Loader resistance prediction method and system fusing transfer learning and incremental learning

Abstract

The invention discloses a loader resistance prediction method and system fusing transfer learning and incremental learning, and the method comprises the steps: training a resistance prediction pre-training model through source domain data, screening out a data segment matched with a target domain, and carrying out the further training, so as to obtain an offline resistance prediction model suitable for the target domain; preprocessing the online resistance data of the target domain, and optimizing the offline and online resistance prediction models of the target domain through incremental learning by using the obtained data segments; based on the physical property similarity between the source domain and the target domain and the data maturity of the target domain, the fusion weight of the two models is adjusted through a dynamic adjustment function, and weight fusion is carried out to obtain a fused resistance prediction model; and inputting online data of the target domain into the fusion model to realize resistance prediction of the target domain. According to the method, dynamic adjustment and weight fusion are performed through historical and online resistance data of the source domain and the target domain, and resistance prediction of the loader during operation in the target domain is realized.

Description

Technical Field

[0001] The present invention relates to the field of operation resistance prediction, and in particular to a loader resistance prediction method and system integrating transfer learning and incremental learning. Background Art

[0002] Loaders are heavy-duty construction machinery primarily used in construction sites, mines, and ports. They are used to load, transport, and unload bulk materials such as soil, fine sand, iron ore, lime, and coal. They are widely used in construction, mining, industrial production, and rescue operations. The loading phase of a loader's operation is the most demanding part of the operator's skill cycle, characterized by the most frequent interaction with the environment and the highest power consumption.

[0003] With technological advancements and developments, technologies related to unmanned and autonomous loader operation are continuously being developed and applied. However, due to the complexity of the loading process, coupled with the requirements for high bucket fill rates, low energy consumption, and high loading efficiency, the autonomous loading technology involved in unmanned loader operation is the least mature, with few reference technologies available. Performance indicators such as loader operating resistance are key operating parameters of the loading process. Effectively measuring these performance indicators is not only crucial for researching autonomous loading technology but also a prerequisite for achieving unmanned and autonomous loader operation.

[0004] Loaders not only have complex operating conditions and harsh working environments, but also have diverse operating objects and different material density and particle size. This leads to large differences in the performance indicators of the loader's shoveling process under different working conditions. The shoveling operation resistance prediction model constructed using traditional methods has poor adaptability.

[0005] Chinese invention patent 202310655034.3 proposes a method for predicting the resistance of loader shoveling operations based on transfer learning. This method has a hidden layer with shared parameters between the source domain model and the target domain model, which reflects the migration of model parameters and is model-based transfer learning. Chinese invention patent 202010000988.7 proposes a real-time test method for loader shoveling operation resistance based on geometric position calculation. Chinese invention patent 202210027732.4 proposes a precise method for measuring the operating resistance of a loader. However, these patents mainly propose methods for measuring operating resistance. Their disadvantages are that they all rely on specific working conditions, are sensitive to working condition changes, have poor environmental adaptability, and require extremely high sensor accuracy and stability. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a loader resistance prediction method and system that integrates transfer learning and incremental learning. By acquiring and processing the historical resistance data of the source domain and the target domain, the transfer learning and incremental learning methods are used to train offline and online resistance prediction models suitable for the target domain. The model fusion weights are dynamically adjusted based on the similarity of physical properties and data maturity. Finally, an accurate fusion resistance prediction model is obtained to predict the shoveling resistance in the target domain.

[0007] The specific plan is as follows:

[0008] On the one hand, the loader resistance prediction method that integrates transfer learning and incremental learning includes:

[0009] S1, obtain the historical resistance data of the source domain and the historical resistance data of the target domain and perform preprocessing to obtain the historical shoveling operation segments of the source domain and the historical shoveling operation segments of the target domain;

[0010] S2: Input the historical shoveling operation segments in the source domain into the resistance prediction pre-training model and train it to obtain a trained resistance prediction pre-training model. Calculate the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain. Based on the matching difference and the set matching difference threshold, select the historical shoveling operation segments with high matching between the domains. Input the historical shoveling operation segments with high matching between the domains and the historical shoveling operation segments in the target domain into the trained resistance prediction pre-training model. By updating the fine-tuning layer parameters of the trained resistance prediction pre-training model, obtain the target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain.

[0011] S3: Preprocess the online resistance data during the loader's autonomous shoveling of target domain materials to obtain target domain online shoveling operation data segments. Input the target domain online shoveling operation data segments into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on historical target domain shoveling operation segments. Update the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model to obtain the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model.

[0012] S4, calculating the physical property similarity between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain and the maturity of the target domain data, and adjusting the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model based on the dynamic adjustment function generated based on the physical property similarity and the maturity of the target domain data, and obtaining a fused resistance prediction model based on the fusion weights;

[0013] S5, inputting the online shoveling operation data segment of the target domain into the fused resistance prediction model to obtain the resistance prediction of the target domain.

[0014] Furthermore, in S1, the data types of the source domain historical resistance data and the target domain historical resistance data include: loader boom cylinder displacement data, loader bucket cylinder displacement data, loader boom cylinder large chamber pressure data, loader boom cylinder small chamber pressure data, loader bucket cylinder large chamber pressure data, loader bucket cylinder small chamber pressure data, loader oil pump pressure data, loader vehicle displacement data, loader vehicle energy consumption data and loader operation resistance data.

[0015] Furthermore, the preprocessing specifically includes: filtering and cleaning the source domain historical resistance data and the target domain historical resistance data, data segmentation and division, and time series processing.

[0016] Furthermore, in S2, the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain is calculated, and the historical shoveling operation segments with high matching between the domains are screened out based on the matching difference and the set matching difference threshold, specifically including:

[0017] S21, using the kernel density method to count the historical shoveling operation segments T in the target domain H The probability density function of is calculated as follows:

[0018]

[0019] in, Indicates The probability density value at ; Represents the historical shoveling operation segment T in the target domain H The random variables in the target domain have a value range of the historical shoveling operation segment T H The minimum to maximum value of the data points in the target domain; n represents the historical shoveling operation section T H The total number of data points in ; h represents bandwidth; G() represents kernel function; Represents the historical shoveling operation segment T in the target domain H The i-th data point in ;

[0020] S22, calculate the matching difference between the probability density function of the historical shovel loading operation segment in the source domain and the probability density function of the historical shovel loading operation segment in the target domain, and the calculation formula is as follows:

[0021]

[0022] Among them, Y Hm represents the mth historical shoveling operation segment in the source domain; represents the probability density function of the mth source domain historical shoveling operation segment; Em represents the matching difference between the mth source domain historical shoveling operation segment and the target domain historical shoveling operation segment; p represents the number of target domain historical shoveling operation segments obtained; q represents the number of source domain historical shoveling operation segments obtained;

[0023] S23, set the matching difference threshold τ, if Em<τ, then it means that the mth segment of the source domain historical shovel operation data Y Hm The target domain history shovel operation segment T H The matching difference is less than the matching difference threshold, and the data segment is selected as the historical shoveling operation segment Y with high inter-domain matching degree. E .

[0024] Furthermore, in S3, the target domain online shoveling operation data segment is input into the target domain offline resistance prediction model, and the incremental learning target domain offline resistance prediction model is obtained by updating the fine-tuning layer parameters of the target domain offline resistance prediction model, which specifically includes:

[0025] Based on the target domain offline resistance prediction model, the target domain online shoveling operation data segment and the historical shoveling operation segment with high matching degree between domains are predicted to obtain the total loss L, which is as follows:

[0026]

[0027] Among them, L new and L old They represent the loss of the online shoveling operation segment of the target domain and the loss of the historical shoveling operation segment with high matching between domains; n new and n old They represent the number of data points of the online shoveling operation segment of the target domain and the historical shoveling operation segment with high matching degree between domains; and They represent the predicted resistance value and the actual resistance value of the online shoveling operation section of the i-th target domain respectively; and They represent the predicted resistance value and the actual resistance value of the j-th historical shoveling operation section with high inter-domain matching degree;

[0028] Based on the total loss L, the gradient of the fine-tuning layer of the target domain offline resistance prediction model is calculated layer by layer to obtain the update direction of the model parameters;

[0029] The target domain offline resistance prediction model is continuously fine-tuned based on the update direction of the model parameters to obtain an incrementally learned target domain offline resistance prediction model.

[0030] Furthermore, in S4, the dynamic adjustment function includes a dynamic adjustment function for the fusion weight of the target domain offline resistance prediction model and a dynamic adjustment function for the fusion weight of the target domain online resistance prediction model. The specific setting method of the dynamic adjustment function is as follows:

[0031] Based on the prior knowledge of the historical shoveling operation segments in the source domain and the target domain, the inter-domain correlation σ representing the similarity of physical characteristics is defined, and the value range of the inter-domain correlation σ is set to (0.6, 0.9);

[0032] The dynamic adjustment function of the fusion weight α of the target domain offline resistance prediction model is defined based on the similarity of physical properties and the maturity of the target domain data:

[0033] α=σ·e -kv ;

[0034] Where σ is the inter-domain correlation, which is used to represent the similarity of physical characteristics between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain; e is a constant; k is the attenuation factor; v is the accumulated value of the online shoveling operation segments in the target domain, which is used to represent the maturity of the target domain data;

[0035] The dynamic adjustment function of the fusion weight β of the target domain online resistance prediction model is:

[0036] β=1-α.

[0037] On the other hand, the loader resistance prediction system that integrates transfer learning and incremental learning includes:

[0038] The shoveling operation segment acquisition module is used to obtain the historical resistance data of the source domain and the historical resistance data of the target domain and perform preprocessing to obtain the historical shoveling operation segments of the source domain and the historical shoveling operation segments of the target domain;

[0039] The target domain offline resistance prediction model acquisition module is used to input the source domain historical shoveling operation segments into the resistance prediction pre-training model and train it to obtain a trained resistance prediction pre-training model, calculate the matching difference between the source domain historical shoveling operation segments and the target domain historical shoveling operation segments, screen out the inter-domain high-matching historical shoveling operation segments based on the matching difference and a set matching difference threshold, input the inter-domain high-matching historical shoveling operation segments and the target domain historical shoveling operation segments into the trained resistance prediction pre-training model, and update the fine-tuning layer parameters of the trained resistance prediction pre-training model to obtain a target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain;

[0040] An incremental learning module is used to preprocess the online resistance data during the loader's autonomous shoveling of target domain materials to obtain target domain online shoveling operation data segments. The target domain online shoveling operation data segments are input into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on historical target domain shoveling operation segments. The incrementally learned target domain offline resistance prediction model and target domain online resistance prediction model are obtained by updating the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model.

[0041] A fusion module is used to calculate the physical property similarity between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain, as well as the maturity of the target domain data. A dynamic adjustment function is generated based on the physical property similarity and the maturity of the target domain data to adjust the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model, and a fused resistance prediction model is obtained based on the fusion weights.

[0042] The resistance prediction module is used to input the online shoveling operation data segment of the target domain into the fused resistance prediction model to obtain the resistance prediction of the target domain.

[0043] The present invention adopts the above technical solution and has the following beneficial effects:

[0044] (1) The present invention dynamically adjusts and weights the historical and online resistance data of the source and target domains, thereby achieving accurate prediction of the resistance of the loader under different operating environments and improving the adaptability of the model to diverse operating environments.

[0045] (2) The present invention combines transfer learning with incremental learning. As the online data of the target domain continues to accumulate, the weights of the source domain and target domain models are dynamically adjusted, so that the prediction model can be continuously optimized, thereby achieving efficient prediction of the loader's autonomous shoveling performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a loader resistance prediction method integrating transfer learning and incremental learning according to an embodiment of the present invention;

[0047] Figure 2 A technical roadmap for a loader resistance prediction method integrating transfer learning and incremental learning according to an embodiment of the present invention;

[0048] Figure 3 This is a diagram of a device for predicting loader resistance by integrating transfer learning and incremental learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0050] like Figure 1 As shown, the loader resistance prediction method of the present invention that integrates transfer learning and incremental learning includes:

[0051] S1: Obtain historical resistance data in the source domain and historical resistance data in the target domain and perform preprocessing to obtain historical shoveling operation segments in the source domain and historical shoveling operation segments in the target domain.

[0052] Specifically, the data types of the source domain historical resistance data and the target domain historical resistance data include: loader boom cylinder displacement data, loader bucket cylinder displacement data, loader boom cylinder large cavity pressure data, loader boom cylinder small cavity pressure data, loader bucket cylinder large cavity pressure data, loader bucket cylinder small cavity pressure data, loader oil pump pressure data, loader vehicle displacement data, loader vehicle energy consumption data and loader operation resistance data.

[0053] Specifically, the preprocessing includes filtering and cleaning the source domain historical resistance data and the target domain historical resistance data, segmenting and dividing the data, and time series. In this embodiment, the operation source domain historical resistance data and the target domain historical resistance data are obtained at a ratio of 95% and 5% respectively; after preprocessing, the source domain historical shoveling operation segment Y is extracted. H =

[0054] {Y H1 、Y H2 ,……,Y Hm}, target domain historical shovel operation segment T H ={T H1 、T H2 ,……,T H(5m / 95)}.

[0055] S2: Input the historical shoveling operation segments in the source domain into the resistance prediction pre-training model and train it to obtain a trained resistance prediction pre-training model. Calculate the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain. Based on the matching difference and the set matching difference threshold, select the historical shoveling operation segments with high matching between domains. Input the historical shoveling operation segments with high matching between domains and the historical shoveling operation segments in the target domain into the trained resistance prediction pre-training model. By updating the fine-tuning layer parameters of the trained resistance prediction pre-training model, obtain the target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain.

[0056] Specifically, the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain is calculated, and based on the matching difference and the set matching difference threshold, the historical shoveling operation segments with high matching between the domains are screened out. Specifically, the following steps are performed:

[0057] S21, using the kernel density method to count the historical shoveling operation segments T in the target domain H The probability density function of is calculated as follows:

[0058]

[0059] in, Indicates The probability density value at ; Represents the historical shoveling operation segment T in the target domainH The random variables in the target domain have a value range of the historical shoveling operation segment T H The minimum to maximum value of the data points in the target domain; n represents the historical shoveling operation section T H The total number of data points in ; h represents bandwidth; G() represents kernel function; Represents the historical shoveling operation segment T in the target domain H The i-th data point in ;

[0060] S22, calculate the matching difference between the probability density function of the historical shovel loading operation segment in the source domain and the probability density function of the historical shovel loading operation segment in the target domain, and the calculation formula is as follows:

[0061]

[0062] Among them, Y Hm represents the mth historical shoveling operation segment in the source domain; represents the probability density function of the mth source domain historical shoveling operation segment; Em represents the matching difference between the mth source domain historical shoveling operation segment and the target domain historical shoveling operation segment; p represents the number of target domain historical shoveling operation segments obtained; q represents the number of source domain historical shoveling operation segments obtained; in this embodiment, the historical resistance data of the source domain and the historical resistance data of the target domain are obtained at a ratio of 95% and 5%, respectively, so q = 95, p = 5;

[0063] S23, set the matching difference threshold τ, if Em<τ, then it means that the mth segment of the source domain historical shovel operation data Y Hm The target domain history shovel operation segment T H The matching difference is less than the matching difference threshold, and the data segment is selected as the historical shoveling operation segment Y with high inter-domain matching degree. E .

[0064] In addition, in this embodiment, by updating the fine-tuning layer parameters of the trained resistance prediction pre-training model, a target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain is obtained. The specific implementation method is the same as the target domain offline resistance prediction model obtained by incremental learning in S3; that is, by inputting the historical shoveling operation segments with high inter-domain matching and the historical shoveling operation segments of the target domain into the trained resistance prediction pre-training model, the fine-tuning layer gradient is calculated based on the total loss L (including the prediction error of the new and old data segments), and the model parameters are updated accordingly, thereby obtaining the target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain.

[0065] S3, preprocess the online resistance data of the loader during the process of autonomously shoveling target domain materials to obtain the target domain online shoveling operation data segment, and input the target domain online shoveling operation data segment into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on the target domain historical shoveling operation segment respectively. By updating the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model, the incremental learning target domain offline resistance prediction model and the target domain online resistance prediction model are obtained.

[0066] Specifically, the target domain online shoveling operation data segment is input into the target domain offline resistance prediction model to obtain the incrementally learned target domain offline resistance prediction model, which specifically includes:

[0067] Based on the target domain offline resistance prediction model, the target domain online shoveling operation data segment and the historical shoveling operation segment with high matching degree between domains are predicted to obtain the total loss L, which is as follows:

[0068]

[0069] Among them, L new and L old They represent the loss of the online shoveling operation segment of the target domain and the loss of the historical shoveling operation segment with high matching between domains; n new and n old They represent the number of data points of the online shoveling operation segment of the target domain and the historical shoveling operation segment with high matching degree between domains; and They represent the predicted resistance value and the actual resistance value of the online shoveling operation section of the i-th target domain respectively; and They represent the predicted resistance value and the actual resistance value of the j-th historical shoveling operation section with high inter-domain matching degree;

[0070] Based on the total loss L, the gradient of the fine-tuning layer of the target domain offline resistance prediction model is calculated layer by layer to obtain the update direction of the model parameters;

[0071] The target domain offline resistance prediction model is continuously fine-tuned based on the update direction of the model parameters to obtain an incrementally learned target domain offline resistance prediction model.

[0072] In this embodiment, the method for implementing the target domain online resistance prediction model of incremental learning is the same as the method for implementing the target domain offline resistance prediction model of incremental learning.

[0073] Specifically, in this embodiment, the specific operation of step S3 is as follows: first, as the loader performs the first operation in the target domain, the online resistance data of the loader in the process of autonomously shoveling the target domain material is obtained and pre-processed to obtain the target domain online shoveling operation data segment T O1Then load the target domain offline resistance prediction model 0, and use the target domain online shoveling operation data segment T based on the target domain offline resistance prediction model 0. O1 and the historical shoveling operation segment Y with high matching degree between domains E Train and update the fine-tuning layer parameters of the target domain offline resistance prediction model to obtain the target domain offline resistance prediction model 1; then load the target domain online resistance prediction model 0, and use the target domain online shoveling operation data segment T based on the target domain online resistance prediction model 0. O1 and the target domain historical shovel operation data segment T H Train and update the fine-tuning layer parameters of the target domain online resistance prediction model to obtain the target domain online resistance prediction model 1;

[0074] As the loader performs the u-th operation in the target domain, the online resistance data of the loader during the process of autonomously shoveling the target domain material is obtained and preprocessed to obtain the target domain online shoveling operation data segment T Ou Then load the target domain offline resistance prediction model u-1, and use the target domain online shoveling operation data segment T based on the target domain offline resistance prediction model u-1. O1 ,……,T Ou and the historical shoveling operation segment Y with high matching degree between domains E The training and update of the fine-tuning layer parameters of the target domain offline resistance prediction model are performed to obtain the target domain offline resistance prediction model u; then the target domain online resistance prediction model u-1 is loaded, and the target domain online shoveling operation data segment T is used based on the target domain online resistance prediction model u-1. O1 ,……,T Ou and the target domain historical shovel operation data segment T H The parameters of the fine-tuning layer of the target domain online resistance prediction model are trained and updated to obtain the target domain online resistance prediction model u. As the loader performs operations from the 1st to the uth time in the target domain, through incremental learning, the target domain offline resistance prediction model 1 is continuously iteratively updated to the target domain offline resistance prediction model u, and the target domain online resistance prediction model 1 is continuously iteratively updated to the target domain online resistance prediction model u.

[0075] S4 calculates the physical property similarity between the historical shoveling operation segments in the source domain and the target domain and the maturity of the target domain data. Based on the dynamic adjustment function generated by the physical property similarity and the maturity of the target domain data, the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model are adjusted, and the fused resistance prediction model is obtained based on the fusion weights.

[0076] Specifically, the dynamic adjustment function includes a weight dynamic adjustment function of the target domain offline resistance prediction model and a weight dynamic adjustment function of the target domain online resistance prediction model. The specific setting method of the dynamic adjustment function is as follows:

[0077] Based on the prior knowledge of the historical shoveling operation segments in the source domain and the target domain, the inter-domain correlation σ representing the similarity of physical characteristics is defined, and the value range of the inter-domain correlation σ is set to (0.6, 0.9);

[0078] The dynamic adjustment function of the weight α of the target domain offline resistance prediction model is defined based on the similarity of physical properties and the maturity of the target domain data:

[0079] α=σ·e -kv ;

[0080] Where σ is the inter-domain correlation, which is used to represent the similarity of physical characteristics between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain; e is a constant; k is the attenuation factor; v is the accumulated value of the online shoveling operation segments in the target domain, which is used to represent the maturity of the target domain data;

[0081] The dynamic adjustment function of the weight β of the target domain online resistance prediction model is:

[0082] β=1-α.

[0083] Specifically, in the initial stage of the operation, the cumulative value v of the target domain online shoveling operation segment is extremely small, α approaches σ, β approaches 1-σ, and the α value is greater than the β value, that is, the fusion weight of the target domain offline resistance prediction model is greater than the fusion weight of the target domain online resistance prediction model, which reflects the early reliance on the target domain offline resistance prediction model and avoids the high variance of the target domain online resistance prediction model under low data maturity affecting the prediction stability; in the mid-term stage of the operation, as the loader performs autonomous shoveling operations in the target domain, the target domain online resistance prediction model obtains more target domain online shoveling operation segments, and the target domain The accumulated value v of the online shoveling operation section in the target domain continues to increase, α continues to decrease, and β continues to increase, that is, the fusion weight of the offline resistance prediction model in the target domain continues to decrease, and the fusion weight of the online resistance prediction model in the target domain continues to increase; at the end of the operation, the accumulated value v of the online shoveling operation section in the target domain is extremely large, α approaches 0, and β approaches 1, and the β value is greater than the α value, that is, the fusion weight of the online resistance prediction model in the target domain is greater than the fusion weight of the offline resistance prediction model in the target domain, reflecting the dependence on the online resistance prediction model in the target domain in the later stage; the fused resistance prediction model is obtained through the fusion weight at each stage.

[0084] In step S5, the online shoveling operation data segments in the target domain are input into the fused resistance prediction model to obtain a resistance prediction for the target domain. With the continuous progress of transfer learning and incremental learning, an accurate resistance prediction model for the target domain and abundant offline training data are ultimately obtained, enabling accurate prediction of autonomous shoveling performance.

[0085] Specifically, in terms of algorithm interpretability verification: First, all source and target domain resistance data should be subjected to preprocessing operations such as filtering and cleaning, data segmentation and time series, in order to adapt to the input specifications of the deep learning time series prediction model. Secondly, when building the model, the processed data set should be divided into training set, validation set and test set. The model is constructed with the help of the Keras interface of TensorFlow in Python. The Adam optimizer is uniformly used in the training process. Model training is implemented on the training set, hyperparameter tuning is completed on the validation set, performance testing is carried out on the test set, and transfer learning and incremental learning are integrated to fine-tune the model parameters. Finally, for the fused resistance prediction model, the mean absolute error (MAE) and the coefficient of determination (R 2 ), mean absolute percentage error (MAPE) and other quantitative indicators, and introduced online data to accurately measure the degree of deviation between the predicted value and the actual resistance value; at the same time, the cross-validation method was used to repeatedly verify the stability of the model on different target domain data to avoid the risk of overfitting or underfitting. For the prediction results, a time series comparison curve between the predicted value and the true value was drawn to intuitively show the model's ability to capture the resistance change trend and mutation point. Through the above-mentioned multi-dimensional verification method, from the perspectives of quantitative indicator analysis, stability testing, and visualization of prediction results, the effectiveness and rationality of the loader resistance prediction method and system that integrates transfer learning and incremental learning are fully verified, effectively ensuring its applicability and reliability in actual engineering scenarios.

[0086] like Figure 2 As shown in the figure, the technical roadmap of the loader resistance prediction method integrating transfer learning and incremental learning in an embodiment of the present invention is explained as follows: ① Acquisition of shoveling operation segment: Data preprocessing is performed on the historical resistance data of the loader operation source domain and target domain to obtain the source domain historical shoveling operation segment Y H and the target domain historical shoveling operation segment T H ② Construction of offline and online resistance prediction model: First, the source domain history shovel operation section Y H As a data set, a resistance prediction pre-training model is obtained by deep learning method, and the Y H The calculation and screening of T H High matching degree between domains with high matching degree historical shoveling operation segment Y E ; Secondly, load the resistance prediction pre-training model and use Y E and T H Fine-tune the model to obtain the target domain offline resistance prediction model 0, and use T HThe target domain online resistance prediction model 0 is trained by deep learning method. Then, as the loader operates in the target domain, the target domain offline resistance prediction model and the target domain online resistance prediction model are continuously updated by using transfer learning and incremental learning technology. Finally, when the loader operates in the target domain for the uth time, the target domain offline resistance prediction model u-1 is loaded and the target domain online resistance prediction model is continuously updated by using Y E and target domain online shoveling operation segment T O1 、T O2 ,……,T Ou Update the target domain offline resistance prediction model u and use T H and T O1 、T O2 ,……,T Ou The target domain online resistance prediction model u is updated; ③ Dynamic weight fusion and accurate resistance prediction: For each updated target domain offline resistance prediction model and target domain online resistance prediction model, dynamic fusion weight calculation is performed based on the physical property similarity and target domain data maturity to obtain the fusion weight α of the target domain offline resistance prediction model and the fusion weight β of the target domain offline resistance prediction model. Dynamic fusion is performed based on α and β to obtain a fused resistance prediction model, and the relatively latest target domain online shoveling operation segment is input into the fused resistance prediction model to obtain an accurate predicted resistance result; the loader obtains the fused resistance prediction model 1 for the first operation in the target domain. As the number of operations increases, the fused resistance prediction model is continuously updated. Finally, the fused resistance prediction model u with the highest accuracy is obtained in the u-th operation in the target domain.

[0087] like Figure 3 As shown, this embodiment also discloses a loader resistance prediction device that integrates transfer learning and incremental learning, including:

[0088] The shoveling operation segment acquisition module 31 is used to acquire the source domain historical resistance data and the target domain historical resistance data and perform preprocessing to obtain the source domain historical shoveling operation segment and the target domain historical shoveling operation segment;

[0089] The target domain offline resistance prediction model acquisition module 32 is used to input the source domain historical shoveling operation segments into the resistance prediction pre-training model and perform training to obtain a trained resistance prediction pre-training model, calculate the matching difference between the source domain historical shoveling operation segments and the target domain historical shoveling operation segments, screen out the inter-domain high-matching historical shoveling operation segments based on the matching difference and a set matching difference threshold, input the inter-domain high-matching historical shoveling operation segments and the target domain historical shoveling operation segments into the trained resistance prediction pre-training model, and update the fine-tuning layer parameters of the trained resistance prediction pre-training model to obtain a target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain;

[0090] The incremental learning module 33 is used to preprocess the online resistance data during the process of the loader autonomously shoveling target domain materials to obtain target domain online shoveling operation data segments, input the target domain online shoveling operation data segments into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on historical target domain shoveling operation segments, and update the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model to obtain the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model;

[0091] A fusion module 34 is configured to calculate the physical property similarity between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain and the maturity of the target domain data, generate a dynamic adjustment function based on the physical property similarity and the maturity of the target domain data, and adjust the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model, thereby obtaining a fused resistance prediction model based on the fusion weights.

[0092] The resistance prediction module 35 is used to input the online shoveling operation data segment of the target domain into the fused resistance prediction model to obtain the resistance prediction of the target domain.

[0093] The specific implementation of the loader resistance prediction system integrating transfer learning and incremental learning is the same as the loader resistance prediction method integrating transfer learning and incremental learning, and will not be repeated in this embodiment.

[0094] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. A loader resistance prediction method integrating transfer learning and incremental learning, characterized in that: include: S1, obtain the historical resistance data of the source domain and the historical resistance data of the target domain and perform preprocessing to obtain the historical shoveling operation segments of the source domain and the historical shoveling operation segments of the target domain; S2: Input the historical shoveling operation segments in the source domain into the resistance prediction pre-training model and train it to obtain a trained resistance prediction pre-training model. Calculate the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain. Based on the matching difference and the set matching difference threshold, select the historical shoveling operation segments with high matching between the domains. Input the historical shoveling operation segments with high matching between the domains and the historical shoveling operation segments in the target domain into the trained resistance prediction pre-training model. By updating the fine-tuning layer parameters of the trained resistance prediction pre-training model, obtain the target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain. S3: Preprocess the online resistance data during the loader's autonomous shoveling of target domain materials to obtain target domain online shoveling operation data segments. Input the target domain online shoveling operation data segments into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on historical target domain shoveling operation segments. Update the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model to obtain the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model. S4, calculating the physical property similarity between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain and the maturity of the target domain data, and adjusting the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model based on the dynamic adjustment function generated based on the physical property similarity and the maturity of the target domain data, and obtaining a fused resistance prediction model based on the fusion weights; S5, inputting the online shoveling operation data segment of the target domain into the fused resistance prediction model to obtain the resistance prediction of the target domain.

2. The loader resistance prediction method integrating transfer learning and incremental learning according to claim 1 is characterized in that: In S1, the data types of the source domain historical resistance data and the target domain historical resistance data include: loader boom cylinder displacement data, loader bucket cylinder displacement data, loader boom cylinder large cavity pressure data, loader boom cylinder small cavity pressure data, loader bucket cylinder large cavity pressure data, loader bucket cylinder small cavity pressure data, loader oil pump pressure data, loader vehicle displacement data, loader vehicle energy consumption data and loader operation resistance data.

3. The loader resistance prediction method integrating transfer learning and incremental learning according to claim 1 is characterized in that: The preprocessing specifically includes: filtering and cleaning the source domain historical resistance data and the target domain historical resistance data, data segmentation and division, and time series conversion.

4. The loader resistance prediction method integrating transfer learning and incremental learning according to claim 1 is characterized in that: In S2, the matching difference between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain is calculated. Based on the matching difference and the set matching difference threshold, the historical shoveling operation segments with high matching between the domains are screened out. Specifically, the following steps are performed: S21, using the kernel density method to count the historical shoveling operation segments T in the target domain H The probability density function of is calculated as follows: in, Indicates The probability density value at ; Represents the historical shoveling operation segment T in the target domain H The random variables in the target domain have a value range of the historical shoveling operation segment T H The minimum to maximum value of the data points in the target domain; n represents the historical shoveling operation section T H The total number of data points in ; h represents bandwidth; G() represents kernel function; Represents the historical shoveling operation segment T in the target domain H The i-th data point in ; S22, calculate the matching difference between the probability density function of the historical shovel loading operation segment in the source domain and the probability density function of the historical shovel loading operation segment in the target domain, and the calculation formula is as follows: Among them, Y Hm represents the mth historical shoveling operation segment in the source domain; represents the probability density function of the mth source domain historical shoveling operation segment; Em represents the matching difference between the mth source domain historical shoveling operation segment and the target domain historical shoveling operation segment; p represents the number of target domain historical shoveling operation segments obtained; q represents the number of source domain historical shoveling operation segments obtained; S23, set the matching difference threshold τ, if Em<τ, then it means that the mth segment of the source domain historical shovel operation data Y Hm The target domain history shovel operation segment T H The matching difference is less than the matching difference threshold, and the data segment is selected as the historical shoveling operation segment Y with high inter-domain matching degree. E .

5. The loader resistance prediction method integrating transfer learning and incremental learning according to claim 1 is characterized in that: In S3, the target domain online shoveling operation data segment is input into the target domain offline resistance prediction model. By updating the fine-tuning layer parameters of the target domain offline resistance prediction model, the incremental learning target domain offline resistance prediction model is obtained. Specifically, it includes: Based on the target domain offline resistance prediction model, the target domain online shoveling operation data segment and the historical shoveling operation segment with high matching between domains are predicted to obtain the total loss L, which is as follows: Among them, L new and L old They represent the loss of the online shoveling operation segment of the target domain and the loss of the historical shoveling operation segment with high matching between domains; n new and n old They represent the number of data points of the online shoveling operation segment of the target domain and the historical shoveling operation segment with high matching degree between domains; and They represent the predicted resistance value and the actual resistance value of the online shoveling operation section of the i-th target domain respectively; and They represent the predicted resistance value and the actual resistance value of the j-th historical shoveling operation section with high inter-domain matching degree; Based on the total loss L, the gradient of the fine-tuning layer of the target domain offline resistance prediction model is calculated layer by layer to obtain the update direction of the model parameters; The target domain offline resistance prediction model is continuously fine-tuned based on the update direction of the model parameters to obtain an incrementally learned target domain offline resistance prediction model.

6. The loader resistance prediction method integrating transfer learning and incremental learning according to claim 1 is characterized in that: In S4, the dynamic adjustment function includes the dynamic adjustment function of the fusion weight of the target domain offline resistance prediction model and the dynamic adjustment function of the fusion weight of the target domain online resistance prediction model. The specific setting method of the dynamic adjustment function is as follows: Based on the prior knowledge of the historical shoveling operation segments in the source domain and the target domain, the inter-domain correlation σ representing the similarity of physical characteristics is defined, and the value range of the inter-domain correlation σ is set to (0.6, 0.9); The dynamic adjustment function of the fusion weight α of the target domain offline resistance prediction model is defined based on the similarity of physical properties and the maturity of the target domain data: a=s·e -kv ; Where σ is the inter-domain correlation, which is used to represent the similarity of physical characteristics between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain; e is a constant; k is the attenuation factor; v is the accumulated value of the online shoveling operation segments in the target domain, which is used to represent the maturity of the target domain data; The dynamic adjustment function of the fusion weight β of the target domain online resistance prediction model is: β＝1-α。 7. A loader resistance prediction system integrating transfer learning and incremental learning, characterized in that: include: The shoveling operation segment acquisition module is used to acquire the historical resistance data of the source domain and the historical resistance data of the target domain and perform preprocessing to obtain the historical shoveling operation segments of the source domain and the historical shoveling operation segments of the target domain; The target domain offline resistance prediction model acquisition module is used to input the source domain historical shoveling operation segments into the resistance prediction pre-training model and train it to obtain a trained resistance prediction pre-training model, calculate the matching difference between the source domain historical shoveling operation segments and the target domain historical shoveling operation segments, screen out the inter-domain high-matching historical shoveling operation segments based on the matching difference and a set matching difference threshold, input the inter-domain high-matching historical shoveling operation segments and the target domain historical shoveling operation segments into the trained resistance prediction pre-training model, and update the fine-tuning layer parameters of the trained resistance prediction pre-training model to obtain a target domain offline resistance prediction model that supports transfer learning from the source domain to the target domain; An incremental learning module is used to preprocess the online resistance data during the loader's autonomous shoveling of target domain materials to obtain target domain online shoveling operation data segments. The target domain online shoveling operation data segments are input into the target domain offline resistance prediction model and the target domain online resistance prediction model trained based on historical target domain shoveling operation segments. The incrementally learned target domain offline resistance prediction model and target domain online resistance prediction model are obtained by updating the fine-tuning layer parameters of the target domain offline resistance prediction model and the target domain online resistance prediction model. A fusion module is used to calculate the physical property similarity between the historical shoveling operation segments in the source domain and the historical shoveling operation segments in the target domain, as well as the maturity of the target domain data. A dynamic adjustment function is generated based on the physical property similarity and the maturity of the target domain data to adjust the fusion weights of the incrementally learned target domain offline resistance prediction model and the target domain online resistance prediction model, and a fused resistance prediction model is obtained based on the fusion weights. The resistance prediction module is used to input the online shoveling operation data segment of the target domain into the fused resistance prediction model to obtain the resistance prediction of the target domain.

Images