Real-time acquisition and monitoring system for whole process data of asphalt mixing station

Abstract

This invention relates to the field of production data monitoring technology for asphalt mixing plants, specifically a real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant. The system includes modules for signal acquisition, data aggregation, feature extraction, state extrapolation, and decision generation. Multiple sensor arrays synchronously acquire operating signals from the raw material warehousing to finished product discharge throughout the entire process of the asphalt mixing plant. These signals are then synchronized and unified via a signal aggregation gateway, forming a time-stamped raw data stream. An improved sparse autoencoder, dynamically adjusting the sparse constraint strength based on equipment start-up and shutdown status, purifies and reduces the dimensionality of the data, generating a compact feature vector sequence. A temporal convolutional network model is used to extrapolate subsequent production state feature predictions from preceding feature vectors, and residual analysis is used to generate equipment control parameter suggestions. This system can filter out redundant data, accurately capture production time-series correlations, and ensure that control parameters match real-time production condition changes.

Description

Technical Field

[0001] This invention relates to the field of production data monitoring technology for asphalt mixing plants, and in particular to a real-time data acquisition and monitoring system for the entire production process of asphalt mixing plants. Background Technology

[0002] Existing asphalt mixing plant production monitoring systems often rely on scattered sensors deployed along the production line to collect operating signals at various stages. Data transmission lacks protocol integration, resulting in the inability to synchronize the time nodes of signals from different stages. Production data feature processing employs conventional sparse autoencoder algorithms, with fixed sparsity constraint parameters that are not adjusted to reflect the actual operating conditions of the production line equipment. Production status prediction utilizes ordinary neural network models without a corresponding extrapolation structure tailored to the time-series characteristics of the mixing plant's production.

[0003] Under conventional monitoring, the entire process operating condition signals suffer from chaotic protocols and inconsistent time bases, with the raw data stream containing a large amount of irrelevant information. Autoencoders with fixed sparse constraints cannot adapt to the operating condition fluctuations caused by equipment start-ups and shutdowns, and key production information is easily lost or redundant data is retained during feature purification. Ordinary neural network models struggle to capture the continuous temporal correlations of various stages of the mixing plant's production, resulting in discrepancies between the projected production status and actual operating conditions, and ultimately, the resulting control parameters fail to meet actual production needs.

[0004] The constraint parameters for extracting production data features from mixing plants cannot be dynamically adapted to the start-up and shutdown status of equipment, and the production status time-series extrapolation model cannot accurately derive the production status features of subsequent time steps. These problems directly affect the accuracy of production monitoring and the rationality of regulation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a real-time data acquisition and monitoring system for the entire production process of asphalt mixing plants.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant, comprising:

[0007] The signal acquisition module deploys multiple sensor arrays on the production line of the asphalt mixing plant to synchronously capture real-time operating signals throughout the entire process, from raw material warehousing, heating and drying, aggregate screening, weighing and batching, asphalt mixture mixing to finished product discharge.

[0008] The data aggregation module sends the real-time operating condition signals captured by the multiple sensor arrays to a signal aggregation gateway for protocol unification and time synchronization, forming a full-process raw data stream with time stamps.

[0009] The feature extraction module applies an improved sparse autoencoder algorithm to the time-stamped raw data stream to perform feature purification and dimensionality reduction. The improved sparse autoencoder algorithm dynamically adjusts the sparse constraint strength according to the start-up and shutdown status of the production line equipment to generate a compact feature vector sequence containing key production features.

[0010] The state inference module inputs the compact feature vector sequence into a production state time series inference model based on a temporal convolutional network. The production state time series inference model uses the preceding feature vector as a condition to deduce the predicted production state feature values ​​for subsequent time steps.

[0011] The decision generation module performs residual analysis on the predicted values ​​of the production status characteristics and generates suggestions for equipment control parameters.

[0012] As a further aspect of the present invention, the improved sparse autoencoder algorithm dynamically adjusts its sparse constraint strength according to the start-up and shutdown status of the production line equipment, including:

[0013] From the time-stamped raw data stream, binary status signals representing the operating status of key equipment are parsed out.

[0014] Based on the binary status signal, the production operation stage of the entire production process is identified, including the batch start-up stage, the stable operation stage, and the batch end stage.

[0015] A baseline sparsity target value is preset for each of the production operation stages, and the baseline sparsity target value for the stable operation stage is higher than the baseline sparsity target values ​​for the batch start-up stage and the batch end stage.

[0016] During the training and online inference of the sparse autoencoder, the binary state signal is monitored in real time;

[0017] When a production operation phase switch is detected, the coefficient of the sparse penalty term in the sparse autoencoder loss function is smoothly adjusted to match the strength of the baseline sparsity target value corresponding to the current production operation phase.

[0018] Under the dynamic adjustment of the sparse penalty term coefficient, the sparse autoencoder retains more features related to the transition process during the batch start-up and end stages, and forces the learning of more concise and discriminative feature representations during the stable operation stage, thereby generating the compact feature vector sequence that adapts to different production rhythms.

[0019] As a further aspect of the present invention, the compact feature vector sequence is input into a production state time series extrapolation model based on a temporal convolutional network, including:

[0020] The temporal convolutional network is trained using historical normal production batch data, enabling it to learn the evolution of production characteristics over time without abnormal perturbations.

[0021] The temporal convolutional network is composed of multiple stacked causal dilated convolutional layers, ensuring that the predicted value of the production state features at each time step depends only on the compact feature vector sequence of its past moments, and does not depend on future information.

[0022] In real-time monitoring, the compact feature vector sequence within the current and past time window is used as the input to the temporal convolutional network;

[0023] The temporal convolutional network outputs predicted values ​​of production status features for one or more fixed time intervals in the future, forming the sequence of predicted production status features.

[0024] As a further aspect of the present invention, residual analysis is performed on the predicted production state characteristics to generate suggested control parameters for the equipment, including:

[0025] The actual feature values ​​in the compact feature vector sequence are compared with the production state feature prediction values ​​at the corresponding time generated by the production state time series extrapolation model in real time, and the feature residual sequence is calculated.

[0026] The feature residual sequence is input into an adaptive threshold decision-maker. When the residual values ​​of multiple consecutive time steps in the feature residual sequence exceed their corresponding dynamic thresholds, an abnormal event is triggered.

[0027] By integrating the current compact feature vector sequence, feature residual sequence, and historical anomaly event markers, a multi-dimensional production status snapshot is constructed.

[0028] The multidimensional production status snapshot is matched with the preset production process knowledge graph. If the match is successful, a status diagnosis entry containing specific process steps and deviation types is generated.

[0029] The state diagnostic entries are sent to the instruction generation engine, which generates a set of control parameter suggestions with clear operational direction based on a predefined rule mapping table.

[0030] As a further aspect of the present invention, the real-time comparison of the actual feature values ​​in the compact feature vector sequence with the corresponding time-series production state feature prediction values ​​generated by the production state time-series deduction model specifically includes:

[0031] The compact feature vector sequence and the production state feature prediction value sequence are time-aligned to ensure that the comparison occurs at the same physical point in time.

[0032] Calculate the absolute difference between the actual feature vector and the predicted feature vector in each feature dimension at each time point after alignment;

[0033] The absolute differences across all feature dimensions are weighted and summed to obtain the comprehensive feature residual value at each time point. The comprehensive feature residual values ​​at all time points are arranged in chronological order to form the feature residual sequence.

[0034] As a further aspect of the present invention, the dynamic threshold determination method of the adaptive threshold decision-maker includes:

[0035] In the initial stage, the production status time series extrapolation model is run using a period of historical normal production data to calculate a baseline feature residual sequence.

[0036] Statistical analysis is performed on the baseline feature residual sequence to calculate its moving mean and moving standard deviation;

[0037] The moving mean is added to a number of times the moving standard deviation to form the initial dynamic threshold baseline;

[0038] In real-time operation, the short-term statistical properties of the feature residual sequence are continuously updated, and the dynamic threshold baseline is slightly smoothed and adjusted according to the updated statistical properties to adapt to the slow time-varying characteristics of the production process.

[0039] As a further aspect of the present invention, the integration of the current compact feature vector sequence, feature residual sequence, and historical anomaly event markers to construct a multi-dimensional production status snapshot includes:

[0040] Extract a feature vector segment of fixed length centered at the current time from the compact feature vector sequence;

[0041] Extract residual value segments from the feature residual sequence that have the same time range as the feature vector segments;

[0042] Extract and mark several recent historical anomalous events that are temporally close to the current time;

[0043] The feature vector fragment, the residual value fragment, and the historical abnormal event marker are concatenated and aligned in the time dimension and jointly encoded into a fixed-dimensional tensor. The fixed-dimensional tensor is the multi-dimensional production status snapshot that represents the current instantaneous and recent production status.

[0044] As a further aspect of the present invention, matching the multi-dimensional production status snapshot with a preset production process knowledge graph includes:

[0045] The production process knowledge graph is stored in the form of a graph structure, where nodes represent different production process links, equipment components or quality parameter indicators, and edges represent material flow, logical sequence or causal and correlation relationships between parameters.

[0046] The multidimensional production status snapshot is input into a graph neural network, the structure of which is aligned with the topological structure of the production process knowledge graph.

[0047] The activation intensity or abnormal contribution of the multidimensional production status snapshot on each node of the knowledge graph is calculated through the message passing and node update mechanism of the graph neural network.

[0048] Nodes whose activation intensity or abnormal contribution exceeds a threshold and edges strongly associated with them are identified. The subgraph structure formed by the nodes and edges is mapped to the specific process step and deviation type.

[0049] As a further aspect of the present invention, the step of generating a set of control parameter suggestions with clear operational direction based on a predefined rule mapping table includes:

[0050] The rule mapping table is indexed by the state diagnosis entries output after knowledge graph matching;

[0051] Each status diagnostic entry is associated with one or more sets of control logic, and each set of control logic defines the calculation rules for the target device, target parameters, adjustment direction, and adjustment range.

[0052] The calculation rule for the adjustment range is based on the specific values ​​of relevant features in the multidimensional production status snapshot, the recent trend of the feature residual sequence, and the effect feedback of historical similar control records.

[0053] As a further aspect of the present invention, the system further includes:

[0054] The instruction generation module is used to encode the suggested control parameters into control instructions that the equipment can recognize, and send them to the corresponding actuators on the asphalt mixing plant production line.

[0055] The proposed encoding of the control parameters into control commands recognizable by the device specifically includes:

[0056] Based on the target device identifier in the control parameter recommendations, query the device communication protocol library to determine the instruction format and communication protocol of the corresponding actuator;

[0057] The target parameter, adjustment direction, and calculated adjustment range in the proposed control parameter are filled in according to the determined instruction format to generate the original instruction data frame.

[0058] Add necessary frame headers, frame trailers, checksums, and timestamps for the generation of this instruction to the original instruction data frame to assemble it into a complete device control instruction message.

[0059] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0060] An improved sparse autoencoder algorithm, which dynamically adjusts the sparse constraint strength based on the start-up and shutdown status of production line equipment, performs feature extraction and dimensionality reduction on the time-stamped raw data stream of the entire process. The sparse constraint strength changes in real time with the start-up and shutdown status of the equipment, which can match the signal feature distribution patterns of different production stages, filter out non-critical interference information in the raw data stream, retain feature information directly related to the production conditions, and generate a compact feature vector sequence that reduces data dimensionality redundancy, making the presentation of production features more consistent with the actual operating characteristics of each production stage of the mixing plant.

[0061] The production state time-series extrapolation model, built based on a temporal convolutional network, receives a compact feature vector sequence and uses the preceding feature vector as a condition to extrapolate the predicted production state features for subsequent time steps. The temporal convolutional network can mine the temporal correlation attributes of features in each stage of the entire mixing plant production process, adapting to the continuous production process changes from raw material warehousing to finished product output. The extrapolated state feature predictions closely match the dynamic trajectory of the production process. After performing residual analysis on these predictions, the resulting control parameter suggestions can match the real-time operating conditions of the entire production process. Attached Figure Description

[0062] Figure 1 This is a timing diagram of the real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant as described in this invention.

[0063] Figure 2 A flowchart for the time-series deduction of the production state of a temporal convolutional network;

[0064] Figure 3 A flowchart for generating residual analysis and control parameter suggestions. Detailed Implementation

[0065] 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.

[0066] 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.

[0067] See Figure 1 The implementation of a real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant is based on a chain of signal flow and data processing. The system deploys multiple sensor arrays on the asphalt mixing plant's production line. These arrays cover all key process stages from raw material warehousing, heating and drying, aggregate screening, weighing and batching, asphalt mixture mixing, to finished product discharge, synchronously capturing real-time operating signals at each stage. The signal acquisition module drives these sensor arrays to acquire data. The data aggregation module receives real-time operating signals from multiple sensor arrays, performs protocol parsing and unification of heterogeneous sensor signals through a signal aggregation gateway, and performs precise time synchronization, ultimately outputting a time-stamped, coherent, full-process raw data stream. The feature extraction module receives this full-process raw data stream and applies an improved sparse autoencoder algorithm for feature purification and dimensionality reduction. This improved sparse autoencoder algorithm can dynamically adjust its sparse constraint strength according to the start-up and shutdown status of the production line equipment, thereby outputting a compact feature vector sequence containing key production information. The state extrapolation module inputs this compact feature vector sequence into a production state time-series extrapolation model built on a temporal convolutional network. This model uses the feature vectors of previous time steps as conditions to extrapolate the predicted production state features for one or more subsequent time steps. The decision generation module then performs residual analysis on the predicted and actual production state feature values ​​and generates control parameter suggestions for production line equipment based on the analysis results, thus completing the closed loop from perception, analysis to decision-making.

[0068] In one embodiment of the present invention, the improved sparse autoencoder algorithm in the feature extraction module dynamically adjusts its sparsity constraint strength according to the start-up and shutdown status of the production line equipment. Binary state signals characterizing the operating status of key equipment, such as motor start-up / stop signals and valve opening / closing signals, are parsed from the time-stamped raw data stream of the entire process. Based on the binary state signals, the production operation stage of the entire production process is identified, with the main production operation stages including the batch start-up stage, the stable operation stage, and the batch end stage. A baseline sparsity target value is preset for each production operation stage, wherein the baseline sparsity target value for the stable operation stage is set higher than the baseline sparsity target values ​​for the batch start-up stage and the batch end stage. During the training and online inference of the sparse autoencoder, the binary state signals are monitored in real time. When a switch in the production operation stage is detected, the coefficient of the sparsity penalty term in the sparse autoencoder loss function is smoothly adjusted to a strength matching the baseline sparsity target value corresponding to the current production operation stage. With the dynamic adjustment of the sparse penalty term coefficient, the sparse autoencoder retains more features related to the transition process during the batch start-up and end stages, and forces the learning of more concise and discriminative feature representations during the stable operation stage, thereby generating a compact feature vector sequence that adapts to different production rhythms.

[0069] The state extrapolation module inputs a compact feature vector sequence into a production state time-series extrapolation model based on a temporal convolutional network. (See also...) Figure 2 The temporal convolutional network is trained using historical normal production batch data, enabling it to learn the evolution of production features over time under normal conditions. This network consists of multiple stacked causal dilated convolutional layers. This structure ensures that the predicted production state features at each time step depend only on a sequence of compact feature vectors from past moments, and not on future information. In real-time monitoring, the compact feature vector sequence within the current and past time windows is used as input to the temporal convolutional network. The network outputs predicted production state features for one or more fixed time intervals in the future, forming a sequence of predicted production state features.

[0070] In practical implementation, the improved sparse autoencoder algorithm dynamically adjusts the sparse constraint strength based on the start-stop status of the production line equipment. This is achieved by real-time parsing of equipment status signals in the original data stream of the entire process with time stamps. In one example scenario, the operating current signal of the cold material supply belt motor and the speed signal of the drying drum drive motor are parsed in real time from the original data stream of the entire process. These continuous signals are then binarized into binary status signals representing whether the equipment is "running" or "stopping". Based on the combination of binary status signals of multiple key equipment, the system identifies the production operation stage of the entire production process. For example, when the cold material supply motor starts but the drying drum has not reached the set speed, the system determines it to be the batch start stage. When all main equipment operating signals are stable and the parameter fluctuations are within the set range, the system determines it to be the stable operation stage. When the mixing cylinder discharge door is closed and the hot aggregate elevator stops, the system determines it to be the batch end stage. In some embodiments, a baseline sparsity target value is pre-configured numerically for each production operation stage. The baseline sparsity target value for the stable operation stage is set to 0.85, and the baseline sparsity target values ​​for both the batch start-up and batch end stages are set to 0.65. During the online inference process of the sparse autoencoder, the system monitors the binary state signal representing the production operation stage in real time. When the system detects a switch from the batch start-up stage to the stable operation stage, the coefficient of the sparsity penalty term in the sparse autoencoder loss function does not immediately jump, but rather smoothly transitions from an intensity corresponding to 0.65 to an intensity corresponding to 0.85 over several time steps according to a predefined smoothing function. Optionally, this smoothing adjustment process can be iteratively calculated according to a formula:

[0071]

[0072] in: This represents the coefficient of the sparse penalty term at the current time step. This represents the coefficient from the previous time step. This represents the baseline coefficient intensity corresponding to the target production and operation phase. It is a smoothing factor between 0 and 1 used to control the adjustment rate. Under the dynamic adjustment of the sparse penalty term coefficient, the sparse autoencoder retains more features related to the transition process during the batch start-up and end stages, and forces the learning of a more concise feature expression during the stable operation stage, thereby generating a compact feature vector sequence that adapts to different production rhythms.

[0073] In practice, a compact feature vector sequence is input into a production state time-series projection model based on a temporal convolutional network. Historical normal production batch data is used to train the temporal convolutional network. This data covers multiple complete production batches without production anomalies or quality deviations. The training objective is to enable the temporal convolutional network to learn the evolution of production features over time under normal conditions. The temporal convolutional network consists of multiple stacked causal dilated convolutional layers. Each layer's convolution operation depends only on the current and past time step inputs. By stacking multiple layers and configuring progressively increasing dilation coefficients, the temporal convolutional network can construct a receptive field covering a long historical period, while ensuring that the predicted production state features at each time step depend only on the compact feature vector sequence of its past moments, and not on future information. In real-time monitoring, the system maintains a fixed-length first-in-first-out queue to cache compact feature vector sequences within the current and past time windows. This sequence is used as input to a temporal convolutional network (TCNN), which outputs predicted production status features for one or more fixed time intervals in the future. For example, if the system sets the prediction step size to 3 and the time interval to 5 seconds, the TCNN outputs predicted production status features 15 seconds later. These predicted values ​​are arranged chronologically to form a sequence of predicted production status features. The causal dilated convolutional structure of the TCNN ensures strict causality in the predictions, avoiding the introduction of future information into real-time monitoring. In some embodiments, the training process of the TCNN uses mean squared error as the loss function, and optimizes the network parameters through backpropagation to minimize the difference between the predicted production status feature sequence and the actual compact feature vector sequence. Optionally, the TCNN can be configured with multiple parallel output heads, each corresponding to a different prediction time step, to generate a sequence of predicted production status features for multiple future time points simultaneously.

[0074] In one embodiment of the present invention, the decision generation module performs residual analysis on the predicted values ​​of production status features to generate suggestions for equipment control parameters. The actual feature values ​​in the compact feature vector sequence are compared in real time with the predicted values ​​of production status features at the corresponding time points generated by the production status time-series extrapolation model to calculate a feature residual sequence. This feature residual sequence is input into an adaptive threshold decision-maker. When the residual values ​​of multiple consecutive time steps in the feature residual sequence exceed their corresponding dynamic thresholds, an abnormal event is triggered. The current compact feature vector sequence, the feature residual sequence, and historical abnormal event markers are integrated to construct a multi-dimensional production status snapshot. This multi-dimensional production status snapshot is matched with a preset production process knowledge graph. If a match is successful, a status diagnosis entry containing specific process steps and deviation types is generated. The status diagnosis entry is sent to the instruction generation engine, which, based on a predefined rule mapping table, generates a set of control parameter suggestions with clear operational direction.

[0075] See Figure 3 The process of comparing the actual feature values ​​in the compact feature vector sequence with the corresponding time-series production state feature prediction values ​​generated by the production state time series extrapolation model includes: real-time alignment of the compact feature vector sequence and the production state feature prediction value sequence to ensure the comparison occurs at the same physical time point; calculation of the absolute difference between the actual feature vector and the predicted feature vector at each aligned time point in each feature dimension; and weighted summation of the absolute differences across all feature dimensions to obtain the comprehensive feature residual value at that time point. The comprehensive feature residual values ​​at all time points are arranged in chronological order to form a feature residual sequence.

[0076] In practical implementation, the process of performing residual analysis on the predicted values ​​of production status characteristics and generating equipment control parameter suggestions is achieved through a closed-loop monitoring and decision-making chain. When comparing the actual feature values ​​in the compact feature vector sequence with the predicted values ​​of production status characteristics at the corresponding time generated by the production status time series extrapolation model in real time, the system first performs strict time alignment on the two sequences to ensure that the comparison occurs at the same physical time point. In an example scenario, the compact feature vector sequence contains features of multiple dimensions such as aggregate temperature, roller pressure, and motor power collected and extracted from sensors, while the production status characteristic prediction value sequence contains the predicted values ​​of the same dimension features at the same time in the future. Before the alignment calculation, the system standardizes the actual and predicted feature values ​​based on historical data of each feature dimension to eliminate the influence of different physical dimensions. At each alignment time point, the system calculates the absolute difference between the standardized actual feature vector and the standardized predicted feature vector in each feature dimension. The absolute differences in all feature dimensions are weighted and summed to obtain the dimensionless comprehensive feature residual value at that time point. The comprehensive feature residual values ​​of all time points are arranged in chronological order to form a feature residual sequence. It is understandable that the weights can be assigned based on the importance of different characteristics to the final product quality or process stability.

[0077] In some embodiments, the feature residual sequence is input into an adaptive threshold decision-maker. The adaptive threshold decision-maker monitors the feature residual sequence, and when the residual values ​​of multiple consecutive time steps in the feature residual sequence exceed their corresponding dynamic thresholds, an anomaly event is triggered. For example, if the dynamic threshold is set to 2.5, and the comprehensive feature residual values ​​of three consecutive time steps are 2.7, 2.9, and 3.2, the adaptive threshold decision-maker will generate a process anomaly mark for that time period. By integrating the compact feature vector sequence, feature residual sequence, and historical anomaly event marks at the current moment, the system constructs a multi-dimensional production status snapshot. The multi-dimensional production status snapshot is matched with a preset production process knowledge graph. If the match is successful, a status diagnosis entry containing specific process steps and deviation types is generated, such as "positive temperature shift in aggregate heating step". The status diagnosis entry is sent to the instruction generation engine, which generates a set of control parameter suggestions with clear operational direction based on a predefined rule mapping table, such as "reduce the opening of the fuel supply valve of the dryer drum burner by 2%".

[0078] In practical implementation, the specific steps for real-time comparison of the actual feature values ​​in the compact feature vector sequence with the corresponding predicted production state features generated by the production state time-series extrapolation model include strict time synchronization, data preprocessing, and difference calculation. The system maintains a unified high-precision clock source, assigning the same baseline timestamp to each data point in both the compact feature vector sequence and the production state feature prediction value sequence. Based on the timestamps, the two sequences are interpolated or resampled to ensure they have corresponding values ​​at equally spaced physical time points within the same series. Before difference calculation, the time-series data for each feature dimension is standardized. The standardization process calculates the mean and standard deviation of each feature dimension based on historical data, converting the original feature values ​​into dimensionless values ​​with a mean of 0 and a standard deviation of 1. The absolute difference between the standardized actual feature vector and the standardized predicted feature vector at each aligned time point is calculated for each feature dimension, forming the difference vector for that time point. Optionally, the standardization process can be replaced by a normalization method that scales the feature values ​​to the [0,1] interval, and then weights and sums the absolute differences across all feature dimensions to obtain the comprehensive feature residual value at each time point. The weighted summation process can be calculated using the following formula:

[0079]

[0080] in: Indicates at a point in time The dimensionless comprehensive characteristic residual value, This represents the total dimension of the feature vector. Indicates the first Preset weight coefficients for each feature dimension, Indicates at a point in time No. The standardized actual feature values ​​of each dimension. Indicates at a point in time No. Standardized predicted feature values ​​for each dimension, all weight coefficients The sum is 1. The comprehensive feature residual values ​​at all time points are arranged in chronological order to form a feature residual sequence. In some embodiments, the feature weight coefficients... The settings are based on prior knowledge from production process experts or determined through correlation analysis of features and final quality indicators in historical data. This can be understood as the comprehensive feature residual value... It provides a unified, dimensionless scalar index to measure the degree of deviation between the current actual production process state and the normal state evolution trajectory predicted by the model.

[0081] In one embodiment of the present invention, the dynamic threshold determination method of the adaptive threshold decision-maker includes: in the initial stage, running a production state time-series extrapolation model using a segment of historical normal production data to calculate a baseline feature residual sequence. Performing statistical analysis on the baseline feature residual sequence to calculate its moving mean and moving standard deviation. Adding a certain multiple of the moving standard deviation to the moving mean to obtain the initial dynamic threshold baseline. During real-time operation, continuously updating the short-term statistical characteristics of the feature residual sequence, and making minor smoothing adjustments to the dynamic threshold baseline based on the updated statistical characteristics to adapt to the slow time-varying characteristics of the production process. The process of integrating the current compact feature vector sequence, feature residual sequence, and historical abnormal event markers to construct a multi-dimensional production status snapshot includes: extracting a feature vector segment of fixed time length centered on the current moment from the compact feature vector sequence; extracting a residual value segment within the same time range as the feature vector segment from the feature residual sequence; and extracting several recently occurring historical abnormal event markers that are temporally adjacent to the current moment. Feature vector fragments, residual value fragments, and historical abnormal event markers are concatenated and aligned along the time dimension and encoded together into a fixed-dimensional tensor. This fixed-dimensional tensor is a multi-dimensional snapshot of the production status, representing the current instantaneous and recent production status.

[0082] In its implementation, the adaptive threshold decision-maker's dynamic threshold determination method is based on statistical learning and dynamic updating of the feature residual sequence. In the initial stage, the adaptive threshold decision-maker uses a historical normal production data set to run a production state time-series extrapolation model, calculating a baseline feature residual sequence. This historical normal production data set is typically selected from multiple fault-free, quality-compliant production batches to ensure that the baseline feature residual sequence represents the inherent random fluctuation level of the steady-state production process. Statistical analysis is then performed on the baseline feature residual sequence, calculating its moving mean and moving standard deviation. The moving mean reflects the recent central trend of the residual sequence, while the moving standard deviation reflects the recent dispersion of the residual sequence. The calculated moving mean is added to a multiple of the moving standard deviation to form the initial dynamic threshold baseline. The dynamic threshold baseline is not a fixed value but a sequence that changes over time, used to describe the upper limit of the normal fluctuation range. In real-time operation, the adaptive threshold decision-maker continuously receives new comprehensive feature residual values ​​and updates the short-term statistical properties of the feature residual sequence based on the new data. That is, it recalculates the moving mean and moving standard deviation of the data in the most recent window and makes a slight smoothing adjustment to the dynamic threshold baseline based on the updated statistical properties. This smoothing adjustment enables the dynamic threshold baseline to slowly track the slow time-varying characteristics of the production process itself, such as the gradual degradation of equipment performance or seasonal changes in environmental conditions, rather than reacting drastically to noise.

[0083] In some embodiments, dynamic threshold baseline At the point of time The calculation is based on the formula. Confirmed, among which Indicates at a point in time The dynamic threshold baseline, Indicates the deadline. The moving average of the characteristic residual sequence within a preset sliding time window This represents the moving standard deviation of the characteristic residual sequence within the same window. It is a preset multiplier coefficient used to control the tightness of the threshold. In the initial stage, a series of... This serves as a baseline for subsequent comparisons. During real-time operation, whenever a new comprehensive feature residual value is obtained... Upon arrival, the system updates the current sliding window. and And calculate the new dynamic threshold baseline according to the formula. :

[0084]

[0085] in: It is the threshold baseline of the previous time step. This is a smoothing factor between 0 and 1; this update mechanism enables minute smoothing adjustments to the dynamic threshold baseline. See Table 1 for a statistical analysis table of the baseline feature residual sequence.

[0086] Table 1: Statistical Analysis of Benchmark Feature Residual Sequences

[0087] Time Index ( ) Comprehensive characteristic residual value ( ) Moving average ( ) Moving standard deviation ( ) Dynamic threshold baseline ( , =3) 1001 0.85 0.87 0.12 1.23 1002 0.91 0.88 0.12 1.24 1003 0.82 0.87 0.11 1.20 1004 1.05 0.89 0.13 1.28 1005 0.88 0.88 0.12 1.24

[0088] In practical implementation, the process of integrating the current compact feature vector sequence, feature residual sequence, and historical anomaly event markers to construct a multi-dimensional snapshot of the production status aims to create a fixed-format instantaneous state representation containing multi-dimensional information. From the compact feature vector sequence, a feature vector segment of fixed length centered on the current moment is extracted; for example, the feature vector sequence from 60 seconds prior to the current moment to the current moment captures the recent evolution trajectory of the production process. From the feature residual sequence, a residual value segment within the same time range as the feature vector segment is extracted; the residual value segment reflects the dynamic deviation between the actual process and the prediction model within the same time period. Several recently occurring historical anomaly event markers that are temporally adjacent to the current moment are extracted; for example, all triggered anomaly event records within the past 5 minutes are extracted, with each record containing information such as event type, trigger time, and duration.

[0089] It is understandable that historical anomaly markers provide background information on recent anomalies. Feature vector fragments, residual value fragments, and historical anomaly markers are concatenated and aligned along the time dimension, collectively encoding a fixed-dimensional tensor. Optionally, the alignment operation may involve padding or truncating time series of unequal length to ensure a uniform final tensor size. This fixed-dimensional tensor represents a multi-dimensional snapshot of the current and recent production status. In some embodiments, historical anomaly markers are converted into binary sequences or multi-dimensional vectors with the same time baseline as the feature vectors before concatenation. This multi-dimensional production status snapshot serves as input for subsequent matching with the production process knowledge graph, and its fixed-dimensional structure facilitates processing by models such as neural networks.

[0090] In one embodiment of the present invention, the process of matching a multi-dimensional production status snapshot with a preset production process knowledge graph includes: the production process knowledge graph is stored in the form of a graph structure, where nodes represent different production process steps, equipment components, or quality parameter indicators, and edges represent material flow, logical order, or causal and correlation relationships between parameters. The multi-dimensional production status snapshot is input into a graph neural network, the structure of which is partially aligned with the topological structure of the production process knowledge graph. Through the message passing and node update mechanism of the graph neural network, the activation intensity or abnormal contribution of the multi-dimensional production status snapshot on each node of the knowledge graph is calculated. Nodes with activation intensity or abnormal contribution exceeding a threshold and their strongly associated edges are identified; the subgraph structure formed by these nodes and edges is mapped to specific process steps and deviation types.

[0091] The process of generating a set of suggested control parameters with clear operational direction based on a predefined rule mapping table includes: the rule mapping table is indexed by state diagnostic entries output after knowledge graph matching. Each state diagnostic entry is associated with one or more sets of control logic, and each set of control logic defines the calculation rules for the target equipment, target parameters, adjustment direction, and adjustment magnitude. The calculation rules for the adjustment magnitude are dynamically calculated based on the specific values ​​of relevant features in the multidimensional production status snapshot, the recent trend of the feature residual sequence, and the effect feedback of historical similar control records.

[0092] In practical implementation, the process of matching multi-dimensional production status snapshots with a pre-defined production process knowledge graph relies on a structured knowledge representation and a dedicated graph neural network. The production process knowledge graph is stored in graph structure form, where nodes represent different production process links, equipment components, or quality parameter indicators. For example, nodes may include "cold aggregate bin," "drying drum," "aggregate temperature," "drum pressure," and "finished product bin." Edges represent material flow, logical order, or causal and correlational relationships between parameters. For instance, there could be a "heating effect" edge between the "drying drum" node and the "aggregate temperature" node, and a "temperature compensation" edge between the "aggregate temperature" node and the "asphalt metering" node. The multi-dimensional production status snapshot is input into a graph neural network whose structure is partially aligned with the topological structure of the production process knowledge graph. That is, the nodes of the graph neural network correspond to the nodes of the knowledge graph, and the edges of the graph neural network and the edges of the knowledge graph have the same connectivity. By using the message passing and node update mechanism of graph neural networks, the activation intensity or abnormal contribution of multi-dimensional production status snapshots on each node of the knowledge graph is calculated. Nodes with activation intensity or abnormal contribution exceeding a preset threshold and edges strongly associated with them are identified. The subgraph structure formed by these nodes and edges is mapped to specific process links and deviation types.

[0093] The process of generating a set of control parameter suggestions with clear operational direction based on a predefined rule mapping table is a process based on rule matching and logical calculation. The rule mapping table is indexed by the state diagnosis entries output after knowledge graph matching. Each state diagnosis entry is associated with one or more sets of control logic. Each set of control logic defines the calculation rules for the target equipment, target parameters, adjustment direction, and adjustment magnitude. The calculation rules for the adjustment magnitude are dynamically calculated based on the specific values ​​of relevant features in the multi-dimensional production status snapshot, the recent trend of the feature residual sequence, and the effect feedback of historical similar control records. For example, for the state diagnosis entry "insufficient aggregate heating," the associated control logic may define the target equipment as "drying drum burner," the target parameter as "fuel flow setpoint," the adjustment direction as "increase," and the adjustment magnitude as calculated based on the difference between the current aggregate temperature and the target temperature, the average value of the temperature residual over a period of time, and the actual temperature change rate after the last similar adjustment.

[0094] In some embodiments, graph neural networks compute nodes through message passing. At any moment Hidden state vector The update formula is:

[0095]

[0096] in: Represents a node In the Hidden state vector of layer, initial state From the multi-dimensional production status snapshot and nodes The relevant feature components are obtained through encoding. Indicates the relationship between nodes in a knowledge graph. The set of directly adjacent nodes. Indicates from node To the node The eigenvectors of the edges, It is a learnable message function used to determine the status of neighboring nodes. State, current node The state and connecting edges The feature-generated message, This represents an aggregation function used to aggregate messages from all neighboring nodes. It is a learnable node update function that combines the node's current state with aggregated messages to generate a new node state. After multiple rounds of message passing, the final hidden state vector of each node is... The node is input into a fully connected layer and a sigmoid activation function, and the activation strength of the node is calculated. An activation intensity calculation table is shown in Table 2:

[0097] Table 2: Calculation Table of Node Activation Strength in Knowledge Graph

[0098] Knowledge graph nodes Activation intensity ( ) Does it exceed the threshold (0.7)? Cold aggregate bins 0.15 no Drying drum 0.92 yes Aggregate temperature 0.88 yes Roller pressure 0.45 no Asphalt weighing scale 0.32 no

[0099] In practice, the status diagnosis entry "Aggregate heating temperature is too low" will be output and used as an index to look up the corresponding control logic in the rule mapping table. The rule mapping table is stored in key-value pairs, where the key is the status diagnosis entry and the value is a list of corresponding control logics. A control logic entry may include a target equipment identifier, a target parameter identifier, an adjustment direction (increase / decrease), and a calculated adjustment range. The rule expression for calculating the adjustment range might be:

[0100]

[0101] in: This indicates the target value for aggregate temperature. This represents the current aggregate temperature value extracted from the current multidimensional production status snapshot. This represents the mean of temperature-related residuals in the characteristic residual sequence over a recent time window. This represents the effect feedback coefficient calculated based on historical records of similar adjustments (e.g., the ratio of the actual rate of temperature change to the expected rate of temperature change after the last adjustment). This is the preset gain coefficient. Optionally, the calculation rule expression can be more complex, including conditional judgments or nonlinear mappings. The specific adjustment range calculated based on the rule expression, combined with the target device, parameters, and direction, constitutes a control parameter recommendation with clear operational direction.

[0102] In one embodiment of the present invention, the system further includes an instruction generation module, used to encode the control parameter suggestions into equipment-recognizable control instructions and issue them to the corresponding actuators on the asphalt mixing plant production line. The process of encoding the control parameter suggestions into equipment-recognizable control instructions specifically includes: querying the equipment communication protocol library based on the target equipment identifier in the control parameter suggestions to determine the instruction format and communication protocol of the corresponding actuator; filling in the target parameters, adjustment direction, and calculated adjustment range in the control parameter suggestions according to the determined instruction format to generate an original instruction data frame; adding necessary frame headers, frame trailers, checksums, and a timestamp of this instruction generation to the original instruction data frame to assemble it into a complete equipment control instruction message.

[0103] In practical implementation, the instruction generation module encodes the suggested control parameters into control instructions that the equipment can recognize, and queries the equipment communication protocol library according to the target equipment identifier to determine the instruction format and communication protocol of the corresponding actuator. The equipment communication protocol library stores the communication protocol details supported by various controllable equipment in the mixing plant, such as the drying drum burner, the cold material supply belt frequency converter, and the asphalt metering valve. These details include, but are not limited to, physical layer interface type, data link layer frame structure, application layer instruction function code, register address mapping, and byte order. In an example scenario, the received control parameter suggestion is "Target device: Dryer drum burner, Target parameter: Fuel flow setpoint, Adjustment direction: Increase, Adjustment magnitude: 2.5%". The instruction generation module first parses the target device identifier "Dryer drum burner", then queries the device communication protocol library to find that the burner controller supports the Modbus RTU protocol, its fuel flow setpoint corresponds to the holding register address 0x0103, and the instruction format is "Slave address + Function code 0x06 + Register address high byte + Register address low byte + Setpoint high byte + Setpoint low byte + CRC check low byte + CRC check high byte". The instruction generation module then extracts the current value of "Fuel flow setpoint" from the control parameter suggestion, calculates the new setpoint based on the adjustment direction and magnitude of "Increase by 2.5%", fills the new setpoint into the corresponding "Setpoint high byte" and "Setpoint low byte" positions in the instruction data frame according to the queried instruction format, and generates the original instruction data frame.

[0104] In some embodiments, a necessary frame header, frame trailer, checksum, and timestamp of the current instruction generation are added to the original instruction data frame to assemble a complete device control instruction message. For devices using the Modbus RTU protocol, the added frame header is typically a silent interval lasting at least 3.5 characters, while the frame trailer does not require special characters. The checksum is a cyclic redundancy check (CRC) generated by calculating all preceding bytes. The process can be carried out according to the formula:

[0105]

[0106] Where: function This represents the standard Modbus CRC-16 checksum algorithm. Indicates the slave address byte. Indicates function code byte, and These represent the high byte and low byte of the register address, respectively. and These represent the high-order and low-order bytes of the set value, respectively. The generated CRC checksum is appended to the original data frame with the low-order byte first and the high-order byte last. The timestamp of this instruction generation can be appended to the application layer data field or recorded in the system's instruction issuance log for tracking. It is understandable that different communication protocols have different frame encapsulation requirements. For other protocols, such as PROFIBUS / USDP, the instruction generation module will add the corresponding start character, length byte, frame check sequence, and end character according to the definitions in the protocol library, ultimately assembling a complete message that conforms to the target device's protocol.

[0107] In practical implementation, the instruction generation module sends complete equipment control instruction messages to the corresponding actuators on the asphalt mixing plant production line via the corresponding industrial network. Optionally, the physical channel for sending can be an RS-485 bus, Ethernet, or fieldbus network. The instruction generation module ensures that the message is sent to the correct device address on the correct logical channel. In a specific example, the assembled dry drum burner fuel flow setting instruction message is sent in binary form via an RS-485 serial bus to slave station 1. After receiving the message, the burner controller verifies the CRC checksum and passes it. Then, it parses the application layer data and writes the new setting value into the internal register, thereby completing the execution of the control instruction. In some embodiments, the target parameter, adjustment direction, and calculated adjustment range in the control parameter suggestion may need to undergo data type conversion or scaling before being filled into the instruction data frame. For example, the floating-point percentage adjustment range may be converted to the integer raw value expected by the device register. This conversion rule is also defined in the device communication protocol library. Optionally, before issuing a command, the command generation module can perform boundary verification on the command value according to preset security rules to ensure that it does not exceed the safe operating range allowed by the device. Only after the verification is passed will the complete device control command message be finally sent.

[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 real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant, characterized in that: include: The signal acquisition module deploys multiple sensor arrays on the production line of the asphalt mixing plant to synchronously capture real-time operating signals throughout the entire process, from raw material warehousing, heating and drying, aggregate screening, weighing and batching, asphalt mixture mixing to finished product discharge. The data aggregation module sends the real-time operating condition signals captured by the multiple sensor arrays to a signal aggregation gateway for protocol unification and time synchronization, forming a full-process raw data stream with time stamps. The feature extraction module applies an improved sparse autoencoder algorithm to the time-stamped raw data stream to perform feature purification and dimensionality reduction. The improved sparse autoencoder algorithm dynamically adjusts the sparse constraint strength according to the start-up and shutdown status of the production line equipment to generate a compact feature vector sequence containing key production features. The state inference module inputs the compact feature vector sequence into a production state time series inference model based on a temporal convolutional network. The production state time series inference model uses the preceding feature vector as a condition to deduce the predicted production state feature values ​​for subsequent time steps. The decision generation module performs residual analysis on the predicted values ​​of the production status characteristics and generates suggestions for equipment control parameters.

2. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 1, characterized in that, The improved sparse autoencoder algorithm dynamically adjusts its sparse constraint strength based on the start-up and shutdown status of the production line equipment, including: From the time-stamped raw data stream, binary status signals representing the operating status of key equipment are parsed out. Based on the binary status signal, the production operation stage of the entire production process is identified, including the batch start-up stage, the stable operation stage, and the batch end stage. A baseline sparsity target value is preset for each of the production operation stages, and the baseline sparsity target value for the stable operation stage is higher than the baseline sparsity target values ​​for the batch start-up stage and the batch end stage. During the training and online inference of the sparse autoencoder, the binary state signal is monitored in real time; When a production operation phase switch is detected, the coefficient of the sparse penalty term in the sparse autoencoder loss function is smoothly adjusted to match the strength of the baseline sparsity target value corresponding to the current production operation phase. Under the dynamic adjustment of the sparse penalty term coefficient, the sparse autoencoder retains more features related to the transition process during the batch start-up and end stages, and forces the learning of more concise and discriminative feature representations during the stable operation stage, thereby generating the compact feature vector sequence that adapts to different production rhythms.

3. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 2, characterized in that, The compact feature vector sequence is input into a production state time series extrapolation model based on a temporal convolutional network, including: The temporal convolutional network is trained using historical normal production batch data, enabling it to learn the evolution of production characteristics over time without abnormal perturbations. The temporal convolutional network is composed of multiple stacked causal dilated convolutional layers, ensuring that the predicted value of the production state features at each time step depends only on the compact feature vector sequence of its past moments, and does not depend on future information. In real-time monitoring, the compact feature vector sequence within the current and past time window is used as the input to the temporal convolutional network; The temporal convolutional network outputs predicted values ​​of production status features for one or more fixed time intervals in the future, forming the sequence of predicted production status features.

4. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 3, characterized in that, Residual analysis is performed on the predicted values ​​of the production status characteristics to generate recommendations for equipment control parameters, including: The actual feature values ​​in the compact feature vector sequence are compared with the production state feature prediction values ​​at the corresponding time generated by the production state time series extrapolation model in real time, and the feature residual sequence is calculated. The feature residual sequence is input into an adaptive threshold decision-maker. When the residual values ​​of multiple consecutive time steps in the feature residual sequence exceed their corresponding dynamic thresholds, an abnormal event is triggered. By integrating the current compact feature vector sequence, feature residual sequence, and historical anomaly event markers, a multi-dimensional production status snapshot is constructed. The multidimensional production status snapshot is matched with the preset production process knowledge graph. If the match is successful, a status diagnosis entry containing specific process steps and deviation types is generated. The state diagnostic entries are sent to the instruction generation engine, which generates a set of control parameter suggestions with clear operational direction based on a predefined rule mapping table.

5. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 4, characterized in that, The real-time comparison of the actual feature values ​​in the compact feature vector sequence with the corresponding time-series production state feature prediction values ​​generated by the production state time-series deduction model specifically includes: The compact feature vector sequence and the production state feature prediction value sequence are time-aligned to ensure that the comparison occurs at the same physical point in time. Calculate the absolute difference between the actual feature vector and the predicted feature vector in each feature dimension at each time point after alignment; The absolute differences across all feature dimensions are weighted and summed to obtain the comprehensive feature residual value at each time point. The comprehensive feature residual values ​​at all time points are arranged in chronological order to form the feature residual sequence.

6. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 4, characterized in that, The dynamic threshold determination method of the adaptive threshold decision-maker includes: In the initial stage, the production status time series extrapolation model is run using a period of historical normal production data to calculate a baseline feature residual sequence. Statistical analysis is performed on the baseline feature residual sequence to calculate its moving mean and moving standard deviation; The moving mean is added to a number of times the moving standard deviation to form the initial dynamic threshold baseline; In real-time operation, the short-term statistical properties of the feature residual sequence are continuously updated, and the dynamic threshold baseline is slightly smoothed and adjusted according to the updated statistical properties to adapt to the slow time-varying characteristics of the production process.

7. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 6, characterized in that, The integration of the current compact feature vector sequence, feature residual sequence, and historical anomaly event markers to construct a multi-dimensional production status snapshot includes: Extract a feature vector segment of fixed length centered at the current time from the compact feature vector sequence; Extract residual value segments from the feature residual sequence that have the same time range as the feature vector segments; Extract and mark several recent historical anomalous events that are temporally close to the current time; The feature vector fragment, the residual value fragment, and the historical abnormal event marker are concatenated and aligned in the time dimension and jointly encoded into a fixed-dimensional tensor. The fixed-dimensional tensor is the multi-dimensional production status snapshot that represents the current instantaneous and recent production status.

8. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 7, characterized in that, Matching the multidimensional production status snapshot with a preset production process knowledge graph includes: The production process knowledge graph is stored in the form of a graph structure, where nodes represent different production process links, equipment components or quality parameter indicators, and edges represent material flow, logical sequence or causal and correlation relationships between parameters. The multidimensional production status snapshot is input into a graph neural network, the structure of which is aligned with the topological structure of the production process knowledge graph. The activation intensity or abnormal contribution of the multidimensional production status snapshot on each node of the knowledge graph is calculated through the message passing and node update mechanism of the graph neural network. Nodes whose activation intensity or abnormal contribution exceeds a threshold and edges strongly associated with them are identified. The subgraph structure formed by the nodes and edges is mapped to the specific process step and deviation type.

9. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 8, characterized in that, Based on a predefined rule mapping table, a set of control parameter suggestions with clear operational direction is generated, including: The rule mapping table is indexed by the state diagnosis entries output after knowledge graph matching; Each status diagnostic entry is associated with one or more sets of control logic, and each set of control logic defines the calculation rules for the target device, target parameters, adjustment direction, and adjustment range. The calculation rule for the adjustment range is based on the specific values ​​of relevant features in the multidimensional production status snapshot, the recent trend of the feature residual sequence, and the effect feedback of historical similar control records.

10. The real-time data acquisition and monitoring system for the entire production process of an asphalt mixing plant according to claim 1, characterized in that, The system also includes: The instruction generation module is used to encode the suggested control parameters into control instructions that the equipment can recognize, and send them to the corresponding actuators on the asphalt mixing plant production line. The proposed encoding of the control parameters into control commands recognizable by the device specifically includes: Based on the target device identifier in the control parameter recommendations, query the device communication protocol library to determine the instruction format and communication protocol of the corresponding actuator; The target parameter, adjustment direction, and calculated adjustment range in the proposed control parameter are filled in according to the determined instruction format to generate the original instruction data frame. Add necessary frame headers, frame trailers, checksums, and timestamps for the generation of this instruction to the original instruction data frame to assemble it into a complete device control instruction message.

Images