CN-121834241-B - Excavator fuel consumption rate prediction method and device based on space-time diagram convolution and multi-working condition comparison learning

CN121834241BCN 121834241 BCN121834241 BCN 121834241BCN-121834241-B

Abstract

The invention discloses a method and a device for predicting the fuel consumption rate of an excavator based on space-time diagram convolution and multi-task comparison learning. In the method, an energy flow space-time diagram of an excavator is built, a core component is defined as a diagram node, physical connection is defined as a weighted edge, a multi-level physical characteristic engineering scheme is designed, original sensor data and energy conversion class derivative characteristics are fused, a space-time diagram convolution network (ST-GCN) and supervision comparison learning fusion model is built again, and prediction precision and generalization capability are improved through double-task training. According to the invention, the node information of the working system of the excavator is associated by combining the space-time diagram convolution network, and the physical mechanism knowledge is injected, so that the adaptability of the model to complex working conditions is obviously improved, the prediction precision and generalization capability are superior to those of the existing model, the accurate guidance can be provided for the energy-saving control of the excavator, and the method has important engineering application value.

Inventors

Su Deying
LIAN XIAOZHEN
YU DEXIN
Chen Jinqu

Assignees

集美大学

Dates

Publication Date: 20260512
Application Date: 20260313

Claims (8)

1. The excavator fuel consumption rate prediction method based on space-time diagram convolution and multi-working condition comparison learning is characterized by comprising the following steps: Constructing an energy flow space-time diagram of the excavator, selecting a core component of the excavator as a diagram node, constructing a diagram edge based on an energy flow path and physical connection among the components, and generating a weighted adjacent matrix, wherein the initial characteristic of each diagram node comprises original sensor data of a corresponding component; the construction of the excavator energy flow time-space diagram comprises the steps of selecting an engine, a main hydraulic pump, a movable arm cylinder, a bucket rod cylinder, a bucket cylinder and a rotary motor as diagram nodes, setting initial weights of an adjacent matrix according to an energy transfer efficiency analysis result, wherein the initial weights reflect connection strength or energy transfer efficiency between the diagram nodes and are optimized through fine adjustment in a subsequent model training process; Designing a physical characteristic engineering scheme, and calculating physical derivative characteristics of an engine side, a hydraulic system executing mechanism side and a system level in real time, wherein the physical derivative characteristics are fused with the original sensor data and then are used as input characteristics of the graph nodes, and the physical derivative characteristics comprise engine indication power of the engine side; constructing a space-time diagram convolution and contrast learning fusion model, taking a space-time diagram convolution network as a main body, adding a projection head and a contrast learning module, and designing a weighted total loss function to train and optimize the fusion model; and inputting the test data into the fusion model after training, and outputting a predicted value of the fuel consumption rate of the next time step through a prediction head.
2. The excavator fuel consumption rate prediction method based on space-time graph convolution and multi-task comparison learning according to claim 1, wherein the space-time graph convolution network is formed by stacking a plurality of space-time convolution blocks, each space-time convolution block comprises a space-dimensional graph convolution and a time-dimensional one-dimensional convolution, the space-dimensional graph convolution is used for aggregating characteristic information of adjacent graph nodes, and the time-dimensional one-dimensional convolution is used for capturing time sequence state changes of the graph nodes.
3. The excavator fuel consumption rate prediction method based on space-time diagram convolution and multi-task comparison learning according to claim 1, wherein the comparison learning module adopts a supervision comparison learning mode and is realized by the following steps: clustering the training set data by adopting a K-Means unsupervised clustering algorithm to obtain a plurality of operation mode labels; Selecting a data segment of a time window from a training set as an anchor point sample, selecting positive samples from other time segments of the same operation mode, selecting a plurality of negative samples from time segments of different operation modes, and constructing a sample pair consisting of the anchor point sample, the positive samples and the negative samples.
4. The excavator fuel consumption rate prediction method based on space-time diagram convolution and multi-task comparison learning according to claim 3, wherein the weighted total loss function is obtained by weighted summation of prediction loss and comparison loss, the prediction loss adopts mean square error regression loss for ensuring prediction accuracy of fuel consumption rate, and the comparison loss adopts InfoNCE loss for pulling up the characteristics of the anchor point sample and the positive sample and pushing away the characteristics of the anchor point sample and the negative sample.
5. The method for predicting the fuel consumption rate of the excavator based on space-time diagram convolution and multi-task comparison learning according to claim 1, wherein the outputting of the predicted fuel consumption rate value of the next time step through the prediction head comprises: Splicing the feature vectors of all the graph nodes output by the space-time convolution blocks; And inputting the spliced feature vectors into a full-connection layer for regression, and outputting the predicted value of the fuel consumption rate of the next time step.
6. Excavator fuel consumption rate prediction device based on space-time diagram convolution and multi-working condition contrast study, which is characterized by comprising: The energy flow time space diagram construction unit is used for constructing an energy flow time space diagram of the excavator, selecting a core component of the excavator as a diagram node, constructing a diagram edge based on an energy flow path and physical connection among the components, and generating a weighted adjacent matrix, wherein the initial characteristic of each diagram node comprises the original sensor data of the corresponding component; The energy flow time-space diagram construction unit is specifically used for selecting an engine, a main hydraulic pump, a movable arm cylinder, a bucket rod cylinder, a bucket cylinder and a rotary motor as diagram nodes, setting initial weights of an adjacent matrix according to an energy transfer efficiency analysis result, reflecting connection strength or energy transfer efficiency between the diagram nodes, and optimizing through fine adjustment in a subsequent model training process; The system comprises a physical characteristic calculation unit, a system-level system total demand power and an engine load rate, wherein the physical characteristic calculation unit is used for calculating physical derivative characteristics of an engine side, a hydraulic system actuator side and a system level in real time, and fusing the physical derivative characteristics with the original sensor data to be used as input characteristics of the graph nodes, wherein the physical derivative characteristics comprise engine indication power of the engine side; The fusion model construction and training unit is used for constructing a space-time diagram convolution and contrast learning fusion model, the fusion model takes a space-time diagram convolution network as a main body, a projection head and a contrast learning module are added, and a weighted total loss function is designed to train and optimize the fusion model; And the fuel consumption rate prediction unit is used for inputting the test data into the trained fusion model and outputting a fuel consumption rate predicted value of the next time step through a prediction head.
7. An electronic device, comprising a processor and a memory; the memory is used for storing instructions; the processor being configured to execute the instructions in the memory and to perform the method of any of claims 1-5.
8. A computer readable storage medium comprising instructions which, when run on a computer, cause the computer to perform the method of any of the preceding claims 1-5.

Description

Excavator fuel consumption rate prediction method and device based on space-time diagram convolution and multi-working condition comparison learning Technical Field The invention relates to the technical field of engineering machinery state monitoring and energy efficiency optimization, in particular to a method and a device for predicting the fuel consumption rate of an excavator based on space-time diagram convolution and multi-task comparison learning. Background The excavator is used as core engineering machinery, the operation environment is complex, the working condition is dynamic and changeable, the fuel consumption rate directly affects the operation cost and the environmental benefit, and therefore, the accurate prediction of the fuel consumption rate is realized and has important significance for energy-saving optimization control. At present, the data driving method is widely applied to the field of prediction of engineering machinery fuel consumption, related researches have been advanced to a certain extent, but a plurality of core technical pain points still exist under the specific application scene of the excavator, and the engineering actual requirements are difficult to meet. In the prior art, partial researches adopt a scheme of combining Variational Modal Decomposition (VMD) and Informer models, although the combined model is proved to be superior to the traditional LSTM and GRU models under complex working conditions, the importance of data preprocessing and attention mechanisms is emphasized, the VMD is only relied on for noise reduction, the characteristics with physical significance are not applied, meanwhile, a compared baseline model is more traditional, and is not compared with models which are more suitable for modeling systems such as a Graph Neural Network (GNN) and the like, and the interpretability analysis is only remained on the level of error analysis and fails to reveal the internal relation among influencing factors. In addition, the influence of different feature combinations is systematically evaluated by adopting CatBoost model and SHAP explanatory framework aiming at the fuel consumption prediction of the heavy truck to obtain the critical conclusion of engine features, but the model can only process static feature data without time sequence dependence, can not effectively capture the strong time sequence dependence and dynamic working condition change features in the working process of the excavator, and the continuous running working condition of the truck corresponding to the data source is different from repeated, intermittent and strong dynamic working cycles of the excavator, so that the direct migration application is difficult. Aiming at the problem that the mechanical oil injection system excavator cannot directly provide oil consumption data, the technology adopts a stacking regression model of a random forest and K adjacent algorithm to conduct oil consumption prediction through attribute data of the same type of electronic oil injection system excavator, and corrects the oil consumption by combining with actual oil filling amount. However, the scheme depends on the migration application of the data of the same type of electronic fuel injection machine, cannot directly capture the system coupling relation of the mechanical fuel injection system excavator, does not consider the influence of the dynamic change of the working condition on the prediction precision, and has limited generalization capability. Some researches attempt to construct a prediction model based on deep learning and support vector machine, and perform dimension reduction processing on multi-dimensional sensor data through principal component analysis so as to simplify the calculation complexity of the model. However, the method only screens the characteristics through statistical means, lacks of integration of physical mechanisms such as energy conversion and hydraulic transmission of the excavator, has insufficient depth of characteristic engineering, does not design an adaptation strategy aiming at the multi-working-condition characteristics of the excavator, and is easy to influence in prediction precision under a complex working scene. In the field of space-time sequence prediction, the prior art constructs a combined space-time correlation adjacency matrix based on a space-time graph neural network, captures space-time correlation characteristics of data, and is suitable for prediction of scenes such as traffic flow. However, the technology does not adapt to the field characteristics of the excavator oil consumption prediction, does not integrate with the physical mechanism knowledge of engineering machinery, does not consider the working condition difference of multiple working modes of the excavator, cannot be directly migrated for the accurate prediction of the fuel consumption rate, and lacks the mechanism level interpretability analysis of the prediction result. In summary, th