An engineering project intelligent task reminding method based on deep learning

Abstract

The present application relates to a kind of engineering project intelligent task reminding method based on deep learning, belong to engineering project information management technical field.The method includes: obtaining engineering project whole life cycle business data, generates engineering task uniform feature table;Constitute space-time double-flow dynamic heterogeneous hypergraph network, generate initial process delay state value;With time-varying process delay propagation flow input space-time double-flow dynamic heterogeneous hypergraph network, execute delay conduction analysis and correlation intensity weighted fusion, output time-varying delay influence degree data;With the task node exceeding preset threshold as risk source, extract high-risk propagation path, carry out buffer duration flexible adjustment, output linkage duration delay estimation;With the engineering task uniform feature table of risk source node carries out feature matching, generates pre-position risk intervention decision, and adaptively configures and pushes hierarchical post-linkage decision chain.The dynamic accurate identification and differentiation of delay risk are realized, and the scientificity of engineering management decision is improved.

Description

Technical Field

[0001] This invention belongs to the field of information management technology for engineering projects, specifically relating to a deep learning-based intelligent task reminder method for engineering projects. Background Technology

[0002] This invention belongs to the field of information management technology for engineering projects, specifically relating to a deep learning-based intelligent task reminder method for engineering projects.

[0003] Intelligent task reminders for engineering projects are a key technology in engineering management. The core challenge lies in accurately predicting the propagation path and impact of process delays, and accordingly pushing targeted intervention suggestions and reminder strategies to the responsible parties. Currently, engineering management practices mainly rely on manual experience to assess process delay risks and issue notifications at each level, which suffers from problems such as untimely warnings, highly subjective assessments of the propagation scope, and a disconnect between reminder strategies and task topology. To address these issues, those skilled in the art have attempted to introduce technologies such as graph networks or digital twins for schedule risk prediction and early warning.

[0004] For example, in the area of ​​graph network-based project progress management and risk prediction, Chinese patent application publication number CN120355373A discloses a construction project full-process monitoring system based on digital intelligence fusion. This system integrates a graph convolutional network (GCN) model and a regularized graph inference model. By collecting and fusing multi-source heterogeneous data, it achieves component status prediction, construction delay scoring, and risk level identification, and generates high-risk node warnings and construction path suggestions based on graph neural networks. However, this existing technology has exposed obvious shortcomings in practical engineering applications: it uses a common graph neural network, which can only establish binary relationships between nodes, making it difficult to simultaneously characterize high-order multivariate relationships formed between processes due to overlapping resource occupation time and overlapping work areas; at the same time, the scheme mainly relies on the score value output by the graph inference model to calculate the signal strength of delay propagation, lacking a quantitative assessment of the project's own schedule flexibility and the ability to absorb on-site resource redundancy, resulting in insufficient accuracy in assessing the degree of delay impact, and its warning method is not differentiated according to the depth of the task's impact in the project topology, making it difficult to achieve accurate hierarchical post-warning.

[0005] Furthermore, in the area of ​​construction progress monitoring based on Building Information Modeling (BIM) technology, Chinese patent application publication number CN115860499A discloses a multi-disciplinary coordinated construction management system based on BIM technology. This system simulates the construction process through a BIM information model and uses a progress monitoring module to monitor and remind workers of the construction progress in real time. When the progress deviation exceeds a preset threshold, a reminder is issued and the construction plan is dynamically adjusted. However, this existing technology also has significant limitations: it essentially relies on a static topology driven by the BIM model, and its reminder triggering mechanism is based on a simple judgment of threshold comparison. It does not consider the complex propagation paths formed by resource competition and process dependence between processes, nor does it establish an adaptive mapping relationship between the cascading depth of task nodes in the engineering topology and the push strategy. This results in a rigid reminder strategy that is difficult to adapt to the multi-hop linkage propagation scenario of process delays in large and complex projects.

[0006] In summary, existing technologies still have significant shortcomings in areas such as high-order correlation modeling of process delay propagation, dynamic quantitative assessment of the risk resistance of task nodes, and generation of differentiated reminder strategies based on task topology correlation status. These shortcomings have become key technical bottlenecks restricting the accuracy and practicality of intelligent task reminders in engineering projects. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a deep learning-based intelligent task reminder method for engineering projects. The objective of this invention can be achieved through the following technical solution: Acquire business data throughout the entire lifecycle of engineering projects, extract semantic association features by modality through a pre-trained language model, and fuse them to generate a unified feature table for engineering tasks; construct a spatiotemporal dual-stream dynamic heterogeneous hypergraph network containing process-dependent hyperedges and resource-competitive hyperedges to generate initial process delay state values; The initial process delay status value is input into the spatiotemporal dual-stream dynamic heterogeneous hypergraph network, and delay transmission analysis and correlation strength weighted fusion are performed to generate delay impact value. Combined with the task schedule elastic correction delay impact value, the time-varying delay impact data of each task is output. Taking the task node whose time-varying delay impact data exceeds the preset threshold as the risk source node, perform adaptive risk gradient traversal to extract high-risk propagation paths, perform buffer period flexible adjustment on the high-risk propagation paths, and output the linked project delay prediction. Feature matching is performed on the unified feature table of engineering tasks of risk source nodes to generate pre-risk intervention decisions. Based on the estimated delay of the linked project period and the topological cascading depth of the tasks in the network, hierarchical post-linkage decision chains are adaptively configured and pushed.

[0008] Specifically, the process of generating the unified feature table for engineering tasks includes: The business data of the entire life cycle of the project is divided into text-based process data, tabular progress data, and IoT time-series data. The semantic association features in the text process data are extracted using a pre-trained language model, the topological node features in the table progress data are extracted according to the row and column structure, and the on-site working condition time series features in the IoT time series data are extracted through a one-dimensional convolutional layer. The cross-modal association weights between the semantic association features, topological node features and on-site working condition temporal features are calculated using a cross-attention mechanism, and projection alignment is performed to generate a unified feature table for engineering tasks in a unified high-dimensional feature space.

[0009] Specifically, the formation process of the spatiotemporal dual-stream dynamic heterogeneous hypergraph network includes: Extract the overlapping duration of different engineering tasks within the same time period and calculate the proportion of their respective planned total duration as the temporal representation of the resource competition superedge. Combine this with the proportion of the intersection of the work areas of different engineering tasks at the same construction site to form a joint spatiotemporal constraint degree. Calculate the domain difference between the planned completion time of the preceding task and the planned start time of the subsequent task, and establish the flow constraint operator of the process-dependent hyperedge. The temporal representation and the joint spatiotemporal constraint are fused using multidimensional spatiotemporal features, and together with the flow constraint operator, they are written into the topological matrix of the heterogeneous hyperedge to form the spatiotemporal dual-flow dynamic heterogeneous hypergraph network.

[0010] Specifically, the specific process of the delay propagation analysis includes: The initial process delay state value is used as the initial energy operator and mapped to the source node of the spatiotemporal dual-stream dynamic heterogeneous hypergraph network; The spatiotemporal flow efficiency of adjacent task nodes in the process-dependent hyperedge is extracted to calculate the dynamic transmission rate, and the spatiotemporal overlap coefficient of the corresponding task nodes in the resource-competitive hyperedge is extracted to calculate the dynamic diffusion probability. By combining the dynamic transmission rate and the dynamic diffusion probability, the temporal transmission process of delay impact in the internal topology path of the network is dynamically simulated to generate delay impact values ​​for each task node.

[0011] Specifically, the process of the correlation strength weighted fusion includes: When the impact of delays propagates along the topological path, the topological adjacency relationship and the degree of association between each task node and its corresponding hyperedge are extracted. Calculate the topological mapping weights between hyperedges and nodes for the topological adjacency relationship and the association tightness, and construct the corresponding association strength weighted matrix; The delay impact values ​​propagated to each hyperedge are back-aggregated and updated to the corresponding central task node according to the topological weights using the correlation strength weight matrix, thus completing the structured weighted fusion of the delay impact values ​​in the network topology.

[0012] Specifically, the delay impact value correction process includes: Real-time acquisition of the remaining compressible elasticity coefficient of the engineering entity corresponding to each engineering task, as well as the real-time available redundant resource ratio corresponding to the current construction stage; The compressible elasticity coefficient of the remaining construction period and the proportion of real-time available redundant resources are solved by linear elastic combination to obtain the independent local absorption damping coefficient of each engineering task node. By applying the local absorption damping coefficient to the weighted and fused delay impact value, the time-varying delay impact data of each engineering task is obtained.

[0013] Specifically, the process of risk gradient traversal and high-risk propagation path extraction includes: Taking the task node whose time-varying delay impact data exceeds a preset threshold as the root node, calculate the risk transfer gradient value between adjacent nodes along the topological direction of the process-dependent hyperedge and the resource-competition hyperedge; A preset gradient traversal step size and flow direction threshold are used to sequentially filter adjacent task nodes whose risk transfer gradient values ​​are greater than the preset threshold and extend them in a chain until the risk transfer gradient values ​​of all adjacent nodes are lower than the preset threshold. The node sequence generated by the chain extension is then used as the high-risk propagation path.

[0014] Specifically, the process of flexible adjustment correction for the buffer period includes: With the goal of minimizing the overall project delay due to high-risk propagation paths, a linear programming model is constructed with the comprehensive scheduling cost as the objective function, the criticality level and real-time elasticity coefficient of each affected engineering task as the weight coefficients of the objective function, and the limit supply load of limited on-site construction resources as the boundary constraint. A multi-dimensional resource scheduling cost table is constructed by calculating the physical span of resource scheduling between each node and the supply-demand gap value, and the multi-dimensional resource scheduling cost table is used as the input parameter of the linear programming model. The optimal resource scheduling scheme is obtained by optimizing the linear programming model using a linear programming algorithm, and the start time and duration of the delay of each affected engineering task are adjusted accordingly, and the estimated delay of the coordinated project period is output.

[0015] Specifically, the process of generating the pre-emptive risk intervention decision includes: Calculate the Euclidean distance between the current unified feature table of the engineering task of the risk source node and each historical feature vector in the pre-stored historical intervention case library; Retrieve and extract the target historical case with the shortest Euclidean distance from the historical intervention case database; The system retrieves the pre-intervention texts recorded in the target historical cases, performs semantic parsing and adaptation, generates pre-risk intervention decisions that include suggestions for supplementing construction resources and deadlines, and pushes them to the corresponding project responsible parties.

[0016] Specifically, the generation process of the hierarchical post-linkage decision chain includes: The proportion of the estimated delay in the coordinated construction period to the total planned construction period of the corresponding project task is obtained, and a preset multi-level warning threshold band is matched to determine the warning level. The warning level includes at least two levels, and each warning level corresponds to a preset delay percentage range. The frequency of reminders and the adaptive configuration of push channels are combined with the topological cascading of the corresponding task nodes in the network. A low-frequency single-channel push strategy is configured for task nodes with a cascading depth greater than a preset depth, and a high-frequency multi-channel overlay push strategy is configured for task nodes with a cascading depth less than a preset depth. The output includes a hierarchical post-linkage decision chain with different push sequences, and is pushed to the terminal devices of the corresponding project responsible entities. The low-frequency single-channel push strategy sends reminders through a single push channel according to a first preset reminder cycle; the high-frequency multi-channel overlay push strategy sends reminders through a combination of multiple push channels according to a second preset reminder cycle shorter than the first preset reminder cycle.

[0017] The beneficial effects of this invention are as follows: (1) This invention generates a unified feature table by extracting full lifecycle data of a project through modal analysis. Based on dynamic elements such as the overlap of process sequence and on-site resource occupation time, and the intersection of work areas, process-dependent hyperedges and resource-competition hyperedges are constructed respectively, forming a spatiotemporal dual-stream dynamic heterogeneous hypergraph network. This hypergraph network can comprehensively capture the high-order multi-element relationships formed by resource competition among multiple processes in the two-dimensional space of "time domain-spatial domain". At the same time, by performing impact transmission analysis and using the association strength weighting matrix to perform structured weighted fusion in the network topology, it can dynamically simulate the temporal evolution process of delays in the hypergraph topology path. Compared with the prior art, this invention can accurately extract high-risk propagation paths with cascading reactions, solving the technical problem of difficulty in quantifying and tracking multi-hop delay propagation in complex engineering.

[0018] (2) This invention constructs a local absorption damping coefficient for each task node by solving the compressible elastic coefficient of the remaining construction period of the engineering entity and the proportion of available redundant resources in the construction stage in real time. It performs scale reduction on the delay data after topology weighted fusion, thereby accurately stripping and deducting the project's own risk resistance attenuation and effectively eliminating the artificially high delay impact value. On this basis, this invention performs flexible adjustment of the buffer construction period through a multi-dimensional resource scheduling cost table and optimization search algorithm, and collaboratively corrects the delay time and duration to provide an optimal resource scheduling scheme. Finally, this invention breaks the rigid mode of traditional single threshold judgment, deeply binds the estimated proportion of linkage delay with the topology cascading depth of the corresponding task node in the hypergraph network, adaptively configures the reminder frequency and push channel, and performs differentiated strategy control for deep nodes and shallow nodes. This enables the adaptive generation of hierarchical decision chains for pre-risk and post-linkage and accurate push, improving the practical value and decision guidance significance of intelligent task reminders. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating an intelligent task reminder method for engineering projects based on deep learning, according to the present invention. Figure 2 This is a schematic diagram illustrating the process of generating a unified feature table for engineering tasks and dynamically constructing a hypergraph in the method of this invention. Figure 3 This is a flowchart of the delay propagation simulation and damping correction in the method of the present invention; Figure 4 This is a flowchart of the adaptive risk traversal and flexible adjustment in the method of this invention; Figure 5 This is a flowchart of the generation and push of the hierarchical decision chain in the method of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0022] This embodiment combines Figure 1 The overall implementation process of the intelligent task reminder method for engineering projects based on deep learning of the present invention is described. Figure 1 The overall technical route of the present invention is shown. The steps are arranged in a manner that involves data transmission at each level, model analysis at each level, and decision optimization at each step, forming a complete risk analysis and intelligent reminder process, which realizes dynamic perception of construction risks in engineering projects, risk propagation analysis, and intelligent scheduling decisions.

[0023] Specifically, step S1 is first executed to acquire multi-source heterogeneous data throughout the entire construction process of the project, including construction logs, construction organization design, technical briefing documents, WBS task decomposition data, construction schedule plan, resource allocation plan, and on-site IoT monitoring data. A pre-trained language model is then used to extract construction semantic features, which are combined with construction schedule features and on-site time-series monitoring features to complete multi-modal feature fusion and generate a unified feature table for the project tasks. Based on this, a spatiotemporal dual-stream dynamic heterogeneous hypergraph network is established according to the process dependencies and resource competition relationships between construction tasks. Simultaneously, the weights of resource competition hyperedges and process dependency hyperedges are calculated, and initial delay status values ​​corresponding to each construction task are generated based on the deviation between planned and actual progress, providing a unified data foundation for subsequent risk propagation analysis.

[0024] Subsequently, step S2 is executed. Using the dynamic hypergraph and initial delay state value as input, the risk propagation process between construction tasks is calculated through a delay propagation dynamics model. A weighted fusion mechanism based on correlation strength is used to comprehensively consider two types of risk propagation paths: process dependency propagation and resource competition propagation. Simultaneously, based on the construction period flexibility, resource redundancy, and construction environment conditions, an absorption damping coefficient is calculated to adaptively correct the propagated delay state, thereby obtaining a corrected delay state value that accurately reflects the operational status of the construction site.

[0025] Next, step S3 is executed. Using the corrected delay status value as input, the risk transfer gradient between construction tasks is calculated based on the dynamic hypergraph. An adaptive risk threshold is established by combining the criticality of each process, and a risk gradient traversal analysis is performed to identify critical risk propagation paths. After identifying the risk propagation paths, the system establishes a flexible scheduling optimization model by comprehensively considering the costs of construction schedule adjustments and resource scheduling, generating corresponding pre-emptive risk intervention schemes and post-emptive coordinated scheduling schemes. Pre-emptive risk intervention refers to construction adjustment measures implemented before the risk propagates to subsequent construction tasks, while post-emptive coordinated scheduling refers to resource coordination adjustment measures implemented for affected construction tasks after the risk has propagated, thereby reducing the overall construction delay risk.

[0026] Finally, step S4 is executed, using the risk propagation path and scheduling optimization results as input to construct the feature vector of the current project case, and performing similarity matching with the historical project case library to select the historical case closest to the current project as an auxiliary decision-making basis; when there is no case in the historical case library that meets the matching conditions, the flexible scheduling optimization result is adopted as the default scheduling scheme for the current project, and the historical case library is automatically updated after the project is completed, so as to realize the continuous expansion and dynamic optimization of the case library, thereby continuously improving the accuracy of subsequent project risk analysis and task reminders.

[0027] Through the above steps, this invention forms a complete technical closed loop based on unified engineering task modeling, dynamic hypergraph risk propagation analysis as the core, flexible scheduling optimization as the decision-making means, and continuous learning of historical cases as the feedback mechanism. This achieves dynamic identification, intelligent analysis, and full-process management of construction risks in engineering projects, including task reminders. Specifically, step S1 corresponds to Embodiment 1, focusing on the unified engineering task modeling and spatiotemporal dual-stream dynamic heterogeneous hypergraph construction process; step S2 corresponds to Embodiment 2, focusing on delay propagation analysis and correlation strength weighted fusion process; step S3 corresponds to Embodiment 3, focusing on risk gradient traversal analysis and flexible scheduling optimization process; and step S4 corresponds to Embodiment 4, focusing on historical case matching and continuous decision optimization process. These embodiments are interconnected, jointly completing the overall implementation process of this invention.

[0028] Example 1: Construction of a unified feature table for engineering tasks and establishment of a spatiotemporal dual-stream dynamic heterogeneous hypergraph network This embodiment combines Figure 2 This paper details the construction of a unified feature table for engineering tasks and the establishment process of a spatiotemporal dual-stream dynamic heterogeneous hypergraph network. This embodiment uses a construction project for the main structure of an urban rail transit station as an application scenario. The project includes multiple construction tasks such as foundation pit excavation, rebar tying, formwork installation, concrete pouring, and support removal. There are strict sequential dependencies between these construction tasks, while they share large construction machinery, temporary materials, work surfaces, and labor resources. This fully reflects the complex engineering characteristics of multi-process collaborative construction, where process coupling and resource competition coexist.

[0029] After the project starts, the system acquires project operation data through the project management platform interface and on-site IoT data acquisition terminals. This includes textual process data such as construction logs, construction organization design, and technical briefing documents; planning management data such as WBS task breakdown table, construction schedule plan, and resource allocation plan; and on-site operation data such as equipment operating status, personnel location information, material inventory information, and environmental monitoring information. In this embodiment, the system acquires project operation data through the project management platform interface and on-site IoT data acquisition terminals. Exemplary methods for acquiring various types of data include: text / structured data such as construction logs, construction organization designs, technical disclosure documents, WBS task breakdown data, construction schedule plans, and resource allocation plans, which can be retrieved from the structured database of the project's BIM collaborative management platform or the enterprise's ERP system, or obtained through standardized API interfaces connecting to the enterprise's historical project archives; IoT time-series data such as tower crane operation status, personnel location information, environmental monitoring information, and material inventory information can be collected through IoT data acquisition terminals deployed on the construction site. Specifically, large equipment operation data is acquired through the equipment's built-in controller interface, personnel location data is acquired through UWB base stations, and environmental data is acquired through on-site sensors. The acquisition frequency can be configured according to the project monitoring needs. After collection, the data is uploaded to edge computing nodes for preprocessing via the MQTT protocol. The above acquisition methods or their equivalent implementations can be flexibly selected based on the actual digital construction situation of the project.

[0030] To achieve a unified representation of multi-source heterogeneous data, the system performs preprocessing on different types of data. For text-based process data, a pre-trained language model is used for semantic encoding to extract semantic features such as construction objects, construction techniques, construction areas, construction conditions, and construction resources. For planning management data, a task index is established based on WBS coding to extract topological node features such as planned start time, planned completion time, planned duration, responsible work team, and relationships with preceding processes. For on-site operation data, time synchronization, outlier removal, and statistical analysis are performed according to a unified time window to extract time-series operation features such as equipment utilization rate, resource occupancy rate, construction intensity, and environmental conditions.

[0031] Preferably, this embodiment employs supervised fine-tuning to train the pre-trained language model. Training samples are derived from historical project construction logs, construction organization design documents, and technical briefing documents. The training objective is semantic classification of construction tasks and recognition of construction entities. The loss function used is the cross-entropy loss function. Training stops when the validation set loss falls below a preset convergence threshold for multiple consecutive training cycles. After model training is complete, the trained model is deployed at the construction site, where only forward inference is performed to generate text semantic features; no further online training is conducted to meet the real-time operational requirements of the construction site.

[0032] Subsequently, the system employs a cross-attention mechanism to complete multimodal feature fusion. Specifically, textual semantic features are used as the query vector, planning management features as the key vector, and on-site operation features as the value vector. Scaled dot product attention is used to calculate the correlation between different modal features, and multimodal feature fusion is completed based on normalized attention weights, ultimately generating a unified feature table for each construction task. This unified feature table serves as the data input for subsequent hypergraph construction and risk analysis.

[0033] After completing the construction of the unified feature table for engineering tasks, the system establishes a spatiotemporal dual-stream dynamic heterogeneous hypergraph network. Unlike traditional graph models, which can only describe pairwise relationships between nodes, the spatiotemporal dual-stream dynamic heterogeneous hypergraph constructed in this embodiment allows multiple engineering task nodes to be connected by the same hyperedge simultaneously. This is used to describe the higher-order relationships formed by shared construction resources among multiple construction tasks, thereby fully expressing the process coupling relationship and resource competition relationship in complex construction scenarios. Its mathematical expression is shown in formula (1): (1) in, V(t) represents the spatiotemporal dual-stream dynamic heterogeneous hypergraph corresponding to time (t); V(t) represents the set of engineering task nodes at time t, with each node corresponding to a construction task; E p (t) represents the set of process-dependent hyperedges at time t, used to describe the sequence constraints between multiple construction tasks; E r W(t) represents the set of resource competition hyperedges at time t, used to describe the high-order correlation formed by multiple construction tasks competing for limited construction resources; W(t) represents the set of weights corresponding to each hyperedge at time t, used to reflect the degree of influence of different hyperedges on risk propagation.

[0034] To describe the dependency relationship between engineering task nodes and hyperedges, the system further establishes a hypergraph association matrix, the mathematical expression of which is shown in formula (2): (2) Among them, H ij (t) represents the element in the i-th row and j-th column of the hypergraph correlation matrix at time t; v i Represents the i-th project task node; e j This represents the j-th superedge. When node v i Belongs to hyperedge e j When the condition is met, the matrix element takes a value of 1; otherwise, it takes a value of 0. The correlation matrix is ​​used to establish the mapping relationship between engineering task nodes and hyperedges, and serves as the data basis for subsequent delay propagation analysis.

[0035] Subsequently, the system calculates the resource competition superedge weights based on the degree of overlap in construction time, the degree of overlap in construction operation areas, and the degree of scarcity of construction resources. The resource competition superedge weights are calculated according to formula (3): (3) Among them, w r (t) represents the resource contention superedge weight at time t; T r (t) represents the proportion of overlapping construction times among multiple construction tasks; S r (t) represents the overlap ratio of construction work areas among multiple construction tasks; Q r (t) represents the degree of scarcity of construction resources, which is calculated based on the ratio between the available resources and the required resources; α, β, and γ represent the weighting coefficients of the time competition factor, spatial competition factor, and resource scarcity factor, respectively, satisfying α+β+γ=1. All of the above weighting coefficients are configurable parameters, which can be adjusted by technical personnel according to the different project priorities regarding schedule, safety, and cost.

[0036] For process-dependent hyperedges, the system calculates the corresponding hyperedge weights based on the process relaxation between adjacent construction tasks, and its mathematical expression is shown in formula (4): (4) Among them, w p (t) represents the process-dependent hyperedge weight; t represents the process slack between the planned completion time of the preceding construction task and the planned start time of the subsequent construction task; λ represents the time decay coefficient, used to adjust the influence of the process slack on the degree of process dependence. The smaller the process slack, the greater the corresponding process dependence weight, indicating that the subsequent process is more easily affected by the delay of the preceding process.

[0037] To reflect the dynamic characteristics of the hypergraph network, this embodiment adopts a dynamic update mechanism that combines event triggering and rolling cycle correction. When events such as the addition of a construction task, completion of a construction task, adjustment of the construction plan, change of key resource status, or abnormal on-site monitoring status are detected, the system immediately triggers a hypergraph update; if no triggering event is detected, a topology correction is performed according to a preset rolling update cycle to ensure that the hypergraph always remains consistent with the construction site status. The hypergraph update process is shown in formula (5): (5) in, This represents the updated dynamic hypergraph; V represents the change in construction task nodes, including newly added nodes and deleted completed nodes; E represents the change in the superedge, including new process dependencies and new resource competition relationships; W represents the change in hyperedge weight; f(·) represents the hypergraph update operator, which regenerates node, hyperedge, and weight information based on the triggering event, thereby achieving dynamic topology update.

[0038] After the hypergraph is constructed, the system calculates the initial delay status value corresponding to each construction task based on the deviation between the planned completion progress and the actual completion progress. Its mathematical expression is shown in formula (6): (6) Where E0(t) represents the initial delay state value corresponding to each construction task at time (t); P plan (t) represents the planned progress; P actual (t) represents the actual progress. When the actual progress is higher than the planned progress, the system automatically corrects the delay status value to 0 to ensure that only delayed construction tasks participate in subsequent risk propagation analysis.

[0039] Through the above steps, the system completed the construction of a unified feature table for engineering tasks, the establishment of a spatiotemporal dual-stream dynamic heterogeneous hypergraph, the calculation of resource competition hyperedges and process dependency hyperedge weights, and generated initial delay state values ​​corresponding to each construction task. The dynamic hypergraph provides the risk propagation topology, and the initial delay state values ​​provide the initial input for risk propagation. Together, they constitute the basic data for delay propagation analysis and correlation strength weighted fusion calculation in Example 2, realizing unified modeling of the topological relationships of engineering tasks and the operational status of the construction site. This provides complete data support for subsequent risk propagation analysis, adaptive risk gradient traversal, and flexible scheduling optimization.

[0040] Example 2: Delay propagation analysis and correlation strength weighted fusion process This embodiment combines Figure 3 This paper details the process of analyzing the propagation of engineering task delays, weighted fusion of correlation strength, and adaptive absorption damping correction based on a spatiotemporal dual-stream dynamic heterogeneous hypergraph. Using the unified feature table of engineering tasks, the spatiotemporal dual-stream dynamic heterogeneous hypergraph, and the initial delay state values ​​of each construction task constructed in Example 1 as input data, this paper dynamically analyzes the risk propagation relationship between construction tasks and generates a damped-corrected delay state. This provides a data foundation for the risk gradient traversal and flexible scheduling optimization performed in Example 3.

[0041] The application scenario is a construction project of the main structure of a city rail transit station. When a critical equipment failure, a shortage of critical construction resources, or a delay in a previous process occurs at the construction site, the system first identifies the delay source node based on the initial delay state value obtained in Implementation Example 1 as the initial energy operator, and then determines the associated nodes that have process dependencies and resource competition relationships with the delay source node based on the dynamic hypergraph, thus establishing the delay propagation calculation range.

[0042] To describe the propagation process of delay status in the dynamic hypergraph, this embodiment uses a delay propagation dynamics model to calculate the delay status of each construction task node, and the calculation method is shown in formula (7): (7) in, : Represents the rate of change of the delay state corresponding to node i over time; E i (t): Represents the delay state value of node i at time t; λ: Represents the delay attenuation coefficient, used to describe the natural attenuation of delay caused by the node's own recovery capability; N dep : Represents the set of nodes that have a process dependency relationship with node i; V pass : Indicates the rate at which process delays are transmitted; N comp : Represents the set of nodes that have resource competition with node i; P diff : Indicates the probability of risk diffusion caused by resource competition.

[0043] The above formula takes into account both process-dependent propagation and resource competition propagation paths, enabling project delays to propagate synchronously along the construction process chain and resource competition chain, which is more in line with the risk evolution law in complex engineering construction.

[0044] To facilitate on-site calculations, this embodiment further employs a discrete-time iterative method to numerically solve the above dynamic model, as shown in formula (8): (8) Among them, E i (t+Δt): Represents the delay state value of node i at the next calculation time; Δt: Represents the propagation time step, which can be configured according to the monitoring frequency at the engineering site; The meanings of the remaining parameters are the same as those in formula (7).

[0045] Discrete iterative calculations can transform the continuous propagation process into an iterative update process suitable for real-time calculations in engineering projects, making it easier for the system to continuously update the delay status of each construction task in conjunction with real-time on-site monitoring data.

[0046] Since the same construction task is usually affected by multiple adjacent construction tasks at the same time, this embodiment further calculates the association strength weight based on the hypergraph association relationship and completes the weighted fusion of multiple propagation paths. The calculation method is shown in formula (9): (9) in, M ij (t): Represents the association strength weight between node j and node i; H ij (t): Represents the corresponding element of the hypergraph association matrix established in Example 1; c ij This represents the degree of association between node i and node j, which is calculated based on the degree of process dependence, the degree of resource competition, and the frequency of historical collaborative construction. : Represents the number of hyperedges associated with node i, used to complete the association weight normalization process.

[0047] The correlation tightness comprehensively reflects the actual coupling degree between construction tasks. When multiple construction tasks share key construction resources and have close process connections, their correlation tightness automatically increases, and the corresponding propagation weight increases accordingly; conversely, when the correlation between construction tasks decreases, their propagation impact automatically weakens, thereby achieving dynamic adjustment of the risk propagation intensity.

[0048] After completing the correlation strength weighted fusion, in order to avoid the risk of some nodes at the construction site continuing to propagate due to their strong recovery ability, this embodiment further introduces an adaptive absorption damping correction mechanism to correct the delay state. The calculation method is shown in formula (10): (10) in, : Represents the delay state value of node i after damping correction; E i (t): Represents the delay state value of the corresponding node before damping correction; γ i : Represents the absorption damping coefficient of node i.

[0049] The damping coefficient is calculated based on the construction period flexibility coefficient, resource redundancy ratio, and environmental interference coefficient. The calculation process can be completed using a preset weighted combination method, allowing the damping coefficient to be adjusted in real time according to the construction site conditions. When the construction task has significant time adjustment flexibility, a high proportion of reserve resources, and a stable on-site construction environment, the damping coefficient increases accordingly, automatically reducing the propagated delays. Conversely, when construction resources are scarce or the construction environment is complex, the damping coefficient automatically decreases, allowing the risk to continue propagating to subsequent construction tasks, thus better reflecting the actual conditions of the engineering construction.

[0050] Preferably, in this embodiment, the correlation strength weight and absorption damping coefficient are updated in real time with the dynamic hypergraph. When events such as construction plan adjustment, reconfiguration of key resources, start or end of construction tasks, and changes in on-site monitoring status occur at the construction site, the system automatically calls the dynamic hypergraph update mechanism in Embodiment 1 to re-update the correlation matrix and propagation weight, and re-executes the calculations of formulas (7) to (10), so that the delay propagation analysis is always based on the latest construction status.

[0051] Example 3: Risk Gradient Traversal Analysis and Flexible Scheduling Optimization Process This embodiment combines Figure 4 This embodiment provides a detailed explanation of the risk gradient traversal analysis, adaptive risk threshold calculation, and flexible scheduling optimization process. Using the corrected delay state value output from the aforementioned embodiment two as input data, it performs gradient traversal analysis on the risk propagation path in the dynamic hypergraph, identifies key risk propagation links, and generates pre-emptive risk intervention strategies and post-emptive linkage scheduling schemes based on the traversal results, providing a decision-making basis for the dynamic optimization of the construction plan.

[0052] After completing the delay propagation analysis, the system first calculates the risk transfer gradient between nodes based on the corrected delay state values ​​corresponding to each construction task. To avoid the omission of critical path risks or excessive propagation of non-critical path risks due to the use of a fixed risk threshold, this embodiment introduces an adaptive risk threshold mechanism, which dynamically adjusts the traversal termination condition according to the importance of different construction tasks. The calculation method is shown in formula (11): (11) Among them, Threshold i : Represents the adaptive risk traversal threshold corresponding to node i; α base : Indicates the basic risk threshold, which is a preset parameter of the system; K i : Represents the criticality coefficient of the process, which is calculated based on whether the construction task is located on the critical path, the duration of the process, the degree of resource consumption, and the number of subsequent related tasks.

[0053] The larger the criticality coefficient of a process, the greater the impact of the corresponding construction task on the overall construction progress, and the risk traversal threshold increases accordingly. When the criticality coefficient of a process is small, the risk traversal threshold decreases accordingly, enabling the system to dynamically adjust the risk propagation range according to the importance of different construction tasks, thereby avoiding the problem of insufficient or excessive risk propagation caused by a fixed threshold.

[0054] Subsequently, the system calculates the risk transfer gradient between adjacent construction tasks based on the corrected delay state value, as shown in formula (12): (12) Among them, G i,j : Represents the risk transfer gradient from node i to node j; : Represents the delay state value of node i after damping correction according to Example 2; : Represents the delay state value of node j after damping correction; L i,j This represents the topological span between node i and node j, which is determined based on the shortest propagation path in the dynamic hypergraph.

[0055] To avoid interruption of risk propagation direction due to negative gradients, this embodiment uses the absolute value of the gradient for calculation, so that the risk propagation intensity only reflects the degree of risk difference and is not affected by the positive or negative direction of the gradient.

[0056] After calculating the risk gradient, the system performs a gradient-first traversal based on the dynamic hypergraph. When the risk transition gradient corresponding to an adjacent node is greater than the adaptive risk threshold of the current node, the propagation continues to the next node; when the risk transition gradient corresponding to all adjacent nodes is less than the adaptive risk threshold of the current node, the current path traversal stops, and the current node is marked as the risk propagation termination node, thus forming a complete critical risk propagation path.

[0057] In this embodiment, proactive risk intervention refers to adjusting the construction plan, resource allocation, and construction sequence in advance based on the identified key risk propagation paths before the risk spreads to subsequent construction tasks, thereby reducing the possibility of further risk spread. Post-implementation coordinated scheduling refers to coordinating adjustments to subsequent construction resources, work teams, and plans based on the actual propagation results after the risk has spread to subsequent construction tasks, thereby reducing the impact of overall construction delays. Through these definitions, proactive risk intervention and post-implementation coordinated scheduling correspond to different construction stages before and after risk propagation, respectively, clarifying the applicable timing of the two scheduling strategies.

[0058] Specifically, the objective of minimizing the overall project delay due to high-risk propagation paths is achieved by constructing an objective function for the comprehensive scheduling cost Z: After obtaining the key risk propagation path, the system further establishes a flexible scheduling optimization model to jointly optimize the construction plan and construction resources. The calculation method is shown in formula (13): (13) The constraints are:

[0059] Where Z represents the overall scheduling cost; w i : Represents the criticality weight of the construction task, where w iThe value of K is related to the value of K in formula (11). i Positive correlation: the higher the criticality coefficient of a process, the greater the weight of its schedule adjustment, ensuring that critical tasks receive priority resource support. ΔT i : Indicates the schedule adjustment amount corresponding to construction task i; C j : Represents the resource adjustment cost coefficient; ΔR j : Indicates the scheduling adjustment amount corresponding to resource j; R j : Indicates the amount of resources invested; R max : Indicates the maximum allowable resources to be invested at the construction site, where R max This refers to the maximum available total amount of various resources at the construction site as statistically analyzed in Example 1, such as the maximum number of tower crane shifts and the maximum number of laborers. By substituting the objective function, weight coefficients, and boundary constraints into the linear programming model, and using the multi-dimensional resource scheduling cost table as the model input parameters, the optimal resource scheduling scheme is obtained through optimization using a linear programming algorithm.

[0060] The above optimization model comprehensively considers the costs of adjusting the construction period and resource scheduling. Under the condition of satisfying the resource constraints at the construction site, it solves the construction adjustment scheme with the minimum comprehensive scheduling cost, so that key construction tasks can obtain resource guarantees first, while reducing the impact of non-critical construction tasks on the overall construction progress.

[0061] After completing the scheduling optimization, the system outputs key risk propagation paths, pre-emptive risk intervention plans, and post-emptive coordinated scheduling plans, and simultaneously updates the construction plan and resource allocation results. The optimized construction scheduling results serve as data input for historical case matching and decision optimization in Example 4, providing a data foundation for subsequent historical case retrieval, adaptive decision recommendation, and continuous optimization of scheduling strategies.

[0062] Example 4: Historical Case Matching and Decision Optimization Process This embodiment combines Figure 5 This embodiment provides a detailed explanation of the historical case matching, adaptive decision recommendation, and dynamic update process of the case library. Using the key risk propagation path, pre-emptive risk intervention plan, and post-emptive linkage scheduling plan output from Embodiment 3 as input data, it generates optimized decisions through intelligent matching of historical engineering cases and continuously updates the historical case library based on the engineering execution results, thereby achieving continuous optimization of risk intervention strategies.

[0063] After completing the flexible scheduling optimization, the system first establishes a historical project case library. The historical project case library consists of multiple completed projects. Each historical case includes information such as construction task characteristics, resource allocation status, risk propagation path, intervention measures, scheduling results, and project execution effects, and is uniformly encoded to form a case feature vector.

[0064] Preferably, the historical cases are derived from the company's historical construction projects and archived engineering data. The system performs data cleaning, outlier removal, and feature standardization according to a unified data format, and establishes a classification index based on construction type, construction stage, and construction organization method, enabling the case library to support rapid retrieval in different engineering scenarios.

[0065] After the case library is established, the system constructs the current case feature vector based on the risk propagation characteristics formed by the current project, and uses Euclidean distance to calculate the similarity between the current case and historical cases. The calculation method is shown in formula (14): (14) Where D represents the Euclidean distance between the current project and historical cases; M: Represents the dimension of the case feature vector; : Represents the m-th eigenvalue of the current project; : Represents the m-th eigenvalue of a historical case.

[0066] The case feature vector includes at least the following features: criticality of construction tasks, intensity of resource competition, degree of process dependence, length of risk propagation path, delay status value, scheduling adjustment amount, and resource allocation ratio, so that the case matching results can simultaneously reflect the construction organization characteristics and risk evolution characteristics.

[0067] The system calculates the Euclidean distance between the current project and each case in the historical case library, sorts them in ascending order of distance, and selects one or more historical cases with the smallest distance as candidate cases. When the difference in the Euclidean distances between multiple candidate cases is less than a preset threshold, the system integrates the intervention strategies corresponding to the multiple candidate cases to generate a final scheduling suggestion to improve decision stability.

[0068] To ensure the historical case library can continuously adapt to different engineering scenarios, this embodiment further establishes a dynamic case update mechanism. Upon completion of the current project, the system obtains the final construction results, including actual construction period, resource consumption, risk management effectiveness, and scheduling execution results, and generates a corresponding case feature vector. Once the case passes integrity and consistency checks, it is automatically written into the historical case library, and the corresponding classification index is updated, enabling subsequent projects to directly utilize the latest engineering experience to make risk decisions.

[0069] To address the issue of insufficient historical cases during the initial deployment of the new system, this embodiment further incorporates a cold start mechanism. When no cases meeting the similarity requirements exist in the historical case library, or when the Euclidean distance between the current project and all historical cases exceeds a preset distance threshold, the system does not directly invoke historical cases. Instead, it uses the flexible scheduling optimization result output in Embodiment 3 as the default scheduling scheme for the current project. After the current project is completed, the execution result of this project is automatically used to generate new historical cases and write them into the case library, gradually expanding the scale of historical cases and enabling continuous learning and dynamic improvement of the case library.

[0070] Specifically, the multi-level early warning threshold bands in this embodiment include: a yellow warning level corresponds to an estimated delay of 5% to 15% of the total planned construction period; an orange warning level corresponds to 15% to 30%; and a red warning level corresponds to more than 30%. The low-frequency single-channel push strategy sends a reminder once every 24 hours via SMS; the high-frequency multi-channel superimposed push strategy sends a reminder every 2 hours via SMS, APP messages, and telephone voice calls. The preset cascading depth threshold is 5 levels. Task nodes with a cascading depth deeper than 5 levels are configured with the low-frequency single-channel push strategy, while task nodes with a cascading depth shallower than or equal to 5 levels are configured with the high-frequency multi-channel superimposed push strategy.

[0071] The specific values ​​mentioned above are only preferred configurations for this embodiment. In actual applications, they can be reasonably adjusted according to the project's schedule tolerance and organizational communication standards. The preset delay impact threshold, first preset reminder period, second preset reminder period, preset cascading depth threshold, etc., are all configurable parameters.

[0072] Through the above steps, the system completes intelligent matching of historical cases, adaptive decision recommendation, dynamic case updates, and cold start processing, and outputs the final risk intervention decision plan. Thus, the system completes a full closed loop from unified modeling of engineering tasks, delay propagation analysis, risk gradient traversal, flexible scheduling optimization to continuous optimization of historical cases, realizing dynamic perception, intelligent analysis, and continuous optimization management of construction risks in engineering projects. The parameter values ​​in the various embodiments described in this invention are only preferred implementation methods, and those skilled in the art can adjust them according to the actual project scale, resource allocation, and construction organization methods.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent task reminders for engineering projects based on deep learning, characterized in that, include: Acquire business data throughout the entire lifecycle of engineering projects, extract semantic association features by modality through a pre-trained language model, and fuse them to generate a unified feature table for engineering tasks; construct a spatiotemporal dual-stream dynamic heterogeneous hypergraph network containing process-dependent hyperedges and resource-competitive hyperedges to generate initial process delay state values; The initial process delay status value is input into the spatiotemporal dual-stream dynamic heterogeneous hypergraph network, and delay transmission analysis and correlation strength weighted fusion are performed to generate delay impact value. Combined with the task schedule elastic correction delay impact value, the time-varying delay impact data of each task is output. Taking the task node whose time-varying delay impact data exceeds the preset threshold as the risk source node, perform adaptive risk gradient traversal to extract high-risk propagation paths, perform buffer period flexible adjustment on the high-risk propagation paths, and output the linked project delay prediction. Feature matching is performed on the unified feature table of engineering tasks of risk source nodes to generate pre-risk intervention decisions. Based on the estimated delay of the linked project period and the topological cascading depth of the tasks in the network, hierarchical post-linkage decision chains are adaptively configured and pushed.

2. The method according to claim 1, characterized in that, The process of generating the unified feature table for the engineering tasks specifically includes: The business data of the entire life cycle of the project is divided into text-based process data, tabular progress data, and IoT time-series data. The semantic association features in the text process data are extracted using a pre-trained language model, the topological node features in the table progress data are extracted according to the row and column structure, and the on-site working condition time series features in the IoT time series data are extracted through a one-dimensional convolutional layer. The cross-modal association weights between the semantic association features, topological node features and on-site working condition temporal features are calculated using a cross-attention mechanism, and projection alignment is performed to generate a unified feature table for engineering tasks in a unified high-dimensional feature space.

3. The method according to claim 1, characterized in that, The formation process of the spatiotemporal dual-stream dynamic heterogeneous hypergraph network specifically includes: Extract the overlapping duration of different engineering tasks within the same time period and calculate the proportion of their respective planned total duration as the temporal representation of the resource competition superedge. Combine this with the proportion of the intersection of the work areas of different engineering tasks at the same construction site to form a joint spatiotemporal constraint degree. Calculate the domain difference between the planned completion time of the preceding task and the planned start time of the subsequent task, and establish the flow constraint operator of the process-dependent hyperedge. The temporal representation and the joint spatiotemporal constraint are fused using multidimensional spatiotemporal features, and together with the flow constraint operator, they are written into the topological matrix of the heterogeneous hyperedge to form the spatiotemporal dual-flow dynamic heterogeneous hypergraph network.

4. The method according to claim 1, characterized in that, The specific process of the delay propagation analysis includes: The initial process delay state value is used as the initial energy operator and mapped to the source node of the spatiotemporal dual-stream dynamic heterogeneous hypergraph network; The spatiotemporal flow efficiency of adjacent task nodes in the process-dependent hyperedge is extracted to calculate the dynamic transmission rate, and the spatiotemporal overlap coefficient of the corresponding task nodes in the resource-competitive hyperedge is extracted to calculate the dynamic diffusion probability. By combining the dynamic transmission rate and the dynamic diffusion probability, the temporal transmission process of delay impact in the internal topology path of the network is dynamically simulated to generate delay impact values ​​for each task node.

5. The method according to claim 1, characterized in that, The specific process of the correlation strength weighted fusion includes: When the impact of delays propagates along the topological path, the topological adjacency relationship and the degree of association between each task node and its corresponding hyperedge are extracted. Calculate the topological mapping weights between hyperedges and nodes for the topological adjacency relationship and the association tightness, and construct the corresponding association strength weighted matrix; The delay impact values ​​propagated to each hyperedge are back-aggregated and updated to the corresponding central task node according to the topological weights using the correlation strength weight matrix, thus completing the structured weighted fusion of the delay impact values ​​in the network topology.

6. The method according to claim 1, characterized in that, The process for correcting the delay impact value specifically includes: Real-time acquisition of the remaining compressible elasticity coefficient of the engineering entity corresponding to each engineering task, as well as the real-time available redundant resource ratio corresponding to the current construction stage; The compressible elasticity coefficient of the remaining construction period and the proportion of real-time available redundant resources are solved by linear elastic combination to obtain the independent local absorption damping coefficient of each engineering task node. By applying the local absorption damping coefficient to the weighted and fused delay impact value, the time-varying delay impact data of each engineering task is obtained.

7. The method according to claim 1, characterized in that, The process of risk gradient traversal and high-risk propagation path extraction specifically includes: Taking the task node whose time-varying delay impact data exceeds a preset threshold as the root node, calculate the risk transfer gradient value between adjacent nodes along the topological direction of the process-dependent hyperedge and the resource-competition hyperedge; A preset gradient traversal step size and flow direction threshold are used to sequentially filter adjacent task nodes whose risk transfer gradient values ​​are greater than the preset threshold and extend them in a chain until the risk transfer gradient values ​​of all adjacent nodes are lower than the preset threshold. The node sequence generated by the chain extension is then used as the high-risk propagation path.

8. The method according to claim 1, characterized in that, The process of flexible adjustment and correction for the buffer period specifically includes: With the goal of minimizing the overall project delay due to high-risk propagation paths, a linear programming model is constructed with the comprehensive scheduling cost as the objective function, the criticality level and real-time elasticity coefficient of each affected engineering task as the weight coefficients of the objective function, and the limit supply load of limited on-site construction resources as the boundary constraint. A multi-dimensional resource scheduling cost table is constructed by calculating the physical span of resource scheduling between each node and the supply-demand gap value. The multi-dimensional resource scheduling cost table is then used as the input parameter of the linear programming model. The optimal resource scheduling scheme is obtained by optimizing the linear programming model using a linear programming algorithm, and the start time and duration of the delay of each affected engineering task are adjusted accordingly, and the estimated delay of the coordinated project period is output.

9. The method according to claim 1, characterized in that, The process of generating the aforementioned pre-emptive risk intervention decision specifically includes: Calculate the Euclidean distance between the current unified feature table of the engineering task of the risk source node and each historical feature vector in the pre-stored historical intervention case library; Retrieve and extract the target historical case with the shortest Euclidean distance from the historical intervention case database; The system retrieves the pre-intervention texts recorded in the target historical cases, performs semantic parsing and adaptation, generates pre-risk intervention decisions that include suggestions for supplementing construction resources and deadlines, and pushes them to the corresponding project responsible parties.

10. The method according to claim 1, characterized in that, The generation process of the hierarchical post-linkage decision chain specifically includes: The proportion of the estimated delay in the coordinated construction period to the total planned construction period of the corresponding project task is obtained, and a preset multi-level warning threshold band is matched to determine the warning level. The warning level includes at least two levels, and each warning level corresponds to a preset delay percentage range. The reminder frequency and push channels are adaptively configured based on the network topology cascading depth of the corresponding task nodes; A low-frequency single-channel push strategy is configured for task nodes with a cascading depth greater than a preset depth, and a high-frequency multi-channel overlay push strategy is configured for task nodes with a cascading depth less than a preset depth. The output includes a hierarchical post-linkage decision chain with different push sequences, and is pushed to the terminal devices of the corresponding project responsible entities. The low-frequency single-channel push strategy sends reminders through a single push channel according to a first preset reminder cycle; the high-frequency multi-channel overlay push strategy sends reminders through a combination of multiple push channels according to a second preset reminder cycle shorter than the first preset reminder cycle.

Images