A prediction method and system for analyzing overbreak data by using artificial intelligence

Abstract

The application discloses a kind of prediction method and system for analyzing overbreak and underbreak data using artificial intelligence, specifically relates to the field of overbreak and underbreak prediction and edge data transmission, including by collecting the section point data, attitude data and working condition data corresponding to current driving cycle, the section point data is divided into multiple section segments according to the same footage interval, and the attitude data and working condition data are corresponded to each section segment according to time sequence, and the data set indexed by section segment is output;By performing deletion reconstruction verification on the data in the section segment under the condition of limited edge data transmission to identify early abnormal evidence that cannot be supplemented, and the identified early abnormal evidence is organized as evidence segment according to the position and time continuous relationship, and combined with the contour, attitude and working condition information for time series prediction analysis, to solve the problem that overbreak and underbreak starting abnormal evidence is easy to lose under the condition of limited edge data transmission, which leads to insufficient basis for forward-looking prediction.

Description

Technical Field

[0001] This invention relates to the field of over-drilling and under-drilling prediction and edge data transmission technology. More specifically, this invention relates to a prediction method and system that uses artificial intelligence to analyze over-drilling and under-drilling data. Background Technology

[0002] In the field of over-excavation and under-excavation prediction after tunnel, roadway, or slope excavation, the mainstream practice in the industry is mainly to solve the problem of whether the subsequent cross-section will expand due to over-excavation or under-excavation, and whether construction correction can be carried out in advance. Usually, cross-section scanning data, equipment attitude data, and propulsion or drilling and blasting condition data are collected on site, and after compression, sampling, or hierarchical screening at the edge side, they are uploaded to the central side, and then artificial intelligence models analyze the subsequent over-excavation and under-excavation trends. Taking the continuous tunneling construction scenario under weak network conditions as an example, the site needs to continuously complete the over-excavation and under-excavation prediction under the hard constraints of limited edge bandwidth, limited storage resources, and the inability to fully transmit the original point cloud and continuous working condition stream for a long time. However, under this constraint, the mainstream approach will consistently reveal such a bottleneck: some early abnormal signs with small amplitude, short duration and only local appearance, although they have appeared before the subsequent obvious over- or under-excavation, are often discarded as low-value data during compression, sampling or priority screening because of their small changes and low frequency. Even if more obvious cross-sectional deviation results are obtained on the central side later, it is impossible to deduce whether these signs have ever appeared or how the deviation initially started to form. In the end, the model can identify the deviation that has been formed, but it is difficult to predict the initial stage of the deviation in a timely manner, and the prediction basis is inconsistent and the results are unstable when different batches of data are replayed. Therefore, the technical problem to be solved by this application is: how to prioritize the preservation of early abnormal evidence that has unique evidentiary significance for the initiation of over- or under-mining under the condition of limited edge data transmission, and which cannot be subsequently recovered, so as to support artificial intelligence in making reliable forward predictions of over- or under-mining risks. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a prediction method and system for analyzing over-excavation and under-excavation data using artificial intelligence. By performing censoring and reconstruction verification on data within a cross-sectional segment under the condition of limited edge data transmission, the method identifies irreparable early abnormal evidence. The identified early abnormal evidence is then organized into evidence segments according to the continuity of position and time, and combined with contour, attitude, and working condition information for time-series prediction analysis, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a prediction method for analyzing over- and under-mining data using artificial intelligence, comprising: S1. By collecting the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle, the cross-section point data is divided into multiple cross-section segments according to the same advance interval, and the attitude data and working condition data are mapped to each cross-section segment in chronological order, and the data set indexed by the cross-section segment is output. S2. Based on each dataset, delete one original data from the current dataset, and call the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. Then, compare the reconstruction result with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data, and the irrecoverable dataset corresponding to each cross-section is output. S3. By connecting the various irretrievable datasets, the irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment are sequentially connected, and the first data time, the last data time, the cross-sectional segment location, and the data source are written into the connection result, and the evidence segment set is output. S4. Based on the evidence segment set, first sort the evidence segments from earliest to latest according to the time of the first data, then sort the evidence segments with the same first data time from most to least according to the number of data sources, and attach the contour summary, attitude summary and working condition summary of the cross section to each sorted evidence segment, and output the evidence packet sequence arranged in a fixed sending order.

[0005] In a preferred embodiment, it further includes: S5. By converting the evidence packet sequence into an evidence time series matrix according to the sending order, and inputting the evidence time series matrix into the time series attention network, the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section is extracted, and the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections are output. S6. After the excavation of the corresponding subsequent tunneling section is completed, the actual cross-sectional deviation data of the subsequent tunneling section is collected, and the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results are written back according to the cross-sectional section position, and an evidence sample set with actual deviation markers is output.

[0006] In a preferred embodiment, S1 includes: S1-1. Read the acquisition time and advance position of each sampling point in the cross-section point data, and calculate the lower limit of time, upper limit of time, lower limit of position and upper limit of position for each cross-section segment, and output the spatiotemporal range corresponding to each cross-section segment. S1-2. Read the acquisition time and advance position of each data in the attitude data and working condition data. Write the attitude data and working condition data whose acquisition time is between the lower limit and the upper limit of the time and whose advance position is between the lower limit and the upper limit of the position into the corresponding section segment to form the intra-segment dataset corresponding to each section segment. S1-3. Read the attitude data and working condition data of the dataset not written to the segment. For each piece of unwritten data, calculate the time difference between its acquisition time and the lower limit of the time of each segment and the time difference between its acquisition time and the upper limit of the time of each segment. Write the unwritten data whose advance position is between the lower limit and the upper limit of the position and whose absolute value of the time difference is first into the corresponding segment. Output the data set corresponding to each segment.

[0007] In a preferred embodiment, S2 includes: S2-1. Read the data set corresponding to each cross section, extract the current raw data one by one, and read the data source, field identifier, acquisition time, advance position, order of occurrence and data value corresponding to the current raw data. Then delete the current raw data from the cross section where it is located, and extract the previous and next neighbor data of the current raw data in this cross section, the same position data in the previous cross section and the same position data in the next cross section, and output the data group to be judged. S2-2. Based on the data group to be judged, construct the time-series reconstruction value according to the preceding and following neighbor data, construct the inter-segment reconstruction value according to the co-position data in the preceding segment and the co-position data in the following segment, form the front increment according to the preceding and following neighbor data and the co-position data in the preceding segment, and form the rear increment according to the following and following neighbor data and the co-position data in the following segment. Then, construct the propagation reconstruction value together with the front increment and the rear increment, and output the three-way reconstruction value corresponding to the current original data.

[0008] In a preferred embodiment, S2 further includes: S2-3. Based on the three reconstructed values, calculate the original value difference, previous neighbor difference, and subsequent neighbor difference between the time-series reconstructed value, inter-segment reconstructed value, and propagation reconstructed value and the current original data, respectively. The original value difference is the difference between the reconstructed value and the data value of the current original data, the previous neighbor difference is the difference between the reconstructed value and the data value of the previous neighbor data, and the subsequent neighbor difference is the difference between the data value of the subsequent neighbor data and the reconstructed value. Then, take the positive and negative results of the original value difference, the previous neighbor difference, and the subsequent neighbor difference to form the time-series symbol group, the inter-segment symbol group, and the propagation symbol group, respectively. Take the corresponding positive and negative results of the current original data relative to the previous neighbor data and the subsequent neighbor data to form the original symbol group, and output the three-way comparison results. S2-4. Based on the three-way comparison results, first sort the original value difference, the previous neighbor difference, and the next neighbor difference in ascending order of absolute value. Then, read the symbol group corresponding to the first reconstructed value and compare it with the original symbol group bit by bit. When they match bit by bit, write the current original data into the compensable dataset. When they do not match bit by bit, continue to read the symbol group corresponding to the second reconstructed value and compare it with the original symbol group bit by bit. When the second reconstructed value still does not match bit by bit, write the current original data into the non-compensable dataset of the section segment and output the non-compensable dataset corresponding to each section segment.

[0009] In a preferred embodiment, S3 includes: S3-1. Read the non-recoverable datasets corresponding to each cross section, extract the acquisition time, cross section coordinates and data source for each non-recoverable data, expand the cross section coordinates along the cross section outline into position numbers, generate position markers corresponding to each non-recoverable data according to the acquisition time and position number, and output the position marker set. S3-2. Based on the location marker set, compare the acquisition time difference, location sequence difference, and data source change of two adjacent non-recoverable data in the same cross section. Connect the non-recoverable data with continuous acquisition time difference, continuous location sequence, and consistent expansion direction to form the same evidence segment. Write the connection start point as the first data time and start position, and the connection end point as the last data time and end position. Write the data sources in the connection process into the source sequence according to the order of appearance, and output the evidence segment set. S3-3. Based on the evidence segment set, the starting and ending values ​​of the position number of each evidence segment are counted, and the evidence segments are written into the side wall area, arch area or invert arch area according to the cross section interval where the starting and ending values ​​are located. The evidence segment set with the first data time, the last data time, the cross section position, the source sequence and the partition mark is output.

[0010] In a preferred embodiment, S4 includes: S4-1. Extract the first data time, data source, section location, duration, and location span of each evidence segment, as well as the contour data, attitude data, and working condition data corresponding to the section segment. Generate the sending order by sorting the first data time in ascending order, the number of data sources in descending order, the duration in descending order, and the location span in descending order, and output the set of evidence segments to be sent. S4-2. Extract the local contour data, attitude change data and operating condition change data within the location range of each evidence segment in the order of transmission. Write the evidence segment, local contour data, attitude change data and operating condition change data into the evidence packet in the order of transmission. Output the evidence packet sequence arranged in a fixed transmission order.

[0011] In a preferred embodiment, S5 includes: S5-1. Expand the evidence package sequence into a timeline queue according to the sending order, extract the evidence segment, local contour data, attitude change data and working condition change data from each evidence package, write them into the corresponding time position according to the sending order, and generate position code, source code and direction code according to the position of the evidence segment, data source and change direction within the same time position, and output the evidence time sequence matrix. S5-2. Input the evidence time series matrix into the time series attention network, calculate the evidence association value, contour association value, attitude association value and working condition association value between each time position and each subsequent time position, then accumulate each association value according to the time position and cross section position and map it to the corresponding subsequent tunneling section, and output the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling section.

[0012] In a preferred embodiment, S6 includes: S6-1. Extract the actual cross-sectional deviation data and over- and under-excavation prediction results for the corresponding subsequent tunneling sections. Match the actual cross-sectional deviation data with the over- and under-excavation prediction results segment by segment according to the cross-sectional section position. Write the cross-sectional section position, predicted deviation value, actual deviation value and deviation direction into each corresponding result and output the corresponding record set. S6-2. Subtract the predicted deviation value from the actual deviation value in each corresponding record to form the deviation difference value. Compare the predicted deviation direction with the actual deviation direction in each corresponding record to form the direction comparison result. Then write the deviation difference value and the direction comparison result back to the corresponding evidence segment according to the cross-section segment position and output the evidence record set with the actual deviation mark. S6-3. Write the evidence record set into the sample storage area in the order of cross-sectional segment position, first data time, and data source, and aggregate them according to cross-sectional segment position to form an evidence sample set with actual deviation markers.

[0013] In a preferred embodiment, a predictive system employing artificial intelligence to analyze over- and under-mining data includes: The data segmentation module is used to collect the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle. It divides the cross-section point data into multiple cross-section segments according to the same advance interval, and maps the attitude data and working condition data to each cross-section segment in chronological order, and outputs a data set indexed by the cross-section segment. The deletion and verification module deletes one original data from the current data set based on each data set. It then calls the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. The reconstruction result is then compared with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data and the irrecoverable dataset corresponding to each cross-section is output. The evidence connection module is used to connect various irretrievable datasets. It sequentially connects irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment, and writes the first data time, the last data time, the cross-sectional segment location, and the data source into the connection result, and outputs a set of evidence segments. The evidence arrangement module, based on the evidence segment set, first sorts the evidence segments from earliest to latest according to the time of the first data, then sorts the evidence segments with the same first data time from most to least according to the number of data sources, and adds the contour summary, attitude summary and working condition summary of the cross section segment to each sorted evidence segment, and outputs a sequence of evidence packets arranged in a fixed sending order. The risk prediction module is used to convert the evidence packet sequence into an evidence time series matrix according to the sending order, and input the evidence time series matrix into the time series attention network to extract the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section, and output the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections. The sample write-back module is used to collect the actual cross-sectional deviation data of the subsequent tunneling section after the excavation of the corresponding subsequent tunneling section is completed, and write back the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results according to the cross-sectional section position, and output the evidence sample set with actual deviation markers.

[0014] The technical effects and advantages of this invention are as follows: By identifying irretrievable data that cannot be stably recovered after deletion at the edge and retaining it as a subsequent object, the omission of abnormal evidence of over- or under-mining in the compression and sampling process is relatively reduced, thereby improving the problem of insufficient forward prediction basis under the condition of limited edge data transmission. By performing censoring and reconstruction on the current cross-sectional data, and combining the adjacent data of this segment with the co-located data of the adjacent cross-sectional segments to form a three-way reconstruction result, and then comparing it layer by layer according to the difference and sign, the ability to distinguish between independent abnormal data and continuously changing data is improved. By connecting non-recoverable data that are temporally continuous and spatially point to the same region into evidence segments, and writing the first and last times, cross-sectional segment locations and data sources, the discrete distribution of early abnormal signs can be improved. By generating a fixed transmission order for the evidence segments according to the first data time and the number of data sources, and attaching contour summaries, attitude summaries and operating condition summaries to form an evidence package sequence, the relevance of the content uploaded at the edge and the consistency of the input for subsequent time series analysis are improved. By converting the evidence package sequence into an evidence time series matrix and inputting it into a time series attention network, combined with the corresponding write-back of the actual cross-sectional deviation, the forward correspondence analysis capability for the risk of over-excavation and under-excavation in subsequent tunneling sections is relatively enhanced. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method steps of the present invention.

[0016] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Refer to the instruction manual appendix Figure 1-2 The present invention provides a predictive method for analyzing over- and under-mining data using artificial intelligence, comprising: S1. By collecting the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle, the cross-section point data is divided into multiple cross-section segments according to the same advance interval, and the attitude data and working condition data are mapped to each cross-section segment in chronological order, and the data set indexed by the cross-section segment is output. In this implementation, the section point data, attitude data, and working condition data within the current tunneling cycle are first processed into segments. The purpose of this step is to first use the section point data to determine the actual time and location boundaries covered by each section segment, and then write the attitude data and working condition data into the corresponding section segment accordingly. Data that is not directly written is then supplemented, thereby forming a data set indexed by section segments. After this processing, the data called for subsequent deletion, reconstruction, and comparison all have a clear section segment affiliation, and there will be no problems such as cross-segment mixing, unclear sources, or time sequence breaks. The implementation process includes the following steps: In S1-1, the acquisition time and advance position of each sampling point in the cross-section point data are first read. These sampling points can be directly taken from the point data within the cross-section scan frame. When the cross-section point data is stored by scan frame, each sampling point within the same scan frame shares the acquisition time and advance position corresponding to that scan frame. For each cross-section segment, all sampling points already assigned to that segment are sequentially traversed, and the earliest acquisition time is used as the lower limit of the time, the latest acquisition time as the upper limit of the time, the smallest advance position as the lower limit of the position, and the largest advance position as the lower limit of the position. The position is used as the upper limit of the position, and the lower limit of the time, the upper limit of the time, the lower limit of the position, and the upper limit of the position are written into the spatiotemporal range record corresponding to the cross section. If a cross section corresponds to only one frame of cross section point data, then the lower limit of the time and the upper limit of the time of the cross section are both taken from the acquisition time of that frame, and the lower limit of the position and the upper limit of the position are both taken from the advance position of that frame. If a cross section contains multiple frames of cross section point data, then all sampling points are counted uniformly. In this way, the spatiotemporal range of each cross section is directly formed by the cross section point data itself, without relying on additional manual boundary setting. In S1-2, the acquisition time and advance position of each data point in the attitude data and working condition data are read, and each data point is compared with the spatiotemporal range of each section segment in turn. For any attitude data or working condition data, it is first determined whether its acquisition time is between the lower and upper time limits of a certain section segment, and then whether its advance position is between the lower and upper position limits of that section segment. When both conditions are met, the data is written to that section segment. During writing, the data source, acquisition time, advance position, data value, and writing order of the data are saved. The writing order is generated in ascending order according to the order of entry into the section segment, which is used to identify the preceding and following data. If a data point only meets the spatiotemporal range of one section segment, it is directly written to that section segment. If a data point does not meet the spatiotemporal range of any section segment, it is retained in the unwritten data queue and is not discarded here. After this processing, the intra-segment dataset corresponding to each section segment can be formed. This part of the data has fallen into the corresponding section segment in terms of both time and position. In S1-3, the attitude data and working condition data not yet written to the data queue are continued to be written. During processing, the advance position of the unwritten data is extracted one by one and compared with the lower and upper limits of the position of each section segment. Only the section segments whose advance position is between the lower and upper limits of the position are retained as candidate section segments. Then, for each candidate section segment, the time difference between the acquisition time of the unwritten data and the lower limit of the time, and the time difference between the acquisition time and the upper limit of the time are calculated, and the smaller of the two absolute values ​​of the time difference is taken as the absolute value of the time difference of the unwritten data relative to the candidate section segment. Then, all candidate section segments are sorted from smallest to largest by the absolute value of the time difference, and the candidate section segment ranked first is determined as the writing target. The data is processed by dividing the cross-section into segments and writing the unwritten data to the corresponding segment. If the absolute value of the time difference between two candidate cross-section segments is the same, the difference between the unwritten data and the median time of the two candidate cross-section segments is compared, and the candidate cross-section segment with the smaller difference is determined as the target cross-section segment for writing. If they are still the same, the data is written to the cross-section segment with the earlier number in the cross-section segment number order. After writing, the data is removed from the unwritten data queue, and its data source, acquisition time, advance position, data value, and writing order are written to the corresponding cross-section segment. Through this process, data whose position has fallen into a certain cross-section segment but whose acquisition time is slightly earlier than the lower limit or slightly later than the upper limit of time can be written to the cross-section segment with the closest time, avoiding data dangling. Through the above processing, each cross-section segment can form a data set that is consistent with the location range of the cross-section point data and is connected with the actual collection time sequence, providing a unified data foundation for subsequent supplementation, deletion and verification. Here, the spatiotemporal range is first formed by the cross-section point data, and then intra-segment writing and supplementation are performed respectively. This can unify the data that falls directly inside the cross-section segment and the data that is near the boundary but still belongs to the location range of the cross-section segment into the same cross-section segment. In the subsequent deletion, reconstruction and comparison, the corresponding data can be directly extracted from the cross-section segment and adjacent cross-section segments, avoiding the breakage of the calculation chain due to unclear segmentation. In practical applications: Taking continuous tunnel excavation as an example, the cross-section scanning equipment generates a set of cross-section point data every time it advances a certain distance. The attitude sensor periodically outputs pitch and yaw angle data, and the working condition acquisition unit periodically outputs thrust and rotational speed data. When the cross-section point data corresponding to a certain cross-section segment covers 24.00 meters to 24.15 meters and the acquisition time covers 14:03:10 to 14:03:18, the system first generates the spatiotemporal range of that cross-section segment based on this, and then simultaneously places the acquisition time and advance position into that range. Attitude and working condition data within the spatiotemporal range are directly written into the cross-section segment. For working condition data acquired at 14:03:19 but with an advance position of 24.08 meters, the system continues to calculate the time difference between the data and the lower and upper time limits of the cross-section segment, and then writes it into the cross-section segment. After completion, the cross-section segment forms a data set containing cross-section point data, attitude data, and working condition data. Subsequently, supplementation, deletion, verification, evidence linking, evidence packaging, risk prediction, and sample write-back can be performed directly on this data set.

[0019] S2. Based on each dataset, delete one original data from the current dataset, and call the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. Then, compare the reconstruction result with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data, and the irrecoverable dataset corresponding to each cross-section is output. In this embodiment, the purpose of this processing is to delete one piece of original data one by one after each cross-sectional segment has formed a data set, and then reconstruct the original data using the neighboring data before and after the original data's location, as well as the data in adjacent cross-sectional segments corresponding to its location. Based on the numerical and directional relationships between the reconstruction results and the original data, it is determined whether the original data can still be recovered from the surrounding data after deletion. Here, reconstruction does not rely solely on a single path, but simultaneously generates temporal reconstruction values, inter-segment reconstruction values, and propagation reconstruction values. These are then compared layer by layer using differences and signs to distinguish data that can still be recovered from the surrounding data from data that cannot be recovered after deletion. After this processing, an irrecoverable dataset corresponding to each cross-sectional segment can be formed for subsequent connection to form evidence segments. The implementation process includes the following steps: In S2-1, the current raw data is first extracted from the data set corresponding to each cross-section segment, and the data source, field identifier, acquisition time, advance position, order of appearance, and data value corresponding to the current raw data are read simultaneously. Here, the data source is used to distinguish the specific source type of attitude data and working condition data, the field identifier is used to distinguish different fields under the same source, such as pitch angle, yaw angle, thrust, or speed, and the order of appearance is used to indicate the order of the data within this cross-section segment. Subsequently, the current raw data is deleted from the cross-section segment, and the preceding and following data are searched in the data set after deletion. The preceding data is the data within the same cross-section segment that has the same data source and the same field identifier as the current raw data, and whose appearance order is immediately before the current raw data. The following data is the data within the same cross-section segment that is the same as the current raw data. The preceding original data has the same data source and the same field identifier, and appears immediately after the current original data. Next, the corresponding data is searched in the previous and subsequent cross-sections. The method for finding the corresponding data is as follows: first, the data source and field identifier are consistent with the current original data. Then, the difference in the advance position between the candidate data and the current original data is compared. The data with the highest absolute value of the advance position difference is taken as the corresponding data in the corresponding cross-section. After the above extraction is completed, the current original data, the previous neighbor data, the next neighbor data, the corresponding data in the previous cross-section, and the corresponding data in the subsequent cross-section are written into the same data group to be judged. After this processing, each data group to be judged is formed around a deleted current original data. Subsequent reconstruction and comparison are directly performed based on this group of data, and there will be no problem of unclear value objects. In S2-2, after forming the data group to be judged, three reconstruction values ​​corresponding to the current original data are constructed respectively. During construction, the time-series reconstruction value is first formed using the preceding and following neighbor data. Specifically, the data values ​​of the preceding and following neighbor data are read, and the median value is taken as the time-series reconstruction value according to the positional relationship between the current original data occurrence order and the occurrence order of the preceding and following neighbor data. When the occurrence order of the current original data is exactly between the two, the average value of the preceding and following neighbor data is directly taken as the time-series reconstruction value. Subsequently, the inter-segment reconstruction value is formed using the corresponding data in the previous section segment and the corresponding data in the next section segment. Specifically, the data values ​​of the corresponding data in the previous section segment and the corresponding data in the next section segment are read, and the median value is taken as the time-series reconstruction value according to the positional relationship between the current original data advance position and the corresponding data advance position in the previous and next section segments. The intermediate value is taken as the inter-segment reconstruction value. When the current position is in the middle between the two, the average value of the two corresponding data points before and after is directly taken as the inter-segment reconstruction value. Then, the propagation reconstruction value is constructed. First, the difference between the data value of the preceding neighboring data and the data value of the corresponding data in the preceding section segment is calculated to form the front increment. Then, the difference between the data value of the following neighboring data and the data value of the corresponding data in the following section segment is calculated to form the back increment. Then, the average increment of the front increment and the back increment is taken, and this average increment is superimposed on the average value of the corresponding data in the preceding section segment and the corresponding data in the following section segment to form the propagation reconstruction value. In this way, the temporal reconstruction value reflects the continuous relationship of the preceding and following adjacent data in this section segment, the inter-segment reconstruction value reflects the data continuity relationship of the corresponding positions in adjacent sections, and the propagation reconstruction value reflects the change transmission relationship formed by the adjacent sections and this section segment, thus obtaining the three reconstruction values ​​corresponding to the current original data. In S2-3, after the three reconstructed values ​​are formed, the original value difference, previous neighbor difference, and subsequent neighbor difference between the three reconstructed values ​​and the current original data are calculated respectively. Here, the original value difference is the difference between the reconstructed value and the current original data, the previous neighbor difference is the difference between the reconstructed value and the data value of the previous neighbor data, and the subsequent neighbor difference is the difference between the data value of the subsequent neighbor data and the reconstructed value. After performing the above calculations on the time-series reconstructed value, inter-segment reconstructed value, and propagation reconstructed value respectively, three sets of difference results can be obtained. Then, the positive and negative directions of the original value difference, previous neighbor difference, and subsequent neighbor difference in each set are determined: a difference greater than zero is recorded as positive, a difference less than zero is recorded as negative, and a difference equal to zero is recorded as zero. The results are written as three-digit symbol results in the order of original value difference, previous neighbor difference, and subsequent neighbor difference, thus forming the time-series symbol group, inter-segment symbol group, and propagation symbol group respectively. At the same time, the current original data is read relative to the... Regarding the difference relationship between the preceding and following neighbor data, the specific method is as follows: calculate the difference between the current original data and the data value of the preceding neighbor data, calculate the difference between the data value of the following neighbor data and the current original data, and combine the relationship between the current original data and its own original value to write the original sign group, where the original value bit is uniformly written as zero. After this processing, each reconstructed value not only has a specific numerical difference result, but also a corresponding directional relationship result. For example, if the data value of a certain current original data is 10, the preceding neighbor data is 8, the following neighbor data is 12, and a certain reconstructed value is 9, then the original value difference is -1, the preceding neighbor difference is 1, and the following neighbor difference is 3, and the corresponding sign group is negative, positive, and positive. If the directional relationship formed by the current original data relative to the preceding and following neighbor data is zero, positive, and positive, then it can be directly used to compare bit by bit with the sign group corresponding to each reconstructed value. In S2-4, after the three-way comparison results are formed, it is then determined whether the current original data belongs to recoverable or non-recoverable data. Specifically, the sum of the absolute values ​​of the original difference, the absolute values ​​of the previous neighbor difference, and the absolute values ​​of the subsequent neighbor difference corresponding to each reconstructed value is calculated, and the three reconstructed values ​​are sorted from smallest to largest sum. If the sums are the same, the absolute values ​​of the original differences are compared, with the smaller absolute value ranking first. After sorting, the symbol group corresponding to the first reconstructed value is read and compared bit by bit with the original symbol group. When all three bits match, the current original data is written to the recoverable dataset, indicating that the data can still be recovered from surrounding data after deletion, and subsequent comparisons are not performed. When the symbol group corresponding to the first reconstructed value is inconsistent with any bit of the original symbol group, the symbol group corresponding to the second reconstructed value is read and compared with the original symbol group. The data is compared bit by bit. If the symbol group corresponding to the second reconstructed value is consistent bit by bit, the current original data is written into the recoverable dataset. If there are still inconsistent bits in the symbol group corresponding to the second reconstructed value, the current original data is written into the non-recoverable dataset of its segment. The reason why the final judgment is not based directly on the first reconstructed value is that different reconstruction paths correspond to different data relationships. Although the first result is closer in value, if its directional relationship is inconsistent with the current original data, it still means that the original data cannot be stably recovered from the surrounding data after deletion. Only when at least one main reconstruction path satisfies both numerical similarity and directional consistency is it recorded as recoverable data. Through this process, each segment can form a corresponding non-recoverable dataset, which retains data that cannot be recovered from the relevant data in the current segment and adjacent segments after deletion. Through the above processing, the data that truly has independent evidentiary significance within the current cross-sectional segment can be separated from the general continuously changing data. Here, a data group to be judged is first constructed around the current original data, then three reconstructed values ​​are formed, and then comparisons are made at the difference and sign levels. This avoids making judgments based solely on the proximity of a single value and also avoids mistakenly writing data that should be considered independent anomalies as recoverable data. After this process, the data used to connect and form the evidence segment is data that cannot be recovered even after deletion and reconstruction verification, and is therefore more suitable as candidate evidence for the initiation of over- or under-excavation. In practical applications: Taking the thrust field in a tunnel excavation scenario as an example, a thrust data point within a certain cross-section is selected as the current raw data. Its preceding and following neighbor data are taken from the thrust records at adjacent moments within the same cross-section. The corresponding data in the preceding and following cross-sections are taken from the thrust record with the closest advance position in the adjacent cross-sections. The system first forms a time-series reconstruction value from the preceding and following thrust records of this segment, then forms an inter-segment reconstruction value from the thrust records at corresponding positions in adjacent cross-sections, and finally forms a propagation reconstruction value from the thrust difference between this segment and the adjacent segment. Subsequently, the differences and symbol groups between the three reconstruction values ​​and the current raw data are calculated respectively, and the values ​​are sorted by summation and compared with the original symbol groups in turn. If the symbol groups corresponding to the first two reconstruction values ​​are inconsistent with the original symbol groups, the thrust data point is written into the non-recoverable dataset of the cross-section. Subsequently, position unfolding, evidence connection, and evidence grouping can continue to be performed on the basis of this non-recoverable dataset.

[0020] S3. By connecting the various irretrievable datasets, the irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment are sequentially connected, and the first data time, the last data time, the cross-sectional segment location, and the data source are written into the connection result, and the evidence segment set is output. In this embodiment, the purpose of this processing is to further organize the irretrievable data identified in each cross-section segment into evidence segments that can continuously characterize the initial signs of over-excavation and under-excavation, according to their position distribution and order of appearance on the cross-section. This is not simply a matter of summarizing the irretrievable data within the same cross-section segment, but rather first converting the cross-section coordinates corresponding to each piece of irretrievable data into a unified position number after unfolding along the cross-section contour. Then, based on the acquisition time and data source, it is determined which irretrievable data are continuous in time, continuous in position, and consistent in the unfolding direction. Thus, irretrievable data with the same evolutionary trend are connected end to end into the same evidence segment. After the evidence segment is formed, it is determined whether the evidence segment corresponds to the sidewall area, the arch area, or the invert area based on the cross-section interval into which its start and end positions fall. After this processing, the originally discretely distributed irretrievable data can be transformed into information units with start and end boundaries, source trajectories, and cross-section partitions, which can be directly called upon for subsequent evidence packaging and trend prediction. The implementation process includes the following steps: In S3-1, the non-recoverable datasets corresponding to each cross-section segment are first read, and the acquisition time, cross-section coordinates, and data source corresponding to each non-recoverable data are extracted one by one. The cross-section coordinates here are the actual position coordinates of the non-recoverable data on the cross-section contour. The cross-section point position corresponding to the non-recoverable data can be directly taken, or the mapping point position of the non-recoverable data on the cross-section contour can be taken. The data source is used to distinguish which type of attitude field or working condition field the non-recoverable data comes from. Then, the cross-section coordinates of each non-recoverable data are expanded along the cross-section contour into position numbers. In specific processing, a unified expansion starting point is selected in the current cross-section segment, and the expansion direction is fixed. The expansion starting point can be the contour point corresponding to the junction of the left side wall of the cross-section and the invert arch. The expansion direction can be fixed as a single direction along the cross-section contour from the left side wall to the arch top and then to the right side wall and continue to the invert arch. After determining the expansion starting point and expansion direction, the expansion starts from the expansion starting point along the cross-section. The contour points are sequentially accumulated, with each contour point corresponding to a position number. Each non-recoverable data point is then mapped to the contour point closest to its cross-sectional coordinates, and the position number of that contour point is recorded as the position number of that non-recoverable data point. After generating the position numbers, the acquisition time and position number are written together into the position marker of that non-recoverable data point, thus forming a position marker set. Through this process, the discrete positions on the cross-section, originally represented by two-dimensional or three-dimensional coordinates, are uniformly converted into sequentially comparable position numbers, allowing for direct comparison of the order of adjacent non-recoverable data points on the cross-sectional contour. For example, if the cross-sectional coordinates corresponding to a certain non-recoverable data point are located on the upper left wall, and after unfolding, it falls to the 126th contour point, then its position number is recorded as 126; if the cross-sectional coordinates corresponding to another non-recoverable data point are located near the arch, and after unfolding, it falls to the 208th contour point, then its position number is recorded as 208. In S3-2, after the location marker set is formed, the non-recoverable data within the same cross-sectional segment is joined. During processing, all location markers within the same cross-sectional segment are first sorted from earliest to latest according to the acquisition time. When the acquisition times are the same, they are then sorted from smallest to largest according to the location sequence number. After sorting, the acquisition time difference, location sequence number difference, and data source change of adjacent non-recoverable data are compared sequentially. Here, the acquisition time difference is the acquisition time of the later non-recoverable data minus the acquisition time of the earlier non-recoverable data; the location sequence number difference is the location sequence number of the later non-recoverable data minus the location sequence number of the earlier non-recoverable data; and the data source change is the data source change of the later non-recoverable data. The data source is determined by whether it is the same as or has changed from the previous irretrievable data source. If two adjacent irretrievable data sets satisfy the following relationships, they are connected as adjacent members of the same evidence segment: First, the collection time of the later irretrievable data set is no earlier than that of the previous irretrievable data set, meaning the time order after sorting is not reversed; second, the position number difference remains the same, meaning the position number always increases or decreases during continuous comparisons to indicate a consistent direction of development; third, the current position number difference does not jump back, meaning the position number of the later irretrievable data set does not cross the previously connected intervals; fourth, although the data source is allowed to change, the changed data source follows the... The data sources appearing sequentially are continuously written into the same source sequence without interrupting the connection individually. At the start of the connection, the first non-recoverable data point is recorded as the connection start point, its acquisition time is written as the first data point's time, and its position number is written as the starting position. As the connection continues, the current connection tail is continuously updated. At the end of the connection, the acquisition time of the last non-recoverable data point is written as the last data point's time, the position number of the last non-recoverable data point is written as the end position, and the data sources appearing throughout the connection process are written into the source sequence in the order of their appearance, thus forming an evidence segment. If a non-recoverable data point does not satisfy the above connection relationship with its subsequent non-recoverable data point, then the previous non-recoverable data point is used. The data serves as the endpoint of the current evidence segment, and a new connection is started from the next irretrievable data point. In this way, multiple evidence segments can be formed within the same cross-sectional segment, each corresponding to a set of data that remains continuous in terms of time progression and cross-sectional position changes. For example, if three irretrievable data points appear sequentially according to their collection times, with position numbers of 126, 129, and 133 respectively, their position numbers will continue to increase, and they can be connected end to end to form the same evidence segment. If the position number of the next irretrievable data point suddenly returns to 118, it indicates that the direction of development has changed, and this irretrievable data point will no longer be merged into the previous evidence segment, but will be used as the starting point for the next evidence segment to reconnect. In S3-3, after the evidence segment set is formed, cross-sectional partitioning is performed on each evidence segment. During processing, the start and end positions of each evidence segment are read first, and these are used as the position boundaries of the evidence segment in the cross-sectional outline unfolding sequence. Then, based on the total length of the cross-sectional outline corresponding to the current cross-sectional segment, all position numbers are divided into cross-sectional intervals corresponding to the sidewall area, the arch area, and the invert area. Specifically, the boundaries of each interval can be determined first based on the structural boundary points of the unfolded cross-sectional design outline. For example, the position number intervals corresponding to the left and right walls are jointly classified as the sidewall area, the position number interval corresponding to the top of the cross-section is classified as the arch area, and the position number interval corresponding to the bottom of the cross-section is classified as the invert area. After determining the cross-sectional intervals, it is then determined which cross-sectional interval each evidence segment's start and end positions fall into: when the start and end positions fall into which interval... When the start and end positions both fall within the same cross-sectional interval, the evidence segment is directly written to the corresponding partition. When the start and end positions span two adjacent cross-sectional intervals, the cross-sectional interval with the larger number of covered position numbers is used as the partition marker for that evidence segment. When the number of covered positions is the same, the cross-sectional interval where the position corresponding to the first data time is located is used as the partition marker. When writing the partition marker, the first data time, the last data time, the cross-sectional segment position, and the source sequence are retained simultaneously, thereby outputting a set of evidence segments with the first data time, the last data time, the cross-sectional segment position, the source sequence, and the partition marker. Through this processing, each evidence segment not only has a time range, a location range, and a source trajectory, but also has a clear cross-sectional structure attribution, which can be directly called by sidewall area, arch area, or invert area during subsequent packaging and prediction. Through the above processing, the originally scattered and irretrievable data in various cross-sections are uniformly converted into evidence segments with continuous boundaries and structural attribution. Here, the cross-section coordinates are first expanded into position numbers, then connected according to the acquisition time and position number in sequence, and finally combined with the cross-section intervals to write partition markers. This can further enhance the single-point irretrievable data into evidence units that reflect the continuous expansion process of local anomalies. The evidence segments formed in this way not only retain the first data time, the last data time, the start position, the end position, and the source sequence, but also clarify whether they are located in the side wall area, the arch area, or the invert arch area. Subsequently, local contour data, attitude change data, and working condition change data can be directly extracted from this data and assembled into evidence packages according to a fixed transmission order. In practical applications: Taking the irretrievable data of a certain cross-section segment as an example, the system first maps the irretrievable data corresponding to different positions on the cross-section contour to position numbers. If the acquisition times of one group of irretrievable data are 14:03:12, 14:03:13, and 14:03:14, and the position numbers are 126, 129, and 133, respectively, and the expansion direction remains increasing, then this group of irretrievable data is connected into the same evidence segment, and 14:03:12 is written as the first data time. 03 minutes 14 seconds is written as the last data time, 126 is written as the start position, and 133 is written as the end position. At the same time, the data sources are recorded in the order of their appearance. Then, based on the position interval corresponding to 126 to 133, it is determined that the evidence segment is located in the side wall area and written into the evidence segment set with partition markers. After completion, the evidence segment can be used as a basic unit in the subsequent evidence package, and enter the subsequent processing together with the local contour data within the corresponding position range, the attitude change data and the working condition change data within the corresponding time period.

[0021] S4. Based on the evidence segment set, first sort the evidence segments from earliest to latest according to the time of the first data, then sort the evidence segments with the same first data time from most to least according to the number of data sources, and attach the contour summary, attitude summary and working condition summary of the cross section to each sorted evidence segment, and output the evidence packet sequence arranged in a fixed sending order. In this embodiment, the purpose of this processing is to further organize the already formed evidence segments into a sequence of evidence packets arranged in a fixed transmission order. This ensures that the input received by the subsequent timing attention network is no longer discrete evidence segments, but rather grouped evidence units containing local contour data, attitude change data, and operating condition change data. Instead of simply sorting the evidence segments themselves and sending them directly, a unified transmission order is first generated based on the time of the first data entry, the number of data sources, the duration, and the location span. Then, local contour data, attitude change data, and operating condition change data are extracted around the location and time range corresponding to each evidence segment, and these contents are written into the evidence packet in the same transmission order. After this processing, each evidence packet retains both the spatiotemporal boundaries and source information of the evidence segment itself, and is supplemented with the contour, attitude, and operating condition content directly corresponding to that evidence segment. It can then be directly expanded into an evidence timing matrix in sequence. The implementation process includes the following steps: In S4-1, the first data time, data source, section location, duration, and position span of each evidence segment are extracted, along with the contour data, attitude data, and working condition data corresponding to the section segment. The first data time is directly taken from the acquisition time corresponding to the first non-recoverable data in the evidence segment. The data source is taken from all source items in the source sequence of the evidence segment, and the number of different source items is counted as the number of data sources. The section location is taken from the position number or advance interval identifier of the section segment to which the evidence segment belongs. The duration is the time length obtained by subtracting the first data time from the last data time. The position span is the position length obtained by subtracting the starting position from the ending position. Subsequently, a transmission order is generated for all evidence segments in a unified order. Specifically, the segments are first sorted from earliest to latest based on the first data time. When the first data times of two evidence segments are the same, they are then sorted from most to least data source. When the number of data sources is still the same, they are then sorted from longest to shortest based on the duration. Sort by length from longest to shortest; if the duration is still the same, sort by position span from largest to smallest; if the position span is still the same, arrange by cross-sectional segment position order; after sorting, write the corresponding transmission sequence number for each evidence segment in sequence, and record the evidence segments with transmission sequence numbers as pending evidence segments, thus forming a set of pending evidence segments; through this process, all evidence segments are organized into the same transmission sequence, and when extracting local contour data, attitude change data, and working condition change data later, the transmission sequence is directly used as the standard, and the sorting is not repeated; for example, if the first data time of two evidence segments is 14:03:12 and 14:03:14 respectively, then the evidence segment with the first data time of 14:03:12 is ranked first; if the first data time of two evidence segments is 14:03:12, but one evidence segment contains 3 data sources and the other evidence segment contains 2 data sources, then the evidence segment containing 3 data sources is ranked first. In S4-2, after the transmission order is determined, the local contour data, attitude change data, and working condition change data within the location range of each evidence segment are extracted sequentially according to the transmission order, and these contents are written into the evidence package along with the evidence segment. During processing, the evidence segment to be transmitted first is taken, its start and end positions are read, and the contour point data with position numbers between the start and end positions are extracted from the contour data corresponding to the cross-section segment where the evidence segment to be transmitted is located. The data are then arranged in ascending order of position number as the local contour data corresponding to the evidence segment. If the start and end positions span the contour expansion... For the boundary, data from the starting position to the boundary position is taken first, followed by data from the starting point to the ending position, and then spliced ​​together in the same unfolding direction. Next, the first and last data moments of the evidence segment to be sent are read. All attitude data acquired between the first and last data moments are extracted from the attitude data corresponding to the segment, and arranged from earliest to latest acquisition time to form attitude change data. Then, all working condition data acquired between the first and last data moments are extracted from the working condition data corresponding to the segment, and arranged from earliest to latest acquisition time to form working condition change data. (Complete) After extraction, the pending evidence segment, local contour data, attitude change data, and operating condition change data are written into the same evidence packet with the same transmission sequence number, and this evidence packet is written into the evidence packet sequence. Then, the next pending evidence segment in the transmission sequence is processed until all pending evidence segments are processed, thus outputting an evidence packet sequence arranged in a fixed transmission order. This writing is not a simple concatenation; rather, each evidence packet separately saves its transmission sequence number, evidence segment body, local contour data, attitude change data, and operating condition change data, so that subsequent extraction of the evidence packet sequence can be done directly according to the transmission sequence number. For example, if a certain pending evidence segment... The starting position of the evidence segment is 126, the ending position is 133, the time of the first data is 14:03:12, and the time of the last data is 14:03:14. First, the contour point data with position numbers 126 to 133 is extracted from the contour data of this section as local contour data. Then, the attitude data between 14:03:12 and 14:03:14 is extracted from the attitude data of this section as attitude change data. At the same time, the working condition data within the same time period is extracted from the working condition data of this section as working condition change data. Finally, these contents are written together with the evidence segment to be sent into the evidence packet with the corresponding sending sequence number. Through the above processing, the evidence segments no longer exist only in the form of start and end times, location range, and source sequence, but also carry local contour data, attitude change data, and working condition change data directly corresponding to their location boundaries and time boundaries, forming a sequence of evidence packets that can be sent sequentially. Here, a unified sending order is first generated, and then local contour data within the location range and attitude change data and working condition change data within the time period range are extracted around each evidence segment. This ensures that the information inside each evidence packet is complete, and the arrangement order between different evidence packets is fixed. Subsequently, when constructing the evidence time sequence matrix, it can be directly expanded according to the sending order, without the problem of input order drift or misalignment of associated objects. In practical applications: Taking multiple evidence segments already formed in a certain cross-section as an example, the system first counts the first data time, the number of data sources, the duration, and the location span of each evidence segment, and generates a transmission order accordingly. For example, evidence segments with an earlier first data time and a larger number of data sources are prioritized for writing to the beginning of the transmission sequence. Subsequently, for the evidence segment that is first in the transmission order, the system extracts the contour point data between the start and end positions of the evidence segment from the contour data of the corresponding cross-section, extracts the attitude data between the first and last data times from the corresponding attitude data, and extracts the working condition data within the same time period from the corresponding working condition data. These contents are then written together with the evidence segment into the same evidence package. After all evidence segments are processed, an evidence package sequence arranged in a fixed transmission order is formed. Subsequently, the system can directly expand the timeline queue, generate the evidence time series matrix, and perform over-excavation and under-excavation prediction based on this evidence package sequence.

[0022] S5. By converting the evidence packet sequence into an evidence time series matrix according to the sending order, and inputting the evidence time series matrix into the time series attention network, the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section is extracted, and the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections are output. In this embodiment, the purpose of this processing is to further convert the evidence packet sequence, which has been arranged in a fixed sending order, into an evidence time-series matrix that can be directly processed by the temporal attention network. Based on this matrix, the correlation between each evidence packet along the time progression direction is calculated to obtain the over-excavation and under-excavation prediction results for the corresponding subsequent tunneling sections. Instead of directly sending the evidence packet sequence into the network, each evidence packet is first expanded into a time-axis queue according to the sending order. Then, the evidence segments, local contour data, attitude change data, and working condition change data are written into the corresponding time positions. Position codes, source codes, and direction codes are added to each time position so that the evidence content at different time positions has formed a unified arrangement and unified field caliber before being input into the temporal attention network. After completing this conversion, the temporal attention network calculates the evidence correlation values, contour correlation values, attitude correlation values, and working condition correlation values ​​between each time position and the subsequent time positions. These correlation results are accumulated and mapped according to the time position and cross-section position, and finally output the over-excavation and under-excavation prediction results for the corresponding subsequent tunneling sections. The implementation process includes the following steps: In S5-1, the evidence packet sequence is first expanded into a timeline queue according to the sending order. During processing, each evidence packet is read sequentially from smallest to largest according to its sending sequence number, and each sending sequence number corresponds to a time bit. When the sending sequence numbers are consecutive, each evidence packet is written into consecutive time bits. When there is a gap between the sending sequence numbers, the time bit order is preserved according to the original value of the sending sequence number, and the empty bits in the middle are not recompressed. Subsequently, the evidence segment, local contour data, attitude change data, and working condition change data in each evidence packet are extracted and written into the corresponding time bits according to the sending order. Among them, the evidence segment part must at least write the first data time, the last data time, the section position, the start position, the end position, the source sequence, and the partition mark. The local contour data is written sequentially within the location range of the evidence segment according to the location number; the attitude change data is written sequentially between the first and last data acquisition times; the working condition change data is written sequentially between the first and last data acquisition times; after the basic data is written, a position code, source code, and direction code are generated for the same time position; the position code is generated based on the evidence segment position, specifically: the position of the cross-section segment to which the evidence segment belongs within the time position, as well as the start and end positions, are read, and the cross-section segment position is written as the segment code, and the start and end positions are written as the interval code, which together form the position code; the source code is generated based on the source... Sequence generation specifically involves: writing different data sources in the source sequence according to a pre-fixed source number order, and forming source codes according to the order of their appearance; direction codes are generated based on the direction of change, specifically: first, the direction of contour change is determined by the difference between the beginning and end of local contour data, then the direction of attitude change is determined by the difference between the beginning and end of attitude change data, and finally the direction of working condition change is determined by the difference between the beginning and end of working condition change data. These three are then written into direction codes in sequence, where a difference between the beginning and end of the data greater than zero indicates upward movement, a difference between the beginning and end of the data less than zero indicates downward movement, and a difference between the beginning and end of the data equal to zero indicates parallel movement; after completing the above writing, the data content corresponding to each time point is arranged according to a unified field order to form an evidence time series matrix; after this processing, each time point corresponds to a... The complete evidence package contains the evidence segment body, local contour data, attitude change data, working condition change data, as well as position code, source code, and direction code within each time bit. It can be directly used as input for the timing attention network. For example, if the first evidence package in the transmission sequence corresponds to the first data time of 14:03:12, the section position is 24.00 meters to 24.15 meters, and the position range is 126 to 133, then this evidence package is written into time bit 1. The corresponding local contour data, attitude change data, and working condition change data are synchronously written into time bit 1. Then, the position code is generated from the section position and position range, the source code is generated from the source sequence, and the direction code is generated from the difference between the first and last values ​​of the three types of data. In S5-2, after the evidence time series matrix is ​​formed, it is input into the time series attention network to calculate the correlation between each time point and its subsequent time points, and output the over-excavation and under-excavation prediction results for the corresponding subsequent tunneling sections. Specifically, the evidence time series matrix is ​​input into the time series attention network sequentially according to the time points, and the network extracts the evidence features, contour features, attitude features, and working condition features corresponding to each time point. The evidence features are formed by the evidence segment body, location encoding, and source encoding; the contour features are formed by local contour data; the attitude features are formed by attitude change data; and the working condition features are formed by working condition change data. Subsequently, for any current time point, its correlation with subsequent time points is calculated. Four types of correlation values ​​exist between time points: Evidence correlation value is generated from the similarity relationship between evidence features at the current time point and evidence features at subsequent time points; contour correlation value is generated from the similarity relationship between contour features at the current time point and contour features at subsequent time points; attitude correlation value is generated from the similarity relationship between attitude features at the current time point and attitude features at subsequent time points; and working condition correlation value is generated from the similarity relationship between working condition features at the current time point and working condition features at subsequent time points. The calculation of each type of correlation value is completed within the same network inference process, and the output is the correlation result between the corresponding time points. After completing the pairwise correlation calculations for all time points, the correlation values ​​are then accumulated according to the time point and cross-sectional segment position. The procedure is as follows: First, the evidence association values, contour association values, attitude association values, and working condition association values ​​related to the same subsequent cross-section location are merged into the same location. Then, the merged results are accumulated sequentially according to time position to form the comprehensive association result corresponding to the location of the subsequent cross-section. Next, the comprehensive association results of all cross-section locations within the same subsequent tunneling section are written into the subsequent tunneling section in the order of cross-section location to form the over-excavation and under-excavation prediction results for the corresponding subsequent tunneling section. The over-excavation and under-excavation prediction results here include at least the prediction deviation value and prediction deviation direction corresponding to the location of each cross-section within the subsequent tunneling section. The prediction deviation direction is output as the over-excavation direction, under-excavation direction, or zero deviation direction according to the direction corresponding to the comprehensive association result. Through this processing, the forward relationship between the earlier evidence segments in the evidence package sequence and the contour changes, attitude changes, and working condition changes in subsequent time positions is uniformly written into the corresponding position of the subsequent tunneling section. Subsequently, the prediction result and the actual cross-section deviation data can be directly called for corresponding back-writing. For example, if the evidence package corresponding to time position 1 forms a high evidence correlation value and working condition correlation value with time position 3 and time position 4 respectively, and these two time positions jointly correspond to the same cross-section position in the subsequent tunneling section, then the system will accumulate the correlation results from time position 1 to time position 3 and time position 4 at the cross-section position, and use the accumulated result as the over-excavation and under-excavation prediction output for the subsequent cross-section position. Through the above processing, the evidence package sequence is further transformed into an evidence time-series matrix with a unified time position structure and a unified coding caliber. The time-series attention network then completes the calculation of multiple correlations between consecutive time positions, enabling the forward connections between discrete evidence packages to be mapped to the over-excavation and under-excavation prediction results on subsequent tunneling sections. Here, the time axis queue is first expanded, and then evidence segments, local contour data, attitude change data, and working condition change data are written in, and position codes, source codes, and direction codes are added to ensure that the input fields entering the time-series attention network are complete and consistent. Then, by calculating the evidence correlation value, contour correlation value, attitude correlation value, and working condition correlation value separately, and accumulating and mapping them according to time position and cross-section position, the common influence of multi-source evidence on subsequent tunneling sections can be uniformly converged into the same prediction result. In practical applications: Taking continuous tunneling as an example, the system first writes the already formed evidence packet sequence into consecutive time bits such as time bit 1, time bit 2, and time bit 3 in the order of transmission. Then, within each time bit, it writes the evidence segment body, the local contour data within the corresponding location range, the attitude change data and working condition change data within the corresponding time period, and simultaneously generates the position code, source code, and direction code. Subsequently, the formed evidence timing matrix is ​​input into the timing attention network, which calculates the time bit 1 and time bit 2, time bit 1 and time bit 3, and time bit 4 respectively. The evidence correlation values, contour correlation values, attitude correlation values, and working condition correlation values ​​between position 2 and time position 3 are then accumulated into the corresponding subsequent tunneling sections according to the cross-section position. When a certain subsequent cross-section position shows an outward bias in multiple sets of forward correlation results, the system outputs the position as an over-excavation direction prediction. When multiple sets of forward correlation results show an inward bias, the system outputs the position as an under-excavation direction prediction, thus forming the over-excavation and under-excavation prediction results for the corresponding subsequent tunneling sections, which are then used to write back the corresponding deviation data from the actual cross-section.

[0023] S6. After the excavation of the corresponding subsequent tunneling section is completed, the actual cross-sectional deviation data of the subsequent tunneling section is collected, and the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results are written back according to the cross-sectional section position, and an evidence sample set with actual deviation markers is output. In this embodiment, this part of the processing is used to match the actual cross-sectional deviation data with the previously obtained over- and under-excavation prediction results according to the cross-sectional segment position after the excavation of the subsequent tunneling section is completed. Then, the corresponding deviation difference and direction comparison results are written back to the evidence segment to form an evidence record with actual deviation markings, and further organized into an evidence sample set. After this processing, the evidence chain formed in the previous step can be closed with the subsequent actual excavation results, which is convenient for subsequent verification and sample accumulation. The implementation process includes the following steps: In S6-1, the actual cross-sectional deviation data and over- and under-excavation prediction results for the corresponding subsequent tunneling sections are first extracted, and then matched segment by segment according to their positions. The actual cross-sectional deviation data is obtained by comparing the measured cross-section after the excavation of the subsequent tunneling section with the design cross-section. The over- and under-excavation prediction results are taken from the prediction results for the corresponding subsequent tunneling sections output earlier. During the matching process, the cross-sectional position in the actual cross-sectional deviation data is read first, and then the cross-sectional position in the prediction results is read. Records with the same position number or the same advance interval are paired as the same corresponding result. If the same cross-sectional position exists... If there are multiple actual cross-sectional deviation data, the average deviation value is taken as the actual deviation value for that cross-sectional segment. If there are multiple prediction results for the same cross-sectional segment, the average predicted deviation value is taken as the predicted deviation value for that cross-sectional segment. Subsequently, the deviation direction is determined based on the predicted deviation value and the actual deviation value, where a deviation value greater than zero is recorded as the over-excavation direction, a deviation value less than zero is recorded as the under-excavation direction, and a deviation value equal to zero is recorded as the zero deviation direction. Then, the cross-sectional segment location, predicted deviation value, actual deviation value, and deviation direction are written into the same corresponding result to form a corresponding record set. In S6-2, after the corresponding record set is formed, the difference calculation and direction comparison are performed on the predicted deviation value and the actual deviation value in each corresponding record, and the results are written back to the corresponding evidence segment. During processing, the deviation difference is first calculated by subtracting the predicted deviation value from the actual deviation value for each corresponding record. When the deviation difference is greater than zero, it means that the actual deviation is greater than the predicted deviation; when the deviation difference is less than zero, it means that the actual deviation is less than the predicted deviation; when the deviation difference is equal to zero, it means that the actual deviation is consistent with the predicted deviation. Subsequently, the predicted deviation direction and the actual deviation direction are compared one by one. When they are the same, it is recorded as consistent direction; when they are different, it is recorded as inconsistent direction. After the difference calculation and direction comparison are completed, the deviation difference and direction comparison results are written back to the corresponding evidence segment according to the cross-section segment position. During the writing back, the predicted deviation value and the actual deviation value are written at the same time, thus forming an evidence record with the actual deviation mark. If the same cross-section segment position corresponds to multiple evidence segments, the same corresponding record is written back to all evidence segments under that cross-section segment position, and the evidence record set with the actual deviation mark is output. In S6-3, after the evidence record set is formed, each evidence record is written into the sample storage area and aggregated according to the cross-sectional segment position to form an evidence sample set with actual deviation markers. During processing, all evidence records are first grouped according to the cross-sectional segment position, and then sorted from earliest to latest according to the first data time within each group. When the first data time is the same, they are sorted according to the data source order. After sorting, each evidence record is written into the sample storage area in sequence. When writing, each evidence record must include at least the cross-sectional segment position, the first data time, the last data time, the start position, the end position, the source sequence, the partition marker, the predicted deviation value, the actual deviation value, the deviation difference, and the direction comparison result. After writing is completed, the evidence records in the sample storage area are aggregated by cross-sectional segment position. All evidence records belonging to the same cross-sectional segment position are grouped together to form the evidence sample corresponding to that cross-sectional segment position, and further summarized into an evidence sample set with actual deviation markers. Through the above processing, a segment-by-segment correspondence is formed between the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results. The deviation difference and direction comparison results are written back to the corresponding evidence segment, thereby forming an evidence sample set that can be directly called. In this way, the original time boundary, location boundary, source sequence and partition mark of the evidence segment are preserved, and the actual deviation value, deviation difference and direction comparison results are added, so that the preceding evidence chain and the subsequent real results form a closure. In practical applications: After a subsequent tunneling section is excavated, the system first compares the measured cross-section with the designed cross-section to obtain the actual deviation value of each cross-section position. Then, it pairs these actual deviation values ​​with the previously output predicted deviation values ​​according to the cross-section position. For example, after pairing the predicted deviation value of 0.05 meters for a certain cross-section position with the actual deviation value of 0.08 meters, the deviation difference is calculated to be 0.03 meters, and it is determined that the predicted direction is consistent with the actual direction. Subsequently, these results are written back to all evidence segments corresponding to that cross-section position. Finally, they are written into the sample storage area in the order of cross-section position, first data time, and data source, and aggregated to form the evidence sample set for the corresponding subsequent tunneling section.

[0024] Furthermore, it also includes a predictive system that uses artificial intelligence to analyze over- and under-mining data, including: The data segmentation module is used to collect the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle. It divides the cross-section point data into multiple cross-section segments according to the same advance interval, and maps the attitude data and working condition data to each cross-section segment in chronological order, and outputs a data set indexed by the cross-section segment. The deletion and verification module deletes one original data from the current data set based on each data set. It then calls the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. The reconstruction result is then compared with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data and the irrecoverable dataset corresponding to each cross-section is output. The evidence connection module is used to connect various irretrievable datasets. It sequentially connects irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment, and writes the first data time, the last data time, the cross-sectional segment location, and the data source into the connection result, and outputs a set of evidence segments. The evidence arrangement module, based on the evidence segment set, first sorts the evidence segments from earliest to latest according to the time of the first data, then sorts the evidence segments with the same first data time from most to least according to the number of data sources, and adds the contour summary, attitude summary and working condition summary of the cross section segment to each sorted evidence segment, and outputs a sequence of evidence packets arranged in a fixed sending order. The risk prediction module is used to convert the evidence packet sequence into an evidence time series matrix according to the sending order, and input the evidence time series matrix into the time series attention network to extract the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section, and output the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections. The sample write-back module is used to collect actual cross-sectional deviation data of the subsequent tunneling section after the excavation of the corresponding section is completed. It then writes back the actual cross-sectional deviation data and the over- and under-excavation prediction results according to the cross-sectional section location, outputting an evidence sample set marked with the actual deviation. Working Principle: This scheme focuses on over- and under-excavation prediction under edge data transmission conditions. First, cross-section point data, attitude data, and working condition data are collected at the edge side within a tunneling cycle, and organized into corresponding data sets according to cross-section segments. Then, each data point is reconstructed after deletion, and a three-way reconstruction result is formed using adjacent data before and after the current segment, corresponding data from adjacent cross-section segments, and the propagation relationship between them. Data that cannot be stably recovered after deletion is filtered out. This unrecoverable data is then connected into evidence segments according to cross-section location and time sequence, and combined with local contour data within the corresponding location range, and corresponding... The attitude change data and working condition change data within a time period are combined into evidence packets. Edge data transmission is performed according to a fixed transmission order. After forming an evidence packet sequence, it is sent into a time-series attention network to output the over-excavation and under-excavation prediction results of subsequent tunneling sections. After the actual excavation of subsequent sections is completed, the measured cross-sectional deviation is matched with the prediction results segment by segment and written back to the previous evidence segments to form evidence samples with real result labels. Overall, the key to this scheme is not just to make predictions, but to identify and prioritize the retention of early abnormal evidence that cannot be recovered later under the condition of limited edge data transmission, and then use this evidence to complete forward prediction. For example, in a continuous tunnel excavation scenario, edge devices near the tunnel face continuously collect cross-sectional contours, equipment attitude, and propulsion conditions. However, the bandwidth of the on-site link is limited, and the raw data is not suitable for long-term full-volume uploading. In this case, this solution first completes cross-sectional segmentation, censoring and reconstruction, and identification of irretrievable data at the edge side. Data that is more likely to represent the initial signs of over- or under-excavation are connected into evidence segments. Then, the evidence segments, along with local contour data, attitude change data, and working condition change data, are combined into evidence packages. The edge data is transmitted in a fixed order and sent to the subsequent prediction link. In this way, what the system uploads is not just a pile of ordinary raw data, but key evidence that has been screened and organized at the edge side. After the section ahead continues to be excavated, the measured cross-sectional results are compared with the previous prediction results segment by segment and written back. This verifies which early signs uploaded by the edge side have actually developed into over- or under-excavation, thus achieving a closed loop between edge data transmission, key evidence preservation, and over- or under-excavation prediction.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A prediction method for analyzing overbreak data using artificial intelligence, characterized by, include: S1. By collecting the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle, the cross-section point data is divided into multiple cross-section segments according to the same advance interval, and the attitude data and working condition data are mapped to each cross-section segment in chronological order, and the data set indexed by the cross-section segment is output. S2. Based on each dataset, delete one original data from the current dataset, and call the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. Then, compare the reconstruction result with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data, and the irrecoverable dataset corresponding to each cross-section is output. S3. By connecting the various irretrievable datasets, the irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment are sequentially connected, and the first data time, the last data time, the cross-sectional segment location, and the data source are written into the connection result, and the evidence segment set is output. S4. Based on the evidence segment set, first sort the evidence segments from earliest to latest according to the time of the first data, then sort the evidence segments with the same first data time from most to least according to the number of data sources, and attach the contour summary, attitude summary and working condition summary of the cross section to each sorted evidence segment, and output the evidence packet sequence arranged in a fixed sending order. S5. By converting the evidence packet sequence into an evidence time series matrix according to the sending order, and inputting the evidence time series matrix into the time series attention network, the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section is extracted, and the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections are output. S6. After the excavation of the corresponding subsequent tunneling section is completed, the actual cross-sectional deviation data of the subsequent tunneling section is collected, and the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results are written back according to the cross-sectional section position, and an evidence sample set with actual deviation markers is output. 2.The prediction method for analyzing overbreak data by using artificial intelligence according to claim 1, wherein: S1 includes: S1-1. Read the acquisition time and advance position of each sampling point in the cross-section point data, and calculate the lower limit of time, upper limit of time, lower limit of position and upper limit of position for each cross-section segment, and output the spatiotemporal range corresponding to each cross-section segment. S1-2. Read the acquisition time and advance position of each data in the attitude data and working condition data. Write the attitude data and working condition data whose acquisition time is between the lower limit and the upper limit of the time and whose advance position is between the lower limit and the upper limit of the position into the corresponding section segment to form the intra-segment dataset corresponding to each section segment. S1-3. Read the attitude data and working condition data of the dataset not written to the segment. For each piece of unwritten data, calculate the time difference between its acquisition time and the lower limit of the time of each segment and the time difference between its acquisition time and the upper limit of the time of each segment. Write the unwritten data whose advance position is between the lower limit and the upper limit of the position and whose absolute value of the time difference is first into the corresponding segment. Output the data set corresponding to each segment. 3.The method of claim 2, wherein the method further comprises: S2 includes: S2-1. Read the data set corresponding to each cross section, extract the current raw data one by one, and read the data source, field identifier, acquisition time, advance position, order of occurrence and data value corresponding to the current raw data. Then delete the current raw data from the cross section where it is located, and extract the previous and next neighbor data of the current raw data in this cross section, the same position data in the previous cross section and the same position data in the next cross section, and output the data group to be judged. S2-2. Based on the data group to be judged, construct the time-series reconstruction value according to the preceding and following neighbor data, construct the inter-segment reconstruction value according to the co-position data in the preceding segment and the co-position data in the following segment, form the front increment according to the preceding and following neighbor data and the co-position data in the preceding segment, and form the rear increment according to the following and following neighbor data and the co-position data in the following segment. Then, construct the propagation reconstruction value together with the front increment and the rear increment, and output the three-way reconstruction value corresponding to the current original data.

4. The prediction method for analyzing overbreak data using artificial intelligence according to claim 3, characterized in that: S2 further includes: S2-3. Based on the three reconstructed values, calculate the original value difference, previous neighbor difference, and subsequent neighbor difference between the time-series reconstructed value, inter-segment reconstructed value, and propagation reconstructed value and the current original data, respectively. The original value difference is the difference between the reconstructed value and the data value of the current original data, the previous neighbor difference is the difference between the reconstructed value and the data value of the previous neighbor data, and the subsequent neighbor difference is the difference between the data value of the subsequent neighbor data and the reconstructed value. Then, take the positive and negative results of the original value difference, the previous neighbor difference, and the subsequent neighbor difference to form the time-series symbol group, the inter-segment symbol group, and the propagation symbol group, respectively. Take the corresponding positive and negative results of the current original data relative to the previous neighbor data and the subsequent neighbor data to form the original symbol group, and output the three-way comparison results. S2-4. Based on the three-way comparison results, first sort the original value difference, the previous neighbor difference, and the next neighbor difference in ascending order of absolute value. Then, read the symbol group corresponding to the first reconstructed value and compare it with the original symbol group bit by bit. When they match bit by bit, write the current original data into the compensable dataset. When they do not match bit by bit, continue to read the symbol group corresponding to the second reconstructed value and compare it with the original symbol group bit by bit. When the second reconstructed value still does not match bit by bit, write the current original data into the non-compensable dataset of the section segment and output the non-compensable dataset corresponding to each section segment.

5. The prediction method for analyzing overbreak data using artificial intelligence according to claim 4, characterized in that: S3 includes: S3-1. Read the non-recoverable datasets corresponding to each cross section, extract the acquisition time, cross section coordinates and data source for each non-recoverable data, expand the cross section coordinates along the cross section outline into position numbers, generate position markers corresponding to each non-recoverable data according to the acquisition time and position number, and output the position marker set. S3-2. Based on the location marker set, compare the acquisition time difference, location sequence difference, and data source change of two adjacent non-recoverable data in the same cross section. Connect the non-recoverable data with continuous acquisition time difference, continuous location sequence, and consistent expansion direction to form the same evidence segment. Write the connection start point as the first data time and start position, and the connection end point as the last data time and end position. Write the data sources in the connection process into the source sequence according to the order of appearance, and output the evidence segment set. S3-3. Based on the evidence segment set, the starting and ending values ​​of the position number of each evidence segment are counted, and the evidence segments are written into the side wall area, arch area or invert arch area according to the cross section interval where the starting and ending values ​​are located. The evidence segment set with the first data time, the last data time, the cross section position, the source sequence and the partition mark is output.

6. The prediction method for analyzing overbreak data using artificial intelligence according to claim 5, characterized in that: S4 includes: S4-1. Extract the first data time, data source, section location, duration, and location span of each evidence segment, as well as the contour data, attitude data, and working condition data corresponding to the section segment. Generate the sending order by sorting the first data time in ascending order, the number of data sources in descending order, the duration in descending order, and the location span in descending order, and output the set of evidence segments to be sent. S4-2. Extract the local contour data, attitude change data and operating condition change data within the location range of each evidence segment in the order of transmission. Write the evidence segment, local contour data, attitude change data and operating condition change data into the evidence packet in the order of transmission. Output the evidence packet sequence arranged in a fixed transmission order.

7. The prediction method for analyzing overbreak data using artificial intelligence according to claim 6, characterized in that: S5 includes: S5-1. Expand the evidence package sequence into a timeline queue according to the sending order, extract the evidence segment, local contour data, attitude change data and working condition change data from each evidence package, write them into the corresponding time position according to the sending order, and generate position code, source code and direction code according to the position of the evidence segment, data source and change direction within the same time position, and output the evidence time sequence matrix. S5-2. Input the evidence time series matrix into the time series attention network, calculate the evidence association value, contour association value, attitude association value and working condition association value between each time position and each subsequent time position, then accumulate each association value according to the time position and cross section position and map it to the corresponding subsequent tunneling section, and output the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling section. 8.The prediction method for analyzing overbreak data by using artificial intelligence according to claim 7, wherein: S6 includes: S6-1. Extract the actual cross-sectional deviation data and over- and under-excavation prediction results for the corresponding subsequent tunneling sections. Match the actual cross-sectional deviation data with the over- and under-excavation prediction results segment by segment according to the cross-sectional section position. Write the cross-sectional section position, predicted deviation value, actual deviation value and deviation direction into each corresponding result and output the corresponding record set. S6-2. Subtract the predicted deviation value from the actual deviation value in each corresponding record to form the deviation difference value. Compare the predicted deviation direction with the actual deviation direction in each corresponding record to form the direction comparison result. Then write the deviation difference value and the direction comparison result back to the corresponding evidence segment according to the cross-section segment position and output the evidence record set with the actual deviation mark. S6-3. Write the evidence record set into the sample storage area in the order of cross-sectional segment position, first data time, and data source, and aggregate them according to cross-sectional segment position to form an evidence sample set with actual deviation markers.

9. A prediction system for analyzing overbreak data using artificial intelligence, comprising: include: The data segmentation module is used to collect the cross-section point data, attitude data and working condition data corresponding to the current tunneling cycle. It divides the cross-section point data into multiple cross-section segments according to the same advance interval, and maps the attitude data and working condition data to each cross-section segment in chronological order, and outputs a data set indexed by the cross-section segment. The deletion and verification module deletes one original data from the current data set based on each data set. It then calls the same type of data from the previous and next cross-sections, as well as the remaining data in the current cross-section, to reconstruct the deleted original data. The reconstruction result is then compared with the deleted original data item by item according to the data content and the order of appearance. If there are items that cannot be recovered, the deleted original data is recorded as irrecoverable data and the irrecoverable dataset corresponding to each cross-section is output. The evidence connection module is used to connect various irretrievable datasets. It sequentially connects irretrievable data that are temporally continuous and spatially point to the same sidewall, arch, or invert within the same cross-sectional segment, and writes the first data time, the last data time, the cross-sectional segment location, and the data source into the connection result, and outputs a set of evidence segments. The evidence arrangement module, based on the evidence segment set, first sorts the evidence segments from earliest to latest according to the time of the first data, then sorts the evidence segments with the same first data time from most to least according to the number of data sources, and adds the contour summary, attitude summary and working condition summary of the cross section segment to each sorted evidence segment, and outputs a sequence of evidence packets arranged in a fixed sending order. The risk prediction module is used to convert the evidence packet sequence into an evidence time series matrix according to the sending order, and input the evidence time series matrix into the time series attention network to extract the forward correspondence between each evidence segment and the over-excavation risk or under-excavation risk of the subsequent section, and output the over-excavation and under-excavation prediction results of the corresponding subsequent tunneling sections. The sample write-back module is used to collect the actual cross-sectional deviation data of the subsequent tunneling section after the excavation of the corresponding subsequent tunneling section is completed, and write back the actual cross-sectional deviation data and the over-excavation and under-excavation prediction results according to the cross-sectional section position, and output the evidence sample set with actual deviation markers.

Images