A method and device for predicting fuel consumption rate of a excavator based on space-time graph convolution and multi-working-condition contrast learning

Abstract

The application discloses a kind of based on spatiotemporal diagram convolution and multi-working condition contrast learning's excavator fuel consumption rate prediction method and device.In the method, construct excavator energy flow spatiotemporal diagram, define core component as graph node, define physical connection as weight edge;And design multi-level physical feature engineering scheme, fusion original sensor data and energy conversion type derived feature;Again construct spatiotemporal diagram convolution network (ST-GCN) and supervised contrast learning fusion model, improve prediction accuracy and generalization ability by double task training.The application combines spatiotemporal diagram convolution network and links excavator working system node information, injects physical mechanism knowledge, significantly improves the adaptability of model to complex working condition, and the prediction accuracy and generalization ability are superior to existing model, can provide accurate guidance for excavator energy-saving control, and has important engineering application value.

Description

Technical Field

[0001] This invention relates to the field of engineering machinery condition monitoring and energy efficiency optimization technology, and in particular to a method and device for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning. Background Technology

[0002] As a core piece of construction machinery, excavators operate in complex environments with dynamic and changing conditions. Their fuel consumption rate directly impacts operating costs and environmental benefits. Therefore, accurate prediction of fuel consumption rate is crucial for energy-saving optimization and control. Currently, data-driven methods are widely used in the field of construction machinery fuel consumption prediction, and related research has made some progress. However, many core technical challenges remain in specific excavator application scenarios, making it difficult to meet actual engineering needs.

[0003] In existing technologies, some studies have adopted a scheme combining variational mode decomposition (VMD) with the Informer model. Although this combined model has been shown to outperform traditional LSTM and GRU models under complex conditions, and emphasizes the importance of data preprocessing and attention mechanisms, it relies solely on VMD for noise reduction and does not apply physically meaningful features. At the same time, the baseline models used for comparison are relatively traditional and have not been compared with models such as graph neural networks (GNNs) that are more suitable for modeling system correlations. Furthermore, the interpretability analysis is limited to error analysis and fails to reveal the intrinsic relationships between influencing factors.

[0004] Another study on heavy truck fuel consumption prediction used the CatBoost model and the SHAP interpretability framework to systematically evaluate the impact of different feature combinations, concluding that engine features are crucial. However, this model can only handle static feature data without time-series dependencies and cannot effectively capture the strong time-series dependencies and dynamic operating condition changes during excavator operation. Furthermore, the continuous driving conditions of the trucks corresponding to its data source are different from the repetitive, intermittent, and highly dynamic operation cycles of excavators, making it difficult to directly transfer and apply the data.

[0005] To address the issue that mechanical fuel injection system excavators cannot directly provide fuel consumption data, some technologies use attribute data from excavators with the same model of electronic fuel injection system to predict fuel consumption using a stacked regression model combining random forest and K-nearest neighbor algorithms, and then adjust the prediction based on actual fuel consumption. However, this approach relies on the transfer and application of data from similar electronic fuel injection systems, fails to directly capture the system coupling relationships within the mechanical fuel injection system excavator itself, and does not consider the impact of dynamic changes in operating conditions on prediction accuracy, thus limiting its generalization ability.

[0006] Some studies have attempted to build predictive models based on deep learning and support vector machines, using principal component analysis to reduce the dimensionality of multi-dimensional sensor data and simplify the computational complexity of the models. However, these methods only select features through statistical means, lacking the integration of physical mechanisms such as energy conversion and hydraulic transmission in excavators. The feature engineering depth is insufficient, and no adaptation strategies are designed for the multi-condition characteristics of excavators, making the prediction accuracy susceptible to impacts in complex operating scenarios.

[0007] In the field of spatiotemporal series prediction, existing technologies construct combined spatiotemporal correlation adjacency matrices based on spatiotemporal graph neural networks to capture the spatiotemporal correlation features of data, which are suitable for prediction in scenarios such as traffic flow. However, this technology has not been adapted to the domain characteristics of excavator fuel consumption prediction, has not incorporated the physical mechanism knowledge of construction machinery, and has not considered the working condition differences of excavators in multiple operating modes. Therefore, it cannot be directly transferred for accurate fuel consumption rate prediction, and lacks mechanistic-level interpretability analysis of the prediction results.

[0008] In summary, existing fuel consumption prediction methods have significant shortcomings in adapting to the complex dynamic operating characteristics of excavators, understanding the physical significance of excavation, and improving the generalization ability of operating conditions. There is an urgent need for a fuel consumption rate prediction method that can accurately capture the spatiotemporal coupling relationship of the excavator system, incorporate domain physics knowledge, and adapt to changes in multiple operating conditions. Summary of the Invention

[0009] This invention provides a method and apparatus for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning, which can solve the problems of existing technologies being unable to adapt to the complex operating conditions of excavators, insufficient feature engineering depth, and limited model generalization ability.

[0010] The first aspect of this invention provides a method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning, comprising:

[0011] Construct a spatiotemporal graph of excavator energy flow, select core components of the excavator as graph nodes, construct graph edges based on energy flow paths and physical connections between components, generate a weighted adjacency matrix, and the initial features of each graph node include the original sensor data of the corresponding component;

[0012] Design a physical feature engineering scheme to calculate the physical derived features of the engine side, hydraulic system actuator side and system level in real time, and use the physical derived features and the original sensor data as the input features of the graph nodes.

[0013] A spatiotemporal graph convolutional network and contrastive learning fusion model is constructed. The spatiotemporal graph convolutional network is the main body, and a projection head and contrastive learning module are added. A weighted total loss function is designed to train and optimize the fusion model.

[0014] The test data is input into the trained fusion model, and the prediction head outputs the fuel consumption rate prediction for the next time step.

[0015] Optionally, a spatiotemporal diagram of the excavator's energy flow is constructed, including:

[0016] The engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor are selected as the drawing nodes;

[0017] The initial weights of the adjacency matrix are set based on the energy transfer efficiency analysis results. The initial weights reflect the strength of the connection between graph nodes or the energy transfer efficiency, and are optimized through fine-tuning during subsequent model training.

[0018] Optionally, physically derived features include:

[0019] Engine power indication on the engine side;

[0020] Instantaneous hydraulic power of each hydraulic cylinder on the actuator side of the hydraulic system and power of the rotary motor;

[0021] System-level total power demand and engine load rate.

[0022] Optionally, the spatiotemporal graph convolutional network is composed of multiple spatiotemporal convolutional blocks stacked together. Each spatiotemporal convolutional block includes spatial graph convolution and temporal one-dimensional convolution. The spatial graph convolution is used to aggregate the feature information of adjacent graph nodes, and the temporal one-dimensional convolution is used to capture the temporal state changes of the graph nodes themselves.

[0023] Optionally, the contrastive learning module employs a supervised contrastive learning approach, implemented through the following steps:

[0024] The K-Means unsupervised clustering algorithm is used to cluster the training set data to obtain multiple job mode labels;

[0025] Select a data segment of a time window from the training set as the anchor sample, select positive samples from other time periods of the same work mode, and select multiple negative samples from time periods of different work modes to construct a sample pair consisting of the anchor sample, positive sample, and negative sample.

[0026] Optionally, the weighted total loss function is obtained by weighted summation of the prediction loss and the contrast loss. The prediction loss adopts the mean squared error regression loss to ensure the prediction accuracy of fuel consumption rate; the contrast loss adopts the InfoNCE loss to bring the features of the anchor sample closer to the positive sample and push the features of the anchor sample further away from the negative sample.

[0027] Optionally, the predicted fuel consumption rate for the next time step is output via the prediction head, including:

[0028] Concatenate the feature vectors of all graph nodes output by multiple spatiotemporal convolutional blocks;

[0029] The concatenated feature vectors are input into a fully connected layer for regression, and the predicted fuel consumption rate for the next time step is output.

[0030] A second aspect of the present invention provides a fuel consumption rate prediction device for excavators based on spatiotemporal graph convolution and multi-condition comparative learning, comprising:

[0031] The energy flow spatiotemporal graph construction unit is used to construct the excavator energy flow spatiotemporal graph. It selects the core components of the excavator as graph nodes, constructs graph edges based on the energy flow paths and physical connections between components, and generates a weighted adjacency matrix. The initial features of each graph node include the original sensor data of the corresponding component.

[0032] The physical feature calculation unit is used to calculate the physical derived features of the engine side, hydraulic system actuator side and system level in real time, and fuse the physical derived features with the original sensor data as the input features of the graph nodes.

[0033] The fusion model construction and training unit is used to build a spatiotemporal graph convolutional and contrastive learning fusion model. The fusion model is based on a spatiotemporal graph convolutional network, with the addition of a projection head and a contrastive learning module. A weighted total loss function is designed to train and optimize the fusion model.

[0034] The fuel consumption rate prediction unit is used to input test data into the trained fusion model and output the fuel consumption rate prediction value for the next time step through the prediction head.

[0035] A third aspect of this invention provides a fuel consumption rate prediction device for excavators based on spatiotemporal graph convolution and multi-condition comparative learning, comprising:

[0036] One or more processors;

[0037] A memory on which one or more programs are stored;

[0038] When the one or more programs are executed by the one or more processors, the one or more processors implement the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning as described above.

[0039] The fourth aspect of the present invention provides a computer storage medium for storing a program, which, when executed, is used to implement the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning as described in any of the preceding claims.

[0040] Beneficial effects:

[0041] Predictive model innovation: By constructing a convolutional network with a spatiotemporal graph using an energy flow graph, the system simultaneously learns the spatial characteristics of the mutual influence between components and the temporal characteristics of the changes in the state of the components themselves, thus accurately capturing the spatiotemporal coupling relationship of the excavator operation system.

[0042] Application of physical features: Real-time calculation of physical features such as engine indicated power, hydraulic cylinder instantaneous power, and total system demand power transforms raw sensor signals into high-level energy-related features, providing physical clues to the "reasons for fuel consumption changes" for the model and improving prediction accuracy and interpretability.

[0043] Enhanced generalization of operating conditions: By employing a supervised contrastive learning strategy, K-Means clustering is used to obtain operating mode labels, construct positive and negative sample pairs, guide the model to learn the essential characteristics of different operating conditions, reduce the sensitivity to small fluctuations within the operating conditions, and significantly improve the robustness to unknown operating modes.

[0044] High engineering applicability: Based on mature technologies, it combines innovation with prediction accuracy and engineering feasibility, providing precise guidance for energy-saving control of excavators and has significant engineering application value. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating a method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning, provided in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the excavator energy flow spatiotemporal diagram provided in an embodiment of the present invention;

[0048] Figure 3 This is a block diagram of the spatiotemporal graph convolution and contrastive learning fusion model provided in an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of an excavator fuel consumption rate prediction device based on spatiotemporal graph convolution and multi-condition comparative learning, provided in an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0051] This invention provides a method and device for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning. By introducing key component node weight information, physical feature engineering deepening, and comparative learning strategies, it achieves accurate prediction of excavator fuel consumption rate, improving the accuracy and reliability of excavator fuel consumption rate prediction.

[0052] See Figure 1 This figure is a flowchart illustrating a method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning, provided by an embodiment of the present invention. Raw sensor data from various core components of the excavator are collected, including engine speed and torque, main hydraulic pump outlet pressure and flow rate, rodless chamber pressure, rod chamber pressure and displacement of each hydraulic cylinder, and inlet and outlet pressure and flow rate of the swing motor; the collected data are then linearly interpolated to fill missing values. Outliers were removed and min-max standardization was performed to obtain a standardized raw dataset. The engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor were selected as graph nodes; initial weights for the adjacency matrix were set based on energy transfer efficiency analysis results. Physically derived features were calculated in real-time based on the preprocessed raw data, including engine indicated power, instantaneous hydraulic power of each hydraulic cylinder, swing motor power, total system power demand, and engine load rate. The original sensor features were fused with the calculated physically derived features to form the input feature matrix for each graph node, which served as the input data for subsequent models. Core physically derived features such as total system power demand, engine load rate, and instantaneous hydraulic power of each actuator were selected to form a clustering input matrix. The K-Means unsupervised clustering algorithm was used to cluster the training data, and the optimal number of clusters was determined through silhouette coefficient verification, resulting in labels for four operating modes: heavy-load excavation, light-load excavation, swing, and idling. Anchor samples were randomly selected from time windows in the training set. One positive sample was selected from other time periods of the same work mode, and two negative samples were selected from each of the three other work modes, constructing a sample pair with an anchor sample:positive sample:negative sample ratio of 1:1:6 to provide training data for contrastive learning. A fusion model based on a spatiotemporal graph convolutional network was constructed, with a projection head and a contrastive learning module added. The node feature matrix was input into ST-GCN for spatiotemporal feature learning, and the features were mapped to the contrastive learning space through the projection head. A total loss function composed of a weighted average mean squared error prediction loss and an InfoNCE contrastive loss was designed to train and optimize the model. Test data was input into the trained model. The feature vectors of all graph nodes output from multiple spatiotemporal convolutional blocks were concatenated and input into a fully connected layer for regression, outputting the predicted fuel consumption rate for the next time step. The Informer model and the CatBoost model were selected as baselines for comparison, and mean squared error, mean absolute error, and coefficient of determination were used as validation metrics. At the same time, the predictive performance of the model under unseen combined work conditions was tested to verify the model's generalization ability.

[0053] The excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning provided in this embodiment of the invention can be implemented, for example, through the following steps S101-104.

[0054] S101: Construct a spatiotemporal diagram of excavator energy flow.

[0055] In this embodiment of the invention, core components of an excavator, including the engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor, are selected as graph nodes. Graph edges are constructed based on the energy flow paths and physical connections between components, generating a weighted adjacency matrix. The initial features of each graph node include the original sensor data of the corresponding component. The initial weights of the adjacency matrix are set according to the energy transfer efficiency analysis results. The initial weights reflect the connection strength or energy transfer efficiency between graph nodes and are optimized through fine-tuning during subsequent model training.

[0056] Specifically, a sensor system is deployed on the excavator to collect operational data from various core components: engine speed, torque, and fuel consumption; main hydraulic pump outlet pressure and flow rate; rodless chamber pressure, rod chamber pressure, and displacement of the boom cylinder, stick cylinder, and bucket cylinder; and inlet and outlet pressure and flow rate of the swing motor. Simultaneously, operational action signals and engine fuel consumption rate data are collected. The collected data undergoes preprocessing: missing values ​​are filled using linear interpolation, and... Outliers are removed using the criteria, and the data is mapped to the [0,1] interval through min-max standardization to obtain a standardized original dataset. This invention defines each core component of the excavator's working system as a graph node, and the physical connections between components as graph edges, constructing an excavator energy flow graph. The specific implementation is as follows:

[0057] (1) Node definition: The engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder and swing motor are selected as core nodes. The initial characteristics of each node include the original sensor data of the corresponding components, specifically the engine speed and torque, the outlet pressure and flow of the main hydraulic pump, the rodless chamber pressure, rod chamber pressure and displacement of each hydraulic cylinder, and the inlet and outlet pressure and flow of the swing motor.

[0058] (2) Edge and Adjacency Matrix Definition: Edges are constructed based on the energy flow paths and physical connections between components, including the mechanical transmission connection between the engine and the main hydraulic pump, and the hydraulic pipeline connection between the main hydraulic pump and each hydraulic cylinder; the adjacency matrix is ​​represented by weights, with the initial weights set according to the energy transfer efficiency during excavator operation (these weights can be adjusted according to the specific machine model and working mode, and this invention does not specifically limit them). The energy transfer efficiency is obtained based on relevant empirical values ​​and physical calculations. For example, the ratio of the main pump input power to the engine output power is the energy transfer efficiency between them, and the efficiency value is related to the performance of the engine and the main pump. Figure 2 As shown, Figure 2The diagram above is a schematic diagram of the excavator energy flow spatiotemporal diagram provided in the embodiment of the present invention. Based on the energy transfer efficiency analysis results, the initial weights of the adjacency matrix are set as follows: the weight of engine → main hydraulic pump is set to 0.8, the weights of main hydraulic pump → boom cylinder, stick cylinder, bucket cylinder and swing motor are set to 0.58, 0.50, 0.48 and 0.62 respectively, and the weights of non-directly connected components are set to 0.

[0059] S102: Design a physical characteristic engineering scheme.

[0060] In this embodiment of the invention, physical derived features are calculated in real time at the engine side, hydraulic system actuator side, and system level. These physical derived features are then fused with the original sensor data and used as input features for graph nodes. The physical derived features include: engine indicated power at the engine side; instantaneous hydraulic power of each hydraulic cylinder and rotary motor power at the hydraulic system actuator side; and total system power demand and engine load rate at the system level.

[0061] Specifically, by embedding real-time computed physical derivative features into the spatiotemporal graph node features, the depth of feature engineering is deepened, transforming raw sensor signals into high-level semantic features reflecting energy conversion states. Based on the preprocessed raw data, the physical derivative features of each node are computed in real time, including:

[0062] (1) Calculate the physical characteristics of the engine side: calculate the indicated power and fuel consumption rate. The formula for calculating the indicated power is: indicated power = 2π × speed (rpm) / 60 × torque (N·m). It is the output power of the engine, which is mainly converted from the internal energy of fuel combustion. It is a physical characteristic of the engine side that is closely related to the fuel consumption rate of the engine.

[0063] (2) Calculate the physical characteristics of the hydraulic system actuator side: calculate the instantaneous hydraulic power of each hydraulic cylinder, where the instantaneous hydraulic power calculation formula is: instantaneous hydraulic power = |(rodless chamber pressure - rod chamber pressure) × flow rate, rotary motor power = rotary motor inlet and outlet pressure difference × flow rate. They are the actuators of the engine power and also have an implicit correlation with the fuel consumption rate.

[0064] (3) Calculate the physical characteristics of the system: calculate the total power demand and load rate of the system. The formula for calculating the total power demand is: Total power demand = instantaneous hydraulic power of boom cylinder + instantaneous hydraulic power of stick cylinder + instantaneous hydraulic power of bucket cylinder + power of swing motor + power of auxiliary system (auxiliary system refers to the engine cooling, hydraulic system cooling and other related systems during the operation of the excavator) estimated power; The formula for calculating the load rate is: Load rate = actual engine torque / maximum engine torque at the corresponding speed.

[0065] This invention, based on preprocessed raw data, calculates the physical derived features of each node in real time. It employs a fusion method that directly concatenates feature dimensions (i.e., original sensor features and physical derived features are concatenated column-wise, with unused feature dimensions filled with 0 to ensure uniform feature dimensions across all nodes). The fusion result gives each node an input feature matrix with a dimension of "time step × total number of features." All nodes are then aggregated to form a three-dimensional input data set with a dimension of "number of nodes × time step × total number of features" (in this invention, the total number of features is set to 15, including 8 original sensor features and 7 physical derived features; dimension uniformity ensures consistent model input format). An example of the input feature matrix for 6 nodes is shown in Table 1 (taking a time step of 50 as an example, dimension: 50 × 15).

[0066] Table 1 Example of node input feature matrix

[0067]

[0068] S103: Construct a fusion model of spatiotemporal graph convolution and contrastive learning.

[0069] In this embodiment of the invention, a spatiotemporal graph convolutional network is used as the main body, with the addition of a projection head and a contrastive learning module. A weighted total loss function is designed to train and optimize the fusion model. The spatiotemporal graph convolutional network is composed of multiple stacked spatiotemporal convolutional blocks. Each spatiotemporal convolutional block includes spatial graph convolution and temporal one-dimensional convolution. The spatial graph convolution is used to aggregate the feature information of adjacent graph nodes, and the temporal one-dimensional convolution is used to capture the temporal state changes of the graph nodes themselves.

[0070] The contrastive learning module adopts a supervised contrastive learning approach, which is implemented through the following steps: the K-Means unsupervised clustering algorithm is used to cluster the training set data to obtain multiple job mode labels; a data segment of a time window is selected from the training set as the anchor sample, positive samples are selected from other time periods of the same job mode, and multiple negative samples are selected from time periods of different job modes to construct sample pairs consisting of anchor samples, positive samples, and negative samples.

[0071] The weighted total loss function is obtained by weighted summation of prediction loss and contrast loss. The prediction loss adopts mean squared error regression loss to ensure the prediction accuracy of fuel consumption rate. The contrast loss adopts InfoNCE loss to bring the features of anchor samples closer to those of the positive samples and push the features of anchor samples further away from those of the negative samples.

[0072] Specifically, the main body of the model adopts a spatiotemporal graph convolutional network (ST-GCN), combined with a supervised contrastive learning strategy to improve the generalization ability under various conditions. The structural diagram of the spatiotemporal graph convolutional and contrastive learning fusion model is as follows: Figure 3 As shown, the specific structure and training method are as follows:

[0073] The ST-GCN core module consists of multiple stacked spatiotemporal convolutional blocks, each including spatial and temporal convolutions. The initial weights of the adjacency matrix are set to the energy transfer efficiency of the excavator (these weights can be fine-tuned and optimized during model training; they are not fixed values ​​or purely data-driven). The ST-GCN core module includes:

[0074] Spatial convolution uses graph convolution (GCN) to aggregate the feature information of adjacent nodes and simulate the mutual influence between components. For example, when calculating the features of the main hydraulic pump node, the features of the engine and each hydraulic cylinder node are aggregated. Temporal convolution uses one-dimensional convolution (1D-CNN) to convolve the historical features of each node along the time axis to capture the temporal state changes of the node itself.

[0075] The process of this step includes inputting the feature matrix of each node at T time steps, updating the node features through spatial dimensional graph convolution, capturing temporal evolution through temporal convolution, and outputting the updated node features. This process is repeated to complete multiple rounds of spatiotemporal feature learning.

[0076] The contrastive learning module automatically obtains job pattern labels through K-Means unsupervised clustering and configures a customized 2-layer MLP projection head to adapt to the high-dimensional spatiotemporal features output by ST-GCN. Supervised contrastive learning is employed, and the specific implementation is as follows:

[0077] Data preparation and augmentation: A time window segment in the training data is selected as the anchor sample. Positive samples are selected from other time periods with the same job mode, and multiple negative samples are selected from time periods with different job modes. The job mode labels are automatically obtained through K-Means unsupervised clustering.

[0078] Projection head design: A small MLP is added after ST-GCN as a projection head to map the features output by ST-GCN to the contrast learning space;

[0079] Loss function construction: The total loss function is a weighted average of the prediction loss and the contrastive loss, as shown in the formula:

[0080] ;

[0081] in, The MSE regression loss is used to ensure prediction accuracy; The InfoNCE loss is used to bring anchor samples closer to positive sample features and push them further away from negative sample features.

[0082] This invention employs a dual-branch parallel architecture, sharing the ST-GCN feature extraction layer. The features output by ST-GCN flow to the prediction branch (after feature flattening and a fully connected layer to achieve fuel consumption regression) and the contrast branch (after system-level feature pooling and projection head mapping to the contrast space). The total loss function is the weighted sum of the MSE prediction loss and the InfoNCE contrast loss. During backpropagation, all parameters of the ST-GCN, prediction head, and projection head are updated synchronously to achieve synergistic optimization of prediction accuracy and generalization ability. Six core components, including the engine and main hydraulic pump, are selected as graph nodes. The original sensor data of each component and physical derived features such as engine indicated power, hydraulic cylinder instantaneous hydraulic power, and total system demand power are spliced ​​and fused according to dimensions to form an input matrix of [number of nodes × time step × number of features]. In the ST-GCN layer, this fused feature aggregates the physical coupling relationship between nodes through spatial convolution and captures temporal dynamic changes through temporal convolution. At the same time, the physical derived features inject domain mechanism knowledge into the model, improving the interpretability of predictions. In the contrastive learning branch, the working condition discriminativeness of the physical derived features (such as similar load rate and total system demand power distribution of the same working condition sample) is used to construct positive / negative sample pairs to guide the model to learn the essential features of the working condition, effectively improving the generalization ability and adapting to the multi-working condition changes of the excavator.

[0083] In this embodiment of the invention, model training and optimization includes dataset partitioning, job model clustering, construction of contrastive learning samples, model parameter setting, and training process, as detailed below:

[0084] Dataset partitioning: Divide the fused feature data into training set, validation set and test set in a 7:2:1 ratio. The time window can be set to, for example, 5 seconds (sampling frequency 100Hz, i.e. 500 time steps).

[0085] Job mode clustering: The K-Means algorithm is used to cluster the training set data to obtain four job mode labels: "Heavy load mining", "Light load mining", "Rotation", and "Idle". The number of clusters is determined by silhouette coefficient verification. The K-Means unsupervised clustering algorithm is also used to cluster the training set data, focusing on "physically derived features". The specific clustering process, category differences, and job mode label adaptation logic are as follows:

[0086] (1) Specific clustering process

[0087] 1. Clustering feature selection: Five core physical derivative features were selected, including total system power demand, engine load rate, and instantaneous hydraulic power of each actuator (boom cylinder, stick cylinder, etc.), to form a clustering input feature matrix (dimension: number of training set samples × 5). The basis for this selection is that physical derivative features can directly map the working intensity (e.g., the total system power demand is significantly higher when under heavy load than when idling), and have better working condition differentiation than the original sensor data.

[0088] 2. Data preprocessing: Standardize the clustering feature matrix to eliminate the impact of dimensional differences on the clustering results.

[0089] 3. Clustering execution: The K-Means algorithm is used to iteratively optimize the cluster centers. The initial cluster centers are initially determined by the "elbow rule". The optimal number of clusters is then verified by the silhouette coefficient. When the number of clusters is 4, the silhouette coefficient is the highest (close to 0.8), indicating that the clustering effect is optimal. Finally, four operation mode labels are obtained: "heavy-load mining", "light-load mining", "rotation", and "idle".

[0090] 4. Clustering result verification: By statistically analyzing the physical feature distribution thresholds of each pattern, we ensure that the differences between categories are significant and there is no overlap, providing a reliable basis for distinguishing working conditions for subsequent comparative learning.

[0091] (2) The core differences between the four operating modes

[0092] As shown in Table 2, this table shows the core differences between the four operating modes based on the distribution of physical characteristics.

[0093] Table 2. Core differences among the four operating modes

[0094]

[0095] (3) Adaptation logic for different work modes

[0096] 1. Provide "operating condition differentiation criteria" for comparative learning: Labels are directly used to construct comparative learning samples - data with the same label are "same operating conditions" (consistent distribution of physical characteristics) and are used as positive samples; data with different labels are "different operating conditions" (significant differences in distribution of physical characteristics) and are used as negative samples, ensuring that comparative learning can specifically learn the "essential characteristics of operating conditions" (such as the difference in energy demand between heavy load and idling), rather than the minor fluctuations within the operating conditions.

[0097] 2. Adapting to dynamic changes in multiple working conditions of excavators: The model indirectly perceives the energy demand patterns of different working modes through labels. During the training process, the contrastive loss guides ST-GCN to prioritize the extraction of "core features that are invariant across working conditions" (such as energy transfer efficiency correlation features) and reduce the sensitivity to "noise specific to a single working condition" (such as instantaneous pressure fluctuations under heavy load), thereby improving the generalization ability to unknown working conditions (such as compound operations).

[0098] 3. Strengthen the synergy between physical mechanisms and data-driven approaches: The physical feature thresholds corresponding to the labels allow the construction of comparative learning samples to no longer rely on human experience, but to be based on the energy conversion law of excavators. This ensures that the features learned by the model not only conform to the data distribution but also fit the physical mechanism, further improving the interpretability of the prediction results.

[0099] In this invention, the work mode clustering is not an independent step, but a core pre-step of the "multi-work condition comparative learning" strategy of this invention. It transforms dynamic and changing work scenarios into quantifiable and distinguishable work condition labels through unsupervised clustering, and then uses comparative learning to adapt the model to the feature differences of different work conditions, ultimately solving the core pain point of existing models that are "limited in generalization ability and difficult to adapt to changes in multiple work conditions".

[0100] Contrast learning sample construction: Randomly select time window data from the training set as anchor samples, select 1 positive sample from the data of the same operation mode, and select 2 negative samples from the data of the other 3 operation modes to form contrast learning sample pairs with an anchor sample: positive sample: negative sample = 1:1:6, to ensure that the samples cover all working conditions.

[0101] Model parameter settings: ST-GCN contains 3 spatiotemporal convolutional blocks, each with a spatial dimension graph convolutional kernel size of 3 and a temporal 1D-CNN convolutional kernel size of 5; the projection head is a 2-layer MLP (128 hidden neurons); the loss function weight λ=0.1; the optimizer is Adam, with an initial learning rate of 0.001, which decays to 0.8 every 10 epochs; the training epochs are set to 100, and the batch size is 32.

[0102] Training process: Input the node feature matrix into ST-GCN for spatiotemporal feature learning, map it to the contrast learning space through the projection head, calculate the total loss function, and optimize the model parameters through backpropagation; adjust the hyperparameters using the validation set, and stop training when the validation set MSE loss does not decrease for 5 consecutive epochs, and save the optimal model.

[0103] S104: Input the test data into the trained fusion model and output the fuel consumption rate prediction value for the next time step through the prediction head.

[0104] In this embodiment of the invention, the feature vectors of all graph nodes output by multiple spatiotemporal convolutional blocks are concatenated; the concatenated feature vectors are input into a fully connected layer for regression, and the predicted fuel consumption rate for the next time step is output.

[0105] Specifically, the feature vectors of all nodes output by multiple ST-GCN blocks are concatenated and input into a fully connected layer for regression, outputting the predicted fuel consumption rate for the next time step.

[0106] In one implementation of this invention, the model performance is verified by selecting the Informer model and the CatBoost model as the comparison baseline. The verification indicators include MSE (mean squared error), MAE (mean absolute error), and R² (coefficient of determination). At the same time, the predictive performance of the model under the unseen "composite operation" (digging + rotation) is tested to verify the generalization ability.

[0107] Beneficial effects:

[0108] 1. Innovative predictive model for accurately capturing system coupling relationships: By constructing energy flow maps and modeling with ST-GCN, it learns both spatial features (inter-component influences) and temporal features (component state changes), which better matches the working mechanism of the complex mechanical system of excavators and solves the problem that traditional methods ignore the physical connection and spatiotemporal coupling relationship of the system;

[0109] 2. Deepen feature engineering and inject domain physics knowledge: By calculating the physical derivative features of the engine, hydraulic system and whole vehicle in real time, the original sensor signals are transformed into high-level energy-related features, providing the model with physical clues to the "reasons for fuel consumption changes", improving prediction accuracy and interpretability, and solving the problems of shallow feature engineering and lack of physical meaning in existing methods;

[0110] 3. Enhance generalization ability and adapt to actual engineering needs: By using supervised comparative learning strategies, the model is guided to learn the essential characteristics of different working conditions, reducing the sensitivity to small fluctuations within the same working condition, and improving the robustness to unknown operating modes. This solves the problem that the existing model has limited generalization ability and is difficult to adapt to the changes in multiple working conditions of excavators.

[0111] 4. Practical and efficient technical solutions: Based on mature technologies such as ST-GCN and contrastive learning, the solutions combine and innovate to avoid the accumulation of obscure models, balance prediction accuracy and engineering feasibility, and specifically address the challenges of balancing "input data selection" and "model selection" pointed out in existing reviews.

[0112] Based on the methods provided in the above embodiments, this invention also provides a fuel consumption rate prediction device for excavators based on spatiotemporal graph convolution and multi-condition comparative learning. The following describes the fuel consumption rate prediction device for excavators based on spatiotemporal graph convolution and multi-condition comparative learning with reference to the accompanying drawings.

[0113] See Figure 4 The figure is a schematic diagram of the structure of an excavator fuel consumption rate prediction device based on spatiotemporal graph convolution and multi-condition comparative learning provided in an embodiment of the present invention.

[0114] The excavator fuel consumption rate prediction device 400 based on spatiotemporal graph convolution and multi-condition comparative learning provided in this embodiment of the invention includes: an energy flow spatiotemporal graph construction unit 401, a physical feature calculation unit 402, a fusion model construction and training unit 403, and a fuel consumption rate prediction unit 404.

[0115] The energy flow spatiotemporal graph construction unit 401 is used to construct the excavator energy flow spatiotemporal graph. It selects the core components of the excavator as graph nodes, constructs graph edges based on the energy flow paths and physical connections between components, and generates a weighted adjacency matrix. The initial features of each graph node include the original sensor data of the corresponding component.

[0116] The physical feature calculation unit 402 is used to calculate the physical derived features of the engine side, hydraulic system actuator side and system level in real time, and fuse the physical derived features with the original sensor data as the input features of the graph nodes.

[0117] The fusion model construction and training unit 403 is used to construct a spatiotemporal graph convolutional and contrastive learning fusion model. The fusion model is based on a spatiotemporal graph convolutional network, with the addition of a projection head and a contrastive learning module. A weighted total loss function is designed to train and optimize the fusion model.

[0118] The fuel consumption rate prediction unit 404 is used to input test data into the trained fusion model and output the fuel consumption rate prediction value for the next time step through the prediction head.

[0119] In one possible implementation, the energy flow spacetime graph construction unit 401 is specifically used for:

[0120] The engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor are selected as the drawing nodes;

[0121] The initial weights of the adjacency matrix are set based on the energy transfer efficiency analysis results. The initial weights reflect the strength of the connection between graph nodes or the energy transfer efficiency, and are optimized through fine-tuning during subsequent model training.

[0122] In one possible implementation, the physically derived features include:

[0123] Engine power indication on the engine side;

[0124] Instantaneous hydraulic power of each hydraulic cylinder on the actuator side of the hydraulic system and power of the rotary motor;

[0125] System-level total power demand and engine load rate.

[0126] In one possible implementation, the spatiotemporal graph convolutional network is composed of multiple spatiotemporal convolutional blocks stacked together. Each spatiotemporal convolutional block includes spatial-dimensional graph convolution and temporal one-dimensional convolution. The spatial-dimensional graph convolution is used to aggregate the feature information of adjacent graph nodes, and the temporal one-dimensional convolution is used to capture the temporal state changes of the graph nodes themselves.

[0127] In one possible implementation, the fusion model building and training unit 403 is specifically used for:

[0128] The K-Means unsupervised clustering algorithm is used to cluster the training set data to obtain multiple job mode labels;

[0129] Select a data segment of a time window from the training set as the anchor sample, select positive samples from other time periods of the same work mode, and select multiple negative samples from time periods of different work modes to construct a sample pair consisting of the anchor sample, positive sample, and negative sample.

[0130] In one possible implementation, the weighted total loss function is obtained by weighted summation of the prediction loss and the contrast loss. The prediction loss adopts mean squared error regression loss to ensure the prediction accuracy of fuel consumption rate; the contrast loss adopts InfoNCE loss to bring the features of anchor samples closer to those of positive samples and push the features of anchor samples further away from those of negative samples.

[0131] In one possible implementation, the fuel consumption rate prediction unit 404 is specifically used for:

[0132] Concatenate the feature vectors of all graph nodes output by multiple spatiotemporal convolutional blocks;

[0133] The concatenated feature vectors are input into a fully connected layer for regression, and the predicted fuel consumption rate for the next time step is output.

[0134] Since the excavator fuel consumption rate prediction device 400 based on spatiotemporal graph convolution and multi-condition comparative learning is a device corresponding to the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning provided in the above method embodiments, the specific implementation of each unit of the excavator fuel consumption rate prediction device 400 based on spatiotemporal graph convolution and multi-condition comparative learning is based on the same concept as in the above method embodiments. Therefore, for the specific implementation of each unit of the excavator fuel consumption rate prediction device 400 based on spatiotemporal graph convolution and multi-condition comparative learning, please refer to the description of the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning in the above method embodiments, and will not be repeated here.

[0135] This invention also provides a device for predicting the fuel consumption rate of excavators based on spatiotemporal graph convolution and multi-condition comparative learning, the device comprising: a processor and a memory;

[0136] The memory is used to store instructions;

[0137] The processor is configured to execute the instructions in the memory to perform the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning mentioned in the above embodiments.

[0138] It should be noted that the excavator fuel consumption rate prediction device based on spatiotemporal graph convolution and multi-condition comparative learning provided in the embodiments of the present invention can all have the following hardware structure: Figure 5 The structure shown, Figure 5 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention.

[0139] Please see Figure 5 As shown, device 500 includes: processor 510, communication interface 520, and memory 530. The number of processors 510 in device 500 can be one or more. Figure 5 Taking a processor as an example, in this embodiment of the invention, the processor 510, communication interface 520, and memory 530 can be connected via a bus system or other means. Figure 5 Taking the connection between China and Israel via the 540 bus system as an example.

[0140] Processor 510 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 510 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0141] Memory 530 may include volatile memory, such as random-access memory (RAM); memory 530 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); memory 530 may also include a combination of the above types of memory.

[0142] Optionally, the memory 530 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic business operations and handling hardware-based tasks. The processor 510 can read the programs in the memory 530 to implement the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning provided in this embodiment of the invention.

[0143] The bus system 540 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system 540 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0144] This invention also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning mentioned in the above embodiments.

[0145] This invention also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning mentioned in the above embodiments.

[0146] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning, characterized in that, The method includes: A spatiotemporal graph of excavator energy flow is constructed by selecting core components of the excavator as graph nodes, constructing graph edges based on the energy flow paths and physical connections between components, and generating a weighted adjacency matrix. The initial features of each graph node include the original sensor data of the corresponding component. The construction of the excavator energy flow spatiotemporal graph includes: selecting the engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor as graph nodes; setting the initial weights of the adjacency matrix based on the energy transfer efficiency analysis results. The initial weights reflect the connection strength or energy transfer efficiency between the graph nodes and are optimized through fine-tuning during subsequent model training. The physical feature engineering scheme is designed to calculate the physical derived features on the engine side, hydraulic system actuator side, and system level in real time. The physical derived features are then fused with the original sensor data and used as the input features of the graph nodes. The physical derived features include the engine indicated power on the engine side; the instantaneous hydraulic power of each hydraulic cylinder and the power of the rotary motor on the hydraulic system actuator side; and the total system demand power and engine load rate at the system level. A spatiotemporal graph convolutional and contrastive learning fusion model is constructed. The spatiotemporal graph convolutional network is the main body, and a projection head and a contrastive learning module are added. A weighted total loss function is designed to train and optimize the fusion model. The test data is input into the trained fusion model, and the prediction head outputs the fuel consumption rate prediction for the next time step.

2. The method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning according to claim 1, characterized in that, The spatiotemporal graph convolutional network is composed of multiple spatiotemporal convolutional blocks stacked together. Each spatiotemporal convolutional block includes a spatial-dimensional graph convolution and a temporal one-dimensional convolution. The spatial-dimensional graph convolution is used to aggregate the feature information of adjacent graph nodes, and the temporal one-dimensional convolution is used to capture the temporal state changes of the graph nodes themselves.

3. The method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning according to claim 1, characterized in that, The contrastive learning module employs a supervised contrastive learning approach, implemented through the following steps: The K-Means unsupervised clustering algorithm is used to cluster the training set data to obtain multiple job mode labels; Select a data segment of a time window from the training set as the anchor sample, select positive samples from other time periods of the same work mode, and select multiple negative samples from time periods of different work modes to construct a sample pair consisting of the anchor sample, the positive sample, and the negative sample.

4. The excavator fuel consumption rate prediction method based on spatiotemporal graph convolution and multi-condition comparative learning according to claim 3, characterized in that, The weighted total loss function is obtained by weighted summation of prediction loss and contrast loss. The prediction loss adopts mean squared error regression loss to ensure the prediction accuracy of fuel consumption rate. The contrast loss adopts InfoNCE loss to bring the features of the anchor sample closer to those of the positive sample and push the features of the anchor sample further away from those of the negative sample.

5. The method for predicting excavator fuel consumption rate based on spatiotemporal graph convolution and multi-condition comparative learning according to claim 1, characterized in that, The step of outputting the predicted fuel consumption rate for the next time step via the prediction head includes: Concatenate the feature vectors of all graph nodes output by multiple spatiotemporal convolutional blocks; The concatenated feature vectors are input into a fully connected layer for regression, and the predicted fuel consumption rate for the next time step is output.

6. A fuel consumption rate prediction device for excavators based on spatiotemporal graph convolution and multi-condition comparative learning, characterized in that, include: The energy flow spatiotemporal graph construction unit is used to construct the excavator energy flow spatiotemporal graph. It selects the core components of the excavator as graph nodes, constructs graph edges based on the energy flow paths and physical connections between components, and generates a weighted adjacency matrix. The initial features of each graph node include the original sensor data of the corresponding component. The energy flow spatiotemporal graph construction unit is specifically used to select the engine, main hydraulic pump, boom cylinder, stick cylinder, bucket cylinder, and swing motor as graph nodes; the initial weights of the adjacency matrix are set according to the energy transfer efficiency analysis results. The initial weights reflect the connection strength or energy transfer efficiency between the graph nodes and are optimized through fine-tuning during subsequent model training. The physical feature calculation unit is used to calculate the physical derived features of the engine side, hydraulic system actuator side, and system level in real time, and fuse the physical derived features with the original sensor data as the input features of the graph node. The physical derived features include the engine indicated power on the engine side; the instantaneous hydraulic power of each hydraulic cylinder and the power of the rotary motor on the hydraulic system actuator side; and the total system demand power and engine load rate at the system level. The fusion model construction and training unit is used to construct a spatiotemporal graph convolutional and contrastive learning fusion model. The fusion model is based on a spatiotemporal graph convolutional network, with the addition of a projection head and a contrastive learning module. A weighted total loss function is designed to train and optimize the fusion model. The fuel consumption rate prediction unit is used to input test data into the trained fusion model and output the fuel consumption rate prediction value for the next time step through the prediction head.

7. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-5 above.

Images