Method for predicting performance degradation of engineering mechanical equipment based on multi-model stacking fusion

Abstract

An engineering machinery equipment performance degradation prediction method based on multi-model stacking fusion is characterized in that design calculation, a simulation model and a data driving model are combined, physical test data under different carrying equipment and different working conditions are used as training materials, confidence weight distribution of sub-models is completed, and a fusion total model is obtained. And when real object live test data under certain conditions are lacked, the environmental conditions are used as input, a final prediction value is obtained through a model stacking weight distribution mechanism, and the prediction result of the performance degradation of the engineering mechanical equipment is optimized.

Description

Technical Field

[0001] The present invention relates to a technology in the field of engineering machinery equipment, and specifically to a method for predicting the performance degradation of engineering machinery equipment based on multi-model stacked fusion. Background Art

[0002] Currently, in the performance degradation test of engineering machinery equipment, the physical test cost is high and the conditions are harsh under certain working conditions. In addition, after the engineering machinery is actually installed on the equipment, some performance parameters are also difficult to measure. Therefore, the physical in-situ test is greatly limited and it is necessary to rely on the digital twin model to complete the prediction. In the existing prediction models, the design calculation index model, the simulation physical model, and the data-driven model run independently, and there is a problem that they cannot complement each other, which limits the application of the digital twin technology to the actual scenario test. It is difficult for the existing technology to solve the prediction deviation of different models for the same working condition, and it is impossible to fuse multiple models to complete the prediction. At the same time, it often ignores the influence of the different whole machine environments carried by the engineering machinery equipment on the performance degradation, and lacks an integrated and targeted model. Summary of the Invention

[0003] In view of the above deficiencies of the existing technology, the present invention proposes a method for predicting the performance degradation of engineering machinery equipment based on multi-model stacked fusion, which combines the design calculation, the simulation model, and the data-driven model, uses the physical test data under different installed equipment and different working conditions as the training materials, completes the confidence weight assignment of the sub-models, and obtains the fused total model. When the physical in-situ test data is lacking under certain conditions, the environmental conditions are used as the input, and the final prediction value is obtained through the model stacking weight assignment mechanism, so as to optimize the prediction result of the performance degradation of the engineering machinery equipment.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a method for predicting the performance degradation of engineering machinery equipment based on multi-model stacked fusion, including:

[0006] Step 1, for the situation where there is no direct reference to the actual working condition test data, input the working conditions such as working pressure, rotational speed, temperature, and load.

[0007] The situation where there is no direct reference to the actual working condition test data includes, but is not limited to: insufficient actual working condition acquisition data, inability to directly measure key parameters after the equipment is installed, that is, there is no space left for sensor placement or there is significant noise or vibration interference at the required sensor placement position, and it is difficult to directly measure relevant parameters under extreme working conditions, etc.

[0008] Step 2: After separately establishing a performance degradation design calculation model for construction machinery and equipment under specific working conditions, establishing a physical simulation model of construction machinery and equipment using finite element analysis software, and constructing a dual-channel residual attention network guided by physical simulation as a data-driven model, input the working conditions obtained in Step 1 respectively to obtain the prediction results corresponding to each model.

[0009] The dual-channel residual attention network guided by physical simulation includes: a dual-channel input processing layer containing an experimental data channel and a simulation data channel, a physical knowledge embedding module, a dynamic cross-modal fusion layer, and a fully connected network, where: the experimental data channel uses a bidirectional long short-term memory network (BiLSTM) combined with a multi-head attention mechanism to process the time-series experimental data collected by sensors, bidirectionally capture the dynamic features in the time series and extract global relevant information, highlighting the impact of key working condition parameters on performance degradation; the simulation data channel extracts features from the data output by the simulation model using a depthwise separable convolutional neural network, and then incorporates the physical constraint knowledge in the construction machinery field into the network in the form of residuals through the physical knowledge embedding module; the physical knowledge embedding module adds the physical laws contained in the simulation model as constraints to the training process of the dual-channel residual attention network guided by physical simulation, so that the network prediction results are guided by physical priors; the dynamic cross-modal fusion layer fuses the features extracted from the experimental data channel and the simulation data channel based on learnable weights, adaptively balances the experimental observation information and the simulation prior information according to the current working conditions, and outputs the fused features to the fully connected network, and the fully connected network obtains the prediction results of the performance degradation of construction machinery and equipment.

[0010] The construction machinery and equipment mentioned above include but are not limited to: power transmission devices such as hydraulic pumps, hydraulic motors, and reducers.

[0011] The data sources of the data-driven model include: the dataset output by the simulation physical model and the experimental dataset obtained by testing the construction machinery and equipment in the laboratory. The data output by the physical simulation model is used to mix with the existing laboratory test data for model establishment and training.

[0012] The described dual-channel residual attention network guided by physical simulation is trained using a hybrid data augmentation strategy, which specifically includes: applying the sliding window interception technique to the measured time-series data to generate more effective training samples to expand the scale of the experimental dataset; screening representative samples under typical working conditions and extreme conditions from the simulated data as auxiliary training data to expand the model's prediction ability for unseen working conditions; and introducing pre-trained model parameters through transfer learning to transfer the knowledge pre-learned on similar devices or simulation data to the initial weights of the dual-channel residual attention network model guided by physical simulation to alleviate the problem of insufficient real data. During the training process, reasonable hyperparameters are set (such as learning rate 0.001, batch size 32, training epochs 1000, etc.), the mean squared error (MSE) is used as the loss function, and the Adam optimization algorithm is used to iteratively update the network weights to ensure that the model converges effectively and achieves the expected performance.

[0013] Step 3: Initially fuse the three prediction results obtained in Step 2, and select the confidence weights of each model according to the classification of input conditions, such as the object carried, rotational speed magnitude, temperature level, etc.

[0014] Step 4: Based on the selected confidence weights, complete the weight allocation for the design calculation model, physical simulation model, and data-driven model. After linear weighting, obtain the final output of the performance degradation prediction result. Technical effects

[0015] The present invention uses multi-model fusion and weight allocation techniques to predict the performance degradation of engineering construction machinery, fuses the prediction results of the design calculation model, simulation model, and data-driven model, and realizes the output of a higher-precision prediction result through weight allocation. Compared with the prior art, the present invention combines the advantages of each sub-model, comprehensively considers the overall machine environment in which the engineering construction machinery is located during weight allocation, making the prediction result of performance degradation more targeted and more in line with the actual situation. Description of the drawings

[0016] Figure 1 It is the flow chart of the present invention;

[0017] Figure 2 It is the flow chart for obtaining the confidence weights;

[0018] Figure 3 It is the flow chart of the embodiment.

[0019] Figure 4 It is the dual-channel residual attention network diagram guided by physical simulation when constructing the data-driven model. Specific implementation manner

[0020] Such as Figure 1 And Figure 3As shown in the figure, this embodiment relates to a method for predicting the performance degradation of engineering machinery based on multi-model stacking fusion. Taking a hydraulic pump as an example, it includes:

[0021] Step 1: According to the inputs required by the sub-models, input the working conditions of pressure and speed, and establish sub-models, specifically including:

[0022] 1.1 Establish a design calculation model and predict the life using the volumetric efficiency degradation formula based on the hydraulic pump, specifically: η(t) = η0 - k1·P·t - k2·N·t, where: η0 is the initial volumetric efficiency, P is the working pressure, N is the speed, k1 and k2 are the wear coefficients of pressure and speed respectively, and t is the working life.

[0023] 1.2 Establish a physical simulation model, set the finite element simulation conditions. The pump body material is cast iron (elastic modulus 210 GPa, Poisson's ratio 0.3). When considering plastic deformation, use the bilinear isotropic hardening model. The global element size ≤ 0.5 mm, and the key areas (such as the port plate clearance and bearing seat) are encrypted to 0.05 mm. Use hexahedron-dominated meshes, and set an inflation layer in the transition area to capture the stress gradient. The total duration covers the typical working condition cycle, and predict the life based on the Miner linear cumulative damage theory and combined with the S-N curve.

[0024] 1.3 Establish a data-driven model using a dual-channel residual attention network guided by physical simulation. The input layer includes an experimental data channel: a 5-dimensional time series of working condition parameters (pressure, speed, temperature, load, working time), the time window length is 60, and a simulation data channel: a 128×128 grid feature map (including pressure field, temperature field, stress field). The number of layers of bidirectional LSTM: 3 layers, the number of hidden units: 128, the time step expansion: 60, the number of multi-head attention heads: 4 heads, the key / value dimension: 32, the number of depthwise separable convolution layers: 4 layers, the kernel size: 3×3, the physical constraint residual strength: λ = 0.5, the activation function: tanh, the initial fusion weight: α = 0.5, the learning rate: 0.001, the batch size: 32, the training period: 1000, the optimizer: Adam, the loss function: mean square error (MSE).

[0025] Step 2: The design calculation model, the simulation physical model, and the data-driven model respectively complete the performance degradation prediction according to the input conditions obtained in Step 1.

[0026] Step 3: As Figure 2 shown, call the average value of the confidence weights of the category or approximate category where the working conditions are located to complete the weight allocation of the output results of the sub-models, and finally obtain the output performance degradation prediction result, specifically including:

[0027] Step 3.1: Conduct on-site tests on the construction machinery and equipment already installed on specific engineering equipment to obtain experimental data under actual working conditions, including data collected by pressure sensors, temperature sensors, vibration sensors, acoustic sensors, etc.

[0028] The on-site test mentioned above refers to the test conducted on the construction machinery and equipment after actual installation, which is different from the physical test simulated in the laboratory.

[0029] Step 3.2: Use the conditions during the on-site test as inputs and input them into each sub-model. The inputs include the installed equipment and working condition data. Each sub-model completes result prediction based on the inputs, uses the initial weight assignment to complete the fusion, and obtains the preliminary prediction result.

[0030] When inputting data into each model, including when using prediction methods later, there may be situations where the inputs are not completely consistent. For example, the on-site test may be completed at a certain pressure and rotational speed, while other conditions such as temperature need to be input when using the simulation model. Or, if the performance degradation degree corresponding to pressure and flow rate is determined in the hypothetical design calculation model, there will be a situation where the input conditions are not completely consistent with what the model requires. In such cases, consider using a regression model to align the inputs and convert the input conditions into what the model needs. The regression model refers to the random forest regression model. First, complete data processing, normalize the features, determine the parameters of the number of trees, the maximum number of features, the maximum depth, and the minimum number of samples required for leaf nodes, use the root mean square error as the error metric, complete model training on the training set, and use cross-validation to ensure generalization ability.

[0031] Step 3.3: Use the test data set obtained in Step 3.2 as the training material, compare the preliminary prediction result with the actual data, and complete one weight optimization for each input based on the gradient decision tree to obtain the new confidence weight after optimization, and store the weight assignment result. Specifically: set the parameters of the gradient boosting model, including the learning rate, the number of trees, and the depth of the trees. In each iteration, the model will try to predict the residuals of the previous round of the model. After each iteration, calculate the residuals between the current model prediction result and the real data, and update the weights of the trees by changing the gradient of the loss function.

[0032] Step 3.4: Classify the weights corresponding to each input condition under different conditions. According to different input conditions, such as the installed object, rotational speed, and temperature, classify them according to the degree of condition approximation, and store the confidence weights. The confidence weights are stored by classifying according to different input conditions, and when calling the confidence weights during the prediction process, they are also selected based on the category of the input conditions.

[0033] The confidence weight mentioned above refers to: different weights assigned to the prediction results of each model under different working conditions based on historical test data. The confidence weight is determined by calculating the prediction error (MSE) of each model under different working conditions. Specifically: Where: ω i is the confidence weight of the i-th sub-model (i = 1, 2, 3, corresponding to the design calculation model, simulation physical model, and data-driven model respectively), and MSE i is the mean square error of the i-th model under a specific input condition category, and the denominator is the sum of the reciprocals of the mean square errors of the three sub-models.

[0034] The confidence weight classification rules classify categories according to input conditions, such as the type of equipment carried (such as crane, excavator), rotational speed (low / medium / high), and pressure (<20MPa, 20 - 40MPa, >40MPa). Under each category, the corresponding confidence weight is used as the basis for the final weight assignment (example weight range: design model 0.2 - 0.3, simulation model 0.3 - 0.4, data-driven model 0.4 - 0.5).

[0035] Step 4: According to the weight calculation and assignment in Step 3, obtain the final prediction result. Specifically: T pred = w design ·T design + w sim ·T sim + w data ·T data , where: T pred is the final predicted working life, w design is the weight of the design calculation model, T design is the working life predicted by the design calculation model, w sim is the weight of the physical simulation model, T sim is the working life predicted by the physical simulation model, w data is the weight of the data-driven model, T data is the working life predicted by the data-driven model.

[0036] For example Figure 3As shown, through specific experiments, first input the equipment equipped with the hydraulic pump, operating speed, liquid temperature, and pressure environment conditions, such as setting: equipped on a crawler crane, operating speed 1800 rpm, liquid temperature maintained at 50 °C, pressure 30 Mpa. Subsequently, set the performance degradation index and prediction requirements, such as: define failure when the leakage flow rate reaches 2.5 L / min, and predict the working life of the hydraulic pump. Input the above conditions into the fusion prediction model. Each sub-model obtains the predicted value of the working life as the output according to the input. For example, for the design calculation model, the design index shows that the working life of the hydraulic pump is 10000 h at an operating speed of 1800 rpm and a pressure of 30 Mpa; for the simulation physical model, after inputting the boundary conditions, load conditions, and environmental conditions, assume that the simulated working life of the simulation model is 8874 h; for the data-driven model, after inputting the boundary conditions, load conditions, and environmental conditions, the model predicts the working life to be 9016 h. According to the classification of input conditions, the confidence weight should be in the category of crawler crane, medium to low speed, relatively high pressure, and medium liquid temperature. Call the confidence weight value of this category, such as: the confidence weight of the design calculation is 0.22, the confidence weight of the simulation physical model is 0.35, and the confidence weight of the data-driven model is 0.43. Finally, complete the output of the predicted working life result according to the confidence weight distribution result: 10000 * 0.22 + 8874 * 0.35 + 9016 * 0.43 = 9182.78 h. The above is a complete prediction process.

[0037] Compared with the prior art, the present invention establishes a prediction method of multi-model stacked fusion for the performance degradation of engineering machinery equipment, and uses the test data after actual installation as the training material, making the final prediction model more targeted. The first three models are only related to the engineering machinery equipment itself, and it is difficult to consider the actual situation of specific installation when making predictions. The prediction results of the hydraulic pumps installed on equipment A and equipment B are the same, but in fact, different installation methods may affect the actual performance degradation of the hydraulic pump. This method can effectively optimize this problem. In addition, the present invention combines the attention mechanism and physical residual constraint and applies them to the prediction of engineering machinery performance degradation, innovatively realizing the dynamic cross-modal fusion based on simulation prior knowledge, and supplementing with a small sample transfer learning mechanism to improve the adaptability of the model to limited data. This data-driven model scheme takes into account both physical credibility and prediction accuracy, overcomes the limitation that traditional pure data-driven models rely on a large amount of experimental data and ignore physical mechanisms, making the performance degradation prediction results more accurate and reliable, and having good engineering application feasibility.

[0038] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation. All implementation solutions within its scope are subject to the constraints of the present invention.

Claims

1. A method for predicting the performance degradation of engineering machinery equipment based on multi-model stacked fusion, characterized in that, Including: Step 1: For the situation where there is no direct reference to actual working condition test data, input the working conditions. Step 2: After separately establishing a performance degradation design calculation model of construction machinery and equipment under specific working conditions, establishing a physical simulation model of construction machinery and equipment using finite element analysis software, and constructing a dual-channel residual attention network guided by physical simulation as a data-driven model, input the working conditions obtained in Step 1 respectively to obtain the prediction results corresponding to each model. Step 3: Initially fuse the three prediction results obtained in Step 2, and select the confidence weights of each model according to the classification of the input conditions. Step 4: Based on the selected confidence weights, complete the weight assignment for the design calculation model, physical simulation model, and data-driven model. After linear weighting, obtain the final output performance degradation prediction result.

2. The method for predicting the performance degradation of engineering machinery equipment based on multi-model stacked fusion according to claim 1, characterized in that, The dual-channel residual attention network guided by physical simulation includes: a dual-channel input processing layer containing an experimental data channel and a simulation data channel, a physical knowledge embedding module, a dynamic cross-modal fusion layer, and a fully connected network, where: the experimental data channel uses a bidirectional long short-term memory network (BiLSTM) combined with a multi-head attention mechanism to process the time series experimental data collected by sensors, bidirectionally capture the dynamic features in the time series and extract global relevant information, highlighting the impact of key working condition parameters on performance degradation; the simulation data channel uses a depthwise separable convolutional neural network to extract features from the data output by the simulation model, and then incorporates the physical constraint knowledge in the field of construction machinery into the network in the form of residuals through the physical knowledge embedding module; the physical knowledge embedding module adds the physical laws contained in the simulation model as constraints to the training process of the dual-channel residual attention network guided by physical simulation, so that the network prediction results are guided by physical priors; the dynamic cross-modal fusion layer fuses the features extracted from the experimental data channel and the simulation data channel based on learnable weights, adaptively balances the experimental observation information and simulation prior information according to the current working conditions, and outputs the fused features to the fully connected network, and the fully connected network obtains the prediction result of the performance degradation of construction machinery and equipment.

3. The method for predicting the performance degradation of engineering machinery equipment based on multi-model stacking fusion according to claim 2, characterized in that, The data source of the data-driven model includes: the data set output by the simulation physical model and the test data set obtained by the laboratory testing of construction machinery and equipment. The data output by the physical simulation model is used to mix with the existing laboratory test data for model establishment and training.

4. The method for predicting the performance degradation of engineering machinery equipment based on multi-model stacking fusion according to claim 2, characterized in that, The dual-channel residual attention network guided by physical simulation adopts a hybrid data augmentation strategy for training.

5. The method for predicting the performance degradation of engineering machinery equipment based on multi-model stacking fusion according to claim 1, characterized in that, The confidence weight mentioned above refers to: different weights assigned to the prediction results of each model under different working conditions based on historical test data. The confidence weight is determined by calculating the prediction error (MSE) of each model under different working conditions, specifically as follows: Where: ω i is the confidence weight of the i-th sub-model (i = 1, 2, 3, corresponding to the design calculation model, simulation physical model, and data-driven model respectively), and MSE i is the mean square error of the i-th model under a specific input condition category, and the denominator is the sum of the reciprocals of the mean square errors of the three sub-models.

6. The method for predicting the performance degradation of engineering machinery equipment based on multi-model stacking fusion according to claim 5, wherein, The confidence weight classification rule classifies according to the input conditions. Under each category, the corresponding confidence weight is taken as the basis for the final weight assignment.

Images