Oil consumption prediction method based on local-global spatio-temporal information decomposition of multivariable excavator operation data
Through the local-global spatiotemporal information decomposition method and the use of multivariate excavator operation data for feature learning, the problem of unstable excavator fuel consumption prediction was solved, high-precision fuel consumption prediction was achieved, and production efficiency and energy saving effects were improved.
Patent Information
- Application Number
- CN202510711607.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-29
- Publication Date
- 2025-09-12
AI Technical Summary
Existing technologies make it difficult to achieve high-precision fuel consumption prediction on excavators, resulting in unstable energy consumption and affecting environmental protection and resource-saving production.
Through the local-global spatiotemporal information decomposition method based on multivariate excavator operation data, sensor data is used for feature learning to extract equipment operation information, and fuel consumption prediction is performed by combining the local multi-scale window attention module, the global attention hybrid module, the time delay and similarity measurement module.
The accuracy of excavator fuel consumption prediction is improved, the delay of prediction results is reduced, and production efficiency and energy conservation and emission reduction effects are improved.
Smart Images
Figure CN120632620A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent prediction of excavator fuel consumption, and in particular to a fuel consumption prediction method based on local-global spatiotemporal information decomposition of multivariable excavator operating data. Background Art
[0002] Excavators, as widely used construction equipment, are a crucial piece of machinery in construction and mining operations due to their diverse functionality, wide range of applications, and mature production technology. However, due to their discrete control methods and variable power requirements, excavators suffer from unstable fuel consumption, which can lead to increased energy consumption and compromise environmental protection and resource-saving production requirements.
[0003] Current research on excavator energy consumption prediction focuses on factory bench experiments and pollutant index estimation methods. However, the existing methods rely too much on precise emission models and a large number of experimental results statistics, which still cannot meet the actual production requirements in terms of research costs and evaluation effects. Therefore, research on excavator energy consumption prediction methods still has great research significance.
[0004] Data-driven deep learning methods offer new research insights for energy consumption prediction. Using multi-sensor data reflecting comprehensive operational information, they effectively learn the spatiotemporal correlations between multiple actuators in an excavator, improving the ability to identify periodic fuel consumption information across the entire operation cycle. Ultimately, they achieve highly accurate fuel consumption predictions for excavators, which has significant engineering implications for energy conservation, emission reduction, and improving excavator production efficiency. Summary of the Invention
[0005] In response to the above content, the present invention provides a fuel consumption prediction method based on the local-global spatiotemporal information decomposition of multivariable excavator operation data, performs feature learning on the excavator multivariable sensor data and extracts relevant equipment operation information, thereby improving the fuel consumption prediction accuracy of the excavator operation process, and has practical engineering value.
[0006] To achieve the above objectives, the present invention provides the following technical solutions:
[0007] The fuel consumption prediction method based on the local-global spatiotemporal information decomposition of multivariate excavator operation data includes the following steps:
[0008] Step 1: Use sensor signal acquisition equipment to obtain the original multi-dimensional operation information of the excavator during operation, and perform data processing to obtain a multivariate time series with standardized features. The above standardized multivariate time series is input into the constructed local multi-scale window attention module, local time correlation information is extracted along the time axis dimension, and multi-head window attention calculation is performed to obtain the local correlation attention feature matrix.
[0009] Step 2: Input the obtained local correlation attention feature matrix into the constructed global attention hybrid module, perform discrete Fourier transform on the obtained feature matrix, and obtain a global feature attention matrix with global hybrid feature weights. After multiplying the obtained global feature attention matrix with the time series input, a spatiotemporal feature matrix with global information is obtained.
[0010] Step 3: Input the spatiotemporal feature matrix with global information obtained in step 2 into the designed time delay and similarity measurement module for processing. The time delay and similarity measurement module includes a time delay measurement module and a similarity measurement module. Input the spatiotemporal feature matrix with global information obtained in step 2 into the prediction regressor for prediction to obtain the final fuel consumption prediction result.
[0011] A further improvement of the technical solution of the present invention is that step 1 includes the following steps:
[0012] Step 11: Use the zero-mean normalization method to normalize the original multidimensional operation information to obtain a multivariate time series with standardized characteristics;
[0013] The calculation method is as follows:
[0014]
[0015] Among them, χ′ represents the standardized value, χ represents the value of the original sample point, represents the mean value of the original data, and σ represents the standard deviation of the original data;
[0016] Step 12: Input the multivariate time series with standardized features obtained in step 1 into the constructed local multi-scale window attention module, extract local time correlation information along the time axis dimension, calculate the multi-head window attention, and obtain the local correlation attention feature matrix;
[0017] The calculation method is as follows:
[0018]
[0019] Among them, Softmax represents the activation function layer, represents the number of windows, L represents the length of the time series, ω represents the window length, q i represents the i-th query vector, represents the transpose of the i-th key vector, v i represents the i-th value vector, γ L Represents the scaling factor.
[0020] A further improvement of the technical solution of the present invention is that step 2 includes the following steps:
[0021] Step 21: Input the local correlation attention feature matrix obtained in step 1 into the channel mixing layer for correlation feature fusion;
[0022] It is calculated as follows:
[0023] Attn mix =Matrix mix (Attn local )
[0024] Among them, Matrix mix It is a non-parameterized mixed diagonal difference matrix, and the mixed attention matrix is obtained after the correlation features are fused;
[0025] Step 22: Perform discrete Fourier transform on the obtained mixed attention matrix to obtain a global feature matrix with global mixed features;
[0026] It is calculated as follows:
[0027]
[0028] Among them, Attn mix [m] represents the mth feature of the mixed mutual difference matrix, j is the imaginary unit, and s represents the potential periodic feature coefficient.
[0029] Step 23: Multiply the global feature attention matrix with global mixed feature weights extracted in step 22 above by the original multivariate input sequence to obtain a spatiotemporal feature matrix with global information for final regression prediction;
[0030] It is calculated as follows:
[0031] h(χ i )=Attn global ·χ i
[0032] Among them, h(χ i ) represents the i-th spatiotemporal prediction value, χ i Represents the i-th original input value;
[0033] A further improvement of the technical solution of the present invention is that step 3 includes the following steps:
[0034] Step 31: Input the spatiotemporal feature matrix with global information obtained in step 23 into the designed time delay and similarity measurement module for processing.
[0035] Define the similarity metric loss function as one of the algorithms for learning high-level fuel consumption feature information features. The specific algorithm is as follows:
[0036]
[0037] Among them, θ represents the smoothing coefficient, A∈A n,m Represents the self-learning minimum path matrix, y i represents the true value, Δ(h(χ i ),y i ) represents the distance between the predicted value and the true value;
[0038] Define the time delay metric loss function as one of the algorithms for learning high-level fuel consumption feature information features. The specific algorithm is as follows:
[0039]
[0040] in, represents the time delay penalty matrix.
[0041] Step 32: Define the excavator fuel consumption prediction task as a regression prediction problem;
[0042] Step 33: Input the high-level fuel consumption prediction information features obtained in step 31 into a regressor with a loss function of mean absolute error to obtain the excavator fuel consumption prediction information, where the mean absolute error is calculated as:
[0043]
[0044] Where z is the total number of sampling points;
[0045] Step 34: The total loss function of the model is expressed as follows:
[0046] Loss = αLoss MAE +(1-α)(Loss shape +Loss delay )
[0047] Among them, α is the weight coefficient, which is used to control the feature learning effect; the obtained result is used as the output of the network to obtain the final working condition recognition result.
[0048] Compared with the prior art, the fuel consumption prediction method based on the local-global spatiotemporal information decomposition of multivariable excavator operation data provided by the present invention has the following beneficial effects:
[0049] A fuel consumption prediction method based on local-global spatiotemporal information decomposition of multivariable excavator operating data obtains multi-sensor operating data from the excavator's operation process, uses data normalization as preprocessing, and inputs the resulting time series into a constructed local multi-scale window attention module. Local temporal correlation information is extracted along the time axis dimension, and window multi-head attention calculation is performed to obtain local correlation features. The obtained local correlation features are input into a constructed global attention hybrid module, and the obtained local attention is discrete Fourier transformed along the variable axis dimension to obtain a spatiotemporal information attention weight matrix with spatial correlation. The obtained weight matrix is multiplied by the time series input and input into a designed time delay and similarity measurement module for processing to obtain the final fuel consumption prediction result, thereby improving the accuracy of fuel consumption prediction. The present invention effectively extracts excavator local sensor information at different time scales and effectively extracts correlations between different variables at the spatial scale. It also effectively learns the trend characteristics of time series changes, reduces the delay of fuel consumption prediction results, and improves the prediction accuracy of excavator fuel consumption prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0051] Figure 1 This is a flow chart of the excavator working condition identification method based on the spatiotemporal fusion of whole machine information of the present invention.
[0052] Figure 2 for Figure 1 Flowchart of signal acquisition in the process.
[0053] Figure 3 for Figure 1 Flowchart of multimodal signal feature fusion.
[0054] Figure 4 The method shows high prediction accuracy both within the cycle and at the cycle transition moment.
[0055] Figure 5 This is the distribution diagram of the prediction results. DETAILED DESCRIPTION
[0056] The technical solutions of the present invention will be clearly and completely described below through specific implementation methods. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.
[0057] This embodiment describes the present invention in detail with reference to the accompanying drawings:
[0058] like Figure 1 As shown in FIG, a method for identifying an excavator working condition based on spatiotemporal fusion of whole-machine information is provided, which includes the following steps:
[0059] S101: Collect multidimensional raw signals from multiple sensors during the operation of the excavator, normalize the raw signals to obtain a standardized multivariate time series, input the series into a local multi-scale window attention module, and calculate a local correlation attention feature matrix;
[0060] S102: Input the local correlation attention feature matrix into the global attention mixing module, perform discrete Fourier transform calculation, obtain the spatiotemporal information attention weight matrix, perform matrix multiplication operation on the weight matrix and the original multivariate time series input, and obtain the spatiotemporal feature matrix with global information;
[0061] S103: Inputting the spatiotemporal feature matrix with global information into a time delay and similarity measurement module for processing to obtain a high-level feature information matrix, and inputting the high-level feature information matrix into a prediction regressor to obtain a fuel consumption prediction result.
[0062] Among them, S101: Figure 2 As shown in the signal acquisition system, the sensor signal acquisition device is used to obtain the original multi-dimensional operation information of the excavator during operation, and the data is processed to obtain a multivariate time series with standardized features. The above standardized multivariate time series is input into the constructed local multi-scale window attention module, and the local time correlation information is extracted along the time axis dimension. The multi-head window attention calculation is performed to obtain the local correlation attention feature matrix. The specific operation steps are as follows:
[0063] Step 11: Use the zero-mean normalization method to normalize the original multidimensional operation information to obtain a multivariate time series with standardized characteristics;
[0064] The calculation method is as follows:
[0065]
[0066] Among them, χ′ represents the standardized value, χ represents the value of the original sample point, represents the mean value of the original data, and σ represents the standard deviation of the original data;
[0067] Step 12: Input the multivariate time series with standardized features obtained in step 1 into the constructed local multi-scale window attention module, extract local time correlation information along the time axis dimension, calculate the multi-head window attention, and obtain the local correlation attention feature matrix;
[0068] The calculation method is as follows:
[0069]
[0070] Among them, Softmax represents the activation function layer, represents the number of windows, L represents the length of the time series, ω represents the window length, q i represents the i-th query vector, k i T represents the transpose of the i-th key vector, v i represents the i-th value vector, γ L Represents the scaling factor.
[0071] S102: Input the obtained local correlation attention feature matrix into the constructed global attention hybrid module, perform discrete Fourier transform on the obtained feature matrix, and obtain a spatiotemporal information attention weight matrix with spatial correlation. After multiplying the obtained weight matrix with the time series input, a spatiotemporal feature matrix with global information is obtained. The specific operation steps are as follows:
[0072] Step 21: Input the local correlation attention feature matrix obtained in step 1 into the channel mixing layer for correlation feature fusion;
[0073] It is calculated as follows:
[0074] Attn mix =Matrix mix (Attn local )
[0075] Among them, Matrix mix It is a non-parameterized mixed diagonal difference matrix, and the mixed attention matrix is obtained after the correlation features are fused;
[0076] Step 22: Perform discrete Fourier transform on the obtained hybrid attention matrix to obtain a global feature attention matrix with global hybrid feature weights;
[0077] It is calculated as follows:
[0078]
[0079] Among them, Attn mix [m] represents the mth feature of the mixed mutual difference matrix, j is the imaginary unit, and s represents the potential periodic feature coefficient.
[0080] Step 23: Multiply the global feature attention matrix with global mixed feature weights extracted in step 22 above by the original multivariate input sequence to obtain the predicted value for the final regression prediction;
[0081] It is calculated as follows:
[0082] h(χ i )=Attn global ·χ i
[0083] Among them, h(χ i ) represents the i-th spatiotemporal prediction value, χ i Represents the i-th original input value;
[0084] S103: The predicted value for the final regression prediction obtained in S102 is input into the designed time delay and similarity measurement module for processing. The time delay and similarity measurement module includes a time delay measurement module and a similarity measurement module. The spatiotemporal feature matrix with global information obtained in S102 is input into the prediction regressor for prediction to obtain the final fuel consumption prediction result. The specific operation steps are as follows:
[0085] Step 31: Input the spatiotemporal fusion feature matrix obtained in step 23 above into the designed time delay and similarity measurement module for processing.
[0086] Define the similarity metric loss function as one of the algorithms for learning high-level fuel consumption feature information features. The specific algorithm is as follows:
[0087]
[0088] Among them, θ represents the smoothing coefficient, A∈A n,m Represents the self-learning minimum path matrix, y i represents the true value, Δ(h(χ i ),y i ) represents the distance between the predicted value and the true value;
[0089] Define the time delay metric loss function as one of the algorithms for learning high-level fuel consumption feature information features. The specific algorithm is as follows:
[0090]
[0091] in, represents the time delay penalty matrix.
[0092] Step 32: Define the excavator fuel consumption prediction task as a regression prediction problem;
[0093] Step 33: Input the high-level fuel consumption prediction information features obtained in step 31 into a regressor with a loss function of mean absolute error to obtain the excavator fuel consumption prediction information, where the mean absolute error is calculated as:
[0094]
[0095] Where z is the total number of sampling points;
[0096] Step 34: The total loss function of the model is expressed as follows:
[0097] Loss = αLoss MAE +(1-α)(Loss shape +Loss delay )
[0098] Among them, α is the weight coefficient, which is used to control the feature learning effect; the obtained result is used as the output of the network to obtain the final working condition recognition result.
[0099] Specific application areas: Excavator fuel consumption prediction field, used to assist in adjusting excavator parameters to achieve energy conservation and emission reduction requirements.
[0100] The claimed method was compared with existing methods and the best results were achieved.
[0101] from Figure 4 As can be seen in Figure 2, the method shown shows high prediction accuracy both within the cycle and at the cycle transition moment, while the other methods have a certain degree of misidentification and fluctuation.
[0102] The predicted results are distributed as follows Figure 5 As shown, the better distribution is near the true value, showing higher accuracy.
[0103] like Figure 4 As shown in the figure, the proposed method is compared with various existing mainstream time series forecasting models (including Transformer, Autoformer, FEDformer, Crossformer, Informer, Reformer, Pyraformer, and VMD Informer). It can be clearly observed that the proposed method can closely track the changing trend of the true value in each time period. Whether in stable intervals or during rapid fluctuations and period transitions, it can better capture the dynamic characteristics of the data, thereby obtaining a more accurate forecast series.
[0104] In contrast, other methods exhibit varying degrees of misidentification or significant deviations between the predicted curve and the true value in certain periodic segments. Phase misalignment or amplitude discrepancies in some prediction results demonstrate their inability to cope with complex periodic variations. Particularly at the boundaries of period transitions, error amplification caused by accumulated errors can more easily lead to sudden changes or offsets in the predicted curve. The method proposed in this paper, however, achieves a relatively smooth transition.
[0105] like Figure 5 As shown in the figure, the method of the present invention shows extremely high accuracy in fitting the distribution of the predicted results and the true value. 2 The accuracy reached 0.996, with an RMSE of only 0.050 and a MAPE of 0.291, indicating that the prediction results are concentrated and distributed in an area close to the true value, with low error dispersion. This stable and high-precision prediction capability is due to both the expressive power of the proposed deep symbolic regression model and the effective capture and optimization of key features during training.
[0106] Overall, the proposed method not only demonstrates high fidelity for full-cycle predictions when forecasting time series data with strong periodicity and non-stationarity, but also maintains good adaptability and robustness at critical moments such as period transitions and extreme value intervals. Compared to other methods that fluctuate at high frequencies and transition boundaries, the proposed model is able to more accurately characterize data features, laying a solid foundation for subsequent multi-scenario applications.
[0107] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.
Claims
1. A fuel consumption prediction method based on local-global spatiotemporal information decomposition of multivariate excavator operation data, characterized in that: The following steps are involved: Step 1: Collect multidimensional raw signals from multiple sensors during the operation of the excavator, normalize the raw signals to obtain a standardized multivariate time series, and input the series into the local multi-scale window attention module to calculate the local correlation attention feature matrix; Step 2: Input the local correlation attention feature matrix into the global attention mixing module, perform discrete Fourier transform calculation, obtain the spatiotemporal information attention weight matrix, perform matrix multiplication operation on the weight matrix and the original multivariate time series input to obtain the spatiotemporal feature matrix with global information; Step 3: Input the spatiotemporal feature matrix with global information into the time delay and similarity measurement module for processing to obtain a high-level feature information matrix, and input the high-level feature information matrix into the prediction regressor to obtain a fuel consumption prediction result.
2. The method according to claim 1, characterized in that The standardization process described in step 1 is zero-mean standardization, and the specific method is as follows: Subtract the mean of all original data samples from each sample value of the original data, and then divide it by the standard deviation of all original data samples to obtain the data after zero mean standardization.
3. The method according to claim 1, characterized in that The local multi-scale window attention module in step 1 uses the multi-head window attention mechanism to calculate the local correlation attention feature matrix. The specific method is: First, the matrix product of each query vector and the transposed key vector is calculated, and the result of the matrix product is divided by the scaling factor, and then the attention weight matrix is obtained through the Softmax activation function; Then, matrix multiplication is performed on the attention weight matrix and the corresponding value vector to obtain the local correlation attention feature matrix.
4. The method according to claim 1, wherein The global attention mixing module in step 2 includes a channel mixing layer, and the specific method is as follows: The mixed attention matrix is obtained by performing matrix multiplication of the non-parameterized mixed diagonal difference matrix and the input local correlation attention feature matrix.
5. The method according to claim 4, characterized in that The specific method of discrete Fourier transform in step 2 is: Each eigenvector in the mixed attention matrix is transformed and calculated based on the imaginary unit exponent, and a Fourier transform coefficient matrix representing the global periodic characteristics is obtained by performing discrete Fourier transform on each eigenvector in the complex domain.
6. The method according to claim 1, characterized in that The time delay and similarity measurement module in step 3 uses a similarity measurement loss function and a time delay measurement loss function. The specific method is: The similarity metric loss function calculates the square of the prediction error by the distance between the predicted value and the true value, divides the result by the sum of the smoothing coefficient and the square of the prediction error, and finally multiplies it by the self-learning minimum path matrix and takes the average value of all sampling points; The time delay metric loss function calculates the prediction error by squaring the difference between the predicted value and the true value, and then multiplies it by the time delay penalty matrix and takes the average value of all sampling points.
7. An excavator fuel consumption prediction system based on multivariable operating data, characterized in that: include: The multivariable sensor data acquisition and preprocessing unit is used to obtain the multidimensional raw sensor signals of the excavator under different operating conditions in real time, and construct a multivariable time series input matrix with a unified scale through the zero-mean normalization method; A local multi-scale window attention calculation unit is connected to the multivariate sensor data acquisition and preprocessing unit, and uses a multi-head attention mechanism to calculate the local correlation of time series data and generate a local attention feature matrix that represents short-term time dependence; A global attention hybrid feature extraction unit is connected to the local multi-scale window attention calculation unit, and mines potential global periodic patterns in time series data through channel mixing and discrete Fourier transform (DFT) method to obtain a global feature matrix that characterizes the overall working condition change law; A time delay and similarity feature fusion unit is connected to the global attention hybrid feature extraction unit, adopts similarity measurement and time delay loss function, and uses dynamic time warping (DTW) and adaptive path optimization strategy to capture the time delay and trend similarity characteristics of the feature matrix; The regression prediction unit is connected to the time delay and similarity feature fusion unit, and realizes the numerical prediction output of the excavator fuel consumption based on the regression prediction model trained by Mean Absolute Error (MAE).
8. The system according to claim 7, characterized in that The local multi-scale window attention calculation unit specifically includes: Multiple parallel local scale window attention submodules, each of which includes three parameter units: query matrix (Query), key matrix (Key) and value matrix (Value); In each window, the correlation strength between features in the local window is calculated through multi-head scaled dot-product attention (Scaled Dot-Product Attention) to obtain the local attention weight matrix; The features of multiple scale windows are fused through weighted summation to improve the ability to capture the sensitivity of short-term feature changes and output the local correlation attention feature matrix.
9. The system according to claim 7, wherein: The global attention hybrid feature extraction unit includes: The channel mixing layer subunit uses a non-parameterized diagonal difference matrix and a local correlation attention feature matrix to perform matrix transformation to reduce the redundant correlation between feature channels; The frequency domain feature extraction subunit transforms the mixed attention matrix through Fast Fourier Transform (FFT) to explore the periodicity and trend features hidden in the frequency domain; The feature matrix and the original input data multiplication operation unit maps the frequency domain periodic features to the time domain space, and obtains the global spatiotemporal feature matrix for subsequent modeling and analysis through matrix element-by-element multiplication operation.
10. The system according to claim 7, wherein: The time delay and similarity feature fusion unit includes: The similarity measurement subunit constructs a similarity loss matrix based on the self-learning minimum path to evaluate the trend consistency between the predicted value and the true value, and introduces a smoothing adjustment factor to suppress the overfitting tendency of the feature matrix; The time delay measurement subunit captures the time lag effect of the predicted sequence relative to the true sequence by constructing a loss matrix with a delay penalty factor; The above two metric matrices are jointly optimized into a composite loss function, and an adaptive weighted fusion strategy is used to achieve comprehensive optimization of the spatiotemporal feature matrix, obtaining a fused feature matrix that can characterize high-level fuel consumption feature information, thereby improving the generalization performance of the fuel consumption prediction model.