An ultra-thin asphalt intelligent production management method and system

Abstract

The present application relates to the field of intelligent production management and control of road engineering, in particular to an intelligent production management method and system based on ultra-thin asphalt, comprising: separating the real-time data of the production line into a core process field and a background interference field. Through the process knowledge graph, discrete data is constructed into a process state network, and periodic disturbance vectors of weather and equipment vibration are extracted. The two are antagonistically fused to form a comprehensive state representation, driving the digital mirror image to deduce the future quality control point prediction trajectory cluster. Analyze the fluctuation extreme value and frequency, compare the process tolerance to identify risks, trace the abnormal factors, and generate control strategies for aggregate supply, asphalt heating, mixing timing, and paving speed. The present application can fully reflect the process correlation, accurately quantify external disturbances, predict quality fluctuations in advance, and improve the accuracy of production control.

Description

Technical Field

[0001] This invention relates to the field of intelligent production control technology for road engineering, and in particular to an intelligent production management method and system based on ultra-thin asphalt. Background Technology

[0002] Current ultrathin asphalt production processes mostly rely on directly collecting real-time production data, conducting condition monitoring through simple statistics and numerical fitting, and adjusting production parameters based on manual experience. Production data is not divided into domains, process parameters are stored in discrete form, and no logical relationships are established between parameters based on process knowledge. External disturbances such as weather and equipment vibration are only processed using conventional filtering, without specifically extracting the periodic characteristics of the disturbances.

[0003] Discretized storage of production data fails to reflect the intrinsic connections between process steps, and the characteristics of external interference are vaguely represented, making it impossible to fully reflect internal and external influences on the process status. Conventional digital models can only perform single-value state predictions and cannot form multi-trajectory forward-looking extrapolations. Quality control can only perform threshold comparisons on real-time data and cannot predict the extreme values ​​and frequencies of fluctuations in future periods, resulting in a lag in process risk identification. Production control lacks a process for tracing the source of abnormal factors, and adjustments to aggregate supply, asphalt heating, mixing sequence, and paving speed lack specificity.

[0004] Production data is separated into different domains, and a logical process state network is constructed using a process knowledge graph. External periodic disturbances are extracted and counteracted through fusion. Based on the synthetic state-driven digital mirror, multi-trajectory predictions are generated, fluctuation characteristics are analyzed, and abnormal factors are traced to form targeted production adjustment strategies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent production management method and system based on ultra-thin asphalt.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent production management based on ultra-thin asphalt, comprising:

[0007] Collect real-time production data streams from the production line and separate the core process field and background interference field;

[0008] For the core process field, a pre-set process knowledge graph is introduced to enhance semantic association, transforming discrete data points in the core process field into a process state network with process logic relationships.

[0009] Wave pattern recognition is performed on the background interference field to extract periodic disturbance pattern vectors related to weather and equipment vibration;

[0010] The process state network and the periodic perturbation mode vector are adversarially fused to generate a synthetic process state representation that includes internal and external influencing factors.

[0011] Based on the state characterization of the synthesis process, the pre-generated digital image of the production line is driven to perform extrapolation calculations, and output the state prediction trajectory cluster of multiple key quality control points of the production line in the future preset time period;

[0012] The predicted fluctuation extreme values ​​and predicted fluctuation frequencies of each key quality control point are extracted from the state prediction trajectory cluster.

[0013] The predicted extreme values ​​and frequencies of fluctuations at key quality control points are compared point by point with the pre-stored process tolerance boundaries of key quality control points to identify the set of process risk points where the tolerance boundaries have been breached.

[0014] Based on the set of process risk points, we trace back to the abnormal influencing factors in the synthetic process state characterization, and generate a set of production adjustment strategies for aggregate supply, asphalt heating, mixing sequence or paving speed according to the weight of the abnormal influencing factors.

[0015] As a further aspect of the present invention, the step of collecting real-time production data streams from the production line and separating the core process field and the background interference field includes:

[0016] Collect real-time production data streams from the production line regarding aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction index;

[0017] The real-time production data stream is cleaned and synchronized and marked, and the cleaned real-time production data stream is mapped into a spatiotemporally correlated multidimensional data field, which includes time dimension, spatial location dimension and data attribute dimension.

[0018] Feature decoupling is performed on the spatiotemporally correlated multidimensional data field to separate the core process field that is strongly related to the production process and the background interference field that is related to the production environment.

[0019] The process of decoupling the spatiotemporally correlated multidimensional data field to separate the core process field strongly related to the production process and the background interference field related to the production environment specifically includes:

[0020] Construct a learnable decoupling filter, the input of which is a spatiotemporally correlated multidimensional data field;

[0021] The spatiotemporally correlated multidimensional data field is input into the decoupling filter. The decoupling filter calculates the correlation weight between each data unit in the multidimensional data field and the preset process standard vector through an internal multi-layer adaptive attention mechanism.

[0022] Based on the calculated correlation weights, all data units in the multidimensional data field are softly divided. Data units with correlation weights higher than the decoupling threshold are assigned to the core process field, and data units with correlation weights lower than the decoupling threshold are assigned to the background interference field.

[0023] The data units assigned to the core process field and the data units assigned to the background interference field are spatially reorganized to form a core process field and a background interference field that are topologically independent but have strictly synchronized timestamps.

[0024] As a further aspect of the present invention, for the core process field, a preset process knowledge graph is introduced for semantic association enhancement, transforming discrete data points in the core process field into a process state network with process logical relationships, including:

[0025] Obtain a process knowledge graph describing the relationship between material transfer, heat exchange, chemical reaction, and quality control standards in the asphalt mixture production process;

[0026] Each data point in the core process field is mapped to an entity node or a relation edge in the process knowledge graph based on its data attribute dimension information.

[0027] During the mapping process, if a numerical transition occurs between data points of the same type with adjacent timestamps in the core process field, a dynamic relationship edge representing the state transition is added between the corresponding entity nodes in the process knowledge graph.

[0028] When the data points representing different data attributes in the core process field meet the triggering conditions defined in the process knowledge graph, a logical relationship edge representing the process chain reaction is activated between the corresponding entity nodes that meet the conditions.

[0029] By using entity nodes, dynamic relationship edges, and logical relationship edges, the core process field is woven into a dynamically evolving process state network.

[0030] As a further aspect of the present invention, wave pattern recognition is performed on the background interference field to extract periodic disturbance pattern vectors related to weather and equipment vibration, including:

[0031] The background interference field is decomposed into intrinsic modes using mode decomposition technology, resulting in a series of intrinsic mode components arranged from high to low frequency.

[0032] From the intrinsic mode components, target intrinsic mode components with an energy percentage exceeding a preset energy threshold and a stable oscillation period are selected;

[0033] For each target intrinsic mode component, its oscillation period, oscillation amplitude, and phase shift are extracted as the basic features for pattern recognition;

[0034] The extracted basic features are matched with a predefined interference pattern feature library, which includes typical weather change pattern features and typical equipment vibration pattern features.

[0035] The basic features that successfully match the interference pattern feature library are classified and aggregated according to their corresponding interference source types to form periodic disturbance pattern vectors that respectively characterize weather disturbances and equipment vibrations.

[0036] As a further aspect of the present invention, the process state network and the periodic disturbance mode vector are adversarially fused to generate a synthetic process state characterization that includes internal and external influencing factors, including:

[0037] Construct an adversarial fusion structure that includes a generation unit and a discrimination unit, inputting the process state network into the generation unit and the periodic perturbation mode vector into the discrimination unit;

[0038] The generation unit is based on the process state network, simulates the process state change path caused by external disturbances, and generates a derived process network with disturbance assumptions.

[0039] The discrimination unit evaluates the authenticity probability of the state change path caused by external disturbance in the derived process network based on the periodic disturbance mode vector, and feeds back the evaluation results to the generation unit.

[0040] Based on the feedback from the discrimination unit, the generation unit adjusts its internal simulation strategy and iteratively optimizes the generation process of the derived process network until the discrimination unit determines that the probability of the authenticity of the state change path of the derived process network caused by external disturbance exceeds the preset confidence threshold.

[0041] The derived process network generated after the last iteration of optimization is combined with the original process state network through feature concatenation and normalization, and the output is used as the synthetic process state representation.

[0042] As a further aspect of the present invention, based on the characterization of the synthesis process state, a pre-generated digital image of the production line is driven to perform extrapolation calculations, outputting a cluster of predicted trajectories for the state of multiple key quality control points of the production line within a preset future time period, including:

[0043] The production line digital mirror includes a geometric model, a physical property parameter library, and a production process rule engine;

[0044] The synthesis process state characterization is used as initial and boundary conditions and input into the production process rule engine of the digital mirror of the production line.

[0045] The production process rule engine, based on its internally encapsulated mixing kinetics rules, heat transfer rules, and rheological rules, combined with the geometric model and physical property parameters of the production line digital mirror, performs multiple rounds of forward state deduction at a preset time step.

[0046] In each round of state simulation, the state parameter values ​​of multiple key quality control points are recorded. The key quality control points include the uniformity of asphalt and aggregate mixing, the temperature of the mixture leaving the plant, and the uniformity of the paving surface temperature.

[0047] Connect the state parameter values ​​of the same key quality control point obtained from all rounds of simulation in chronological order to form the state prediction trajectory of the key quality control point.

[0048] The state prediction trajectories of all key quality control points are collected to form a state prediction trajectory cluster.

[0049] As a further aspect of the present invention, the predicted fluctuation extreme values ​​and predicted fluctuation frequencies of each key quality control point are extracted from the state prediction trajectory cluster, including:

[0050] For each state prediction trajectory in the state prediction trajectory cluster, a sliding time window is used to traverse all state parameter value points on the state prediction trajectory.

[0051] Within each sliding time window, calculate the local maxima and local minima of the state parameter values ​​within the window, and record their magnitudes and the times they occur.

[0052] The maximum value among all the local maxima identified by the sliding time windows within a preset future time period is used as the upper extreme value of the predicted fluctuation, and the minimum value among the local minima is used as the lower extreme value of the predicted fluctuation.

[0053] At the same time, the number of times the state parameter value crosses its trajectory average value within a preset future time period is counted as the predicted fluctuation frequency.

[0054] As a further aspect of the present invention, based on the set of process risk points, the abnormal influencing factors in the characterization of the synthetic process state are traced back, and a set of production adjustment strategies for aggregate supply, asphalt heating, mixing sequence, or paving speed is generated according to the weight of the abnormal influencing factors, including:

[0055] Establish a backpropagation link model from process risk points to data nodes in the characterization of the synthesis process status;

[0056] Using the backpropagation link model, the influence weight value of each data node in the synthetic process state characterization on the occurrence of process risk points is calculated.

[0057] Data nodes whose influence weight values ​​exceed a preset weight threshold are selected and marked as abnormal influence factors.

[0058] The abnormal influencing factors are classified to distinguish the factors originating from aggregate supply, asphalt heating, mixing sequence or paving speed.

[0059] For each category of abnormal influencing factors, corresponding adjustment instructions are generated based on the sum of their influence weights and their positive or negative directions. These adjustment instructions include increasing or decreasing the aggregate supply rate, raising or lowering the asphalt heating temperature, advancing or delaying the mixing sequence, and accelerating or slowing down the paving speed.

[0060] All categories of adjustment instructions are collected to form a set of production adjustment strategies.

[0061] As a further aspect of the present invention, the process of constructing the production line digital mirror includes:

[0062] Three-dimensional laser scanning and measurement were performed on the ultra-thin asphalt production line to obtain point cloud data and geometric dimensions of the production line equipment, pipelines and silos;

[0063] Based on point cloud data and geometric dimensions, a hierarchical 3D geometric model of the production line is constructed, including the assembly level, component level and feature level.

[0064] For different components in the 3D geometric model of the production line, match and bind the corresponding material properties, thermal properties and kinematic properties in the physical property parameter library;

[0065] The production process rules, including formula ratio rules, temperature control rules, and time sequence rules, are encoded into computer-executable state transition logic to form a production process rule engine.

[0066] The three-dimensional geometric model of the production line, the physical attribute parameter library, and the production process rule engine are integrated and verified to ensure that the simulation results of the digital image are consistent with the historical operating data of the real production line within the allowable error range.

[0067] As a further aspect of the present invention, the present invention also includes an intelligent production management system based on ultra-thin asphalt, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent production management method based on ultra-thin asphalt as described above.

[0068] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0069] The production data stream is separated into a core process field and a background disturbance field. The core process field utilizes a process knowledge graph to enhance semantic associations, transforming discrete data points into a process state network with logical relationships. Discrete parameters form a structured relational system, fully presenting the constraints and linkages between process stages, and the description of process states aligns with actual production operation logic. The background disturbance field uses fluctuation pattern recognition to extract periodic disturbance pattern vectors corresponding to weather and equipment vibration. The changing patterns of external disturbances are independently quantified, improving the completeness and recognizability of disturbance features.

[0070] The process state network and periodic disturbance pattern vectors are antagonistically fused to form a synthetic process state representation encompassing internal and external influencing factors. This synthetic representation drives the production line digital mirror to perform extrapolation calculations, outputting a cluster of predicted trajectories for key quality control points in future timeframes. These trajectory clusters reflect state changes under multiple operating conditions, and the predicted extreme values ​​and frequencies of fluctuations at each control point are analyzed. The extreme values ​​and frequencies of fluctuations are compared point-by-point with the process tolerance boundaries to pinpoint process risk points where these boundaries are breached. These risk points are traced back to the abnormal influencing factors in the synthetic representation. Based on the factor weights, adjustment strategies are generated for aggregate supply, asphalt heating, mixing sequence, and paving speed. The control content directly corresponds to the source of the anomaly, ensuring precise targeting of production parameter adjustments. Attached Figure Description

[0071] Figure 1 This is a flowchart of an intelligent production management method based on ultra-thin asphalt as described in this invention;

[0072] Figure 2 A flowchart for transforming semantic association enhancement of core process fields into a process state network;

[0073] Figure 3 This is a flowchart for extracting periodic disturbance pattern vectors from a background interference field using wave pattern recognition. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0076] See Figure 1 This invention provides a smart production management method based on ultra-thin asphalt, the overall implementation scheme of which is as follows:

[0077] Real-time production data streams from the production line are collected, including time-series data on indicators such as aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction degree. The real-time production data streams are cleaned and time-synchronized, and mapped into a spatiotemporally correlated multidimensional data field. Feature decoupling is performed on this multidimensional data field to separate the core process field strongly related to the production process and the background disturbance field related to the production environment. For the core process field, a pre-defined process knowledge graph is introduced for semantic association enhancement, transforming discrete data points into a process state network with process logical relationships. Fluctuation pattern recognition is performed on the background disturbance field, extracting periodic disturbance pattern vectors related to weather and equipment vibration. The process state network and periodic disturbance pattern vectors are input into an adversarial fusion structure to generate a synthetic process state representation that includes internal and external influencing factors. Based on this synthetic process state representation, a pre-generated digital image of the production line is driven to perform forward state extrapolation calculations, outputting a cluster of predicted state trajectories for multiple key quality control points of the production line within a pre-defined future time period. The predicted extreme values ​​and frequencies of fluctuations for each key quality control point are extracted from the state prediction trajectory cluster. These predicted extreme values ​​and frequencies are then compared point-by-point with the pre-stored process tolerance boundaries of key quality control points to identify the set of process risk points where the tolerance boundaries have been breached. Based on this set of process risk points, the abnormal influencing factors in the synthetic process state characterization are traced back to their source, and a set of production adjustment strategies targeting aggregate supply, asphalt heating, mixing sequence, or paving speed is generated according to the weights of these abnormal influencing factors.

[0078] In one embodiment of the invention, real-time production data streams regarding aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction indexes are collected from the production line. The real-time production data streams are cleaned and synchronized, outliers are removed and missing values ​​are filled, and all data points are aligned to a unified timestamp sequence. The cleaned real-time production data streams are mapped into a spatiotemporally correlated multidimensional data field, which includes time, spatial location, and data attribute dimensions. A learnable decoupling filter is constructed, with the spatiotemporally correlated multidimensional data field as input. The spatiotemporally correlated multidimensional data field is input into the decoupling filter, which calculates the correlation weight between each data unit in the multidimensional data field and a preset process standard vector through an internal multi-layer adaptive attention mechanism. Based on the calculated correlation weights, all data units in the multidimensional data field are softly divided, with data units having correlation weights higher than a decoupling threshold assigned to the core process field, and data units having correlation weights lower than the decoupling threshold assigned to the background interference field. The data units assigned to the core process field and the data units assigned to the background interference field are spatially reorganized to form a core process field and a background interference field that are topologically independent but have strictly synchronized timestamps.

[0079] In practical implementation, taking the production period of ultra-thin asphalt mixture on a section of highway as an example, real-time production data streams regarding aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction index are collected from the production line. These data originate from gradation sensors installed at the dryer drum outlet, electromagnetic flow meters on the asphalt delivery pipeline, thermocouples inside the mixing drum, GPS and onboard timers on transport vehicles, infrared temperature sensors on the paver, and compaction measuring instruments on the roller. In practice, the real-time production data streams are cleaned and synchronized. The cleaning operation includes identifying and removing zero-value records from the gradation sensors caused by momentary material obstruction, and filling in missing thermocouple temperature data at specific second-level timestamps due to network latency using linear interpolation results from valid readings at adjacent times. The synchronization tagging operation unifies all data from different sensors, each with its own native acquisition timestamp, into a discrete time series with a one-second interval, based on the millisecond-level clock of the central control system and using a timestamp alignment algorithm. This ensures that each second corresponds to a complete set of data points containing all monitored indicators.

[0080] In some embodiments, the cleaned real-time production data stream is mapped to a spatiotemporally correlated multidimensional data field. The time dimension is the aligned one-second interval sequence described above, and the spatial location dimension is encoded according to the physical installation location of the data source sensors. Aggregate gradation and mixing temperature data are mapped to the "mixing plant" spatial region, asphalt flow rate data is mapped to the "asphalt supply pipeline" spatial region, paving temperature data is mapped to the "paver screed" spatial region, and compaction data is mapped to the "roller steel wheel" spatial region. The data attribute dimension directly corresponds to six index categories: aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction degree. In a specific implementation, a learnable decoupling filter is constructed. The decoupling filter is a fully connected neural network with three hidden layers. Its input is the aforementioned spatiotemporally correlated multidimensional data field, and its output is the probability weight of each data unit belonging to the core process field. In practical implementation, a spatiotemporally correlated multidimensional data field is input into a decoupling filter. The decoupling filter's internal multi-layer adaptive attention mechanism calculates feature correlations layer by layer. The final output layer calculates the correlation weight between each data unit in the multidimensional data field and a predefined process standard vector. The process standard vector is a high-dimensional vector predefined based on the target mix design of the asphalt mixture, the ideal temperature curve, and the compaction range required by specifications.

[0081] The formula for calculating the correlation weight is: Weight = Cosine Similarity (Data Unit Vector, Process Standard Vector) × Temporal Consistency Factor. In the formula, the data unit vector is the feature representation after processing by the first few layers of the network, the process standard vector is a fixed high-dimensional reference vector, and the temporal consistency factor is a measure of the smoothness of numerical changes in data points with the same spatial location and data attribute within adjacent time windows. In practical implementation, based on the calculated correlation weight, all data units in the multidimensional data field are softly divided. Data units with a correlation weight higher than the decoupling threshold are assigned to the core process field, such as aggregate gradation data with a deviation of less than 2% from the target mix ratio, and mixing and paving temperature data within ±5°C of the target temperature range. Data units with a correlation weight lower than the decoupling threshold are assigned to the background interference field, such as environmental temperature and humidity sensor data, slight vibration acceleration data of the main structure of the mixing plant, and fuel consumption fluctuation data generated by transport vehicles during non-uniform speed travel. The data units assigned to the core process field and the background interference field are spatially restructured and organized in the form of a graph structure. Nodes represent data units, and edges represent spatiotemporal adjacencies, thus forming a core process field and a background interference field that are topologically independent but have strictly synchronized timestamps. In the restructured core process field, nodes mainly contain highly correlated process parameter clusters such as "aggregate gradation - mixing plant," "asphalt flow rate - supply pipeline," and "mixing temperature - mixing plant." In the background interference field, nodes contain parameter clusters related to the environment or equipment status, such as "ambient temperature - plant area" and "building vibration X-direction - mixing plant." Both fields have their own graph structure snapshots at every identical timestamp.

[0082] In one embodiment of the present invention, see [reference] Figure 2 This method acquires a process knowledge graph describing the relationships between material transfer, heat exchange, chemical reactions, and quality control standards in the asphalt mixture production process. Each data point in the core process field is mapped to an entity node or a relation edge in the process knowledge graph based on its data attribute dimension information. During the mapping process, if a numerical transition occurs between data points of the same type at adjacent time stamps in the core process field, a dynamic relation edge representing the state transition is added between the corresponding entity nodes in the process knowledge graph. When data points representing different data attributes in the core process field meet the triggering conditions defined in the process knowledge graph, a logical relation edge representing a process chain reaction is activated between the corresponding entity nodes that meet the conditions. Through entity nodes, dynamic relation edges, and logical relation edges, the core process field is woven into a dynamically evolving process state network.

[0083] In practical implementation, a process knowledge graph describing the relationships between material transfer, heat exchange, chemical reactions, and quality control standards in the asphalt mixture production process is obtained. This process knowledge graph is stored in the form of a resource description framework, containing entity categories such as "coarse aggregate," "fine aggregate," "mineral filler," "asphalt," "mixing tank," "transport vehicle," and "paver," as well as relationship categories such as "supply to," "heated to," "mixed in," "transported to," and "paved at." It also includes quality control standard attributes such as "asphalt-aggregate ratio," "temperature range," "mixing time," and "compaction degree," along with their logical constraints. In practical implementation, each data point in the core process field is mapped to an entity node or a relationship edge in the process knowledge graph based on its data attribute dimension information. For example, an aggregate gradation data point representing the 4.75mm sieve passing rate is mapped to the "gradation composition" attribute under the "coarse aggregate" entity node in the process knowledge graph; an asphalt heating data point representing the outlet temperature of the thermal oil furnace is mapped to the "heated to" relationship edge between the "asphalt" entity node and the "heating device" entity node.

[0084] During the mapping process, if a numerical transition occurs between data points of the same type at adjacent timestamps in the core process field, a dynamic relationship edge representing the state transition is added between the corresponding entity nodes in the process knowledge graph. For example, at time t1, the value of the "mixing temperature" data point in the core process field is 165°C. At the next synchronization timestamp t2, the value of the "mixing temperature" data point jumps to 172°C. This change exceeding a preset threshold triggers the addition of a dynamic relationship edge from state "temperature A" to state "temperature B" on the "mixing cylinder" entity node in the process knowledge graph, with the transition value marked as +7°C. When data points representing different data attributes in the core process field meet the triggering conditions defined in the process knowledge graph, a logical relationship edge representing a process chain reaction is activated between the corresponding entity nodes that meet the conditions. The process knowledge graph predefines a triggering condition rule: if the "asphalt heating temperature" is below 160°C and the "mixing time" is less than 30 seconds, then the "asphalt-aggregate ratio uniformity" is at risk of not meeting the standard. In practice, when the data point in the real-time core process field shows that the "asphalt temperature" is 158°C and the "mixing time" is 28 seconds, the above conditions are met. Then, a logical relationship edge is activated between the "asphalt temperature" attribute node, the "mixing time" attribute node and the "asphalt-aggregate ratio uniformity" standard node in the process knowledge graph. This logical relationship edge is marked as "insufficient temperature and insufficient time lead to uniformity risk".

[0085] The formula for determining whether to activate a logical relation edge is: If the following conditions are met... Then activate . In the formula, This represents the numerical value of real-time asphalt temperature data points in the core process area. This represents the minimum permissible asphalt temperature threshold defined in the process knowledge graph. This represents the value of the real-time mixing time data point in the core process field. This represents the minimum required mixing time threshold defined in the process knowledge graph. It is the logical AND operator. This refers to the operation of creating a logical relationship edge marked "risk" between nodes that meet certain conditions. In some embodiments, the core process field is woven into a dynamically evolving process state network through entity nodes, dynamic relationship edges, and logical relationship edges. This process state network is a time-series graph that expands continuously with a timestamp sequence. The nodes in the graph are process knowledge graph entities or attributes with attached real-time data values. The edges between nodes include static relationships inherited from the original process knowledge graph, dynamic relationship edges generated by data transitions, and logical relationship edges generated by condition triggers. It can be understood that in a snapshot of the process state network at a certain moment, there may be a "coarse aggregate-gradation A" node, which is connected to the "mixing cylinder" node through a "supply to" edge. The "mixing cylinder" node is connected to its previous state "temperature A" through the "temperature B" dynamic relationship edge activated at the current moment. At the same time, the "mixing cylinder" node is also connected to the "oil-aggregate ratio uniformity-early warning" node through a "mixing insufficiency" logical relationship.

[0086] In one embodiment of the present invention, see [reference] Figure 3The background interference field is decomposed using mode decomposition (MDD) technology, yielding a series of MMD components arranged from high to low frequency. From these MMD components, target MMD components with an energy percentage exceeding a preset energy threshold and stable oscillation periods are selected. For each target MMD component, its oscillation period, amplitude, and phase shift are extracted as basic features for pattern recognition. These extracted basic features are matched against a predefined interference pattern feature library, which includes typical weather change pattern features and typical equipment vibration pattern features. The basic features that successfully match the interference pattern feature library are categorized and aggregated according to their corresponding interference source types, forming periodic disturbance pattern vectors representing weather disturbances and equipment vibrations, respectively. An adversarial fusion structure is constructed, comprising a generation unit and a discriminator unit. The process state network is input into the generation unit, and the periodic disturbance pattern vectors are input into the discriminator unit. Based on the process state network, the generation unit simulates the process state change path caused by external disturbances, generating a derived process network with disturbance assumptions. The discrimination unit evaluates the probability of the authenticity of state change paths caused by external disturbances in the derived process network based on the periodic perturbation pattern vector, and feeds the evaluation results back to the generation unit. Based on the feedback from the discrimination unit, the generation unit adjusts its internal simulation strategy and iteratively optimizes the generation process of the derived process network until the discrimination unit determines that the probability of the authenticity of the state change paths caused by external disturbances in the derived process network exceeds a preset confidence threshold. The derived process network generated after the last iteration of optimization is then feature-concatenated and normalized with the original process state network, and the output is used as the synthetic process state representation.

[0087] In the specific implementation, mode decomposition (MDD) technology is used to perform intrinsic mode decomposition (IMD) on the background interference field. The background interference field includes vibration acceleration time-series data collected from the foundation of the mixing plant and the bearing housing of the drying drum, as well as ambient temperature and humidity time-series data collected from the plant's meteorological station. Specifically, empirical mode decomposition (EMD) is performed on the vibration acceleration data to obtain a series of intrinsic mode components arranged from high to low frequency. Similarly, EMD is performed on the ambient temperature and humidity data. From all the obtained intrinsic mode components, target intrinsic mode components with an energy percentage exceeding a preset energy threshold (e.g., 5%) and stable oscillation periods are selected. One intrinsic mode component from the vibration data exhibits a stable 0.02-second oscillation period, and one intrinsic mode component from the ambient temperature data exhibits a stable 24-hour oscillation period. For each target intrinsic mode component, its oscillation period, oscillation amplitude, and phase shift are extracted as basic features for pattern recognition. For example, for the intrinsic mode component with the above 24-hour cycle, its oscillation period is extracted to be 86400 seconds, its oscillation amplitude is ±8°C, and its phase shift corresponds to the highest temperature of the day occurring at 14:00.

[0088] The extracted basic features are matched with a predefined interference pattern feature library, which includes typical weather change pattern features and typical equipment vibration pattern features. Typical weather change pattern features include diurnal temperature range cycles and seasonal slow change cycles, while typical equipment vibration pattern features include the power frequency of the drying drum rotation and the passing frequency of the fan blades. In specific implementation, the extracted "24-hour cycle, ±8°C amplitude" feature is matched successfully with the "diurnal temperature range" pattern in the interference pattern feature library; the extracted "0.02-second cycle" feature is also matched successfully with the "drying drum rotation power frequency harmonics" pattern in the feature library. See Table 1 for a simplified matching process.

[0089] Table 1: Matching Table of Target Intrinsic Mode Component Features and Interference Mode Feature Database

[0090] Ambient temperature data 24 hours ±8°C Day and night temperature difference mode Weather changes Horizontal vibration data 0.02 seconds 0.15g Second harmonic of the rotating dryer drum Equipment vibration Vertical vibration data 0.04 seconds 0.08g Wind turbine blade passing frequency Equipment vibration

[0091] After a successful match with the interference pattern feature library, the matched basic features are categorized and aggregated according to their corresponding interference source types. In specific implementation, all features matching the "weather change" category are aggregated into a multi-dimensional vector, representing a periodic disturbance pattern vector characteristic of weather disturbances; all features matching the "equipment vibration" category are aggregated into another multi-dimensional vector, representing a periodic disturbance pattern vector characteristic of equipment vibrations. It can be understood that the weather disturbance pattern vector may contain diurnal cycle components, seasonal trend components, etc., while the equipment vibration pattern vector may contain power frequency components, harmonic components, etc.

[0092] In practical implementation, an adversarial fusion structure is constructed, comprising a generation unit (a graph neural network) and a discriminator unit (a multilayer perceptron). The process state network is input to the generation unit, and the periodic perturbation pattern vector is input to the discriminator unit. Based on the process state network, the generation unit simulates the process state change path caused by external perturbations, generating a derived process network with perturbation assumptions. For example, the generation unit reads the information "temperature is about to drop by 5°C" from the weather perturbation pattern vector and simulates how this perturbation propagates along the "temperature-viscosity-paving uniformity" logical link in the process state network, thus generating a derived process network containing the "predicted value of paving temperature uniformity node decreasing." Based on the periodic perturbation pattern vector, the discriminator unit evaluates the probability of the realism of the state change path caused by external perturbations in the derived process network. The discriminator unit compares the simulated "paving temperature change trend" in the derived process network with the "temperature change trend" implied in the weather perturbation pattern vector, outputting a probability value between 0 and 1 to indicate realism, and feeds this evaluation result back to the generation unit. Based on feedback from the discriminator, the generation unit adjusts its internal simulation strategy, iteratively optimizing the generation process of the derived process network. The strategy adjustment follows these rules: if the probability of authenticity output by the discriminator is lower than a preset confidence threshold, the generation unit will correct the edge weights of the graph neural network used to simulate perturbation propagation. The strategy update formula for adjusting the edge weights is:

[0093]

[0094] in: This represents the updated edge weights in the graph neural network within the generating unit. This indicates the edge weights before the update. The learning rate is a small positive scalar. This represents the probability value of authenticity given by the discrimination unit in the current round of evaluation. This represents the preset confidence threshold. This represents the gradient of the loss function with respect to the edge weights. The iterative optimization process continues until the discrimination unit determines that the probability of the state change path of the derived process network caused by external disturbance exceeds a preset confidence threshold. It can be understood that the simulated disturbance impact path of the derived process network generated after the last iteration of optimization has been recognized by the discrimination unit with a high probability as conforming to historically observed periodic disturbance patterns. In specific implementation, the derived process network generated after the last iteration of optimization is combined with the original process state network through feature concatenation and normalization. The specific concatenation operation involves connecting the corresponding node feature vectors of the two networks along the feature dimension, and then performing max-min normalization on the long feature vector after concatenation to ensure that the values ​​of each dimension fall within the [0,1] interval. The output serves as the synthetic process state representation.

[0095] In one embodiment of the present invention, a three-dimensional laser scan and measurement are performed on an ultra-thin asphalt production line to acquire point cloud data and geometric dimensions of the production line equipment, pipelines, and silos. Based on the point cloud data and geometric dimensions, a hierarchical three-dimensional geometric model of the production line is constructed, including assembly level, component level, and feature level. For different components in the three-dimensional geometric model of the production line, corresponding material properties, thermal properties, and kinematic properties are matched and bound in a physical property parameter library. Production process rules, including formula ratio rules, temperature control rules, and time sequence rules, are encoded into computer-executable state transition logic to form a production process rule engine. The three-dimensional geometric model of the production line, the physical property parameter library, and the production process rule engine are integrated and verified to ensure that the simulation results of the digital image are consistent with the historical operating data of the real production line within the allowable error range, thus completing the construction of the digital image of the production line. This digital image of the production line includes a geometric model, a physical property parameter library, and a production process rule engine. The synthetic process state characterization is used as the initial and boundary conditions and input into the production process rule engine of the digital image of the production line. The production process rule engine, based on its internally encapsulated mixing kinetics, heat transfer, and rheological rules, combined with the geometric model and physical property parameters of the production line's digital mirror image, performs multiple rounds of forward state simulation at preset time steps. In each round of simulation, the state parameter values ​​of several key quality control points are recorded. These key quality control points include the uniformity of asphalt and aggregate mixing, the outlet temperature of the mixture, and the uniformity of paving surface temperature. The state parameter values ​​of the same key quality control point obtained from all simulation rounds are connected in chronological order to form the state prediction trajectory for that key quality control point. The state prediction trajectories of all key quality control points are then aggregated to form a state prediction trajectory cluster.

[0096] In the implementation, a three-dimensional laser scan and measurement of the ultra-thin asphalt production line was conducted. A multi-station three-dimensional laser scanner was used to perform a comprehensive scan of the mixing plant area, asphalt tank area, aggregate bins, conveyor belt corridor, pavers, and rollers, acquiring high-density point cloud data and precise geometric dimensions of key components of the production line equipment, pipelines, and bins. Based on the point cloud data and geometric dimensions, a hierarchical three-dimensional geometric model of the production line was constructed. In computer-aided design software, the point cloud data was first converted into triangular mesh surfaces, and then precise 3D models of the components were generated through solidification operations. The hierarchical structure includes assembly level, component level, and feature level. The assembly level describes the layout relationship of the entire mixing plant and paving site; the component level describes independent equipment such as drying drums, mixing cylinders, and asphalt pumps; and the feature level describes details such as the shape of the blades inside the drums and the angle of the mixing cylinder blades. For different components in the 3D geometric model of the production line, corresponding material properties, thermal properties, and kinematic properties are matched and bound in the physical property parameter library. The physical property parameter library is a relational database containing the density, specific heat capacity, and thermal conductivity of steel, the specific heat and viscosity-temperature curves of asphalt mixtures, and the rated speed and power information of each motor and conveyor belt. For example, if the "drying drum" component is selected in the 3D geometric model, the "carbon steel" material is matched for it from the physical property parameter library, and properties such as density 7850 kg / m³, specific heat capacity 0.46 kJ / (kg·K), and thermal conductivity 45 W / (m·K) are bound; if the "asphalt conveying pipeline" component is selected, the "insulation layer" material and corresponding thermal properties are bound for it.

[0097] The production process rules, including formula ratio rules, temperature control rules, and time sequence rules, are encoded into computer-executable state transition logic to form a production process rule engine. The formula ratio rule is encoded as "If (the material level in coarse aggregate bin A is lower than the set value) then (start the feeding conveyor belt of aggregate bin A)"; the temperature control rule is encoded as "If (the asphalt outlet temperature < the target temperature - tolerance) then (increase the set temperature ΔT of the thermal oil furnace)"; and the time sequence rule is encoded as "Event (aggregate is added to the mixing drum) triggers (start the mixing drum mixing timer), and after (the timer ≥ the minimum mixing time) (event (open the mixing drum discharge gate)) is allowed)". In practical implementation, the 3D geometric model of the production line, the physical property parameter library, and the production process rule engine are integrated and verified within a digital twin platform that supports multiphysics simulation. The digital twin is driven by importing real data recorded in historical production (e.g., complete production data from 8:00 AM to 10:00 AM on a certain morning). Key parameters derived from the digital twin (such as mixing temperature and output) are compared with recorded data from a specific time period in the historical database of the real production line. The verification standard is that the error between the derived results and historical data is consistent within an acceptable range. For example, the average absolute error between the simulated mixing temperature and the historical record is less than ±3°C, and the error between the simulated hourly output and the historical record is less than ±5%. Meeting these conditions completes the construction of the digital twin of the production line.

[0098] The production line digital mirror comprises a geometric model, a physical property parameter library, and a production process rule engine. In implementation, the synthesized process state representation is used as the initial and boundary conditions, input into the production process rule engine of the production line digital mirror. This representation includes information such as the aggregate gradation distribution, asphalt temperature and flow rate, ambient temperature, and equipment vibration status at the current moment. This information is converted into initial values ​​for corresponding logical variables in the production process rule engine. Based on its internally encapsulated mixing kinetics rules, heat transfer rules, and rheological rules, the production process rule engine, combined with the geometric model and physical property parameters of the production line digital mirror, performs multiple rounds of forward state deduction at a preset time step. The mixing kinetics rules describe the collision and coating process between aggregate and asphalt in the mixing drum; the heat transfer rules calculate the heat exchange between asphalt, aggregate, and the environment based on the heat transfer area provided by the geometric model and the thermal conductivity provided by the physical property parameters; the rheological rules simulate the flow and deformation behavior of the mixture during paving and compaction based on the viscosity-temperature curve of the asphalt. An example calculation of the heat transfer rule is shown in the following formula:

[0099]

[0100] in: It represents the heat flow rate conducted through a component within a single time step. This represents the thermal conductivity of the component material, obtained from the physical property parameter library. This represents the heat transfer area obtained from the geometric model. This represents the internal and external temperature difference of the component, initialized by the state of the synthesis process. This represents the wall thickness obtained from the geometric model. In each round of state simulation, the state parameter values ​​of multiple key quality control points are recorded. These key quality control points include the mixing uniformity of asphalt and aggregate, the plant temperature of the mixture, and the temperature uniformity of the paving surface. The mixing uniformity is quantified by simulating the variance of the asphalt film thickness at different sampling points within the mixing drum, and the temperature uniformity of the paving surface is quantified by simulating the dispersion coefficient of the temperature at virtual grid points on the paving surface. The state parameter values ​​of the same key quality control point obtained from all simulations are connected in chronological order to form the state prediction trajectory of the key quality control point. All the state prediction trajectories of the key quality control points are collected to form a state prediction trajectory cluster. The trajectory cluster is stored in memory as a time series array or curve for subsequent analysis. Refer to Table 2, which shows the geometric model hierarchy and some bound physical properties in the construction of the production line digital mirror.

[0101] Table 2: Binding Table of Geometric Model Hierarchy and Some Physical Attributes of Production Line Digital Mirroring

[0102] General Assembly Level Overall layout of the mixing plant Spatial coordinates GPS coordinates of each device Component level Drying drum Material properties / thermal properties Material: Carbon steel; Density: 7850 kg / m³; Specific heat capacity: 0.46 kJ / (kg·K) Component level Asphalt Insulated Pipes Thermal properties Thermal conductivity: 0.03 W / (m·K); Wall thickness: 50 mm Component level Mixing tank agitator motor Kinematic properties Rated speed: 30 rpm; Power: 90 kW Feature level Blades inside the drying drum Geometric properties / material properties Blade tilt angle: 45 degrees; Wear-resistant layer thickness: 10 mm

[0103] In one embodiment of the present invention, for each state prediction trajectory in the state prediction trajectory cluster, a sliding time window is used to traverse all state parameter value points on the state prediction trajectory. Within each sliding time window, the local maxima and local minima of the state parameter values ​​within the window are calculated, and their magnitudes and occurrence times are recorded. The maximum value among all local maxima identified by the sliding time windows within a preset future time period is taken as the upper extreme value of the predicted fluctuation, and the minimum value among the identified local minima is taken as the lower extreme value of the predicted fluctuation. Simultaneously, the number of times the state parameter values ​​cross their trajectory average value within the preset future time period is counted as the predicted fluctuation frequency. A backpropagation link model is established from the process risk point to the data nodes in the synthetic process state representation. Using the backpropagation link model, the influence weight value of each data node in the synthetic process state representation on the occurrence of the process risk point is calculated. Data nodes whose influence weight values ​​exceed a preset weight threshold are selected and marked as abnormal influence factors. Abnormal influence factors are classified to distinguish factor categories originating from aggregate supply, asphalt heating, mixing sequence, or paving speed processes. For each category of abnormal influencing factors, corresponding adjustment instructions are generated based on the sum of their influence weights and their positive or negative direction. These instructions include increasing or decreasing the aggregate supply rate, raising or lowering the asphalt heating temperature, advancing or delaying the mixing sequence, and accelerating or slowing down the paving speed. All adjustment instructions are then compiled to form a set of production adjustment strategies.

[0104] In practical implementation, for each state prediction trajectory in the state prediction trajectory cluster, a sliding time window is used to traverse all state parameter value points on the state prediction trajectory. Taking the state prediction trajectory of "paving surface temperature uniformity" as an example, this trajectory is a temperature uniformity coefficient sequence of one point per second for the next two hours, derived from the digital mirror of the production line. In practical implementation, the length of the sliding time window is set to 300 seconds, and the window slides for 60 seconds each time. Within each sliding time window, the local maxima and local minima of the state parameter values ​​within the window are calculated, and their values ​​and the time of occurrence are recorded. For example, within the time window [10:00:00, 10:05:00], the local maximum of the temperature uniformity coefficient sequence is 0.92 (occurring at 10:02:31), and the local minimum is 0.87 (occurring at 10:03:45). The values ​​0.92 and 10:02:31, and the values ​​0.87 and 10:03:45 are recorded. The maximum value among all local maxima identified by all sliding time windows within a preset future time period (e.g., the next two hours) is used as the upper extreme value of the predicted fluctuation. For example, the maximum value of 0.95 is found among all local maxima recorded by all windows. The minimum value among the identified local minima is used as the lower extreme value of the predicted fluctuation. For example, the minimum value of 0.82 is found among all local minima recorded by all windows. In some embodiments, the number of times the state parameter value crosses its trajectory average value within the preset future time period is also counted as the predicted fluctuation frequency. The trajectory average value is the arithmetic mean of all value points on the entire state prediction trajectory. Crossing means that the state parameter value changes from greater than the average value to less than the average value, or from less than the average value to greater than the average value. For example, the average value of the temperature uniformity coefficient trajectory is 0.89, and the number of times the coefficient value fluctuates around 0.89 within two hours is counted as 8 crossings.

[0105] In practical implementation, a backpropagation link model is established from the process risk point to the data nodes in the synthesized process state representation. The backpropagation link model is a directed graph. The endpoint of the graph is the identified process risk point, and the starting and intermediate nodes are the entity and attribute nodes in the synthesized process state representation. The directed edges between nodes represent direct influence relationships, and the edge weights are initialized using a neural network trained on historical data. Through the backpropagation link model, the influence weight value of each data node in the synthesized process state representation on the occurrence of the process risk point is calculated. The calculation adopts a gradient-based contribution analysis method. Starting from the process risk point node, the error signal is backpropagated along the edges of the backpropagation link model, and the sensitivity of each data node's features to changes in the risk point's state is calculated as the influence weight value. The formula for calculating the influence weight value is expressed as:

[0106]

[0107] in: This represents the weight value of a data node's impact on a process risk point. This represents the set of all directed paths from this data node to the process risk point. Indicates one of the paths, Representing a path A directed edge on, Represents a directed edge The weight, Representing a path The attenuation factor for signal propagation is a function of path length. Data nodes whose influence weight values ​​exceed a preset weight threshold are selected and marked as anomalous influence factors. For example, with a preset weight threshold of 0.1, the calculated influence weight value for the "Ambient Temperature - Low Temperature Disturbance" node is 0.25, and the influence weight value for the "Asphalt Temperature - Current Value" node is 0.15. These two nodes are marked as anomalous influence factors. Anomalous influence factors are classified to distinguish factors originating from aggregate supply, asphalt heating, mixing sequence, or paving speed processes. For example, "Ambient Temperature - Low Temperature Disturbance" is classified as external environmental interference, which does not directly correspond to the above four processes, while "Asphalt Temperature - Current Value" can be classified as the "Asphalt Heating" process. In some embodiments, for abnormal influencing factors directly originating from the aforementioned processes, corresponding adjustment instructions are generated based on the sum and positive / negative direction of their influence weight values. For example, for abnormal influencing factors categorized as "asphalt heating," if the sum of their influence weight values ​​is positive and large, indicating that the current asphalt temperature has a positive and aggravating effect on risk, then an adjustment instruction to "increase the asphalt heating temperature" is generated. The instruction includes a specific suggested increase in temperature, which is proportional to the magnitude of the influence weight value. Optionally, all categories of adjustment instructions are aggregated to form a production adjustment strategy set. The production adjustment strategy set is a structured list, where each record contains the process category, adjustment instruction content, suggested adjustment amount, and the influence factor weight information on which it is based. It can be understood that the generation of the production adjustment strategy set depends entirely on the quantified influence weight values ​​calculated by the backpropagation link model and the preset logical mapping rules.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications 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 protection scope of the present invention.

Claims

1. A method for intelligent production management based on ultra-thin asphalt, characterized in that, include: Collect real-time production data streams from the production line and separate the core process field and background interference field; For the core process field, a pre-set process knowledge graph is introduced to enhance semantic association, transforming discrete data points in the core process field into a process state network with process logic relationships. Wave pattern recognition is performed on the background interference field to extract periodic disturbance pattern vectors related to weather and equipment vibration; The process state network and periodic perturbation pattern vectors are adversarially fused to generate a synthetic process state representation that includes internal and external influencing factors. This process includes: constructing an adversarial fusion structure containing a generation unit and a discriminator unit; inputting the process state network into the generation unit and the periodic perturbation pattern vectors into the discriminator unit; the generation unit, based on the process state network, simulates process state change paths caused by external perturbations, generating a derived process network with perturbation assumptions; the discriminator unit, based on the periodic perturbation pattern vectors, evaluates the probability of the authenticity of state change paths caused by external perturbations in the derived process network and feeds back the evaluation results to the generation unit; the generation unit adjusts its internal simulation strategy based on the feedback from the discriminator unit, iteratively optimizing the generation process of the derived process network until the discriminator unit determines that the probability of the authenticity of state change paths caused by external perturbations in the derived process network exceeds a preset confidence threshold; the derived process network generated after the last iteration of optimization is feature-concatenated and normalized with the original process state network, and the output is used as the synthetic process state representation. Based on the state characterization of the synthesis process, the pre-generated digital image of the production line is driven to perform extrapolation calculations, and output the state prediction trajectory cluster of multiple key quality control points of the production line in the future preset time period; The predicted fluctuation extreme values ​​and predicted fluctuation frequencies of each key quality control point are extracted from the state prediction trajectory cluster. The predicted extreme values ​​and frequencies of fluctuations at key quality control points are compared point by point with the pre-stored process tolerance boundaries of key quality control points to identify the set of process risk points where the tolerance boundaries have been breached. Based on the set of process risk points, we trace back to the abnormal influencing factors in the synthetic process state characterization, and generate a set of production adjustment strategies for aggregate supply, asphalt heating, mixing sequence or paving speed according to the weight of the abnormal influencing factors.

2. The intelligent production management method based on ultra-thin asphalt as described in claim 1, characterized in that, The process of collecting real-time production data streams from the production line and separating the core process field and background interference field includes: Collect real-time production data streams from the production line regarding aggregate gradation, asphalt flow rate, mixing temperature, transportation time, paving temperature, and compaction index; The real-time production data stream is cleaned and synchronized and marked, and the cleaned real-time production data stream is mapped into a spatiotemporally correlated multidimensional data field, which includes time dimension, spatial location dimension and data attribute dimension. Feature decoupling is performed on the spatiotemporally correlated multidimensional data field to separate the core process field that is strongly related to the production process and the background interference field that is related to the production environment. The process of decoupling the spatiotemporally correlated multidimensional data field to separate the core process field strongly related to the production process and the background interference field related to the production environment specifically includes: Construct a learnable decoupling filter, the input of which is a spatiotemporally correlated multidimensional data field; The spatiotemporally correlated multidimensional data field is input into the decoupling filter. The decoupling filter calculates the correlation weight between each data unit in the multidimensional data field and the preset process standard vector through an internal multi-layer adaptive attention mechanism. Based on the calculated correlation weights, all data units in the multidimensional data field are softly divided. Data units with correlation weights higher than the decoupling threshold are assigned to the core process field, and data units with correlation weights lower than the decoupling threshold are assigned to the background interference field. The data units assigned to the core process field and the data units assigned to the background interference field are spatially reorganized to form a core process field and a background interference field that are topologically independent but have strictly synchronized timestamps.

3. The intelligent production management method based on ultra-thin asphalt as described in claim 2, characterized in that, For the core process field, a pre-defined process knowledge graph is introduced to enhance semantic association, transforming discrete data points in the core process field into a process state network with process logical relationships, including: Obtain a process knowledge graph describing the relationship between material transfer, heat exchange, chemical reaction, and quality control standards in the asphalt mixture production process; Each data point in the core process field is mapped to an entity node or a relation edge in the process knowledge graph based on its data attribute dimension information. During the mapping process, if a numerical transition occurs between data points of the same type with adjacent timestamps in the core process field, a dynamic relationship edge representing the state transition is added between the corresponding entity nodes in the process knowledge graph. When the data points representing different data attributes in the core process field meet the triggering conditions defined in the process knowledge graph, a logical relationship edge representing the process chain reaction is activated between the corresponding entity nodes that meet the conditions. By using entity nodes, dynamic relationship edges, and logical relationship edges, the core process field is woven into a dynamically evolving process state network.

4. The intelligent production management method based on ultra-thin asphalt as described in claim 1, characterized in that, Wave pattern recognition is performed on the background interference field to extract periodic disturbance pattern vectors related to weather and equipment vibration, including: The background interference field is decomposed into intrinsic modes using mode decomposition technology, resulting in a series of intrinsic mode components arranged from high to low frequency. From the intrinsic mode components, target intrinsic mode components with an energy percentage exceeding a preset energy threshold and a stable oscillation period are selected; For each target intrinsic mode component, its oscillation period, oscillation amplitude, and phase shift are extracted as the basic features for pattern recognition; The extracted basic features are matched with a predefined interference pattern feature library, which includes typical weather change pattern features and typical equipment vibration pattern features. The basic features that successfully match the interference pattern feature library are classified and aggregated according to their corresponding interference source types to form periodic disturbance pattern vectors that respectively characterize weather disturbances and equipment vibrations.

5. The intelligent production management method based on ultra-thin asphalt as described in claim 4, characterized in that, Based on the state characterization of the synthesis process, the pre-generated digital image of the production line is driven to perform extrapolation calculations, outputting a cluster of predicted trajectories for multiple key quality control points of the production line within a preset future time period, including: The production line digital mirror includes a geometric model, a physical property parameter library, and a production process rule engine; The synthesis process state characterization is used as initial and boundary conditions and input into the production process rule engine of the digital mirror of the production line. The production process rule engine, based on its internally encapsulated mixing kinetics rules, heat transfer rules, and rheological rules, combined with the geometric model and physical property parameters of the production line digital mirror, performs multiple rounds of forward state deduction at a preset time step. In each round of state simulation, the state parameter values ​​of multiple key quality control points are recorded. The key quality control points include the uniformity of asphalt and aggregate mixing, the temperature of the mixture leaving the plant, and the uniformity of the paving surface temperature. Connect the state parameter values ​​of the same key quality control point obtained from all rounds of simulation in chronological order to form the state prediction trajectory of the key quality control point. The state prediction trajectories of all key quality control points are collected to form a state prediction trajectory cluster.

6. The intelligent production management method based on ultra-thin asphalt as described in claim 1, characterized in that, The predicted fluctuation extremes and predicted fluctuation frequencies of each key quality control point are extracted from the state prediction trajectory cluster, including: For each state prediction trajectory in the state prediction trajectory cluster, a sliding time window is used to traverse all state parameter value points on the state prediction trajectory. Within each sliding time window, calculate the local maxima and local minima of the state parameter values ​​within the window, and record their magnitudes and the times they occur. The maximum value among all the local maxima identified by the sliding time window within a preset future time period is used as the upper extreme value of the predicted fluctuation, and the minimum value among the local minima is used as the lower extreme value of the predicted fluctuation. At the same time, the number of times the state parameter value crosses its trajectory average value within a preset future time period is counted as the predicted fluctuation frequency.

7. The intelligent production management method based on ultra-thin asphalt as described in claim 1, characterized in that, Based on the set of process risk points, the system traces back to the abnormal influencing factors in the synthetic process state characterization, and generates a set of production adjustment strategies for aggregate supply, asphalt heating, mixing sequence, or paving speed according to the weights of these abnormal influencing factors. These strategies include: Establish a backpropagation link model from process risk points to data nodes in the characterization of the synthesis process status; Using the backpropagation link model, the influence weight value of each data node in the synthetic process state characterization on the occurrence of process risk points is calculated. Data nodes whose influence weight values ​​exceed a preset weight threshold are selected and marked as abnormal influence factors. The abnormal influencing factors are classified to distinguish the factors originating from aggregate supply, asphalt heating, mixing sequence or paving speed. For each category of abnormal influencing factors, corresponding adjustment instructions are generated based on the sum of their influence weights and their positive or negative directions. These adjustment instructions include increasing or decreasing the aggregate supply rate, raising or lowering the asphalt heating temperature, advancing or delaying the mixing sequence, and accelerating or slowing down the paving speed. All categories of adjustment instructions are collected to form a set of production adjustment strategies.

8. The intelligent production management method based on ultra-thin asphalt as described in claim 5, characterized in that, The process of building the digital image of the production line includes: Three-dimensional laser scanning and measurement were performed on the ultra-thin asphalt production line to obtain point cloud data and geometric dimensions of the production line equipment, pipelines and silos; Based on point cloud data and geometric dimensions, a hierarchical 3D geometric model of the production line is constructed, including the assembly level, component level and feature level. For different components in the 3D geometric model of the production line, match and bind the corresponding material properties, thermal properties and kinematic properties in the physical property parameter library; The production process rules, including formula ratio rules, temperature control rules, and time sequence rules, are encoded into computer-executable state transition logic to form a production process rule engine. The three-dimensional geometric model of the production line, the physical attribute parameter library, and the production process rule engine are integrated and verified to ensure that the simulation results of the digital image are consistent with the historical operating data of the real production line within the allowable error range.

9. A smart production management system based on ultra-thin asphalt, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent production management method based on ultrathin asphalt as described in any one of claims 1 to 8.

Images