providing accurate data support for subsequent prevention strategy recommendations.

[0102] In this embodiment, S6 specifically includes: for each historical case in the candidate set, reading the treatment measure field and the corresponding effect label field, standardizing and rewriting the treatment measure text, converting the effect label into a numerical representation, such as setting "completely resolved" to 1.0, "partially alleviated" to 0.5, and "no effect" to 0.0, calculating the weighted average of the matching score and effect score of each historical case with the current fault risk characteristics, sorting the results from high to low according to the weighted results, selecting the top three highest-scoring treatment strategies, and concatenating them into structured suggestion content for output to the system platform as the final prevention suggestion.

[0103] This implementation method reads the processing measures and effect tags for each historical case in the candidate set. Semantic consistency is improved through standardized rewriting of the processing measures, and effect tags are converted into uniform numerical levels to ensure the objectivity and comparability of the scoring process. A weighted strategy is used to comprehensively evaluate the matching degree and expected effect between the current fault risk characteristics and historical cases, selecting the optimal strategy combination. Finally, the top-ranked processing solutions are automatically assembled into structured prevention suggestions and accurately output to the system platform, providing reliable and actionable intelligent decision-making basis for maintenance scheduling and fault prevention.

[0104] Example 1

[0105] To verify the feasibility and industrial applicability of this invention, the machine learning-based fault prediction and prevention method for engineering machinery proposed in this invention was applied to a high-frequency loading and unloading operation scenario at a large port machinery maintenance center in a certain province. This maintenance center operates a total of 24 tracked cranes and 12 telescopic loaders, with an average daily continuous operating time exceeding 19 hours, undertaking high-intensity tasks such as container stacking, bulk material transfer, and terminal cargo handling. Due to the operating environment being close to a salt-sprayed sea area, the equipment is constantly exposed to high humidity, high temperature, impact, and vibration conditions, leading to accelerated component aging and frequent control path interference, particularly affecting hydraulic actuators, main electrical control modules, and critical bearing systems. Traditional maintenance methods rely on periodic manual inspections and post-failure repairs, resulting in delayed fault response, lack of early warning, and high equipment idling rates. Especially during high-load scheduling in summer, controller overheating or hydraulic cylinder lag often causes the entire machine to stop, placing significant pressure on operation scheduling.

[0106] To address the above pain points, the operations and maintenance team, in conjunction with this invention, implemented an intelligent fault prediction and prevention system. Deployed on 10 main tracked cranes, the system collects data from 38 channels of existing pressure, temperature, current, voltage, load, and vibration sensors, uploading the data at 1Hz to an edge computing node for graph structure modeling and embedded convolutional analysis. During the data fusion phase, the system constructs a heterogeneous topology graph of components, identifying approximately 184 sensor nodes, 102 functional component nodes, and 36 controller nodes, totaling 742 physical connection edges, energy path edges, and control signal edges. Subsequently, a multi-scale dynamic feature extraction process based on an improved dilated temporal convolutional network is run on each node, and edge attention fusion is performed at the graph neural network layer to output a risk score for key nodes.

[0107] The system automatically performs fault scoring daily and correlates it with historical data. Once the score exceeds the warning threshold, it will coordinate with the control center to issue scheduling suggestions or shutdown maintenance instructions. Combining existing fault handling records and equipment response strategies, it will generate the optimal preventive measures.

[0108] The data collection period was 28 consecutive days, covering 10 test devices and 10 control devices of the same model. The method of this invention and the manual inspection strategy were used for operation and maintenance management. A comparison table of fault response between the method of this invention and the traditional strategy in engineering machinery operation and maintenance scenarios is as follows:

[0109] Table 1. Comparison of the effects of the method of the present invention and traditional maintenance strategies in the operation and maintenance of engineering machinery.

[0110]

[0111] As can be seen from the data in Table 1 above, in terms of fault warning accuracy, the method group of the present invention reached 94.6%, which is 22.1 percentage points higher than the traditional group's 72.5%, significantly improving the accuracy of identifying equipment anomalies; in terms of response timeliness, the warning response time was greatly shortened from the original 7.3 hours to 1.8 hours, and the response speed was improved by more than 75%, providing support for timely intervention and precise operation and maintenance.

[0112] In terms of equipment downtime management, the method of this invention significantly improves the planning and controllability of downtime. The proportion of planned downtime increased from 52.8% to 78.4%, while the frequency of unplanned downtime decreased dramatically from an average of 4.2 times per month to 1.1 times, a reduction of 73.8%, effectively reducing the economic losses caused by sudden downtime. Regarding energy consumption, the method of this invention, through optimized scheduling and early warning control, reduced the average idling energy consumption from 1482 kWh to 1046 kWh, a reduction of 29.4%, achieving the goal of energy conservation and consumption reduction. This invention also demonstrates significant advantages for high-level faults and core component problems. The incidence of high-level faults decreased from 3.6% to 0.8%, and the proportion of control system-related faults decreased from 41.3% to 17.5%, reductions of 2.8 and 23.8 percentage points respectively, fully demonstrating the capabilities of the side-attention heterogeneous graph neural network in key node modeling and fault tracing.

[0113] Furthermore, in terms of labor costs, the method of this invention significantly improves inspection efficiency, reducing monthly manual inspection hours from 184 hours to 66 hours, saving over 64% of labor input. In summary, this invention demonstrates excellent performance in improving prediction accuracy, enhancing operational efficiency, and reducing energy consumption and failure losses, exhibiting promising prospects and practical application value.

[0114] In summary, the comparative data comprehensively verifies the efficiency, accuracy, and practicality of this invention in the prediction and prevention of engineering machinery failures. It not only improves the stability and reliability of equipment operation but also brings significant benefits in cost control and energy management, and has extremely high promotional value.

[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and preventing engineering machinery faults based on machine learning, characterized in that, Includes the following steps: S1. Collect multi-source operation data of engineering machinery and equipment to generate structured multi-dimensional time series data; S2. Based on structured multidimensional time series data and the component connection relationships, control signal paths and energy transmission paths of the equipment, construct a heterogeneous graph; S3. For the time series information attached to each node in the heterogeneous graph, an improved dilated temporal convolutional network is used to extract dynamic features, and the feature channels are weighted through a channel attention mechanism to output the time series embedding representation of the node. S4. Input the time series embedding representation of the node into the edge attention-enhanced heterogeneous graph neural network, introduce the time series embedding representation of the node and the edge type embedding, use the time bias gating aggregation method when fusing neighbor information, and output the fault risk score of each key node. S5. Structure and encode existing fault events, build a historical case library and retrieval vector and matching index mechanism, and perform similarity matching between fault risk features and case features in the historical case library based on fault risk scores to generate a candidate set. S6. Calculate the weighted average of the matching score and effect score of historical cases with the current failure risk characteristics, select the top three highest-scoring processing strategies, and generate structured prevention suggestions. The improved dilated temporal convolutional network specifically includes: In the structured multidimensional time series input corresponding to each node, the time interval sequence between two adjacent time points is calculated based on the sensor sampling timestamp. The maximum, minimum, average and standard deviation of the time interval for each channel are statistically analyzed to obtain the sampling stability index of the channel. The change value between two adjacent points is calculated for the numerical sequence of each channel. The average increase or decrease of the change value sequence and the standard deviation of the rate of change are statistically analyzed to form a set of parameters reflecting the dynamic fluctuation intensity and frequency characteristics of the channel. The time interval statistics and rate of change statistics of all channels are arranged in order of channel number, and a time feature matrix is ​​constructed with channels as rows and features as columns. In each convolutional layer, multiple dilation factors with different values ​​are preset to represent the receptive field length ranges in the short, medium, and long term, respectively. Then, for time series inputs with the same channel dimension, three sets of one-dimensional dilation convolution operations are performed using the dilation factors. Each set of convolutions achieves different receptive field spans by inserting equally spaced holes in the convolution kernel, thereby expanding the time-aware range while keeping the number of convolution kernel parameters unchanged. All convolution operations are completed in parallel, outputting multiple sets of feature representations corresponding to different receptive field scales, and constructing receptive field attention vectors. Read the weight coefficients and temporal feature matrices corresponding to each scale in the receptive field attention vector, perform channel-level alignment on the same time step for all convolution outputs with different dilation factors in the previous layer, multiply the outputs at the corresponding positions of each scale by the corresponding weights and add them element by element, summarize and fuse the results according to the channel direction, and generate the dynamic weighted output of the current layer. Before each convolutional layer input, the mean, maximum, standard deviation and sign trend change of the difference between adjacent time slices at the current time slice are calculated for all channels of the input tensor. These are combined to generate channel-level statistical feature vectors. The input is a weight calculation network containing a two-layer fully connected structure. The first layer performs linear transformation and ReLU activation, and the second layer outputs a scalar as the attention score of the channel. Softmax normalization is performed on all channel scores to generate channel attention weight vectors. Each dimension of the channel attention weight vector is multiplied by the value of all pixels in the corresponding channel intermediate feature tensor. Channel-wise numerical scaling is performed. The mean of all dimensions of the channel attention weight vector is calculated. Channels with a value lower than the mean are considered to have lower weights. The corresponding intermediate feature tensor is multiplied by a scaling factor of less than 1. Channels with a value higher than the mean are considered to have higher weights. The intermediate feature tensor retains its original value to preserve the response amplitude. Finally, the channel-weighted input tensor is generated. Repeatedly perform dynamic expansion factor selection and channel weighting operations to extract embedding vectors containing short-term variation features and long-term dependency structures layer by layer, complete multi-scale dynamic modeling, and output node time series embedding representations with consistent dimensions. The edge-attention-enhanced heterogeneous graph neural network specifically includes: For each adjacent edge of the current node, the time series embedding representation of the current node, the time series embedding representation of the neighboring nodes, and the edge type embedding vector of the edge are concatenated in sequence to form a triplet feature vector of a single edge, which is used as the attention input sample of the edge. The summaries constitute the attention input batch of all adjacent edges of the current node. The triplet feature input vector of each edge is input into the edge attention score calculation network. An affine transformation layer is used followed by a LeakyReLU activation function to generate an intermediate score vector. The original attention score is output through a projection layer. Softmax normalization is performed on the scores of all adjacent edges of the current node to generate edge attention weights. For all adjacent edges of the current node, multiply the time series embedding representation of each neighboring node by the corresponding edge attention weight coefficient, perform a weighted operation and sum them to generate the neighbor feature aggregation vector of the current node; Extract the latest time segments from the original time series corresponding to the current node, and encode the trend of statistical value changes and mean fluctuation characteristics into a fixed-length time state vector as the local time context representation of the current node. The fused neighbor embedding vector and its time state vector of the current node are input into two sets of fully connected layers respectively. After generating representations with the same dimension, vector addition is performed and the result is fed into the Sigmoid activation function to obtain the gating coefficient vector. The gating coefficient vector is multiplied element by element with the neighbor embedding vector to suppress and amplify each feature dimension. The gating adjusted node update representation is output and used as the state of the current layer of the current node. The message passing and update process is repeated to advance the state evolution of all nodes layer by layer. Finally, in the last output stage of the edge attention-enhanced heterogeneous graph neural network, the fully connected regression layer is input to perform weighted summation and bias superposition, and the final failure risk score of each key node is output.

2. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S1 specifically includes: collecting various raw sensor data of engineering machinery and equipment during operation, aligning the collected data according to a unified timestamp, removing missing values ​​and outliers, and using an interpolation algorithm to complete the missing segments, unifying the dimensions of the collected data before normalization according to physical unit standards, performing standardization and normalization operations, and obtaining structured multidimensional time series data.

3. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S2 specifically includes: S21. Read the structured multidimensional time series data, associate each piece of structured multidimensional time series data with the corresponding sensor number, physical installation location, and connection component information, and establish a correspondence table between sensor data and equipment components; S22. Obtain the component structure diagram and control system configuration file of the engineering machinery equipment, analyze the physical connection method, signal control path and energy transmission channel between each functional component, and extract the three types of connection relationships that form physical connection relationship, control path relationship and energy path relationship; S23. Set each sensor in the structured multidimensional time series data as a sensor node, set the execution component connected to the sensor as a functional component node, and set the controller, driver, logic circuit or communication interface connected to the functional component as a control node. S24. Based on the physical connection relationship, add edges between sensor nodes and functional component nodes and mark the edge type as physical connection edge; based on the control path relationship, add edges between functional component nodes and control nodes and mark them as control signal edges; based on the energy path relationship, add edges between functional component nodes and mark them as energy transmission edges. S25. For each node, read the corresponding structured multidimensional time series data and embed it into the graph structure as the node's attribute information. Combine all the constructed sensor nodes, functional component nodes and control nodes, physical connection edges, control signal edges and energy transmission edges into a heterogeneous graph.

4. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S3 specifically includes: S31. Read the structured multidimensional time series data attached to each node in the heterogeneous graph in sequence, align the various raw sensor data in time order and stack them in the channel direction to generate a two-dimensional matrix with the dimension of the number of channels and the time step, and combine the inputs of multiple nodes to form a three-dimensional input tensor. S32. Input the input tensor into the improved dilated temporal convolutional network, extract the short-term and long-term change features in the time series layer by layer to form an intermediate representation. During the convolution process, the time series features of each channel are weighted and fused through the channel attention mechanism. S33. Output the fused feature representation as a time series embedding representation into the graph neural network structure.

5. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S4 specifically includes: S41. Read the time series embedding representation of each node and the structural information of the adjacent edges of the node, find the edge type number of each edge, extract the corresponding edge type embedding, combine the time series embedding representation of the current node with the time series embedding representation of each neighboring node, and construct the neighboring nodes. S42. Embed the time series of the current node into the input edge attention-enhanced heterogeneous graph neural network, combine the edge type embedding of each neighboring edge, calculate the corresponding edge attention weights and perform weighted summation to form the fused neighbor features; S43. Input the neighbor features into a gating function with a time bias. Together with the time series context information of the current node, the gating node features are used as the output of the current propagation layer and passed into the next layer of the graph neural network structure. Repeat the processing flow and finally output the fault risk score of each key node.

6. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S5 specifically includes: S51. Read the archived historical fault data records, extract the original text or table content containing the fault occurrence time, equipment number, target component name, sensor abnormal parameters and corresponding processing actions, perform field mapping and type conversion according to the preset field format, and organize the extracted information into a structured coded representation. S52. For each structured fault record, extract the component number from the target component field and associate it with the equipment topology node. Extract the abnormal channel and numerical features from the abnormal parameters to generate a fault feature vector. Map the corresponding handling measures and effect labels to preset label categories and construct a multi-field combination vector. S53. Input the structured multi-field combination vector into the feature encoder, perform numerical standardization and dimension alignment, generate a fixed-length historical case vector, and establish a historical case library containing case vectors, label identifiers and field indexes; S54. Using a vector similarity-based retrieval method, calculate the Euclidean distance and cosine similarity between the new fault risk feature vector and the historical case vector, construct a retrieval vector set that supports multiple indexing methods, and establish a multi-level indexing mechanism to accelerate the matching process, and select cases with similarity scores higher than a preset threshold as candidate sets.

7. The method for predicting and preventing engineering machinery faults based on machine learning according to claim 1, characterized in that, S6 specifically includes: for each historical case in the candidate set, reading the processing measure field and the corresponding effect label field, standardizing and rewriting the processing measure text, converting the effect label into a numerical representation, calculating the weighted average of the matching score and effect score of each historical case with the current fault risk characteristics, sorting the weighted results from high to low, selecting the top three processing strategies with the highest scores, and splicing them into structured suggestion content to output to the system platform as the final prevention suggestion.

Images