Strip mine dump slope stable state discrimination method based on transfer learning

Abstract

The invention discloses a method for judging the stable state of a slope of a strip mine dump based on transfer learning. The method comprises the following steps: step 1, acquiring monitoring data of a strip mine dump slope, and dividing the monitoring data into source domain data and target domain data; 2, processing the collected monitoring data through a TCA algorithm, and mapping a processed data set to a shared feature space; step 3, in the feature space after TCA mapping, training target domain data by using a GBDT classifier, updating a sample weight at the same time, and using migration of a high-weight sample in source domain data; and step 4, utilizing the trained TCA-GBDT discrimination model to accurately discriminate the slope stability state of the target domain data, outputting a stability evaluation result, and providing a scientific basis for risk prediction and decision making of the refuse dump. The method is suitable for a small sample data set environment, the classification precision is improved, and the problem that a traditional method is low in accuracy under the condition of insufficient data is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of determining the stability state of an open-pit mine dump slope, and in particular to a method for determining the stability state of an open-pit mine dump slope based on transfer learning. Background Art

[0002] Traditional slope stability analysis methods, such as geological analogy and limit equilibrium method, can reflect the stability status of the slope to a certain extent. However, when faced with complex and changeable geological conditions and multiple influencing factors, their judgment accuracy and efficiency are often difficult to meet actual needs.

[0003] With the rapid development of machine learning technology, the GBDT algorithm has been widely used in the field of discriminant analysis due to its powerful feature combination capabilities, efficient computational efficiency, and excellent ability to handle nonlinear relationships. The GBDT algorithm also shows great potential in determining the stability of open-pit mine dump slopes.

[0004] Traditional machine learning models typically rely on large amounts of sample data to establish mapping relationships between data, thereby achieving high-accuracy classification of test samples. Machine learning models are insufficient in handling complex nonlinear problems and are extremely sensitive to data volume and quality. These models typically require a large amount of valid data to establish an accurate mapping relationship between input features and output features. However, when slope instability occurs in open-pit mine dumps, the disaster is often short-lived, severe, and wide-ranging, making it difficult to obtain sufficient and valid slope instability sample data. Therefore, machine learning models are prone to misjudgment and inaccurate recognition in a small sample data environment, resulting in poor classification performance, which in turn affects the accuracy of slope stability state judgment. Summary of the Invention

[0005] In order to overcome the above technical problems, the purpose of the present invention is to provide a method for judging the stability state of open-pit mine spoil dump slopes based on transfer learning. This method is suitable for small sample data set environments, improves classification accuracy, and overcomes the problem of low accuracy of traditional methods in the case of insufficient data.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for determining the stability state of an open-pit mine dump slope based on transfer learning comprises the following steps:

[0008] Step 1: Obtain monitoring data of the slope of the open-pit mine dump and divide the monitoring data into source domain data and target domain data;

[0009] Step 2: Process the collected monitoring data using the TCA algorithm and map the processed dataset into a shared feature space, thereby reducing inter-domain distribution differences and enhancing inter-class separability.

[0010] Step 3: In the feature space after TCA mapping, use the GBDT classifier to train the target domain data and update the sample weights at the same time. By migrating high-weight samples from the source domain data, the target domain’s discrimination ability is improved and the performance of the target domain model is further optimized.

[0011] Step 4: Use the trained TCA-GBDT discriminant model to accurately judge the slope stability state of the target domain data and output the stability assessment results to provide a scientific basis for risk prediction and decision-making of the spoil dump.

[0012] The step 1 is specifically as follows:

[0013] Monitoring data is obtained through the monitoring data system:

[0014] The monitoring data system includes a soil pressure sensor, a pore water pressure sensor, a soil moisture sensor and a data acquisition instrument;

[0015] The soil pressure sensor (model BW16-1) and pore water pressure sensor (model BWK13-1) provided by Liyang Jincheng Monitoring Instrument Co., Ltd. were used to monitor the stress changes of the spoil dump slope.

[0016] The measuring range of the soil pressure sensor is -50 to 50 kPa. It is designed so that the smooth surface is used as the load-bearing surface. When in use, the smooth surface should be facing upwards and covered with fine soil to ensure the accuracy of the measurement.

[0017] The pore water pressure sensor also has a range of -50 to 50 kPa. When using it, make sure that the water hole is facing vertically downwards.

[0018] The soil moisture sensor (model VMS-3000-TR) has a range of 0 to 100%, a probe length of 7 cm, and outputs a voltage signal of 0 to 10 V;

[0019] The YB-2001 dynamic and static strain collector is equipped with 20 measuring point channels and an adjustable sampling frequency ranging from 1 to 10 Hz to meet different monitoring needs.

[0020] When conducting monitoring experiments, we set up monitoring points at three key locations: the top, middle, and foot of the slope, and installed corresponding sensors.

[0021] The key monitoring data of moisture content, soil pressure and pore water pressure are collected through soil pressure sensors, pore water pressure sensors and soil moisture sensors respectively; data collection is carried out at a frequency of 10 Hz, that is, 10 data points are recorded per second.

[0022] The sensor converts the physical signal into an electrical signal, while the data logger converts the electrical signal into a pressure value, simplifies the data by calculating the minute average, and fills in the data points that are invalid due to sensor movement with the last valid value.

[0023] The data is divided into source domain and target domain according to the different rock contents of the dump slope. Specifically, the stability data of the dump slope with 10% rock content is used as the source domain, and the stability data of the dump slope with 20% rock content is used as the target domain.

[0024] The specific implementation method of step 2 is as follows:

[0025] (1) Input data preparation:

[0026] Source domain data:

[0027] Target domain data:

[0028] Where: is the characteristic vector of the i-th sample in the target area (slope stability data of the spoil dump with 10% rock content); is the label corresponding to the target domain (slope stability data of the spoil dump with 10% stone content); n s The number of samples in the source field data (slope stability data of the spoil dump with 10% rock content); is the characteristic vector of the i-th sample in the target area (slope stability data of the spoil dump with 20% rock content); n t The number of samples in the source field data (slope stability data of the spoil dump with 20% rock content);

[0029] (2) Domain Alignment

[0030] Domain alignment is achieved by minimizing the maximum mean difference between source and target domain features;

[0031]

[0032] (3) Feature transformation and mapping

[0033] By minimizing MMD, TCA learns a transformation function T that maps the data in the source and target domains into a shared feature space; in this shared feature space, the distribution difference between the source and target domains is minimized; the feature transformation is expressed as:

[0034] x ′ =T(x)(4)

[0035] Where: x ′ are the features after transformation.

[0036] The specific implementation method of step 3 is as follows:

[0037] (1) Initialize weak learner:

[0038]

[0039] Where: f0(x) is the decision tree fitting function; L(y i ,c) is the loss function; y i is the sample category; i is the i-th sample; c is a constant;

[0040] (2) Construct m weak learners, for each sample (x i ,y i ), i = 1, 2, ..., n, calculate the negative gradient of the loss function of the i-th sample at the m-th iteration:

[0041]

[0042] Where: f m-1 is the cumulative estimated result up to the m-1th independent decision tree; x i is the feature vector of the i-th sample;

[0043] Use negative gradient to train weak learners, and finally iteratively construct m independent decision trees f1(x),f2(x),...,f m (x), m=1,2,...M. Assume that the number of leaf nodes of the mth regression tree is j (j=1,2,3,...,J), and the leaf nodes divide the input space of each tree into j independent sub-regions R m1 ,R m2 ,...,R mj , calculate the area R mj The best fitting value C mj :

[0044]

[0045] (3) Update the weak learner f m (x):

[0046]

[0047]

[0048] Where fm is the cumulative estimated result up to the mth independent decision tree;

[0049] (4) After M rounds of iteration, the final gradient boosting regression tree expression is:

[0050]

[0051] Where F(x) is the final gradient boosting regression tree.

[0052] The specific implementation method of step 4 is as follows:

[0053] The maximum mean difference is used to map the data of the source and target domains into a shared low-dimensional feature space. In this space, the distributions of the source and target domains are aligned as much as possible to reduce the distribution differences between the domains. On the other hand, the weights of weak classifiers with large classification error rates are reduced so that they play a smaller role in voting. The classification results of the base classifiers are integrated to generate the final TCA-GBDT spoil dump slope stability discrimination model based on the idea of ​​transfer learning.

[0054] Beneficial effects of the present invention:

[0055] The present invention obtains monitoring data from the slopes of open-pit mine dumps and divides it into a source domain and a target domain. The collected data is processed using the TCA algorithm, mapping the dataset into a shared feature space. After mapping the data into a high-dimensional space using a kernel function, the present invention minimizes the distribution differences between the source and target domains, aligning the statistical characteristics of the two domains in the shared space. This reduces the distribution offset of the monitoring data caused by different soil-rock mixture ratios. During the mapping process, TCA preserves the principal component variance and class separability of the data through constraints, ensuring that the mapped features eliminate domain differences while containing key information for determining the stability of the slope. This reduces inter-domain distribution differences and enhances inter-class separability.

[0056] The present invention uses the GBDT classifier to train the target domain data in the feature space after TCA mapping, and updates the sample weights at the same time. With the help of the migration of high-weight samples in the source domain, the discrimination ability of the target domain is improved, and the performance of the target domain model is further optimized; in each GBDT iteration, the weight of the target domain samples with classification errors is reduced according to the classification error of the current model in the target domain to avoid overfitting, and the weight of the source domain samples with large distribution differences from the target domain is reduced to suppress negative migration. The high-weight source domain samples that are consistent with the target domain distribution are retained so that they can play a greater role in subsequent training. GBDT gradually fits the residuals of the target domain data through gradient boosting, and the addition of high-weight source domain samples is equivalent to providing the model with a priori common laws across working conditions, making up for the lack of information of small samples in the target domain and improving the generalization of the model.

[0057] The present invention utilizes the trained TCA-GBDT discriminant model. When new data of the target domain is input, TCA maps it to an aligned shared space to eliminate the feature distribution offset caused by differences in stone content. GBDT is based on the common features in the shared space and integrates a voting mechanism through a tree model to calculate the probability that a sample belongs to the "stable" or "unstable" category. The model outputs a binary classification label, which, combined with the threshold judgment, provides a quantitative risk assessment basis for engineering personnel. For example, if the model determines that a certain area is in a "critical state", it can trigger an early warning and guide reinforcement measures to avoid landslide accidents. The stability state of the slope in the target domain is accurately judged, and the stability assessment results are output, providing a scientific basis for risk prediction and decision-making in the spoil dump. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of the overall scheme of the open-pit mine dump slope stability judgment method based on transfer learning of the present invention.

[0059] Figure 2 This is a regional diagram for monitoring point arrangement in the present invention.

[0060] Figure 3 This is the structure diagram of the GBDT classifier used in the present invention.

[0061] Figure 4 This is a structural diagram of the TCA-GBDT model based on transfer learning used in the present invention. DETAILED DESCRIPTION

[0062] The present invention will be described in further detail below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, a method for determining the stability state of an open-pit mine dump slope based on transfer learning includes the following steps:

[0064] Step 1: Obtain monitoring data of the slope of the open-pit mine dump and divide the monitoring data into a source domain and a target domain;

[0065] The monitoring data system consists of earth pressure sensors, pore water pressure sensors, soil moisture sensors, and data acquisition equipment. Earth pressure sensors (Model BW16-1) and pore water pressure sensors (Model BWK13-1) provided by Liyang Jincheng Monitoring Instrument Co., Ltd. are used to monitor stress changes in the spoil dump slope. These earth pressure sensors have a measurement range of -50 to 50 kPa and an accuracy of ±0.5% FS. They are designed with the smooth surface as the load-bearing surface. When used, the smooth surface should face upward and be covered with fine soil to ensure accurate measurements. The pore water pressure sensor also has a range of -50 to 50 kPa and an accuracy of ±0.5% FS. When used, the permeable holes should be facing vertically downward. The soil moisture sensor (Model VMS-3000-TR) has a range of 0 to 100%, an accuracy of ±3%, and a resolution of 10 μs / cm. The probe is 7 cm long and outputs a 0 to 10 V voltage signal, which can be converted to a moisture content value using a specific formula. The YB-2001 dynamic and static strain collector is equipped with 20 measuring point channels and an adjustable sampling frequency ranging from 1 to 10 Hz to meet different monitoring needs.

[0066] Considering that rainfall poses a threat to the stability of the spoil dump slope, it mainly manifests in the following aspects: First, after the spoil dump soil absorbs rainwater, the moisture content increases, resulting in an increase in soil weight, thereby increasing the sliding force; second, the increase in soil moisture will increase the earth pressure and reduce the shear strength of the soil-rock bulk, which will reduce the stability of the spoil dump slope; third, the increase in pore water pressure will further weaken the shear strength of the slope, causing the slope to suffer varying degrees of erosion, thereby increasing the risk of slope sliding and collapse. Therefore, when conducting the monitoring experiment, three locations were selected as monitoring points: the top, middle and foot of the slope, and sensors were installed, namely, moisture content sensor, earth pressure sensor and pore water pressure sensor, such as Figure 2 There are 3 sensors at each location, for a total of 9 sensor monitoring points.

[0067] Sensors collect key monitoring data on moisture content, earth pressure, and pore water pressure. To accurately capture real-time changes in these data, data acquisition is performed at a 10Hz frequency, recording 10 data points per second. Data preprocessing is crucial because rainfall affects slope stability with a time delay. Sensors convert physical signals into electrical signals, while the data acquisition system converts these electrical signals into pressure values. However, monitoring data can be subject to missing values, noise, redundancy, and outliers. Analysis revealed a high incidence of subtle variations and redundancy in the data, leading to the choice of calculating minute-by-minute averages to enhance data accuracy. In practice, slope failure may cause some sensors to shift position, rendering their readings invalid. To address this, a strategy of using fixed values ​​to fill in missing data points is implemented: the last valid reading before sensor failure is used to replace the data after the failure. In short, calculating minute-by-minute averages simplifies the data, and using the last valid value to fill in data points lost due to sensor movement.

[0068] The data is divided into source domain and target domain according to the different rock contents of the dump slope. Specifically, the stability data of the dump slope with 10% rock content is used as the source domain, and the stability data of the dump slope with 20% rock content is used as the target domain.

[0069] Slopes with different rock contents (10% and 20%) exhibit significant differences in physical properties (for example, the soil-rock ratio affects shear strength and permeability), leading to different distributions of monitoring data. By using the 10% condition with a lower rock content as the source domain and the 20% condition with a higher rock content as the target domain, we can simulate cross-condition transfer learning scenarios and verify the model's adaptability to different soil-rock ratio conditions.

[0070] Only part of the data in the target domain (20% stone content) is used in order to construct a small sample condition to verify the classification performance of the TCA-GBDT model when the target domain data is scarce, which is in line with its core research goal of "small sample improvement".

[0071] In actual open-pit mine dumps, the soil-rock mix ratio in different areas may change dynamically (for example, due to filler gradation adjustments during construction). However, obtaining complete monitoring data for all working conditions is costly. This partitioning approach simulates the need to transfer knowledge from known working conditions (source domain) to unknown / new working conditions (target domain), closely resonating with real-world engineering application scenarios.

[0072] By migrating the common feature rules between the source domain (10% stone content) and the target domain (20% stone content), the underfitting problem of the model caused by insufficient samples in the target domain is compensated.

[0073] Differences in soil-rock mix ratios can introduce domain shifts (e.g., a slope with a 20% rock content may have a higher failure threshold). Transfer learning dynamically adjusts the weights of source and target domain samples to suppress interference from inconsistent data distributions in the source and target domains, enhancing the model's robustness to changes in rock content.

[0074] This division can verify: when the target working condition (such as a new construction area) lacks sufficient monitoring data, whether historical data of similar working conditions (such as monitored areas) can be reused to assist in modeling and reduce data collection costs.

[0075] It is verified that the proposed TCA-GBDT model can still maintain the accuracy of slope stability judgment under extreme conditions with large data distribution differences and few target domain samples;

[0076] Provide methodological support for the reuse of monitoring data of slopes with different soil-rock mixture ratios in actual engineering.

[0077] Step 2: Process the collected data using the TCA algorithm to map the dataset into a shared feature space, thereby reducing inter-domain distribution differences and enhancing inter-class separability. The specific implementation method is as follows:

[0078] (1) Input data preparation:

[0079] Source domain data:

[0080] Target domain data:

[0081] Where: is the feature vector of the i-th sample in the target domain; is the corresponding label; n s is the number of samples in the source domain dataset; is the feature vector of the i-th sample in the target domain; n t is the number of samples in the source domain dataset.

[0082] (2) Domain Alignment

[0083] Domain alignment is achieved by minimizing the maximum mean difference between source and target domain features.

[0084]

[0085] (3) Feature transformation and mapping

[0086] By minimizing MMD, TCA learns a transformation function T that maps the data in the source and target domains into a shared feature space. In this shared feature space, the distribution difference between the source and target domains is minimized. The feature transformation can be expressed as:

[0087] x ′ =T(x)(4)

[0088] Where: x ′ are the features after transformation.

[0089] Step 3: In the feature space after TCA mapping, use the GBDT classifier (such as Figure 3 As shown in Figure 2 , GBDT (Gradient Boosted Decision Tree) uses multiple regression trees as base learners and gradually improves model performance through a forward distribution algorithm. Its primary goal is to reduce the residual error of the weak learner from the previous iteration. In each iteration, a new weak learner is constructed based on the gradient direction of the residual error. The weights of the sample data are adjusted to enable the new learner to more comprehensively learn the data characteristics, thereby improving prediction accuracy. Finally, the results of all weak learners are additively combined to produce a more accurate prediction.

[0090] The specific implementation methods are as follows:

[0091] (1) Initialize weak learner:

[0092]

[0093] Where: f0(x) is the decision tree fitting function; L(y i ,c) is the loss function; y i is the sample category; i is the i-th sample; c is a constant.

[0094] (2) Construct m weak learners, for each sample (x i ,y i ), i = 1, 2, ..., n, calculate the negative gradient of the loss function of the i-th sample at the m-th iteration:

[0095]

[0096] Where: f m-1 is the cumulative estimated result up to the m-1th independent decision tree; x i is the feature vector of the i-th sample.

[0097] Use negative gradient to train weak learners, and finally iteratively construct m independent decision trees f1(x),f2(x),...,f m (x), m=1,2,...M. Assume that the number of leaf nodes of the mth regression tree is j (j=1,2,3,...,J), and the leaf nodes divide the input space of each tree into j independent sub-regions R m1 ,R m2 ,...,R mj , calculate the area R mjThe best fitting value C mj :

[0098]

[0099] (3) Update the weak learner f m (x):

[0100]

[0101] Where f m is the cumulative estimated result up to the mth independent decision tree.

[0102] (4) After M rounds of iteration, the final gradient boosting regression tree expression is:

[0103]

[0104] Where F(x) is the final gradient boosting regression tree.

[0105] Step 4: Use the trained TCA-GBDT discriminant model (such as Figure 4 As shown in Figure 2 ), the GBDT model can effectively learn the relationship between feature variables and label categories in the spoil dump slope stability dataset. Therefore, drawing on the concept of transfer learning, we first determine the datasets for the source and target domains. During the iterative training of samples, the TCA algorithm is used to dynamically update the weights of samples from the two domains, and the GBDT model is used as a weak learner in the training process. Ultimately, based on the classification results of the weak learner, a voting strategy is used to generate a transfer learning-based TCA-GBDT model for determining the stability of the spoil dump slope. The specific implementation method is as follows:

[0106] Using techniques such as maximum mean difference, the data of the source and target domains are mapped to a shared low-dimensional feature space. In this space, the distributions of the source and target domains are aligned as much as possible to reduce the distribution differences between the domains. On the other hand, the weights of weak classifiers with large classification error rates are reduced so that they play a smaller role in voting. The classification results of the base classifiers are integrated to generate the final TCA-GBDT spoil dump slope stability discrimination model based on the idea of ​​transfer learning.

[0107] In order to evaluate the effectiveness of the TCA-GBDT model in discriminating the stability state of the spoil dump slope, the GBDT, AdaBoost and SVM models were compared and analyzed, and the accuracy, precision, recall and AUC values ​​were used as evaluation indicators to measure the model performance.

[0108] Evaluation metrics are key tools for measuring model learning effectiveness, prediction accuracy, and generalization. They help understand model performance, select the optimal model, adjust hyperparameters, and assess potential risks. When evaluating models, particular attention should be paid to their performance in the task of determining the stability of spoil dump slopes. To this end, four metrics are used to measure model performance: accuracy, precision, recall, and area under the curve (AUC). The closer the AUC value is to 1, the stronger the model's classification ability.

[0109] Table 1 Performance evaluation indicators of different models

[0110]

[0111]

[0112] Comprehensive analysis results show that according to the data in Table 1, the TCA-GBDT model performed best in terms of accuracy, with a score of 0.857, which shows that the model can correctly identify samples at a high ratio. In terms of precision, the TCA-GBDT model scored 0.785, showing high accuracy in identifying slope instability. In addition, the recall rate of the model was 0.917, which means that it can effectively capture more real instability situations and reduce the risk of missed judgments. The AUC value reached 0.866, showing that the model has good classification ability and can effectively distinguish between stable and unstable states of the slope. Overall, the TCA-GBDT model performed outstandingly in the assessment of slope stability in open-pit mine spoil dumps, especially when dealing with small sample data sets in actual engineering projects, showing excellent classification ability and strong robustness.

Claims

1. A method for determining the stability of open-pit mine dump slopes based on transfer learning, characterized in that: The following steps are included: Step 1: Obtain monitoring data of the slope of the open-pit mine dump and divide the monitoring data into source domain data and target domain data; Step 2: Process the collected monitoring data using the TCA algorithm and map the processed data set into a shared feature space; Step 3: In the feature space after TCA mapping, use the GBDT classifier to train the target domain data and update the sample weights at the same time, using the migration of high-weight samples in the source domain data; Step 4: Use the trained TCA-GBDT discriminant model to accurately judge the slope stability state of the target domain data and output the stability assessment results to provide a scientific basis for risk prediction and decision-making of the spoil dump.

2. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 1, characterized in that: The step 1 is specifically as follows: Monitoring data is obtained through the monitoring data system: The monitoring data system includes a soil pressure sensor, a pore water pressure sensor, a soil moisture sensor and a data acquisition instrument; Use soil pressure sensors and pore water pressure sensors to monitor stress changes in the dump slope; The measuring range of the soil pressure sensor is -50 to 50 kPa. Its smooth surface is used as the load-bearing surface. When in use, the smooth surface should face upwards and be covered with fine soil to ensure the accuracy of the measurement. The pore water pressure sensor has a range of -50 to 50 kPa. When using it, make sure that the water hole is facing vertically downwards; When conducting monitoring experiments, monitoring points are set up at the top, middle and foot of the slope, and corresponding sensors are installed; The key monitoring data of moisture content, soil pressure and pore water pressure are collected through soil pressure sensors, pore water pressure sensors and soil moisture sensors respectively; Data acquisition was performed at a frequency of 10 Hz, i.e., 10 data points were recorded per second.

3. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 2, characterized in that: The sensor is responsible for converting physical signals into electrical signals, and the data logger converts the electrical signals into pressure values. It simplifies the data by calculating the minute average and uses the last valid value to fill in the data points that are invalid due to sensor movement.

4. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 1, characterized in that: The monitoring data are divided into source domain data and target domain data according to the different stone content of the spoil dump slope; the stability data of the spoil dump slope with 10% stone content is used as the source domain, and the partial stability data of the spoil dump slope with 20% stone content is used as the target domain.

5. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 3, characterized in that: The specific implementation method of step 2 is as follows: (1) Input data preparation: Source domain data: Target domain data: Where: is the feature vector of the i-th sample in the target domain; is the label corresponding to the target domain; n s is the number of samples in the source domain dataset; is the feature vector of the i-th sample in the target domain; n t is the number of samples in the source domain dataset; (2) Domain Alignment Domain alignment is achieved by minimizing the maximum mean difference between source and target domain features; (3) Feature transformation and mapping By minimizing MMD, TCA learns a transformation function T that maps the data in the source and target domains into a shared feature space; in this shared feature space, the distribution difference between the source and target domains is minimized; the feature transformation is expressed as: x′=T(x) (4) Where: x′ is the feature after transformation.

6. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 1, characterized in that: The specific implementation method of step 3 is as follows: (1) Initialize weak learner: Where: f0(x) is the decision tree fitting function; L(y i ,c) is the loss function; y i is the sample category; i is the i-th sample; c is a constant; (2) Construct m weak learners, for each sample (x i ,y i ), i = 1, 2, ..., n, calculate the negative gradient of the loss function of the i-th sample at the m-th iteration: Where: f m-1 is the cumulative estimated result up to the m-1th independent decision tree; x i is the feature vector of the i-th sample; Use negative gradient to train weak learners, and finally iteratively construct m independent decision trees f1(x),f2(x),...,f m (x), m = 1, 2, ... M; set the number of leaf nodes of the mth regression tree to j (j = 1, 2, 3, ..., J), and the leaf nodes divide the input space of each tree into j independent sub-regions R m1 ,R m2 ,...,R mj , calculate the area R mj The best fitting value C mj : (3) Update the weak learner f m (x): Where f m is the cumulative estimated result up to the mth independent decision tree; (4) After M rounds of iteration, the final gradient boosting regression tree expression is: Where F(x) is the final gradient boosting regression tree.

7. The method for determining the stability of an open-pit mine dump slope based on transfer learning according to claim 1, characterized in that: The specific implementation method of step 4 is as follows: The maximum mean difference is used to map the source domain data and the target domain data to a shared low-dimensional feature space. In this space, the distribution of the source domain data and the target domain data is aligned as much as possible to reduce the distribution difference between the domains. On the other hand, the weight of the weak classifier with a large classification error rate is reduced so that it plays a smaller role in the voting. The classification results of the base classifier are integrated to generate the final TCA-GBDT spoil dump slope stability discrimination model based on the transfer learning idea.

Images