Unsupervised learning slope monitoring method, device and equipment based on optical fiber sensing and medium

Abstract

The invention discloses an unsupervised learning slope monitoring method, device and equipment based on optical fiber sensing and a medium, and belongs to the technical field of open pit coal mine slope monitoring, and the method comprises the steps: S1, obtaining optical fiber sensing data; s2, preprocessing the optical fiber sensing data obtained in the step S1 to obtain an effective data set; s3, dividing the effective data set obtained in the step S2 into a plurality of groups of sample data according to sampling points; and S4, inputting each group of sample data obtained in the step S3 into the pre-trained isolated forest model of each sampling point to obtain a monitoring report. According to the method, the data is divided according to the sampling points, then the divided data is input into the pre-trained isolated forest model of each sampling point to obtain the monitoring report, and independent modeling is performed for each sampling point, so that the accuracy of anomaly positioning is effectively improved; by adopting the method, the abnormal positioning error is less than or equal to the optical fiber sampling interval (0.5 m-1m), the abnormal area of the slope can be accurately locked, and compared with GNSS monitoring (the positioning error is 1m-3m), the precision is improved by more than 50%.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of open-pit coal mine slope monitoring, and particularly relates to a non-supervised learning slope monitoring method, device, equipment and medium based on optical fiber sensing. BACKGROUND

[0002] As an important infrastructure for mine production, the stability of the open-pit mine slope is directly related to the safety of mine production and personnel and equipment. The open-pit mine slope monitoring mainly relies on traditional monitoring technology, which has obvious limitations: ① GNSS positioning monitoring: only discrete point monitoring can be achieved, which cannot cover the whole slope, is easily affected by weather (such as heavy rain and thick fog) and electromagnetic interference, and the monitoring accuracy is easily reduced in complex mine environment, and it is difficult to capture local small strain anomalies of the slope.

[0003] ② Inclinometer monitoring: the equipment needs to be buried by drilling, which causes damage to the slope rock mass, and the monitoring range is limited to the surrounding of the drill hole, which cannot achieve large-scale continuous monitoring, and the equipment maintenance cost is high, and the response sensitivity to the strain change of the deep slope is insufficient.

[0004] ③ Manual inspection: relying on on-site observation of personnel, which is highly subjective and low in efficiency, cannot achieve real-time dynamic monitoring, and there is a risk of personnel safety in high and steep slopes and adverse weather conditions, and it is difficult to discover hidden slope anomalies in time.

[0005] Distributed optical fiber sensing technology (such as BOTDR) is gradually applied to slope monitoring, which can realize continuous strain monitoring of tens of kilometers through laid optical fibers, and obtain a continuous data set of "optical fiber length-strain", solving the problems of "discrete monitoring" and "limited coverage" of traditional technology. However, the existing monitoring schemes based on BOTDR still have defects in data processing and anomaly identification: ① Threshold method (such as setting a fixed strain threshold) or simple statistical analysis method (such as mean ± standard deviation) is mostly used to identify anomalies, which cannot adapt to the complex geological environment of the open-pit mine slope (such as non-uniformity of rock mass, temperature change interference, and construction disturbance), and is prone to "false positives" (such as judging normal strain changes caused by temperature as anomalies) or "missed reports" (such as being insensitive to slowly accumulated small strain anomalies).

[0006] ② There is a lack of special anomaly detection model for the "continuous multi-sampling point" data set of BOTDR (Brillouin Optical Time-Domain Reflectometer), and existing models are mostly designed for discrete point data, which cannot fully utilize the spatial correlation of continuous monitoring of optical fibers, resulting in low accuracy of anomaly positioning and difficulty in accurately locating the specific position of the slope anomaly strain (i.e. the slope area corresponding to the optical fiber length). SUMMARY

[0007] Based on the technical problems existing in the prior art, the present application provides a non-supervised learning slope monitoring method, device, equipment and medium based on optical fiber sensing.

[0008] According to the first aspect of the technical scheme of the present application, a non-supervised learning slope monitoring method based on optical fiber sensing is provided, which comprises the following steps: S1: obtaining optical fiber sensing data; S2: preprocessing the optical fiber sensing data obtained in step S1 to obtain an effective data set; S3: dividing the effective data set obtained in step S2 into a plurality of groups of sample data according to sampling points; S4: inputting each group of sample data obtained in step S3 into a pre-trained isolation forest model of each sampling point to obtain a monitoring report.

[0009] Further improvement of the present application is that in S2, the following steps are included: S21: obtaining original optical fiber sensing data, and performing noise reduction processing on the original optical fiber sensing data using a wavelet transform algorithm to obtain noise reduction data; S22: obtaining temperature change data, and correcting the noise reduction data according to the temperature change data to obtain corrected data; S23: removing outliers in the corrected data to obtain optical fiber sensing data.

[0010] Further improvement of the present application is that the establishment process of the pre-trained isolation forest model of each sampling point comprises the following steps: S31A: obtaining historical strain data of each sampling point; S32A: establishing an isolation forest model at each sampling point, and the initial abnormal threshold values of a plurality of isolation forest models are the same; S33A: inputting the historical strain data of each sampling point into the corresponding isolation forest model; S34A: randomly extracting a fixed magnitude of samples from the historical strain data of sampling point i to obtain a sub-sample set; S35A: randomly selecting two data points from the sub-sample set, and calculating the normal vector of the two data points as the slope vector of the isolated hyperplane; S36A: generating an intercept according to the slope vector and the two data points; S37A: dividing the sub-sample set into a left subset and a right subset according to the slope vector and the intercept; S38A: according to the left subset and the right subset, recursively dividing until the termination condition is met to obtain a first isolated tree of sampling point i; S39A: repeating S35A-S38A to obtain all isolated trees in the sampling point i, and establishing the isolated forest model of the sampling point i; S310A: repeating S34A-S39A to obtain the isolated forest model of all sampling points.

[0011] Further improvement of the present application is that the training process of the pre-trained isolated forest model of each sampling point comprises the following steps: S31B: dividing the historical strain data into a training set and a test set; S32B: inputting the training set of the sampling point i into the isolated forest model of the sampling point i to obtain the path length of each sample, and obtaining the anomaly score of each sample by normalizing the path length; S33B: comparing the anomaly score of each sample with the initial anomaly threshold to obtain the anomaly judgment result of the sample, wherein the anomaly judgment result comprises anomaly candidate and normal; S34B: judging the data with anomaly candidate in the anomaly judgment result according to the historical strain data to obtain the false alarm rate; S35B: when the false alarm rate does not belong to the preset reasonable interval, adjusting the initial anomaly threshold according to the false alarm rate and the preset false alarm rate; S36B: repeating steps S32B-S35B until the false alarm rate belongs to the preset reasonable interval to obtain the pre-trained isolated forest model of the sampling point i; S37B: repeating steps S32B-S36B to obtain the pre-trained isolated forest model of all sampling points.

[0012] Further improvement of the present application is that the reasonable interval is [1%, 5%].

[0013] Further improvement of the present application is that in S35B, the following steps are included: When the false alarm rate is greater than the maximum value of the reasonable interval, the initial anomaly threshold is increased by a preset first fixed value; When the false alarm rate is less than the minimum value of the reasonable interval, the initial anomaly threshold is decreased by a preset second fixed value, and the absolute value of the preset first fixed value is greater than the absolute value of the preset second fixed value.

[0014] Further improvement of the present application is that in S4, the following steps are included: S41: inputting each group of sample data into the pre-trained isolated forest model of the corresponding sampling point to obtain the anomaly judgment result of each sampling point; S42: obtaining the fiber length data of each sampling point; S43: Generate a mapping relationship between abnormal strain and fiber length based on the anomaly judgment results of each sampling point and the fiber length data of each sampling point; S44: A monitoring report is obtained based on the mapping relationship between the abnormal strain and the fiber length. The monitoring report includes the abnormal strain value and the location of the abnormal strain.

[0015] According to a second aspect of the technical solution of the present invention, an unsupervised learning slope monitoring device based on fiber optic sensing is provided, which employs the aforementioned unsupervised learning slope monitoring method based on fiber optic sensing, and includes: The data acquisition module is used to acquire fiber optic sensing data; The preprocessing module is used to preprocess the fiber optic sensing data to obtain a valid dataset; The partitioning module is used to divide the effective dataset into several groups of sample data according to the sampling points; The monitoring module is used to input the sample data of each group into the pre-trained isolated forest model of each sampling point to obtain a monitoring report.

[0016] According to a third aspect of the present invention, an electronic device is provided, comprising: One or more processors; Storage device for storing one or more computer programs. When one or more computer programs are executed by one or more processors, the one or more processors implement the above-described unsupervised learning slope monitoring method based on fiber optic sensing.

[0017] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the above-described unsupervised learning slope monitoring method based on fiber optic sensing.

[0018] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: 1. This invention divides data according to sampling points and then inputs the divided data into pre-trained isolated forest models for each sampling point to obtain a monitoring report. Independent modeling for each sampling point effectively improves the accuracy of anomaly location. The anomaly location error is kept ≤ fiber optic sampling interval (0.5m-1m), allowing for precise identification of minor anomaly areas on slopes (such as sudden strain changes in cracks within a 5m range). Compared to GNSS discrete monitoring (location error 1m-3m), the accuracy is improved by more than 50%.

[0019] 2. The data in this invention has undergone temperature compensation and wavelet noise reduction processing, and the strain misjudgment rate for temperature changes (±10℃) is ≤3%, which is more than 80% lower than the uncompensated threshold method (misjudgment rate 25%-30%), and can adapt to the environment of large day and night temperature difference in open mines. Attached Figure Description

[0020] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein: Figure 1 This is a flowchart of an unsupervised learning slope monitoring method based on fiber optic sensing according to the present invention; Figure 2 This is a structural block diagram of an unsupervised learning slope monitoring device based on fiber optic sensing according to the present invention. Detailed Implementation

[0021] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] Example One This invention provides an unsupervised learning slope monitoring method based on fiber optic sensing, which includes the following steps: S1: Acquire fiber optic sensing data; S2: Preprocess the fiber optic sensing data obtained in step S1 to obtain a valid dataset; S3: Divide the valid dataset obtained in step S2 into several groups of sample data according to the sampling points; S4: Input the sample data obtained in step S3 into the pre-trained isolated forest model of each sampling point to obtain the monitoring report.

[0023] Specifically, in S1, the fiber optic sensing data is obtained by laying sensing fibers in key areas of the open-pit mine slope, such as potential sliding surfaces, steep areas, and fracture development zones, using a combination of surface bonding and borehole installation. This ensures that the fiber optic cables cover the entire slope area, and each fiber segment corresponds to a specific spatial location on the slope (i.e., the fiber length maps one-to-one with the actual slope area). The fiber optic sensing data is collected using a BOTDR device, with a sampling interval set (recommended 0.5m-1m, adjusted according to the slope monitoring accuracy requirements). Strain data is collected at each sampling point on the fiber optic cable, forming a continuous dataset with strain values ​​on the vertical axis and fiber length on the horizontal axis, denoted as fiber optic sensing data D={(L1,ε1),(L2,ε2),...,(L m,ε m )}, where L i ε is the fiber length (in km) of the i-th sampling point. i Let be the strain value (in με) at the i-th sampling point, and m be the total number of fiber sampling points.

[0024] Specifically, S2 includes the following steps: S21: Acquire the original optical fiber sensing data, and use the wavelet transform algorithm to perform noise reduction processing on the original optical fiber sensing data to obtain noise-reduced data; Specifically, in S21, random noise (e.g., device circuit noise or optical signal attenuation noise) in the original optical fiber sensing data is removed by wavelet transform algorithm.

[0025] S22: Acquire temperature change data, and correct the noise reduction data based on the temperature change data to obtain corrected data; Specifically, in S22, the temperature change data is acquired via optical fiber. The calculation formula for the correction data is as follows: ε comp,i =ε i -α*ΔT; In the formula, ε comp,i To correct the data, με; ε i The original fiber optic sensing data is με; α is the temperature strain coefficient of the fiber (a known constant, such as α≈11*10 for ordinary single-mode fiber). -6 / ℃), ΔT is the temperature change data of the area where the optical fiber is located during the monitoring period (obtained by temperature sensors deployed near the optical fiber).

[0026] S23: Remove outliers from the corrected data to obtain fiber optic sensing data.

[0027] Specifically, in S23, the outlier values ​​are invalid data, including data such as abrupt changes in strain values ​​caused by fiber breakage. The fiber optic sensing data D obtained after preprocessing... valid ={(L1,ε comp,1 ),(L2,ε comp,2 ),...,(L m ,ε comp,m )}.

[0028] Specifically, in S3, the effective dataset obtained in step S2 is divided into several groups of sample data according to the principle of independent modeling of fiber optic sampling points. Each sampling point is monitored independently, thereby improving the accuracy of monitoring. Furthermore, the location of each sampling point is known, so the problem sampling point can be quickly located when the data is judged to be abnormal.

[0029] Specifically, the process of establishing the pre-trained isolated forest model for each sampling point includes the following steps: S31A: Obtain historical strain data for each sampling point; Specifically, in S31A, the historical strain data refers to data from the most recent 30-90 days under normal operating conditions. The 30-90 days range is just an example; it can be freely set in actual use. Normal operating conditions mean that the data covers the slope's daily minor deformations without any abnormal events.

[0030] S32A: An isolated forest model is established at each sampling point, and the initial anomaly thresholds of several isolated forest models are the same; Specifically, in S32A, each sampling point corresponds to one isolated forest model (denoted as IF). i The parameters of the isolated forest model include: number of isolated trees T=100 (balancing model stability and computational efficiency; experimental verification shows that when T=100, the AUC value is ≥0.9 and the variance is ≤0.05), number of subsamples ψ=256 (avoiding the "curse of dimensionality" and adapting to the one-dimensional characteristics of single-sample point strain data), and initial value S for the outlier score threshold. th0 =0.5 (to be adjusted later through voting). The isolated forest model was constructed using unsupervised training.

[0031] S33A: Input the historical strain data of each sampling point into the corresponding isolated forest model; S34A: Randomly extract a fixed number of samples from the historical strain data of sampling point i to obtain a subsample set; S35A: Randomly select two data points from the subsample set, and calculate the vertical normal vector of the two data points as the slope vector of the isolated hyperplane; S36A: Generate the intercept based on the slope vector and the two data points; S37A: Divide the subsample set into a left subset and a right subset according to the slope vector and the intercept; S38A: Based on the left subset and the right subset, perform recursive partitioning until the termination condition is met to obtain the first isolated tree of sampling point i; S39A: Repeat S35A-S38A to obtain all isolated trees at sampling point i and establish an isolated forest model for sampling point i; S310A: Repeat S34A-S39A to obtain an isolated forest model for all sampling points.

[0032] Specifically, the implementation methods of S34A-S310A are as follows: For the training set P of the i-th sampling point... train,i An isolated forest model (IF) is constructed using unsupervised training. i The process is as follows: From P train,iA subset X is formed by randomly selecting ψ=256 samples from the sample. sub,i Randomly select a subset X sub,i Two strain data points ε a ,ε b Calculate the perpendicular normal vector between the two points as the slope vector W of the isolated hyperplane (formula: W=GetNormal(ε)). a ,ε b )); Specifically, in S36A, the generation of the intercept includes the following steps; S36A1: Calculate the product of the strain value and the slope vector at the two data points respectively, and denote it as the first value and the second value, where the first value is greater than the second value; S36A2: Randomly generate the intercept b within the interval [second value, first value].

[0033] Specifically, the first value is represented as W*ε a The second value is represented as W*ε b , formula: b=RandomIntercept(W*ε a ,W*ε b ).

[0034] Specifically, in S37A, a hyperplane is generated based on the slope vector and the intercept. The expression for the hyperplane is: W*ε+b=0; The subset X is determined according to the hyperplane. sub,i Divide into left subset X L,i (W*ε+b≤0) and the right subset X R,i (W*ε+b>0); Repeat the above steps for X L,i With X R,i Recursively partition the tree until the number of samples in the subset is ≤1 or the tree height reaches a threshold (a threshold of log2ψ=8 is recommended), forming one isolated tree (iTree); repeat the above steps to generate T=100 isolated trees, forming the isolated forest model IF for the i-th sampling point. i Traverse all sampling points to complete the total training set P. train The corresponding set of isolated forest models IF total ={IF1,IF2,...,IF m The construction of}.

[0035] Specifically, the training process of the pre-trained isolated forest model for each sampling point includes the following steps: S31B: Divide the historical strain data into a training set and a test set; Specifically, in S31B, the historical strain data is extracted from historical fiber optic sensing data, and for each sampling point i (corresponding to fiber length L)... i Extract its historical strain sequence, i.e., historical strain data ε. comp,i (t1),ε comp,i (t2),...,ε comp,i (t n ), t1-t n For historical monitoring times, n≥200 to ensure sufficient sample size for model training, the historical strain data is divided into a training set P for sampling point i in a 7:3 ratio. train,i The test set P with sampling point i test,i That is, P train,i ={ε comp,i (t1),...,ε comp,i (t 0.7n )}, P test,i ={ε comp,i (t 0.7n+1 ),...,ε comp,i (t n The training set of all sampled points constitutes the total training set P. train ={P train,1 ,P train,2 ,...,P train,m The test set of all sampled points constitutes the total test set P. test ={P test, 1,P test,2 ,...,P test,m}

[0036] S32B: Input the training set of sampling point i into the isolated forest model of sampling point i to obtain the path length of each sample, and then normalize the path length to obtain the anomaly score of each sample. Specifically, in S32B, the path length h(ε) of the sample in each iTree is first calculated (i.e., the number of edges from the root to the leaf node). Then, the path length is normalized according to the following formula to calculate the anomaly score S(ε,ψ): ; In the formula, E(h(ε)) is the average path length of the sample in 100 iTrees; ; In the formula, γ=0.5772156649 is the Euler constant, and when ψ=256, c(ψ)≈8.485.

[0037] S33B: Compare the anomaly score of each sample with the initial anomaly threshold to obtain the anomaly judgment result of the sample, the anomaly judgment result including anomaly candidates and normal; Specifically, if S(ε,ψ)>S th0 If the value is 0.5, the sample is considered abnormal; otherwise, it is considered normal.

[0038] S34B: Based on historical strain data, the data that are candidates for anomaly are judged according to the anomaly judgment result to obtain the false alarm rate; Specifically, in S34B, the abnormal judgment results of all sampling point training sets are voted on, and the proportion of abnormal samples that are actually normal data is calculated to obtain the false alarm rate. Since the historical strain data is past data, it is known whether each data has an anomaly. Therefore, by comparing the data in the anomaly candidates with the actual situation, it can be determined whether it is a false alarm. The false alarm rate is obtained by dividing the number of false alarm data by the number of anomaly candidate data.

[0039] S35B: When the false alarm rate is not within a preset reasonable range, adjust the initial anomaly threshold according to the false alarm rate and the preset false alarm rate; Specifically, S35B includes the following steps: When the false alarm rate is greater than the maximum value of the reasonable range, the initial anomaly threshold is increased by a first fixed value; When the false alarm rate is less than the minimum value of the reasonable range, the initial anomaly threshold is reduced by a second fixed value, wherein the absolute value of the first fixed value is greater than the absolute value of the second fixed value.

[0040] Specifically, the reasonable range is [1%, 5%]. If the false positive rate is >5%, the initial anomaly threshold is increased by +0.05 each time (i.e., the first fixed value). If the false positive rate is <1%, the initial anomaly threshold is decreased (-0.03 each time (i.e., the second fixed value)). This continues until the false positive rate stabilizes between 1% and 5%, at which point the final initial anomaly threshold is determined (experiments have shown that the final initial anomaly threshold is typically between 0.6 and 0.7). Setting a larger first fixed value allows the false positive rate to quickly fall into the target range, reducing the number of iterations. Setting a smaller second fixed value allows for the precise release of suspicious samples, preventing a sharp drop in recall. By setting different first and second fixed values, the threshold calibration process converges earlier, shortening the overall training time.

[0041] S36B: Repeat steps S32B-S35B until the false alarm rate falls within a preset reasonable range, to obtain the isolated forest model pre-trained for sampling point i. S37B: Repeat steps S32B-S36B to obtain a pre-trained isolated forest model for all sampling points.

[0042] Specifically, by using the test set P of the sampling points test,i Input the corresponding isolated forest model IF i The S(ε) of each test sample is obtained according to the above anomaly score calculation method. test ,ψ); if S(ε test ,ψ)>S th The sampling point corresponding to the test sample was determined to have "abnormal strain change", and its fiber length L was recorded. i With strain value ε test Otherwise, it is judged as "normal response".

[0043] Specifically, S4 includes the following steps: S41: Input each set of sample data into the pre-trained isolated forest model corresponding to the sampling point, and the anomaly judgment result of each sampling point; S42; Obtain the fiber optic length data for each sampling point; S43: Generate a mapping relationship between abnormal strain and fiber length based on the anomaly judgment results of each sampling point and the fiber length data of each sampling point; S44: A monitoring report is obtained based on the mapping relationship between the abnormal strain and the fiber length. The monitoring report includes the abnormal strain value and the location of the abnormal strain.

[0044] Specifically, in S4, information from all sampling points identified as exhibiting "abnormal strain" is compiled to form an "abnormal strain-fiber length" lookup table. Combined with the "fiber length-slope area" mapping relationship established during fiber optic cable laying, the specific spatial location of abnormal slope changes is precisely pinpointed (e.g., "fiber length 3.2km corresponds to the crack area of ​​platform 2 on the east side of the slope"). The strain change trend of the abnormal sampling points is output (e.g., "strain value at 3.2km increases from 20με to 85με, anomaly score 0.72"), providing a quantitative basis for slope safety early warning.

[0045] The method in this embodiment has the following effects: 1. Improved accuracy in anomaly detection: ① Accuracy index: The method of this invention was used to test BOTDR data of open-pit mine slopes (fiber optic length 0-4km, 800 monitoring points). The accuracy of the model training set (train_accuracy) was stable at 0.96-0.98, and the accuracy of the test set (test_accuracy) was stable at 0.92-0.94, which is 15%-20% higher than the traditional threshold method (accuracy 0.75-0.80).

[0046] ② Anomaly location accuracy: Each monitoring point of the optical fiber is independently modeled, and the anomaly location error is ≤ optical fiber sampling interval (0.5m-1m). It can accurately locate small anomaly areas on the slope (such as sudden changes in crack strain within a 5m range), which is more than 50% more accurate than GNSS discrete monitoring (location error 1m-3m).

[0047] ③ Environmental interference resistance: After temperature compensation and wavelet denoising, the strain misjudgment rate of this method for temperature changes (±10℃) is ≤3%, which is more than 80% lower than the uncompensated threshold method (misjudgment rate 25%-30%), and can adapt to the environment of large day and night temperature difference in open mines.

[0048] 2. Real-time performance meets engineering requirements: ① Data processing efficiency: The BOTDR data acquisition interval can be set to 5-10 minutes. The isolated forest model adopts unsupervised learning (no manual sample labeling required). The training time for a single monitoring point model is ≤2 seconds (based on Intel Core i7-12700 processor). The total processing time for 800 monitoring points is ≤30 minutes, which meets the "hourly" safety early warning requirements of open-pit mines.

[0049] ② Abnormal response speed: When a sudden abnormal strain occurs on the slope (such as strain rate increase > 5με / h), this method can identify and output the abnormality within one acquisition cycle (≤10 minutes), which is more than 12 times faster than manual inspection (response time ≥ 2 hours).

[0050] Reliability and engineering suitability: ① Model stability: After 500 repeated tests on the same slope BOTDR data, the range of outlier scores of this method is ≤0.05, which is more than 60% more stable than the traditional isolated forest algorithm (range 0.11-0.19).

[0051] ② Engineering cost advantage: The deployment cost of BOTDR fiber optic is only 1 / 3 of that of inclinometer, and this method is based on unsupervised algorithm (no need for a large number of labeled samples), and subsequent maintenance does not require manual intervention, reducing the annual operation and maintenance cost by 40%-50% compared with traditional monitoring technology.

[0052] Example Two like Figure 2 As shown, an unsupervised learning slope monitoring device based on fiber optic sensing is provided, which includes: The data acquisition module is used to acquire fiber optic sensing data; The preprocessing module is used to preprocess the fiber optic sensing data to obtain a valid dataset; The partitioning module is used to divide the effective dataset into several groups of sample data according to the sampling points; The monitoring module is used to input the sample data of each group into the pre-trained isolated forest model of each sampling point to obtain a monitoring report.

[0053] Example Three This embodiment provides an example of the practical application of an unsupervised learning slope monitoring method based on fiber optic sensing, as follows: 1. Equipment deployment and data acquisition: ① Fiber optic cable laying: Single-mode sensing fiber optic cable (model: G.652D) is laid along the west side of the slope (fracture development zone) using the "surface concrete bonding" method. The total length of the fiber optic cable is 3km, the sampling interval is 1m, and there are a total of 3000 monitoring points. A temperature sensor (accuracy ±0.5℃) is deployed every 500m next to the fiber optic cable for temperature compensation.

[0054] ②BOTDR parameter settings: Start the BOTDR device (model: AQ8603), set the laser power to 10dBm, sampling time to 10s, acquisition cycle to 10 minutes, and continuously acquire data for 30 days (normal working conditions, no slope anomalies), obtaining the total dataset D={(L1,ε1),...,(L 3000 ,ε 3000 )}, where L i =0.001ikm (i=1-3000), ε i Range 5με-30με.

[0055] ③ Data preprocessing: Wavelet transform denoising: Using the db4 wavelet basis, decomposed into 3 levels, with a threshold of 0.02, to remove random noise; Temperature compensation: Temperature compensation is performed on the collected data using the following formula, according to ε comp,i =ε i -11×10 -6 ×8 Calculate the compensated strain to eliminate the influence of temperature; Invalid data removal: Remove the strain abrupt change point (ε=150με) at the fiber optic connector (L=1.5km) to obtain the valid dataset D. valid .

[0056] 2. Dataset partitioning and model building: ① Dataset partitioning: The 30-day historical data for each monitoring point (144 data points per day, totaling 4320 data points) was partitioned into a training set P at a ratio of 7:3. train,i (3024 samples) and test set P test,i (1296 samples) form the total training set P train With the total test set P test .

[0057] ② Training of the isolated forest model: Parameter settings: IF of the model for each monitoring point i The number of isolated trees is T=100, the number of subsamples is ψ=256, and the initial threshold is S.th0 =0.5; Model training: Taking the 1000th monitoring point (L=1.0km) as an example, from P train,1000 256 samples were randomly selected from the data to generate 100 iTrees: ε was randomly selected from the samples. a =12με、ε b =18με, calculate the normal vector W=(-1,1) (simplified one-dimensional strain scenario); the intercept b is randomly selected as 15 between W*12=6 and W*18=6 (simplified to b between 12-18 in the one-dimensional scenario); divide the subset according to ε+15≤0 (simplified formula), recursively construct iTree; traverse 3000 monitoring points to complete IF total Build.

[0058] 3. Voting decisions and anomaly detection: ① Training set voting: voting on all P train,i Calculate the outlier score, initial S th0 When the value is 0.5, the false alarm rate is 8%, improving S th With a value of 0.65, the false alarm rate decreased to 3%, and the final threshold S was determined. th =0.65.

[0059] ② Test set anomaly prediction: Input P test,i At the 2000th monitoring point (L=2.0km), 10 test samples were found to have anomaly scores >0.65 (maximum 0.73), corresponding to strain values ​​increasing from 20με to 75με, which were determined to be "abnormal strain".

[0060] 4. Anomaly location and result output: ① Anomaly location: Based on the optical fiber laying mapping relationship, L=2.0km corresponds to platform No. 3 on the west side of the slope (crack number F12).

[0061] ② Output anomaly report: For example, "Abnormal strain has occurred in the F12 crack zone of Platform 3 on the west side of the slope (fiber length 2.0km), strain value 75με, anomaly score 0.73, it is recommended to strengthen the inspection and support of this area."

[0062] 5. Implementation effect verification: ① On-site investigation: At L=2.0km, the width of F12 crack was found to have increased from 5mm to 15mm, which is consistent with the results of abnormal monitoring.

[0063] ②Accuracy verification: The model test set accuracy is 93.5%, and the abnormal positioning error is 0.8m, which meets the engineering requirements.

[0064] ③ Real-time verification: It takes 25 minutes from data acquisition to abnormal output, which meets the "hourly" warning requirement.

[0065] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described unsupervised learning slope monitoring methods based on fiber optic sensing.

[0066] The electronic device of this invention includes: one or more processors; and a storage device for storing one or more computer programs, wherein when the one or more computer programs are executed by the one or more processors, the one or more processors implement the unsupervised learning slope monitoring method based on fiber optic sensing provided by this invention.

[0067] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A non-supervised learning slope monitoring method based on optical fiber sensing, characterized in that, The method comprises the following steps: S1: obtaining optical fiber sensing data; S2: preprocessing the optical fiber sensing data obtained in step S1 to obtain an effective data set; S3: dividing the effective data set obtained in step S2 into a plurality of group sample data according to sampling points; S4: inputting each group sample data obtained in step S3 into a pre-trained isolation forest model of each sampling point to obtain a monitoring report. 2.The unsupervised learning slope monitoring method based on fiber sensing of claim 1, wherein, In S2, the following steps are included: S21: obtaining original optical fiber sensing data, and performing noise reduction processing on the original optical fiber sensing data by using a wavelet transform algorithm to obtain noise reduction data; S22: obtaining temperature change data, and correcting the noise reduction data according to the temperature change data to obtain corrected data; S23: removing outliers in the corrected data to obtain optical fiber sensing data. 3.The unsupervised learning slope monitoring method based on fiber sensing of claim 1, wherein, The establishment process of the pre-trained isolation forest model of each sampling point comprises the following steps: S31A: obtaining historical strain data of each sampling point; S32A: establishing an isolation forest model at each sampling point, and the initial abnormal threshold of a plurality of isolation forest models is the same; S33A: inputting the historical strain data of each sampling point into the corresponding isolation forest model; S34A: randomly extracting a fixed magnitude sample from the historical strain data of sampling point i to obtain a sub-sample set; S35A: randomly selecting two data points from the sub-sample set, and calculating the normal vector of the two data points as the slope vector of the isolated hyperplane; S36A: generating an intercept according to the slope vector and the two data points; S37A: dividing the sub-sample set into a left subset and a right subset according to the slope vector and the intercept; S38A: recursively dividing according to the left subset and the right subset until a termination condition is met to obtain a first isolated tree of sampling point i; S39A: repeating S35A-S38A to obtain all isolated trees in sampling point i, and establishing an isolation forest model of sampling point i; S310A: repeating S34A-S39A to obtain isolation forest models of all sampling points. 4.The unsupervised learning slope monitoring method based on fiber sensing of claim 3, wherein, The training process of the pre-trained isolation forest model of each sampling point comprises the following steps: S31B: dividing the historical strain data into a training set and a test set; S32B: inputting the training set of sampling point i into the isolation forest model of sampling point i to obtain the path length of each sample, and then performing normalization processing on the path length to obtain the abnormal score of each sample; S33B: comparing the abnormal score of each sample with the initial abnormal threshold to obtain an abnormal judgment result of the sample, and the abnormal judgment result includes abnormal candidates and normal; S34B: judging the data with the abnormal judgment result of the abnormal candidate according to the historical strain data to obtain a false alarm rate; S35B: when the false alarm rate does not belong to a preset reasonable interval, adjusting the initial abnormal threshold according to the false alarm rate and a preset false alarm rate; S36B: repeating steps S32B-S35B until the false alarm rate belongs to the preset reasonable interval to obtain a pre-trained isolation forest model of sampling point i; S37B: repeating steps S32B-S36B to obtain pre-trained isolation forest models of all sampling points.

5. The unsupervised learning slope monitoring method based on fiber sensing according to claim 4, characterized in that, The reasonable interval is [1%, 5%]. 6.The unsupervised learning slope monitoring method based on fiber sensing of claim 4, wherein, In S35B, the following steps are included: When the false positive rate is greater than the maximum value of the reasonable interval, the initial abnormal threshold is increased by a preset first fixed value; When the false positive rate is less than the minimum value of the reasonable interval, the initial abnormal threshold is reduced by a preset second fixed value, and the absolute value of the preset first fixed value is greater than the absolute value of the preset second fixed value. 7.The unsupervised learning slope monitoring method based on fiber sensing of claim 1, wherein, In S4, the following steps are included: S41: input the sample data of each group into the isolated forest model pre-trained for the corresponding sampling point, and obtain the abnormal judgment result of each sampling point; S42: obtain the optical fiber length data of each sampling point; S43: generate the mapping relationship between abnormal strain and optical fiber length according to the abnormal judgment result of each sampling point and the optical fiber length data of each sampling point; S44: obtain a monitoring report according to the mapping relationship between abnormal strain and optical fiber length, and the monitoring report includes abnormal strain value and abnormal strain position.

8. A fiber-optic sensing based unsupervised learning slope monitoring device employing the fiber-optic sensing based unsupervised learning slope monitoring method of any one of claims 1-7, characterized in that, It includes: a data acquisition module for acquiring optical fiber sensing data; a preprocessing module for preprocessing the optical fiber sensing data to obtain an effective data set; a division module for dividing the effective data set into a plurality of groups of sample data according to sampling points; a monitoring module for inputting each group of sample data into an isolated forest model pre-trained for each sampling point to obtain a monitoring report.

9. An electronic device, comprising: It includes: one or more processors; a storage device for storing one or more computer programs, When the one or more computer programs are executed by the one or more processors, the one or more processors implement a non-supervised learning slope monitoring method based on optical fiber sensing as claimed in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement a non-supervised learning slope monitoring method based on optical fiber sensing as claimed in any one of claims 1-7.

Images