CN-122022770-A - Causal constraint type maintenance decision method for multi-granularity knowledge aided modeling

CN122022770ACN 122022770 ACN122022770 ACN 122022770ACN-122022770-A

Abstract

The invention discloses a causal constraint type maintenance decision method for multi-granularity knowledge aided modeling, which comprises the following steps of step 1 of multi-granularity knowledge extraction, step 2 of multi-granularity knowledge analysis and internalization mechanism, step 3 of causal constraint two-channel gating fusion network and step 4 of multi-index collaborative hierarchical maintenance decision. The method is mainly used for carrying out deep analysis based on road surface multisource detection data, environment and traffic load data, maintenance standard texts and expert experiences to construct a maintenance knowledge map with causal quantitative weights, realizing deep fusion of knowledge logic and data characterization by using a designed double-channel gating fusion network, executing accurate prediction of road surface performance and identification of high-risk road sections, and finally generating a closed-loop maintenance scheme considering risks, costs and scene benefits by relying on multi-index collaborative hierarchical maintenance decision logic. The invention is especially suitable for the fine maintenance planning and budget optimizing distribution scene of the high-grade highway asphalt pavement.

Inventors

HAO JUN
PEI LILI
XING ZHENZHEN
LI NINGXIN
ZHU YANGYANG
Wu yexing
LI LE

Assignees

长安大学

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (9)

1. The causal constraint maintenance decision method for multi-granularity knowledge aided modeling is characterized by comprising the following steps of: Step 1, multi-granularity knowledge extraction Extracting core information containing explicit constraint and implicit rule to form an original knowledge data set; step 2, multi-granularity knowledge analysis and internalization mechanism (2A) Extracting a hard index through semantic analysis, and taking the hard index as direct mapping of explicit constraint; (2b) Clustering analysis is carried out by utilizing a potential dirichlet allocation model, so that hidden rules are mined; (2c) The extracted rules are unified into a computable form by adopting a traditional fuzzy reasoning method, the (2 a) and (2 b) are unified into the computable form, the semantic ambiguity is solved by membership function, and the unified rules are constructed ; (2D) Mapping the input/output variables in (2 c) to entity nodes, mapping logical associations to edges, traversing all rules Establishing a basic knowledge graph topology; (2e) Introducing a structural causal model, and calculating the average causal effect of a processing variable on a result variable through the influence of the tendency score matching balance environment confusion variable; (2f) Mapping the calculated average causal effect value to the [0,1] interval to generate normalized weight of the map edge The knowledge graph construction with causal logic and numerical strength is completed; Step 3, causal constraint two-channel gating fusion network (3A) Computing dynamic weights using knowledge-aware gating According to the time sequence state characteristics of the real-time pavement Adjusting a predefined rule set The injection proportion of the knowledge rule is realized; (3b) The causality of the causality model mining is used as a priori weight by utilizing causality attention gating, traffic quantity and weather exogenous variables are weighted, and non-causality interference noise is restrained; (3c) Synthesizing the outputs of the steps (3 a) and (3 b), generating a cross-modal fusion characteristic with knowledge constraint and data characterization capability through characteristic splicing and weighted fusion, and calculating the final cross-modal fusion characteristic ; (3D) Fusing the features obtained in (3 c) Converting into a time sequence, inputting the time sequence into a prediction layer, and directly outputting PCI, RQI, RDI pavement performance prediction values in the future 1 year; Step 4, grading maintenance decision of multi-index cooperation (4A) Setting basic thresholds as PCI is more than or equal to 85, RQI is more than or equal to 80 and RDI is more than or equal to 80 as preventive maintenance thresholds based on the prediction result in the step (3 d), and dividing the maintenance range into the repairable maintenance range if the basic thresholds are lower than any threshold; (4b) Calculating specific risk sub-items for each index by combining key influence factors such as month average axle load Li, annual rainfall Ri, annual average temperature Ti and the like, and finally, constructing a risk index by weighting and summarizing weights determined by an analytic hierarchy process; (4c) Performing descending order arrangement according to the risk index calculated in the step (4 b), screening out 30% -40% of high-risk road sections before ranking, and entering a candidate maintenance pool; (4d) Adopting a causal effect-based matching strategy to identify a risk sub-item with the highest proportion in the road section, and combining the causal effect value quantified by the structural causal model in the step (2 f) to perform matching; (4e) Performing secondary optimization on the general measures generated in the step (4 d) aiming at rainy and high-load special region scenes; (4f) Constructing a multi-layer perceptron neural network model, inputting 16-dimensional features including current PCI/RQI/RDI values (3-dimensional), predictive performance values (3-dimensional), traffic load levels (1-dimensional), environmental climate factors (2-dimensional), measure cost normalization values (1-dimensional), measure history success rates (1-dimensional) and 5-dimensional causal association features of road sections, and calculating the adoption probability of each measure Screening out schemes with infeasible technology or poor expected benefit by setting probability threshold; (4g) Calculating a priority score Pi of the measure obtained by the verification of (4 f); (4h) Adding the curing sets S one by one from high to low according to the Pi score obtained in (4 g) according to budget constraint until And stopping, summarizing the finally selected road segments, and outputting a final maintenance decision scheme comprising the pile numbers of the specific road segments, recommended maintenance measure types (M1-M3), construction priorities and estimated cost.
2. The causal constraint maintenance decision method of multi-granularity knowledge aided modeling according to claim 1, wherein in the step 1, core information containing explicit constraints and implicit rules is extracted from maintenance specifications, local rules, engineering files, expert interview records and maintenance cases to form an original knowledge data set.
3. The causal constraint maintenance decision method of multi-granularity knowledge aided modeling according to claim 1, wherein in the step (2 c), semantic ambiguity is resolved by membership function, and a unified rule is constructed: Wherein, the Is an input variable that is used to determine the state of the object, Is a fuzzy set of the values of the set, Is an output variable Is the corresponding weight.
4. The causal constraint maintenance decision method of multi-granularity knowledge-aided modeling of claim 1, wherein in step (3 c), features are fused across modes Calculated by the following formula: Wherein, the The feature vectors are embedded for the atlas, Is the exogenous variable characteristic vector of traffic load and meteorological environment, Is a time sequence state characteristic vector of the historical performance data of the road surface, The dynamic weight obtained by knowledge sensing gating calculation in the step (3 a) is used for adjusting the knowledge injection proportion; a causal weight calculated for causal attention gating in step (3 b); Is that Is a mean value of (c).
5. The causal constraint maintenance decision method of multi-granularity knowledge-aided modeling according to claim 1, wherein in the step (4 b), pci=0.4, rqi=0.3, rdi=0.4, and the risk index calculation formula is constructed as follows: wherein, each index risk sub-item is calculated as follows: PCI risk sub-item Reflecting the road surface damage risk, the calculation formula is: Wherein, the For the future 1-year PCI prediction value, For the annual rainfall level, For curing history factors, let 3 years without curing=3, curing 1 time=2, curing 2 times or more=1; RQI risk sub-items Reflecting the degradation risk of flatness, the calculation formula is Wherein, the For the RQI predictor of the next 1 year, The average axle load grade is the month; RDI risk sub-items Reflecting the rut development risk, the calculation formula is: Wherein, the For the RDI prediction value of the next 1 year, Is an annual average temperature grade.
6. The causal constraint maintenance decision method of multi-granularity knowledge-aided modeling of claim 1, wherein in said step (4 d), when a road section And if the ratio is highest, the ACE value of the measure M2 to the PCI is more than or equal to 0.75, the crack sealing M2 is recommended preferentially.
7. The causal constraint type maintenance decision method based on multi-granularity knowledge aided modeling according to claim 1 is characterized in that in the step (4 e), in a rainy scene, a general measure M2 is updated to be M2 'with better waterproofness, in a high-load scene, M1 is replaced by M3 with stronger rut resistance, and finally, a scene measure set S' is generated by each road section.
8. The causal constraint maintenance decision method of multi-granularity knowledge aided modeling according to claim 1, wherein in the step (4 g), a priority score Pi is calculated, and a calculation formula is as follows: Wherein, the For the calculated link coupling risk index, For the cost of a single measure, Representative measures And road section The matching degree of the dominant risk.
9. The causal constraint maintenance decision method of multi-granularity knowledge aided modeling according to claim 1, wherein in the step (4 h), measure costs of the finally selected road segments are summarized, and the formula is as follows: Wherein, the The final selected set of hundred meters of piles, Represents the first The measure cost of the hundred-meter pile is that, For the batch discount coefficient, when the number of covered road sections of the single maintenance type measures exceeds n, I.e. enjoy % Batch discount.

Description

Causal constraint type maintenance decision method for multi-granularity knowledge aided modeling Technical Field The invention relates to the field of traffic infrastructure operation and maintenance management and intelligent maintenance, in particular to a causal constraint maintenance decision method for multi-granularity knowledge aided modeling. Background In the case of aging of global traffic infrastructure, the core difficulty is that the degradation of road surface performance is affected by multi-factor nonlinear coupling of climate, traffic, materials, etc., and has strong uncertainty. The existing asphalt pavement maintenance decision method mainly extends around two core paths, namely a pure experience driving method and a pure data driving method, and the two methods have emphasis in technical logic and practical application. The essence of the method is that the long-term accumulated expert experience, engineering practice result and maintenance technical specifications specified by plaintext are converted into static decision rules and threshold standards. In practical application, the method compares the field detection data with a preset standard threshold value, and when certain index data reaches or exceeds a threshold value range, a corresponding fixed maintenance scheme is triggered. Hu A, bai Q, chen L et al in the literature 'A review on empirical methods of pavement performance modeling' systematically comb various empirical methods for modeling the pavement performance, summarize the advantages and disadvantages of different methods in the application scene and the prediction precision, and provide theoretical references for the subsequent model selection. The core logic of the pure data driving method depends on the potential rule of data mining, and the method abandons the preset physical model and fixed expert rules and focuses on the value mining of massive historical data. Complex correlations and statistical laws between inputs and outputs are automatically mined from massive historical data (e.g., road performance, traffic volume, maintenance records), and predictions and decisions are made based on these patterns. Han C, ma T, chen S et al in literature "Asphalt pavement maintenance plans intelligent decision model based on reinforcement learning algorithm" propose a reinforcement learning-based asphalt pavement maintenance plan for subjective issues of traditional maintenance plan decisions. The intelligent decision model realizes the dynamic adaptation of maintenance measures and road section performance states by constructing a reward function for quantifying maintenance benefits, and provides a new idea for autonomous iterative optimization of a maintenance scheme. However, the existing method still has the following defects in terms of intelligent maintenance decision: (1) The problem is that the pavement performance prediction method faces the specific extension of technical defects, and when a model cannot analyze professional expressions such as longitudinal cracks of more than or equal to 5mm and needing milling in a detection report, even if the model has accurate detection data, a decision system still has difficulty in automatically generating a disposal scheme; (2) In the knowledge level, most of expert experience is implicit knowledge, and the lack of unified expression standards makes the expert experience difficult to formalize, so that popularization and application of a knowledge driving model are limited; (3) In the benefit level, the existing decision system can only rely on static cost accounting, is difficult to adapt to the full life cycle benefit balance requirement under the scenes of novel material application, traffic flow increase and the like, and has structural imbalance of short-term cost and long-term performance, static planning and dynamic requirement. Therefore, the method for realizing the end-to-end intelligent decision from pavement performance evaluation to maintenance scheme generation through multi-granularity knowledge expression, causal double-channel gating and multi-index collaborative hierarchical maintenance decision is a technical problem to be solved by the technicians in the field. Disclosure of Invention In view of the above, the invention provides a causal constraint type maintenance decision method for multi-granularity knowledge aided modeling, which converts unstructured domain knowledge into a maintenance causal knowledge map by adopting a multi-granularity knowledge analysis mechanism, designs a causal constraint two-channel gating fusion network, and simultaneously fuses a multi-index hierarchical maintenance decision mechanism, thereby solving the problems of static knowledge structure, cross-modal data fusion semantic separation and redundancy interference, decision process rupture and low resource release accuracy in the existing method, finally realizing closed loop intelligent decision from risk identification