CN115585767A

CN115585767A - Landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining

Info

Publication number: CN115585767A
Application number: CN202211076359.8A
Authority: CN
Inventors: 殷跃平; 赵文祎; 常啸寅; 黄喆; 李俊峰; 朱赛楠
Original assignee: China Institute Of Geological Environment Monitoring
Current assignee: China Institute Of Geological Environment Monitoring
Priority date: 2022-09-05
Filing date: 2022-09-05
Publication date: 2023-01-10

Abstract

The invention discloses a landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining. The generation method influences the mining association rule in the key attention time period of the landslide displacement factor, is low in cost and strong in real-time performance, and can be widely applied to numerous monitoring sites.

Description

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

Technical Field

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

Background

The landslide refers to a natural phenomenon that a soil body or a rock body on a slope slides downwards along the slope along a certain weak surface or a weak zone under the action of gravity under the influence of natural or artificial factors, and the natural phenomenon comprises collapse and debris flow. Landslide is a type of geological disaster that is extremely hazardous. Landslide disasters are mostly seen in low mountains and hilly areas. In the high-speed period of economic development and urban construction of China, along with the transformation of landforms and landforms by human activities, manual slope cutting is increased, and landslide risk is increased. Monitoring and early warning are increasingly gaining attention as one of important measures for reducing the risk of geological disasters.

In the prediction and early warning work, the method for releasing the disaster risk level according to a certain early warning rule is widely applied. At present, most landslides still adopt fixed trigger conditions provided by geological teams as unique judgment rules, and the indexes have strong specialties. The specific implementation scheme is that engineering geologists approximately go to the site to perform geological exploration to obtain the lithologic structure of the landslide body, and then through work experience, threshold values of trigger conditions are set, such as indexes of rainfall, displacement speed, displacement acceleration and the like. When the trigger index reaches a certain level, a corresponding early warning grade (red orange yellow blue) is sent out.

The existing early warning method is high in input cost, long in updating period and poor in timeliness, prediction requirements of a large number of landslide sites are difficult to meet, meanwhile, the early warning rule generated by the traditional mode is difficult to well combine with a plurality of observation indexes to judge, and early warning results have certain one-sidedness. Aiming at the general prediction and early warning work of a large number of landslide sites, due to the limitation of input cost and the requirement of a working target, each slope body does not need to carry out extremely accurate exploration and evaluation, but needs to carry out relatively quick displacement prediction and landslide danger early warning on the premise of ensuring certain accuracy. It is on this premise that the present application is presented.

Disclosure of Invention

The invention aims to provide a landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining, which is low in cost, capable of quickly obtaining early warning rules and strong in instantaneity.

In order to realize the purpose of the invention, the landslide early warning rule generation method based on the abnormal deformation mining of the multi-source monitoring data comprises the following steps:

the method comprises the following steps: establishing a historical monitoring database, acquiring historical monitoring data of a certain historical period, wherein the historical monitoring data comprises landslide displacement and rainfall, forming the historical monitoring data into a plurality of groups of sample data by taking a day as a unit, and each group of sample data comprises a date of the day, displacement of the day and rainfall;

step two: acquiring deformation points, and deriving the displacement of each group of sample data to obtain a first derivative representing the landslide displacement speed and a second derivative representing the landslide displacement acceleration of the displacement of the current day relative to the current day; taking dates corresponding to the first derivative result and the second derivative result which are larger than a preset threshold value as deformation points;

step three: establishing a rapid deformation database, extracting displacement and rainfall corresponding to the deformation point determined in the second step from a historical monitoring database, and establishing the rapid deformation database by taking a day as a unit with a first derivative result and a second derivative result corresponding to the displacement and the rainfall;

step four: determining a key attention time period influencing the landslide displacement factor, inputting data in the rapid deformation database into a landslide displacement prediction model, outputting the distribution condition of displacement and rainfall attention intensity along with the time distance, and determining the key attention time period influencing the landslide displacement factor according to the distribution condition;

step five: establishing a data set, wherein the data set comprises a plurality of item sets, and each item set comprises the landslide relative displacement distance of a deformation point and the daily rainfall in a key attention time period;

step six: generating an early warning rule by adopting one of the following modes:

1) Determining a frequent item set and the support degree of the frequent item set in the established data set, and setting each risk level threshold value according to the support degree of the frequent item set to form an early warning rule;

2) And obtaining a landslide rapid displacement rainfall mode by utilizing an FP-growth frequent item set algorithm, and setting each risk level threshold according to the support degree of the frequent item set in the landslide rapid displacement rainfall mode to form an early warning rule.

In some embodiments, the landslide displacement prediction model comprises:

the time sequence convolution network model is used for carrying out convolution operation on the input quantity to obtain respective characteristic vectors of the input quantity and splicing the characteristics of displacement and rainfall passing through different time sequences at corresponding moments;

the full connection layer is used for performing feature mixing on the output of the time sequence convolution network module and then adding position coding to ensure that the output of the time sequence convolution network module meets the input format requirement of the Transformer model data; and

and a Transformer model decoder for obtaining an important influence factor influencing the landslide displacement.

In some embodiments, the time-series convolutional network model includes a causal convolution to ensure that features at a current time are only relevant to samples at historical times.

In some embodiments, when the historical monitoring database is established in the first step, the data is cleaned, abnormal values are removed, the burr data are smoothed, and then the corresponding displacement and rainfall are counted by taking a day as a unit to establish the historical monitoring database; the accuracy of the generated result is guaranteed.

It is another object of the present invention herein to provide a landslide displacement prediction model comprising:

the full connection layer is used for adding position codes after the characteristics of the output of the time sequence convolutional network module are mixed, so that the output of the time sequence convolutional network model meets the input format requirement of the Transformer model data; and

In some embodiments, the time series convolution network model includes a causal convolution to ensure that features at a current time instance are only relevant to samples at historical time instances.

By adopting the technical scheme of the invention, the beneficial effects at least comprise:

1) The method influences the mining association rule in the key focus time period of the landslide displacement factor, is low in cost and strong in real-time performance, and can be widely applied to a large number of monitoring sites.

2) The generating method takes the deformation point and the data of a period of time before the deformation point as the basis for finding the important influence factors influencing the landslide displacement, considers a plurality of influence factors, and effectively avoids the one-sidedness of the empirical method.

3) The method can quickly obtain the early warning rule on the premise of ensuring certain accuracy.

4) According to the landslide short-term displacement prediction model provided by the invention, the causal convolution in the model can ensure that the characteristics of the current moment are only related to the samples of the historical moment, and the model is suitable for a time sequence autoregressive scene; and the time sequence convolution network module introduces residual error linkage, thereby reducing the time consumption of data processing and improving the model effect.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention. It is obvious that the drawings in the following description are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort. In the drawings:

FIG. 1 is a schematic diagram of a landslide displacement curve according to the present invention;

FIG. 2 is a flowchart of a landslide warning rule generation method provided by the present invention;

FIG. 3 is a graph of landslide displacement;

FIG. 4 is a diagram of a model architecture of a transform model according to the present invention;

FIG. 5 is a diagram of a landslide short term displacement prediction model according to the present invention;

FIG. 6 is a schematic illustration of the attention distribution according to the present invention;

FIG. 7 is a graph of the time-distance distribution of attention described in the present invention.

Detailed Description

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

For ease of understanding, the following provides an explanation of concepts related to the present application.

1. Historical monitoring data

Monitoring data in 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, monitoring data from the beginning of the monitoring day to the day before the early warning period and the like, because the data of weather, rainfall, displacement and the like of adjacent months are similar, the data of the adjacent months are used as historical data, and the result is more accurate. The monitoring data comprises displacement and rainfall of landslide of the monitoring point, statistics of the monitoring data is carried out by taking the day as a unit, and data every day is used as a group of sample data.

The early warning period refers to a period of time in which the monitoring point needs early warning monitoring.

2. Item set, single item set, multiple item set and frequent item set

The collection of items is called a set of items. A term set comprising N terms may be referred to as a multi-term set. When N =1, the set of items may be referred to as a single set of items. Frequent items are a large set of items that occur frequently in a database. In the application, the items in the item set are the landslide relative displacement distance of the deformation point and the daily rainfall in the important attention time period.

3. Degree of support

Support refers to the proportion of a term or set of terms in the overall data set. The support degree can be expressed by the following formula:

Y＝X/T

in the formula: y represents the support of an item or set of items, X represents the number of times an item or set of items (A → X) appears in the dataset, and T represents the total number of items or sets of items in the dataset.

4. Rules of early warning

The early warning rule matrix is a rule corresponding to the output risk level under the monitoring data, and is composed of a plurality of early warning rules, as shown in fig. 1. And when the monitoring data is the corresponding condition of the rule early warning, generating a corresponding risk grade. The monitoring data comprises rainfall and displacement.

5. FPgrowth algorithm

The basic idea of the algorithm is that a data set is scanned, the support degree of each item and item set is calculated, and the items and the item sets are sorted from large to small to obtain a support table; constructing FP trees one by one from the minimum support degree at the bottom; the FP tree is associated with the support tables, each item in the support tables is provided with a pointer for storing and pointing to a corresponding node in the FP tree, the FP tree constructed through a few full-table scans changes irregular data of a shopping cart into a tree structure with traces and capable of following, and huge operation of natural connection is omitted; finally, digging out association rules through the FP tree, and conditioning the FP tree and generating frequent patterns according to the rainfall pattern through the FP tree; and performing full permutation and combination according to the conditional FP tree to obtain the mined frequent patterns and the corresponding support degree.

In the prediction and early warning work, the method for releasing the disaster risk level according to a certain early warning rule is widely applied. At present, most landslides still adopt fixed trigger conditions provided by geological teams as unique judgment rules, and the indexes have strong speciality. The specific implementation scheme is that engineering geologists approximately go to the site to perform geological exploration to obtain the lithologic structure of the landslide body, and then through work experience, threshold values of trigger conditions are set, such as indexes of rainfall, displacement speed, displacement acceleration and the like. And when the trigger index reaches a certain level, sending out a corresponding early warning grade (red orange yellow blue).

The existing early warning method is high in input cost, long in updating period and poor in timeliness, prediction requirements of a large number of landslide sites are difficult to meet, meanwhile, an early warning rule generated by the traditional mode is difficult to well combine with a plurality of observation indexes for judgment, and an early warning result has certain one-sidedness. Aiming at the general prediction and early warning work of a large number of landslide sites, due to the limitation of input cost and the requirement of a working target, each slope body does not need to carry out extremely accurate exploration and evaluation, but needs to carry out relatively quick displacement prediction and landslide danger early warning on the premise of ensuring certain accuracy.

Therefore, the invention provides a landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining, as shown in fig. 2, the method identifies a landslide rapid deformation point based on historical monitoring data, determines a key attention time period influencing a landslide displacement factor, and generates an early warning rule after establishing a data set. The method comprises the following specific steps:

the method comprises the following steps: establishing a historical monitoring database

The method comprises the steps of obtaining historical monitoring data of a certain historical period, wherein the historical monitoring data comprise landslide displacement and rainfall, forming a plurality of groups of sample data of the historical monitoring data by taking a day as a unit, and each group of sample data comprises a current day date, a current day displacement and rainfall.

In the landslide monitoring work, various data (displacement and rainfall) are monitored and generated by a monitoring instrument, and the problems of data loss, burrs, noise and the like caused by the fact that the instrument is unstable, faults, line faults and the like are inevitable; meanwhile, the instrument overhaul also brings the monitoring data value to change in a jumping mode, so that the monitoring data has abnormal values, data loss and the like, and in order to ensure the accuracy of the generated result, the data are preprocessed after the displacement and the rainfall are acquired.

The data preprocessing of the disclosure mainly comprises processing such as exception identification and burr smoothing, wherein the exception identification is mainly used for removing monitoring data jump change caused by instability of an instrument, equipment failure, line failure, instrument maintenance and the like. Various monitoring instruments on the landslide body have measurable ranges, and special codes (such as the negative number of a rain gauge and the like) pointing to the abnormity of the instruments in the collected data and exceeding the ranges or being output can be removed and cleaned according to the description of the instruments.

After removing the obvious instrument abnormal value, the abnormal value in the normal observation range needs to be identified. The method comprises the steps of firstly segmenting data according to different displacement rates of landslide displacement data to obtain a landslide stabilization stage and a rapid deformation stage, then identifying abnormal points in each stage, and removing monitoring data which are obviously larger than or smaller than an average value of the stage.

Due to objective observation conditions, a large amount of random fluctuation exists in historical monitoring data, and a burr phenomenon is generated. Therefore, in order to avoid the problem of inaccurate generated results due to the over-fitting phenomenon, the present disclosure uses a moving average method to smooth data for the glitch phenomenon. Averaging 2n +1 observed values of the time before and after the reference datum time (middle time, or time corresponding to the maximum data or time corresponding to the minimum data) selected in the selected historical data period by a sliding average method to obtain a filtering result of the current time; wherein n represents the number of previous and subsequent times, the current time is added with the previous n times and the next n times, and the value of n is set according to the situation, such as 5, 8, 12 or more than or equal to 15. When the real data in the sliding window is not changed much, a large part of noise can be suppressed, and the filtering result is approximate to a real value; when the variation of the real value in the sliding window is large, a part of the accuracy is lost in the filtering mode, and the filtering result is close to the average expectation of the real value.

The history monitoring database in the disclosure is composed of a plurality of records, each record is a time unit, data such as displacement, rainfall and the like are counted, data by day are generated to form sample data, and data by day are used as a group of sample data, such as the sample indication shown in table 1.

TABLE 1 data sample schematic table

Step two: landslide displacement data rapid deformation point identification

In the landslide monitoring work, a time period in which the change of the displacement curve is relatively slow often exists in the landslide displacement curve, as shown in fig. 3; this is because there is not a large amount of external inducing factors such as rainfall in these periods, and the sliding body is in a stable state and moves slowly only by gravity. The slow moving data is not significant for landslide prediction and early warning, but may cause monitoring data fluctuation due to the influence of the monitoring precision of the instrument, and further cause certain interference on the final result.

Compared with the slow change of the landslide displacement, the landslide displacement is obviously changed, the probability of landslide is greatly increased, and the landslide displacement is a period needing important attention during early warning work. Therefore, the method extracts the data of the significant displacement, constructs a corresponding data set, and focuses the emphasis of statistical calculation on the data of the significant displacement.

In order to obtain remarkable landslide displacement data, namely, a curve rapid deformation point is extracted from a landslide displacement curve, for the purpose, a first derivative and a second derivative of displacement are used as standards for judging that the displacement has remarkable change, meanwhile, two indexes adopted by the method respectively represent landslide displacement speed and displacement acceleration, and the specific method is to carry out derivation on the displacement of each group of sample data to obtain a first derivative representing the landslide displacement speed and a second derivative representing the landslide displacement acceleration relative to the current date; and taking the date corresponding to the first derivative result and the second derivative result which are larger than the preset threshold value as deformation points.

Step three: fast deformation database building

And (4) extracting the displacement and rainfall corresponding to the deformation point determined in the second step from a historical monitoring database, and establishing a rapid deformation database by taking the day as a unit with a first derivative result and a second derivative result corresponding to the displacement and rainfall.

Step four: determining a time period of significant concern affecting a landslide displacement factor

The important influence factor in the present disclosure is a factor that has a large influence on the landslide displacement, for example, if the displacement amount in a certain period is large, the rainfall amount in the period is a factor that influences the landslide displacement, and the period is a time period of important attention of the important factor.

The method comprises the steps of inputting data in a rapid deformation database into a landslide displacement prediction model according to the rapid deformation database, outputting distribution conditions of displacement and rainfall attention intensity along with time and distance, and determining a key attention time period influencing a landslide displacement factor according to the distribution conditions.

In the present disclosure, the landslide displacement prediction model includes:

the full connection layer is used for carrying out feature mixing on the output of the time sequence convolution network module and then adding position coding to ensure that the output of the time sequence convolution network module meets the input format requirement of the Transformer model data; and

The Transformer is a deep learning model based on Self-Attention mechanism (Self-Attention) to improve training speed proposed by google research team in recent years. The model has good effect in sequence generation related tasks and shows very excellent characteristic expression capability. The model does not depend on the traditional CNN (Convolutional Neural Network) and RNN, calculates input and output representations by means of an attention-free mechanism, has stronger expression capability on long-term dependence, and can solve the problem of long-distance dependence caused by high time resolution in landslide short-term displacement prediction. Meanwhile, the attention mechanism also has good interpretability, and can help researchers understand the reason why the model makes specific output to a certain extent and the capability of the model for capturing important information.

The transform model is constructed as shown in fig. 4, and is composed of an encoder (left half) and a decoder (right half), and is mainly based on a self-attention mechanism to extract intrinsic features, so that long-term dependency can be learned, and parallelization is easier than that of the RNN-type model.

The Transformer is applied to a machine translation problem of natural language processing when being proposed, obtains high-dimensional characteristics of a single moment sample through an embedding layer, and uses position Encoding (Positional Encoding) to retain position information in an original sequence. Then, the self-Attention layer maps the original features into three features of Q, K and V based on the query-key-value idea, calculates the correlation (Attention score) between the features of different times and the features of the current time by using Q and K, and obtains the features of the next time by weighted summation of V, wherein the specific calculation mode is as follows:

in the formula (I), the compound is shown in the specification,

as scaling factor, d _k Representing the characteristic dimensions of Q and K. According to the market, under the current time, the weight calculation formula for the historical sample at a certain time is as follows:

weight _t-k ＝sof*max(Q _t XK ^T ) _t-k ,k＝0,1,2,....,t

in addition, a Multi-head attentiveness mechanism (Multi-head attention) is also introduced into the Transformer to further improve the model effect. The multi-head attention mechanism projects Q, K and V to a plurality of subspaces through linear transformation, attention characteristics are respectively calculated, the attention characteristics are spliced together and then subjected to linear transformation, and final attention characteristic output is obtained, wherein the specific calculation mode is as follows:

MultiHead(Q,K,V)＝Concat(head ₁ ,...,head _h )W ⁰

in the formula, W represents a living subject performing linear conversion, and different superscripts and subscripts correspond to different conversion targets. The multi-head attention allows the model to focus on information of different positions and different feature subspaces at the same time, the extraction granularity of the information is finer, and the attention degree of the whole feature space in a single attention head is consistent.

In the transform model, an encoder is used for encoding an original input sequence into a high-dimensional feature, and a decoder is responsible for outputting a result. When the decoder outputs, the output samples at the historical time and the characteristics provided by the encoder need to be considered at the same time. For the former, the attention mechanism cannot access sample information at a future moment, and the problem can be solved by adopting a mask, wherein the sequence length is N, the mask is an N multiplied by N0-1 matrix, and the mask is _ij =1 denotes a feature that can depend on the jth time when attention is calculated at the ith time, mask _ij If the matrix is not zero =0, the lower triangular matrix is used as the mask, so that the sample information of the future time can be prevented from being introduced; for the latter, the features extracted by the encoder can be introduced into the decoder by matching Q in the decoder sequence with K, V in the encoder sequence. In the pure sequence generation task, because no original sequence is used for encoding, the sequence generation work of time sequence autoregressive can be completed only by using a decoder.

Considering that a sample at a single moment cannot effectively describe the motion state of the landslide at the current moment, the method combines a time series Convolutional Network (TCN) with a Transformer decoder to construct a landslide short-term displacement prediction model based on the Transformer. TCN is mainly composed of Causal convolution (cause convolution) and hole convolution (related convolution), and TCN introduces residual connection to improve model effect. The causal convolution can ensure that the characteristics of the current moment are only related to the samples of the historical moment, and is suitable for a time sequence autoregressive scene; the hole convolution increases the reception field of convolution on the premise of not changing the control model parameters by adding a hole into a standard convolution kernel.

The method combines a TCN model to replace an embedding layer in a Transformer model, can introduce the historical information with controllable length into the characteristics of the current moment on the premise of keeping the input form unchanged, and keeps the parallelism of the Transformer model. The long-term memory part in the model is completed by a Transformer model self-attention mechanism, the TCN does not need long-time historical information, and the void convolution can not be introduced.

In view of this, the landslide displacement prediction model used in the method is as shown in fig. 5, each input quantity (landslide displacement and rainfall) of a plurality of time sequence convolution network models TCN is respectively input into a corresponding TCN model, and after the sequence data such as landslide displacement and rainfall respectively pass through the TCN, the features at corresponding moments (corresponding moments in different time sequence data (displacement and rainfall)) are spliced (the features are feature vectors obtained by convolution operation of original data through the TCN models, and need to be spliced according to time); the TCN outputs an access right connecting layer Linear, feature mixing additional position coding is carried out through a full connecting layer to generate data which accords with the input format of a transducer, a self-attention layer transducer Casual self-orientation layer of the transducer is accessed to process to obtain a hidden layer vector ht of the self-attention layer, and finally the landslide displacement of the next moment is output through the full connecting layer Linear. The method comprises the steps of carrying out convolution operation on sequence data such as landslide displacement and rainfall respectively through a TCN (transmission control network) model to obtain respective feature vectors, splicing features at corresponding moments, carrying out feature mixing through a full connection layer in order to meet the input format requirement of transform model data, then adding position codes, inputting the position codes to a self-attention layer of a transform, wherein the self-attention layer is a data processing flow of the transform model, carrying out operation to obtain a hidden layer vector (ht) at the current moment of attition, namely a result obtained after calculation on original data, and then carrying out full connection layer again to obtain the landslide displacement at the next moment.

The Transformer model has the capability of outputting self-attention, in the self-attention mechanism, the weights of different time factors in the current time attention composition reflect the attention distribution of the model to historical information at the current time, and a researcher can be assisted to judge the reliability of the model and the importance of monitoring data at different times before. The present disclosure is described herein in terms of a practical modeling of attention distribution.

The displacement data of a certain landslide and the corresponding rainfall are input into the landslide displacement prediction model of the present disclosure as input quantities, wherein the attention intensity is normalized, the sum of 4 attentions is 1, and the finally obtained attention situation is shown in fig. 6. As shown in fig. 6, the attention of heads 1 to 3 is concentrated on the peak of the daily displacement and the concentrated rainfall period, and the attention of head4 is distributed discretely before the rapid deformation period. This shows that the landslide deformation is obviously influenced by rainfall, and the accelerated deformation often occurs after large-scale rainfall, and the attention to the rainfall event conforms to the landslide deformation rule. Each attention is a set of data including the amount of landslide displacement and the amount of rainfall over a period of time.

The present disclosure further shows the change of four attentions with time, as shown in fig. 7, it can be seen that the attention distributions of heads 1-3 mainly focus on the recent information, and decay with time, and show a significant decrease at day 11, and basically do not focus on the information after day 15. The head4 focuses on the long-distance information, and the attention is mainly distributed in 17-20 days. Generally speaking, from the view point of attention distribution, the model can effectively combine long-term information and short-term information, and focuses on key events such as displacement peak values and concentrated rainfall. Important influence shadows can be obtained, and for the landslide point, historical data of the previous 11 days and rainfall data of the previous 17-20 days are mainly considered.

The general landslide monitoring aims at a large number of monitoring points, so that landslide sites can be classified according to geological data, climate information and other related data, enough samples are selected in each class to establish a transform model, and the time range in which landslide influence factors of each class need to pay more attention is obtained.

According to the landslide displacement prediction model, according to data subjected to additional position coding, the format of the landslide displacement prediction model is consistent with the requirement of the data format in transfromer, the subsequent self-attention layer and the full-connection layer are both inherent structures of a transfomer model, and the landslide displacement prediction value at the next moment can be obtained through calculation. After the predicted value is obtained, the predicted value can be compared with the landslide displacement truth value, and then the model is continuously trained to obtain the optimal parameter.

The important influence factors are obtained by the attention mechanism of the transformer, that is, the attention mechanism can output factors for helping the model to make a judgment, for example, rainfall data in the time periods of the previous 15 days and 17-20 days is more important for predicting the displacement, so the important influence factors are rainfall in the previous 1 day, rainfall in the previous 2 days and rainfall in the previous 3 days, and therefore the important influence factors are to be focused, namely the important focusing time period of the production method for influencing the landslide displacement factor.

Step five: building a data set

In the step, after the key attention time period of the landslide influence factor is obtained, a data set is established on the basis of the key attention time period. Each item of the data set should include the relative displacement distance of the landslide on that day (deformation point), the monitored values of the daily impact factors in the time period of important interest, and if the impact factors are accumulatively calculable like rainfall, the accumulated values from day to date should also be considered.

The present disclosure is described herein taking only rainfall as an example and the first 3 days when the time period of heavy attention is only the inflection point. Each item of data in the data set records the relative displacement of the landslide on the day (deformed point), the rainfall on the day of the previous 1 day, the rainfall and the accumulated rainfall on the day of the previous 2 days, and the rainfall and the accumulated rainfall on the day of the previous 3 days. Here, only one morphed point is described, and when there are a plurality of morphed points, there are a plurality of item sets each including the aforementioned items.

After the standard data set is established, the association rule between the relative displacement of the landslide and each influence factor needs to be mined. The traditional association rule mining cannot be directly applied to continuous observed values, so that the observed values need to be graded according to a uniform rule, and the observed values meet the requirement of association rule mining. For example, a daily rainfall of 1ml is defined as a low rainfall, and a daily rainfall of 50ml is defined as a high rainfall.

Step six: early warning rule generation

The step adopts one of the following modes to generate the early warning rule:

1) And determining the frequent item sets and the support degrees of the frequent item sets in the established data set, and setting each risk level threshold value according to the support degrees of the frequent item sets to form an early warning rule.

2) The method adopts FP-growth frequent item set algorithm to mine association rules, and adopts the following divide-and-conquer strategy: the database providing the frequent item set is compressed to a frequent pattern tree (FP-tree), but the item set association information is still retained. After the FP-Tree sorts the transaction data items in the transaction data table according to the support degree, the data items in each transaction are sequentially inserted into a Tree taking NULL as a root node according to the descending order, and the support degree of the node is recorded at each node. Based on the structure, the FP-Growth algorithm accelerates the whole excavation process by continuously iterating the construction and projection process of the FP-tree.

The method uses an FP-growth frequent item set algorithm to sort the established data sets according to the support degree of each frequent item set after the mining of the frequent item sets is finished, and further obtains a representative rainfall pattern which causes the fast displacement of the landslide.

And setting each risk level threshold according to the support degree of the frequent item set obtained in the rainfall mode to form an early warning rule.

According to the generation method, each risk level threshold is set according to the support degree, the support degree of important factors is considered, and the generated early warning rule is high in adaptability. And setting the threshold values of all risk levels according to actual conditions, wherein if the displacement and the rainfall are 25%, 25-50%, 50-75% and 75-100% of the frequent item set, respectively issuing a landslide red early warning, an orange early warning, a yellow early warning and a blue early warning.

According to the method, the risk early warning rule base is obtained by utilizing historical data (data of a continuous period before the predicted period), and compared with the early warning rules obtained by traditional geological teams according to self experience and field investigation, the risk early warning rule base is high in generation efficiency and low in cost, and the generated risk early warning rules have stronger applicability based on the performance of the historical data. The timeliness of the risk early warning rule base is improved by updating the historical data used for training the risk early warning rules at regular time.

According to the method, the early warning rule is automatically generated according to the historical data, and the efficiency is high; the method can be applied to a large number of landslide monitoring points in a large scale, can continuously realize self-optimization in actual operation, and meets the requirement of general landslide monitoring.

After the early warning period is determined, the effective early warning rule can be generated according to the generation method disclosed by the invention, and the adaptability is stronger; and according to the early warning condition, the latest early warning rule can be generated by using the latest historical data pair.

The method is used for rainfall type landslide, and early warning rules are generated based on rainfall type landslide data and are suitable for rainfall type landslide early warning.

The generation method excavates the association rule from the monitoring data, has low cost and strong real-time performance, and can be widely applied to numerous monitoring sites; a plurality of influence factors are comprehensively considered in the data analysis process, so that one-sidedness of an empirical method is avoided; the method can carry out relatively quick displacement prediction and landslide risk early warning on the premise of ensuring certain accuracy.

Those of skill in the art will readily appreciate that the various illustrative elements and method steps described in connection with the embodiments disclosed herein may be implemented as hardware or combinations of hardware and software programs. Whether a function is implemented as hardware or software program driven hardware is determined by the circumstances.

The present disclosure has been described in terms of the above-described embodiments, which are merely exemplary of the implementations of the present disclosure. It must be noted that the disclosed embodiments do not limit the scope of the disclosure. Rather, variations and modifications are possible within the spirit and scope of the disclosure, and these are all within the scope of the disclosure.

Claims

1. A landslide early warning rule generation method based on multi-source monitoring data abnormal deformation mining is characterized by comprising the following steps:

the method comprises the following steps: establishing a historical monitoring database, acquiring historical monitoring data of a certain historical period, wherein the historical monitoring data comprises landslide displacement and rainfall, forming the historical monitoring data into a plurality of groups of sample data by taking a day as a unit, and each group of sample data comprises a current date, a current displacement and rainfall;

step two: acquiring deformation points, and deriving the displacement of each group of sample data to obtain a first derivative representing the landslide displacement speed and a second derivative representing the landslide displacement acceleration of the displacement of the current day relative to the current day; taking the date corresponding to the first derivative result and the second derivative result which are larger than the preset threshold value as deformation points;

step three: establishing a rapid deformation database, extracting displacement and rainfall corresponding to the deformation point determined in the second step from a historical monitoring database, and establishing the rapid deformation database by taking the day as a unit with a first derivative result and a second derivative result corresponding to the displacement and the rainfall;

step four: determining a key attention time period influencing the landslide displacement factor, inputting data in the rapid deformation database into a landslide displacement prediction model, outputting the distribution condition of displacement and rainfall attention intensity along with time distance, and determining the key attention time period influencing the landslide displacement factor according to the distribution condition;

2. The method for generating the landslide early warning rule based on the multisource monitoring data abnormal deformation mining according to claim 1, wherein the landslide displacement prediction model comprises:

3. The method for generating the landslide early warning rules based on multisource monitoring data anomalous deformation mining according to claim 2 wherein the time series convolution network model comprises a causal convolution for ensuring that features at a current time are only relevant to samples at a historical time.

4. The landslide early warning rule generation method based on multisource monitoring data abnormal deformation mining according to claim 1, wherein when the historical monitoring database is established in the first step, the data is firstly cleaned, abnormal values are removed, the burr data are smoothed, and then corresponding displacement and rainfall are counted by taking day as a unit to establish the historical monitoring database.

5. The method for generating the landslide warning rule based on the abnormal deformation mining of the multi-source monitoring data according to claim 4, wherein the abnormal values comprise abnormal values formed by monitoring data jumping changes caused by instrument instability, equipment faults, line faults and instrument overhaul in landslide detection work and abnormal values in a normal monitoring range.

6. The landslide early warning rule generation method based on multisource monitoring data abnormal deformation mining according to claim 5, wherein abnormal value elimination in the normal monitoring range firstly segments displacement data according to different displacement rates of landslide displacement data to obtain a landslide stabilization stage and a rapid deformation stage, and abnormal points are identified in each stage to remove monitoring data which are obviously larger than or smaller than an average value of the stage.

7. The landslide early warning rule generation method based on multisource monitoring data abnormal deformation mining according to claim 4, wherein the burr data is subjected to data smoothing by using a moving average method.

8. A landslide displacement prediction model comprising:

9. The landslide short term displacement prediction model of claim 8 wherein said time series convolution network model comprises a causal convolution for ensuring that features at a current time instant relate only to samples at historical time instants.