Landslide early warning rule generation method based on abnormal deformation mining of multi-source monitoring data

Abstract

The application discloses a landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining, which identifies a landslide rapid deformation point based on historical monitoring data, determines a key attention time period of a landslide displacement factor, and generates an early warning rule after establishing a data set. The generated method mines association rules in the key attention time period of the landslide displacement factor, has low cost, strong real-time performance, and can be applied to a large number of monitoring sites.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of landslide prediction and early warning, and particularly relates to a landslide early warning rule generation method based on abnormal deformation mining of multi-source monitoring data. BACKGROUND

[0002] In the prediction and early warning work, the release of disaster danger level according to certain early warning rules has been widely applied. At present, most landslides still use the fixed trigger conditions provided by the geological team as the only judgment rule. Such indicators are highly professional. The specific implementation scheme is roughly as follows: the engineering geology personnel go to the scene to conduct geological exploration, obtain the lithological structure of the landslide body, and then set the threshold of the trigger condition through work experience, such as rainfall, displacement speed, displacement acceleration and other indicators. When the trigger indicators reach a certain level, the corresponding early warning level (red, orange, yellow, blue) is issued.

[0003] The existing early warning method has high investment cost, long update cycle and poor timeliness, and it is difficult to meet the prediction needs of a large number of landslide sites. At the same time, the early warning rules generated by this traditional method are difficult to combine multiple observation indicators for judgment, and the early warning results have a certain one-sidedness. For the universal prediction and early warning work of a large number of landslide sites, due to the limitation of investment cost and the demand of work goal, each slope body does not need to be extremely accurately explored and evaluated, but needs to be quickly predicted in displacement and warned in landslide danger under the premise of ensuring a certain accuracy. It is under this premise that the present application is proposed. SUMMARY

[0004] The purpose of the present application is to provide a landslide early warning rule generation method based on abnormal deformation mining of multi-source monitoring data, which has low cost, can quickly obtain early warning rules, and has strong real-time performance.

[0005] To achieve the purpose of the present application, the landslide early warning rule generation method based on abnormal deformation mining of multi-source monitoring data comprises the following steps:

[0006] Step 1: Establish a historical monitoring database to obtain historical monitoring data of a certain historical period. The historical monitoring data includes landslide displacement and rainfall. The historical monitoring data is formed into a plurality of sample data in units of days. Each sample data includes the date of the day, the displacement of the day and the rainfall of the day;

[0007] Step 2: Obtain the deformation point. Derive the displacement of each sample data to obtain the first derivative representing the landslide displacement speed and the second derivative representing the landslide displacement acceleration of the displacement of the day with respect to the date of the day. The date corresponding to the first derivative result and the second derivative result greater than the preset threshold is taken as the deformation point;

[0008] Step three: establishing a rapid deformation database, extracting the displacement and rainfall corresponding to the deformation point determined in step two from the historical monitoring database, and establishing a rapid deformation database with the first derivative and the second derivative corresponding to the deformation point as the unit of day;

[0009] Step four: determining the key attention time period of the landslide displacement factor, inputting the data in the rapid deformation database into the landslide displacement prediction model, outputting the distribution of attention intensity of displacement and rainfall with time distance, and determining the key attention time period of the landslide displacement factor according to the distribution;

[0010] Step five: establishing a data set, the data set including a plurality of item sets, each item set including the landslide relative displacement distance of the deformation point and the daily rainfall in the key attention time period;

[0011] Step six: generating an early warning rule, which is generated in one of the following ways:

[0012] 1) determining the frequent item set in the established data set and the support degree of the frequent item set, setting the threshold of each risk level according to the support degree of the frequent item set, and forming the early warning rule;

[0013] 2) obtaining the landslide rapid displacement rainfall pattern by using the FP-growth frequent item set algorithm, setting the threshold of each risk level according to the support degree of the frequent item set in the construction of the landslide rapid displacement rainfall pattern, and forming the early warning rule.

[0014] In some embodiments, the landslide displacement prediction model includes:

[0015] a time series convolution network model for performing convolution operation on the input quantity to obtain the feature vector of each input quantity, and splicing the features at the corresponding time in the displacement and rainfall at different time sequences;

[0016] a fully connected layer for mixing the features of the time series convolution network model output and then adding position coding, so that the output of the time series convolution network model meets the input format requirements of the Transformer model data; and

[0017] a Transformer model decoder for obtaining important influence factors of the landslide displacement.

[0018] In some embodiments, the time series convolution network model includes a causal convolution for ensuring that the features of the current time are only related to the samples of the historical time.

[0019] In some embodiments, the step of establishing the historical monitoring database first cleans the data, removes outliers and smooths burr data, and then constructs the historical monitoring database by taking the corresponding displacement and rainfall as a unit of day.

[0020] Another object of the present application is to provide a landslide displacement prediction model, comprising:

[0021] a time series convolution network model, configured to perform convolution operation on input quantity to obtain respective feature vectors of the input quantity, and splice features at corresponding time points in displacement and rainfall at different time sequences;

[0022] a full connection layer, configured to perform feature mixing on the output of the time series convolution network module and then add position coding, so that the output of the time series convolution network model meets the input format requirements of the Transformer model data; and

[0023] a Transformer model decoder, configured to obtain important influence factors affecting the landslide displacement.

[0024] In some embodiments, the time series convolution network model comprises a causal convolution configured to ensure that the features at the current time are only related to the samples at the historical time.

[0025] The technical scheme of the present application can achieve at least the following beneficial effects:

[0026] 1) The present generation method mines association rules in the key attention time period of the landslide displacement influencing factor, has low cost, high real-time performance, and can be widely applied to a large number of monitoring sites.

[0027] 2) The present generation method takes the deformation point and the data of the deformation point in the previous period of time as the basis for finding important influence factors affecting the landslide displacement, considers many influence factors, and effectively avoids the one-sidedness of the empirical method.

[0028] 3) The present method can obtain the early warning rule relatively quickly under the premise of ensuring a certain accuracy.

[0029] 4) The landslide short-term displacement prediction model provided by the present application can ensure that the features at the current time are only related to the samples at the historical time through the causal convolution in the model, and is suitable for time series autoregressive scenarios; and the time series convolution network module introduces residual links, reduces the data processing time, and improves the model effect. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, further serve to explain the principles of the application. It is to be understood that the drawings are only schematic, and that they do not purport to be to scale with respect to one another. The embodiments will be described with additional specificity and detail to contribute to a full understanding of the application.

[0031] Figure 1 A landslide displacement curve diagram according to the present application;

[0032] Figure 2 A landslide early warning rule generation method flowchart according to the present application;

[0033] Figure 3 A landslide displacement curve diagram;

[0034] Figure 4 A model architecture diagram of a Transformer model according to the present application;

[0035] Figure 5 A landslide short-term displacement prediction model structure diagram according to the present application;

[0036] Figure 6 An attention distribution diagram according to the present application;

[0037] Figure 7 A diagram of attention distribution over time and distance according to the present application. DETAILED DESCRIPTION

[0038] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the example implementations to those skilled in the art.

[0039] For ease of understanding, the following gives an explanation of the concepts related to the present application.

[0040] 1. Historical monitoring data

[0041] The monitoring data within a certain time before the early warning period, such as 30 days of monitoring data before the early warning period, 45 days of monitoring data, 90 days of monitoring data, or monitoring data from the start of monitoring day to the day before the early warning period, etc., are more accurate as the adjacent month data are similar in terms of weather, rainfall, displacement, etc. The monitoring data include the displacement of the landslide of the monitoring point and the rainfall, and the monitoring data are counted in units of days, with daily data as a group of sample data.

[0042] The early warning period refers to a period in which the monitoring point needs to be early warning monitored.

[0043] 2. Item set, single item set, multi-item set and frequent item set

[0044] A set of items is called an item set. An item set containing N items can be called a multi-item set. When N = 1, the item set can be called a single item set. A frequent item is an item set that frequently occurs in a database. In this application, the items in the item set are the landslide relative displacement distances of the deformation points and the daily rainfall in the time period of interest.

[0045] 3. Support

[0046] Support refers to the proportion of an item or item set in the entire data set. The support can be expressed by the following formula:

[0047]

[0048] In the formula: represents the support of the item or item set, represents the item or item set occurs in the data set, represents the total number of items or item sets in the data set.

[0049] 4. Early warning rule

[0050] It refers to a rule that corresponds to the output risk level under the monitoring data, such as Figure 1 The early warning rule matrix as shown in the figure is composed of a plurality of early warning rules. When the monitoring data is the rule early warning corresponding condition, the corresponding risk level is generated. The monitoring data includes rainfall and displacement.

[0051] 5. FPGrowth algorithm

[0052] The basic idea of the algorithm is to scan the data set, calculate the support of each item and item set, and sort them from large to small, to get the support table; start from the bottom minimum support, build FP tree one by one; associate the FP tree and the support table, each item in the support table has a pointer pointing to the corresponding node in the FP tree, through a few full table scans, the FP tree built will change the irregular data of the shopping cart into a traceable tree structure, and save the huge natural connection operation; finally, through the FP tree, the association rules are mined, through the FP tree, the rainfall pattern, the conditional FP tree and the generated frequent pattern can be obtained; according to the conditional FP tree, full permutation combination is carried out to obtain the mined frequent pattern and the corresponding support.

[0053] In the work of early warning, the release of disaster risk level according to certain early warning rules has been widely used. At present, most landslides still use the fixed trigger conditions provided by the geological team as the only judgment rule. Such indicators are more professional. The specific implementation scheme is roughly as follows: engineering geology personnel go to the scene for geological exploration, get the lithological structure of the landslide body, and then set the threshold of the trigger condition through work experience, such as rainfall, displacement velocity, displacement acceleration and other indicators. When the trigger index reaches a certain level, the corresponding early warning level (red, orange, yellow, blue) is issued.

[0054] The existing early warning method has high investment cost, long update cycle and poor timeliness, which is difficult to meet the prediction needs of a large number of landslide sites. At the same time, the early warning rules generated by this traditional method are difficult to combine multiple observation indicators for judgment, and the early warning results have a certain one-sidedness. For the universal prediction and early warning work of a large number of landslide sites, due to the limitation of investment cost and the demand of work goal, each slope body does not need to be extremely accurately explored and evaluated, but needs to be quickly predicted in displacement and early warned in landslide danger under the premise of ensuring a certain accuracy.

[0055] Therefore, the present application proposes a landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining, as shown in Figure 2 The method identifies the rapid deformation point of the landslide based on historical monitoring data, determines the key attention time period of the landslide displacement factor, and generates the early warning rule after establishing the data set. The specific steps are as follows:

[0056] Step 1: Establish a historical monitoring database

[0057] Obtain historical monitoring data of a certain historical period, which includes landslide displacement and rainfall. Form a number of sample data in units of days from the historical monitoring data, and each sample data includes the date and displacement and rainfall of the day.

[0058] In the landslide monitoring work, various data (displacement, rainfall) are generated by monitoring instruments. Instrument instability, failure and line failure are inevitable, which leads to data loss, burr and noise problems; at the same time, instrument maintenance will also bring about a jump change in monitoring data value, leading to abnormal values, data loss and other situations in monitoring data. In order to ensure the accuracy of the generated results, the present application first preprocesses the data after obtaining the displacement and rainfall.

[0059] The data preprocessing of the present disclosure mainly includes abnormality identification, burr smoothing and the like. The abnormality identification is mainly to remove the monitoring data jump changes caused by instrument instability, equipment failure, line failure, instrument maintenance and the like. The various monitoring instruments on the landslide body have their measurable range. For the collected data exceeding the range or outputting special codes (such as negative numbers of rain gauges and the like) of the instrument, the instrument instructions can be used to remove and clean up.

[0060] After removing the obvious instrument abnormal values, the abnormal values in the normal observation range need to be identified. First, the data is segmented according to the different displacement rates of the landslide displacement data to obtain the stable stage and the rapid deformation stage of the landslide, and then the abnormal points are identified in each stage to remove the monitoring data significantly greater or smaller than the average value of the stage.

[0061] Due to the objective observation conditions, there are a large number of random fluctuations in the historical monitoring data, resulting in burr phenomenon. In the actual machine learning model construction process, the model may produce overfitting phenomenon in order to fit these local unstable data. Therefore, in order to avoid the problem of inaccurate generation results caused by overfitting phenomenon, the present disclosure uses the moving average method to smooth the data for burr phenomenon. The moving average method averages the total 2n+1 observation values before and after the reference reference time (middle time, or the time corresponding to the maximum data, or the time corresponding to the minimum data) selected in the selected historical data period to obtain the filtering result of the current time; wherein n represents the number of time before and after, the current time plus the previous n time and the next n time, and the value of n is set according to the situation, such as 5, 8, 12 or greater than or equal to 15. When the real data in the sliding window changes little, a large part of the noise can be suppressed, and the filtering result is close to the true value; when the real value in the sliding window changes greatly, this filtering method will lose part of the accuracy, and the filtering result is close to the average expectation of the true value.

[0062] The historical monitoring database in the present disclosure is composed of a plurality of records. Each record is in time unit of day, and the displacement amount, rainfall and the like are counted to generate day-by-day data to form sample data. The data of each day is taken as a group of sample data, as shown in Table 1.

[0063]

[0064] Step two: landslide displacement data rapid deformation point identification

[0065] In the landslide monitoring work, there are often time periods with slow displacement curve changes in the landslide displacement curve, such as Figure 3It is often because there is no external triggering factor such as heavy rainfall in these time periods, and the landslide body is in a relatively stable state, and only moves slowly under the action of gravity. The data of such slow movement is not meaningful for landslide prediction and warning, but may cause fluctuations in the monitoring data due to the influence of the monitoring accuracy of the instrument itself, thereby interfering with the final results.

[0066] In contrast to the slow change of landslide displacement, the probability of landslide occurrence is greatly increased when the landslide displacement changes significantly, which is the period that needs to be focused on in the early warning work. Therefore, the present method extracts the data of significant displacement, constructs the corresponding data set, and focuses the statistical calculation on the significant displacement data.

[0067] In order to obtain significant landslide displacement data, that is, to extract the rapid deformation points of the curve from the landslide displacement curve, the first derivative and the second derivative of the displacement amount are used as the standard for judging the significant change of displacement in the present method. At the same time, the present method adopts two indexes to represent the landslide displacement speed and displacement acceleration, and the specific method is to take the derivative of the displacement amount of each group of sample data to obtain the first derivative representing the landslide displacement speed and the second derivative representing the landslide displacement acceleration of the displacement amount with respect to the date of the day; The date corresponding to the first derivative result and the second derivative result greater than the preset threshold is taken as the deformation point.

[0068] Step three: rapid deformation database establishment

[0069] The displacement amount and rainfall corresponding to the deformation point determined in step two are extracted from the historical monitoring database, and the first derivative result and the second derivative result corresponding thereto are established as a rapid deformation database with days as the unit.

[0070] Step four: determining the key attention time period of the landslide displacement influencing factor

[0071] The important influencing factor in the present disclosure is the factor that has a greater impact on the landslide displacement, such as a larger displacement amount in a certain period, the rainfall in this period is the factor that affects the landslide displacement, and this period is the key attention time period of the important factor.

[0072] This step takes the rapid deformation database as the basis, inputs the data in the rapid deformation database into the landslide displacement prediction model, and outputs the attention intensity distribution of displacement and rainfall with time distance, and determines the key attention time period of the landslide displacement influencing factor according to the distribution.

[0073] In the present disclosure, the landslide displacement prediction model comprises:

[0074] A time series convolution network model is used to perform convolution operation on the input quantity to obtain the feature vector of each input quantity, and the features at the corresponding time in the displacement and rainfall at different time sequences are spliced.

[0075] a full connection layer, used for mixing features after the output of the time convolution network module and adding position encoding, so that the output of the time convolution network model meets the input format requirements of the Transformer model data; and

[0076] a Transformer model decoder, used to obtain important influence factors affecting landslide displacement.

[0077] Transformer is a deep learning model based on self-attention mechanism proposed by Google research team in recent years to improve training speed. The model has achieved good results in sequence generation related tasks and has shown very excellent feature expression ability. The model no longer relies on traditional CNN (Convolutional Neural Network) and RNN, but relies on self-attention mechanism to calculate input and output representation, has stronger expression ability for long-term dependence, and can cope with the long-distance dependence problem caused by high time resolution in landslide short-term displacement prediction. At the same time, the attention mechanism also has good interpretability, which can help researchers understand the reasons for the model to make a specific output to a certain extent, and the ability of the model to capture key information.

[0078] The model architecture of Transformer is shown in Figure 4 The model is composed of an encoder (left half) and a decoder (right half), mainly based on self-attention mechanism to extract internal features, can learn long-term dependencies, and is more easily parallelized than RNN-based models.

[0079] Transformer was applied to machine translation in natural language processing when it was proposed. Through the embedding layer, the high-dimensional features of the single time sample are obtained, and the position encoding (Positional Encoding) is used to retain the position information in the original sequence. Then, the self-attention layer maps the original features to Q, K and V based on the query-key-value idea, uses Q and K to calculate the correlation (Attention score) between different time features and the current time feature, and obtains the next time feature by weighted sum of V. The specific calculation method is as follows:

[0080]

[0081] In the formula, is a scaling factor, denotes the feature dimension of Q and K. According to the above formula, the weight calculation formula for a historical sample at a certain time at the current time is:

[0082] , k=0,1,2,....,t

[0083] In addition, the multi-head attention mechanism is introduced in the Transformer to further improve the model effect. The multi-head attention mechanism projects Q, K and V into multiple subspaces through linear transformation and calculates the attention features respectively. These attention features are spliced together and then linearly transformed to obtain the final attention feature output. The specific calculation method is as follows:

[0084]

[0085]

[0086] In the formula, W is a matrix for linear transformation, and different subscripts correspond to different transformation objects. Multi-head attention allows the model to simultaneously focus on information from different positions and different feature subspaces, and the extraction granularity of information is finer, while the attention degree of the entire feature space in a single attention head is consistent.

[0087] In the Transformer model, the encoder is used to encode the original input sequence into high-dimensional features, and the decoder is responsible for the result output. When the decoder outputs, it needs to consider both the output samples at the historical time and the features provided by the encoder. For the former, the attention mechanism cannot access the sample information at the future time, and this problem can be solved by using mask. Assuming that the sequence length is N, mask is a 0-1 matrix of N x N, mask ij = 1 indicates that the i-th time can rely on the j-th time feature when calculating attention, and mask ij = 0 is the opposite. Using a lower triangular matrix as mask can avoid introducing sample information at the future time; for the latter, by matching Q in the decoder sequence with K and V in the encoder sequence, the features extracted by the encoder can be introduced into the decoder. In the pure sequence generation task, since there is no original sequence for encoding, only the decoder can be used to complete the time series autoregressive sequence generation work.

[0088] Considering that the samples at a single moment cannot effectively describe the motion state of the landslide at the current moment, a landslide short-term displacement prediction model based on Transformer is constructed by combining a temporal convolutional network (TCN) with a Transformer decoder. The TCN mainly consists of causal convolution and dilated convolution, and the TCN introduces a residual connection to improve the model effect. The causal convolution can ensure that the features at the current moment are only related to the samples at the historical moments, and is suitable for the time series autoregressive scenario; the dilated convolution can increase the receptive field of the convolution by adding holes in the standard convolution kernel, while keeping the model parameters unchanged.

[0089] The method combines the TCN model to replace the embedding layer in the Transformer model, which can introduce controllable length of historical information into the features at the current moment while keeping the input form unchanged, and retains the parallelism of the Transformer model. The long-term memory part in the model is completed by the self-attention mechanism of the Transformer model, and the TCN does not need long-time historical information and can not introduce dilated convolution.

[0090] Therefore, the landslide displacement prediction model used in the method is as shown in Figure 5As shown, several time sequence convolution network models TCN, each input quantity (landslide displacement, rainfall) is input into the corresponding TCN model, and the landslide displacement, rainfall and other sequence data are respectively processed through the TCN, and the features at the corresponding time (through the corresponding time in the different time sequence data (displacement, rainfall)) are spliced (the features are the feature vectors obtained by the original data through the convolution operation of the TCN model, which need to be spliced according to time); The time sequence convolution network model TCN outputs access the right connection layer Linear, and the feature mixing is added through the full connection layer to generate data conforming to the input format of the transformer, and finally access the self-attention layer TransformerCasual self-attention layers for processing to obtain the hidden layer vector ht of the self-attention layer, and finally output the landslide displacement at the next time through the full connection layer. That is, the landslide displacement, rainfall and other sequence data are respectively processed through the convolution operation of the TCN model, and the features at the corresponding time are spliced, in order to meet the input format requirements of the transformer model data, the feature mixing is performed through the full connection layer, and then the position coding is added and input into the self-attention layer of the Transformer, which is the data processing process of the transformer model. After operation, the hidden layer vector (ht) of the attention at the current time is obtained, that is, the result obtained after the original data is calculated, and then the landslide displacement at the next time is obtained through the full connection layer again.

[0091] The transformer model has the ability of output self-attention, in the self-attention mechanism, the weight of different time factors in the current time attention constitutes, which reflects the attention distribution of the model at the current time to the historical information, which can assist the researchers to judge the credibility of the model and the importance of the monitoring data at different times. The present disclosure illustrates this with an actual modeling attention distribution.

[0092] This time, the displacement data of a landslide in a certain place and the corresponding rainfall are input into the landslide displacement prediction model of the present disclosure as input quantities, and the attention intensity is normalized, the sum of the four attentions is 1, and the final attention situation is as shown in Figure 6 As shown. Figure 6 It is found that the attention of head1~3 is concentrated in the peak value of daily displacement and the concentrated rainfall period, and the attention of head4 is scattered in the pre-accelerated deformation period. This shows that the landslide deformation is obviously affected by rainfall, and the accelerated deformation often occurs after large-scale rainfall, and the attention to the rainfall event conforms to the landslide deformation law. Each attention is a group of data, including the landslide displacement and rainfall in a certain period.

[0093] The disclosure further shows the changes of four attention over time, as shown in Figure 7 It can be found that the attention distribution of head 1~3 is mainly concentrated on the recent information and decays with the increase of time, and there is a significant decrease on the 11th day, and basically no attention is paid to the information after the 15th day. Head 4 focuses on long-distance information, and the attention is mainly distributed in 17~20 days. Overall, from the attention distribution, the model can effectively combine long-term information and short-term information, and focus on key events such as displacement peak and concentrated rainfall. Important influence shadow can be obtained, and for this landslide point, the historical data of the previous 11 days and the rainfall data of the previous 17~20 days should be considered.

[0094] The number of monitoring points of the universal landslide monitoring is large, so the landslide site can be classified according to geological data and climate information and other related data, and enough samples are selected in each category to establish the transformer model, and then the time range that the landslide influence factor of each major category should focus on is obtained.

[0095] The landslide displacement prediction model of the disclosure is based on the data after additional position coding, and its format is consistent with the data format requirement in the transformer. The self-attention layer and the fully connected layer are the inherent structures of the transformer model, and the next time landslide displacement prediction value can be obtained through operation. After obtaining the prediction value, the model can be continuously trained by comparing it with the true value of landslide displacement, and the optimal parameters can be obtained.

[0096] The important influence factor is obtained by the attention mechanism of the transformer, that is, the attention mechanism can output which factors help the model to make a judgment. For example, it can be found that the rainfall data in the time period of the previous 15 days and 17~20 days is more important for displacement prediction, so the important influence factor is the rainfall of the previous 1 day, the rainfall of the previous 2 days, the rainfall of the previous 3 days, and thus the factor that should be focused on is obtained, that is, the key attention time period of the influence factor of landslide displacement in the production method.

[0097] Step five: establishing a data set

[0098] After obtaining the key attention time period of the landslide influence factor, the data set is established based on this. Each item of the data set should include the relative displacement distance of the landslide on the day (deformation point), the monitoring value of the influence factor in the key attention time period of each day, and if it is an accumulatively calculable influence factor such as rainfall, the accumulative value from the previous day should also be considered.

[0099] The present disclosure only considers rainfall and focuses on the example that the time period is only the previous 3 days of the deformation point. Each data record in the data set records the landslide relative displacement on the day (deformation point), the rainfall on the day of the previous day, the rainfall on the day and the cumulative rainfall of the previous 2 days, and the rainfall on the day and the cumulative rainfall of the previous 3 days. Only one deformation point is described here, and when there are multiple deformation points, there are multiple item sets, each of which contains the aforementioned items.

[0100] After establishing the standard data set, it is necessary to mine the association rules between the landslide relative displacement and each influencing factor. Traditional association rule mining cannot be directly applied to continuous observation values, so it is necessary to classify the observation values according to uniform rules to meet the needs of association rule mining. For example, rainfall of 1ml on the same day is defined as low rainfall, and rainfall of 50ml on the same day is defined as high rainfall.

[0101] Step six: generating early warning rules

[0102] The early warning rules are generated by one of the following methods in this step:

[0103] The frequent item sets in the established data set and the support of the frequent item sets are determined, and the risk level thresholds are set according to the support of the frequent item sets to form early warning rules.

[0104] 2) The method uses the FP-growth frequent item set algorithm to mine association rules, which adopts the following divide-and-conquer strategy: compress the database providing frequent item sets into a frequent pattern tree (FP-tree), but still retain the association information of the item set. The FP-Tree sorts each transaction data item in the transaction data table according to the support, and inserts the data items in each transaction into a tree with NULL as the root node in descending order, while recording the support of the node at each node. The FP-Growth algorithm is based on the above structure, and the construction and projection process of the FP-tree is iterated to speed up the entire mining process.

[0105] After the frequent item set mining of the established data set is completed using the FP-growth frequent item set algorithm, the support of each frequent item set is sorted according to the support of each frequent item set, and then the representative rainfall pattern that triggers rapid displacement of landslides is obtained.

[0106] According to the support of the frequent item set obtained from the rainfall pattern, the risk level thresholds are set to form early warning rules.

[0107] The generation method sets the risk level threshold according to the support degree, considers the support degree of important factors, and the generated early warning rule is more adaptable. The risk level threshold is set according to the actual situation, such as displacement and rainfall being 25%, 25%-50%, 50%-75%, and 75%-100% of the frequent item set, respectively, to issue red, orange, yellow, and blue early warning of landslides.

[0108] The method uses historical data (data of a continuous period before the prediction period) to obtain a risk early warning rule library. Compared with the early warning rules obtained by the traditional geological team based on their own experience and field investigation, the risk early warning rule generation is efficient, low in cost, and based on the performance of historical data, the generated risk early warning rule has stronger applicability. By regularly updating the historical data used to train the risk early warning rule, the timeliness of the risk early warning rule library is improved.

[0109] The method automatically generates early warning rules based on historical data, which is efficient; can be applied to a large number of landslide monitoring points, and can be continuously optimized in actual operation, meeting the needs of universal landslide monitoring.

[0110] After determining the early warning period, the generation method according to the present disclosure can generate effective early warning rules, which are more adaptable; and according to the early warning situation, the latest historical data can be used to generate the latest early warning rules.

[0111] The method is used for rainfall-induced landslides, and generates early warning rules based on rainfall-induced landslide data, which are applicable to rainfall-induced landslide early warning.

[0112] The generation method mines association rules from monitoring data, which is low in cost, high in real-time performance, and can be applied to a large number of monitoring sites; the data analysis process comprehensively considers many influencing factors, avoiding the one-sidedness of experience methods; the method can perform relatively fast displacement prediction and landslide danger early warning under the premise of ensuring a certain accuracy.

[0113] Those skilled in the art should easily realize that the exemplary units and method steps described in combination with the embodiments disclosed in the present application can be realized in the form of hardware or a combination of hardware and software programs. Whether a certain function is realized in hardware or software program driven hardware depends on the actual situation.

[0114] The present disclosure has been described by the above related embodiments, however, the above embodiments are only examples of implementing the present disclosure. It must be pointed out that the disclosed embodiments do not limit the scope of the present disclosure. On the contrary, variations and modifications made without departing from the spirit and scope of the present disclosure are within the scope of the patent protection of the present disclosure.

Claims

1. A method for generating landslide early warning rules based on anomaly deformation mining from multi-source monitoring data, characterized in that, The method includes the following steps: Step 1: Establish a historical monitoring database and obtain historical monitoring data for a certain historical period. This historical monitoring data includes landslide displacement and rainfall. The historical monitoring data is divided into several sets of sample data on a daily basis. Each set of sample data includes the date, displacement and rainfall of that day. Step 2: Obtain deformation points. Calculate the derivative of the displacement for each set of sample data to obtain the first derivative of the daily displacement relative to the date, representing the landslide displacement velocity, and the second derivative, representing the landslide displacement acceleration. Dates where the first and second derivative results are greater than a preset threshold are used as deformation points. Step 3: Establish a rapid deformation database. Extract the displacement and rainfall corresponding to the deformation points determined in Step 2 from the historical monitoring database, and establish a rapid deformation database with the corresponding first derivative and second derivative results on a daily basis. Step 4: Determine the key time period for influencing landslide displacement factors. Input the data from the rapid deformation database into the landslide displacement prediction model and output the distribution of displacement and rainfall intensity over time and distance. Based on the distribution, determine the key time period for influencing landslide displacement factors. Step 5: Create a dataset, which includes several itemsets. Each itemset includes the relative displacement distance of the landslide at the deformation point and the daily rainfall during the key monitoring period. Step Six: Generate Early Warning Rules. Use one of the following methods to generate early warning rules: 1) Determine the frequent itemsets and their support in the constructed dataset, and set the threshold for each risk level based on the support of the frequent itemsets to form early warning rules; 2) The FP-growth frequent itemset algorithm is used to obtain the landslide rapid displacement rainfall pattern. The threshold of each risk level is set according to the support of the frequent itemsets in the landslide rapid displacement rainfall pattern to form an early warning rule.

2. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 1, characterized in that, The landslide displacement prediction model includes: Temporal convolutional network models are used to perform convolution operations on input quantities to obtain feature vectors of each input quantity, and then concatenate the features at corresponding times in displacement and rainfall at different time sequences. A fully connected layer is used to perform feature mixing on the output of the temporal convolutional network module and then add positional encoding so that the output of the temporal convolutional network model conforms to the input format requirements of the Transformer model data; and The Transformer model decoder is used to obtain the important influencing factors affecting landslide displacement.

3. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 2, characterized in that, The temporal convolutional network model includes causal convolutions to ensure that the features at the current time are only related to samples from historical time periods.

4. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 1, characterized in that, In step one, when establishing the historical monitoring database, the data is first cleaned, outliers are removed, and spike data is smoothed. Then, the corresponding displacement and rainfall are statistically analyzed on a daily basis to construct the historical monitoring database.

5. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 4, characterized in that, The abnormal values ​​include those caused by sudden changes in monitoring data due to instrument instability, equipment failure, line failure, and instrument maintenance during landslide detection, as well as abnormal values ​​within the normal monitoring range.

6. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 5, characterized in that, The outlier removal process within the normal monitoring range first involves segmenting the landslide displacement data based on the different displacement rates, resulting in a stable landslide stage and a rapid deformation stage. Within each stage, outliers are identified, and monitoring data that are significantly greater than or less than the average value of that stage are removed.

7. The landslide early warning rule generation method based on multi-source monitoring data anomaly deformation mining according to claim 4, characterized in that, The spike data was smoothed using a moving average method.

Images