A method for determining the demulsification speed of emulsified asphalt

Abstract

The application discloses an emulsified asphalt demulsification speed determination method, relates to the online detection and intelligent monitoring technical field of road engineering materials, and remarkably improves the practicability and reliability of emulsified asphalt demulsification speed determination through integration of a lightweight time sequence prediction model and a self-adaptive processing mechanism; in terms of portability, an edge model based on a gated recurrent unit structure is adopted, calculation resource dependence is reduced, the method can be run on a handheld device, the problem of clumsiness of traditional laboratory equipment is avoided, and construction personnel can be conveniently and rapidly deployed on site; real-time performance is enhanced, a self-adaptive sliding window dynamically adjusts window length and step length according to data stability, the inference frequency is ensured to match the demulsification process, an online learning mechanism is combined, the model can instantaneously respond to environmental changes such as temperature fluctuations, and a demulsification speed estimation value can be outputted without obvious delay, so that paving and compaction time is guided, and construction interruption is reduced.

Description

Technical Field

[0001] This invention relates to the field of online testing and intelligent monitoring technology for road engineering materials, and in particular to a method for determining the demulsification rate of emulsified asphalt. Background Technology

[0002] In field road construction sites, determining the demulsification rate of emulsified asphalt is crucial for ensuring construction quality. For example, in chip seal or micro-surfacing operations, the demulsification state needs to be quickly determined after the emulsified asphalt is mixed with aggregates to guide the paving and compaction processes. The construction site environment is highly variable, with frequent fluctuations in factors such as temperature and humidity. These conditions directly affect the demulsification process, requiring the measurement method to have real-time response capabilities to avoid material waste or adhesion problems due to delays. However, traditional laboratory methods are difficult to directly transfer to such dynamic scenarios, and monitoring methods adapted to field conditions are urgently needed.

[0003] Common measurement methods in existing technologies include AI-based predictive models and portable conductivity testing techniques. AI models, such as those combining convolutional neural networks and long short-term memory networks, can analyze time-series data during the demulsification process and predict viscosity changes. Conductivity methods infer the demulsification rate by measuring the conductivity properties of emulsified asphalt, and some equipment has been miniaturized. However, these methods rely on high-performance computing resources during construction and cannot run smoothly on portable terminals. Conductivity testing is susceptible to environmental interference; sudden temperature changes or humidity fluctuations can cause data drift, affecting the stability of the results.

[0004] To address these disadvantages, some existing technologies improve adaptability through algorithm simplification or sensor integration; for example, lightweight design reduces computational load, or multi-source data fusion is used to offset environmental impacts. However, these improvements are still limited by hardware resources and algorithm generalization capabilities. Under complex field conditions, the accuracy and efficiency of real-time monitoring have not yet reached ideal levels, and experience-based judgment is often required during construction, increasing uncertainty. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a method for determining the demulsification rate of emulsified asphalt, which solves the problems of variable environment, poor portability and insufficient real-time response of existing methods in monitoring the demulsification of emulsified asphalt in field construction sites, resulting in low construction efficiency.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a method for determining the demulsification rate of emulsified asphalt, comprising:

[0009] Step S1: Collect time-series physical parameter data of the emulsified asphalt sample using a portable sensing unit; the time-series physical parameter data includes at least electrical conductivity.

[0010] Step S2: Perform preprocessing on the time-series physical parameter data, including temperature compensation and baseline normalization;

[0011] Step S3: The preprocessed sequence is segmented using a sliding window and used as input to a lightweight time series prediction model. The lightweight time series prediction model adopts a gated recurrent unit structure and runs on an edge processing unit.

[0012] Step S4: Output the demulsification rate estimate, which is defined as: a quantitative characterization index of the sample demulsification process, including at least one of the following: the estimated time required for the sample to reach the preset demulsification threshold, or the estimated demulsification rate of the sample at the measurement time.

[0013] Step S5: Update the parameters of the lightweight time-series prediction model incrementally based on the online learning mechanism, and display the estimated value of the bursting rate in real time on the embedded display interface.

[0014] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the temperature compensation includes performing temperature normalization processing on the conductivity sequence based on the temperature measured by the temperature sensor, and enabling compensation update when the ambient temperature changes rapidly.

[0015] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the sliding window has a configurable window length and step size. The window length is used to cover the key stages of the demulsification process, and the step size is used to control the inference frequency and is adaptively adjusted based on data stability.

[0016] The adaptive adjustment steps for the window length and step size include:

[0017] a) at time For the preprocessed conductivity and temperature compensation sequences, the short window variance, normalized change point score and normalized value of short-term prediction residual are calculated to form the input metric set;

[0018] b) Based on the comprehensive stability index, the window length and step size are adjusted in conjunction, so that the window is lengthened and the step size is increased during the stable phase, and the window is shortened and the step size is reduced during the fluctuating phase.

[0019] ,

[0020] ,

[0021] in, Indicates time The window length, in units of sample size. Indicates time The step size, in units of sample size. This indicates the lower bound of the window length, expressed in units of the number of samples. This indicates the upper bound of the window length, expressed in units of the number of samples. Represents the exponential compression factor, a dimensionless constant. Indicates time The stability index is dimensionless. This indicates the lower bound of the step size, expressed in units of sample size. The ratio of step size to window length is a dimensionless constant. This represents the floor operator;

[0022] In the formula:

[0023] ,

[0024] in, Indicates time The stability index is dimensionless. Indicates time The short-window variance calculated within the most recent fixed sample segment, in units of , This represents the reference value for stable baseline variance, in units of [unit name missing]. , Indicates time The variable point score, normalized to Dimensionless Indicates time The normalized value of the one-step prediction residual, dimensionless. , , Let each of the three weights be a non-negative dimensionless constant and satisfy the following conditions: ;

[0025] c) When the relative change of the stability index exceeds the hysteresis threshold, or when the short-term prediction error shows a significant decrease in the candidate combination, perform parameter updates; otherwise, remain inert to suppress jitter.

[0026] d) Apply upper and lower bounds and inference time constraints to the candidate window length and step size. If they are close to the resource limit, prioritize the combination with a larger step size to control the call frequency.

[0027] e) Use the updated window length and step size for sliding window slicing, enter edge model inference, and feed the latest residuals back into step a) for the next round of evaluation;

[0028] The execution details and criteria of the adaptive adjustment step are as follows:

[0029] Trigger: If relative change Update if hysteresis exceeds the set upper threshold. If it falls below the lower threshold, it remains unchanged;

[0030] Candidate search: Compare the rolling mean of short-term prediction errors across a discrete set; if the improvement is less than the minimum improvement, do not switch combinations.

[0031] constraint: , When runtime latency is close to the limit, prioritize increasing it. .

[0032] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the online learning mechanism is triggered when data distribution drift is detected or the model confidence is lower than the threshold, and incremental gradient update with regularization constraint is adopted, and the upper limit of learning rate and the upper limit of update number are set to control energy consumption and stability.

[0033] The step of using incremental gradient update with regularization constraints includes:

[0034] f) Calculate the two types of signals after inference: drift score With uncertainty ;when Above the threshold or If the value is above the threshold, the incremental update process begins; if the value is below the threshold at the same time, the update is skipped, and only the monitored value is recorded.

[0035] g) Sample size from the nearest window segment. The samples are used to form incremental batches. Indicates the incremental batch size;

[0036] h) Perform one or more gradient steps with an objective function that includes a penalty term, and apply the regularization term to the joint parameter set of the GRU weights and biases.

[0037] ,

[0038] , ,

[0039] in, Indicates the trigger sequence number The incremental objective function is dimensionless. Indicates predicted value Compared with the labeled value The baseline loss between them is dimensionless. Indicates the trigger sequence number The set of trainable parameters for the GRU and its fully connected output layer, according to norm Weighted, This represents the weight decay coefficient, which is dimensionless. Indicates the trigger sequence number The measured data distribution drift score is dimensionless. The drift penalty weight is dimensionless. Indicates the trigger sequence number The prediction uncertainty score, dimensionless. The uncertainty penalty weight is dimensionless. This represents the set of parameters after one gradient step. Indicates the trigger sequence number The learning rate used, Indicates the upper limit of the learning rate. Indicates the basic learning rate. This represents the amplification factor of drift with respect to the learning rate; it is dimensionless. This represents the inhibition coefficient of the error with respect to the learning rate; it is dimensionless. This represents the dimensionless mean of the most recently predicted residuals calculated using the exponential sliding method. Represents the gradient with respect to the set of parameters.

[0040] i) Employ error-adaptive learning rate and cover Upper limit; the number of gradient steps within a single trigger does not exceed , This indicates the maximum number of gradient steps within a single trigger.

[0041] j) Calculate the validation loss using an independent sliding validation subset. If the improvement relative to the baseline before triggering is less than the threshold, roll back the parameters; if the improvement reaches the threshold, solidify the parameters.

[0042] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the demulsification threshold is set by reference test results or on-site calibration results of historical samples, and can be updated according to the batch of construction materials.

[0043] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the preprocessing further includes baseline estimation and drift detection for the initial conductivity segment; when long-term drift is detected, baseline re-estimation is performed and online updates are paused during the re-estimation period.

[0044] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the lightweight time-series prediction model is an edge model consisting of one or two layers of gated loop units and a fully connected output layer, and the number of model parameters and the inference delay are limited to not exceeding a preset resource threshold.

[0045] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the lightweight time-series prediction model is pre-trained by transfer learning before deployment: the model weights are initialized with historical demulsification data under multiple environmental conditions, and fine-tuned on a small amount of data collected in the target construction scenario.

[0046] As a preferred embodiment of the method for determining the demulsification rate of emulsified asphalt according to the present invention, the embedded display interface displays: an estimated demulsification rate, a demulsification status indicator, a data quality indicator, and an alarm prompt, and supports providing prompt information from the paving / compacting machine based on the estimated value.

[0047] In a second aspect, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for determining the demulsification rate of emulsified asphalt as described in the first aspect of the present invention.

[0048] The beneficial effects of this invention are as follows: By integrating a lightweight time-series prediction model and an adaptive processing mechanism, this invention significantly improves the practicality and reliability of emulsified asphalt demulsification rate determination. Regarding portability, the use of an edge model based on a gated cyclic unit structure reduces computational resource dependence, allowing the method to run on handheld devices, avoiding the bulkiness of traditional laboratory equipment and facilitating rapid on-site deployment by construction personnel. Real-time performance is enhanced; the adaptive sliding window dynamically adjusts the window length and step size based on data stability, ensuring that the inference frequency matches the demulsification process. Combined with an online learning mechanism, the model can respond instantly to environmental changes such as temperature fluctuations, outputting demulsification rate estimates without significant delay, thereby guiding paving and compaction machines and reducing construction interruptions. Regarding environmental adaptability, temperature compensation and baseline normalization in the preprocessing step effectively offset interference from field conditions, while online learning handles data distribution drift through incremental updates, improving the model's robustness in variable scenarios and avoiding prediction bias caused by environmental factors. In addition, the embedded display interface provides intuitive feedback, including the status of demulsification and alarm prompts, which further reduces the risk of human error. Overall, it achieves efficient and stable on-site monitoring, which meets the actual needs of road construction and makes up for the disadvantages of insufficient portability, real-time performance and adaptability in the background technology. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0050] Figure 1 This is a schematic flowchart of a method for determining the demulsification rate of emulsified asphalt in an embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0053] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0054] This application proposes a method for determining the demulsification rate of emulsified asphalt, combined with... Figure 1 As shown, the method includes:

[0055] Step S1: Collect time-series physical parameter data of the emulsified asphalt sample using a portable sensing unit. The time-series physical parameter data includes at least electrical conductivity. In this embodiment, time-series physical parameters refer to discrete measurement point data acquired on a continuous time axis for the same sample. Electrical conductivity is output by a portable conductivity sensor with a fixed sampling period, and temperature is synchronously acquired from the same sampling point for temperature compensation. The default sampling period is 1 second, which can be adjusted within the range of 200 milliseconds to 5 seconds depending on the working conditions, and is determined by verification in a test section based on the on-site construction rhythm and the rate of rheological change of the sample. To avoid misjudgment caused by instantaneous disturbances, the time interval between two adjacent valid samples must be no less than the sampling period, and the sensor self-test must pass. Optionally, if the short-term missing measurement does not exceed three sampling periods, the previous valid value is retained and marked as an interpolation marker; if the continuous missing measurement exceeds this upper limit, subsequent steps are paused and a prompt for resampling is displayed.

[0056] Step S2 involves preprocessing the time-series physical parameter data, including temperature compensation and baseline normalization. Specifically, temperature compensation eliminates the systematic influence of temperature on conductivity, while baseline normalization ensures comparability between different samples or batches. The initial baseline is estimated by default using the stable segment of the most recent 30 to 60 sampling points, and a re-estimation of the compensation parameters is triggered when a short-term temperature change exceeds 2 degrees Celsius. The initial value of the temperature compensation coefficient is derived from indoor control experiments and fine-tuned on-site using short-term control tests, with an update cycle of no less than 30 seconds to suppress frequent fluctuations. If the temperature sensor is malfunctioning or missing, optionally, only baseline normalization can be performed, and the compensation flag can be set to disabled. In case of an anomaly, the minimum feasible approach is baseline normalization only with the compensation parameters frozen.

[0057] Step S3 involves using a sliding window to segment the preprocessed sequence as input to a lightweight time-series prediction model. This model employs a gated recurrent unit (GRU) architecture and runs on an edge processing unit. For example, the sliding window captures local temporal patterns with limited computing resources. The edge processing unit is an integrated processor or an equivalent embedded platform. The single inference latency is limited to no more than 100 milliseconds when running the lightweight GRU model. The default initial values ​​for the window length and step size are set to 30 and 50 sampling points, respectively, and can be adjusted within the range of 10 to 100 sampling points and 1 to 20 sampling points, respectively, based on field verification. The model input consists of preprocessed features arranged chronologically within the window, and the output is an immediate estimate of the demulsification process characteristics or an estimate of the remaining time when a threshold is reached. Optionally, when the resource utilization of the edge processing unit exceeds a preset upper limit, the step size is increased to reduce the inference frequency while maintaining the window length at a minimum.

[0058] Step S4: Output the estimated demulsification rate. The estimated demulsification rate is defined as a quantitative characterization index of the sample's demulsification process, including at least one of the following: the estimated time required for the sample to reach a preset demulsification threshold, or the estimated demulsification rate of the sample at the measurement moment. Further, the estimated time to reach the threshold is used to guide the workable window for paving or compaction, and the rate estimate is used to determine the speed and stability of the demulsification stage. The default threshold is obtained from historical reference tests or short-term field calibration and can be verified according to material batches or environmental changes. If both are calculated simultaneously, the estimated time is displayed as the primary indicator, with the rate as an auxiliary indicator. If the threshold is not set or is under calibration, only the rate estimate is output, and the threshold status is indicated. In abnormal situations, when the model fails to converge or the input data quality is below the minimum requirement, the output is set to temporarily unavailable and the last valid value is maintained for no more than 3 seconds.

[0059] Step S5 involves incrementally updating the parameters of the lightweight time-series prediction model based on an online learning mechanism, and displaying the estimated bursting rate in real time on the embedded display interface. Similarly, the display interface updates the estimated value at a second-level refresh rate, simultaneously providing data quality indicators and alarms. The default refresh cycle is 1 second, which can be adjusted within the range of 500 milliseconds to 5 seconds according to resource consumption. When low data quality or model confidence is detected as being below a threshold, the interface prominently displays a warning and pauses construction suggestions based on the result. To avoid frequent interface jitter, a dual-condition trigger update is used, with a minimum continuous display time of one second and a change threshold of 5%.

[0060] In one embodiment, temperature compensation includes temperature normalization of the conductivity sequence based on the temperature measured by a temperature sensor, and enabling compensation updates when the ambient temperature changes rapidly.

[0061] In one embodiment, the sliding window has a configurable window length and step size, the window length being used to cover key stages of the demulsification process, and the step size being used to control the inference frequency and adaptively adjusted based on data stability.

[0062] The adaptive adjustment steps for window length and step size include:

[0063] a) at time For the preprocessed conductivity and temperature compensation sequences, the short window variance, normalized change point score and normalized value of short-term prediction residual are calculated to form an input metric set for subsequent mapping.

[0064] b) Based on the comprehensive stability index, the window length and step size are adjusted in conjunction, so that the window is lengthened and the step size is increased during the stable phase, and the window is shortened and the step size is reduced during the fluctuating phase.

[0065] ,

[0066] ,

[0067] in, Indicates time The window length, in units of sample size. Indicates time The step size, in units of sample size. This indicates the lower bound of the window length, expressed in units of the number of samples. This indicates the upper bound of the window length, expressed in units of the number of samples. Represents the exponential compression factor, a dimensionless constant. Indicates time The stability index is dimensionless. This indicates the lower bound of the step size, expressed in units of sample size. The ratio of step size to window length is a dimensionless constant. This represents the floor operator;

[0068] In the formula:

[0069] ,

[0070] in, Indicates time The stability index is dimensionless. Indicates time The short-window variance calculated within the most recent fixed sample segment, in units of , This represents the reference value for stable baseline variance, in units of [unit name missing]. , Indicates time The variable point score, normalized to Dimensionless Indicates time The normalized value of the one-step prediction residual, dimensionless. , , Let each of the three weights be a non-negative dimensionless constant and satisfy the following conditions: ;

[0071] c) When the relative change of the stability index exceeds the hysteresis threshold, or when the short-term prediction error shows a significant decrease in the candidate combination, perform parameter updates; otherwise, remain inert to suppress jitter.

[0072] d) Apply upper and lower bounds and inference time constraints to the candidate window length and step size. If they are close to the resource limit, prioritize the combination with a larger step size to control the call frequency.

[0073] e) Use the updated window length and step size for sliding window slicing, enter edge model inference, and feed the latest residuals back into step a) for the next round of evaluation;

[0074] The execution details and criteria for the adaptive adjustment steps are as follows:

[0075] Trigger: If relative change Update if hysteresis exceeds the set upper threshold. If it falls below the lower threshold, it remains unchanged;

[0076] Candidate search: Compare the rolling mean of short-term prediction errors in a discrete set; if the improvement is less than the minimum improvement, do not switch combinations.

[0077] constraint: , When runtime latency is close to the limit, prioritize increasing it. To increase the step size, in this embodiment, the lower and upper bounds of the window length are determined by a combination of on-site computing power and target response speed, and the lower bound of the step size is limited by the smallest resolvable time scale. Default values ​​can be derived from offline stress tests and experimental data fitting before deployment, with an average utilization rate not exceeding 80% of the edge processing unit as the configuration basis. For obtaining stability indicators, the short window variance is based on the statistics of the latest segment of fixed length, the change point score is output by the online change detection algorithm and linearly normalized, and the prediction residual comes from the difference between the model's one-step prediction and the actual observation and is normalized according to the recent scale. Optionally, when resource usage continues to exceed the upper limit, the adaptive window length is temporarily frozen, retaining only the coarse-grained adjustment of the step size; if the latency constraint still cannot be met, the inference frequency is reduced to no less than the minimum safe frequency and a rate-limiting status is indicated.

[0078] Specifically, the mapping adopts an exponential decay structure, enabling the window length to shrink rapidly in response to unstable signals, while smoothly lengthening within the stable range, balancing reaction speed and stability. The step size is linked to the window through scaling and rounding mechanisms, achieving a controllable inference frequency at the edge with constant-level computational cost. The metric aggregation introduces three elements: variance, change point score, and prediction residual, reflecting statistical fluctuations, structural abrupt changes, and model fit, respectively, thereby avoiding distortion by a single indicator. The trigger strategy sets upper and lower hysteresis to reduce jitter caused by frequent switching, and is supplemented by error minimization search of discrete candidate sets to ensure that switching only occurs when real benefits appear. Resource and time constraints are implemented throughout the entire process, and the step size is increased to reduce computational load under stress conditions through upper and lower bounds and scaling coefficient adjustments.

[0079] In one embodiment, the online learning mechanism is triggered when data distribution drift is detected or the model confidence is lower than a threshold. It adopts incremental gradient updates with regularization constraints and sets upper limits on the learning rate and the number of updates to control energy consumption and stability.

[0080] The steps for incremental gradient updates with regularization constraints include:

[0081] f) Calculate the two types of signals after inference: drift score With uncertainty ;when Above the threshold or If the value is above the threshold, the incremental update process begins; if the value is below the threshold at the same time, the update is skipped, and only the monitored value is recorded.

[0082] g) Sample size from the nearest window segment. The samples are used to form incremental batches, avoiding a full rescan of historical data. Indicates the incremental batch size;

[0083] h) Perform one or more gradient steps with an objective function that includes a penalty term, and apply the regularization term to the joint parameter set of the GRU weights and biases.

[0084] ,

[0085] , ,

[0086] in, Indicates the trigger sequence number The incremental objective function is dimensionless. Indicates predicted value Compared with the labeled value The baseline loss between (preferably smoothed) or ), dimensionless, Indicates the trigger sequence number The set of trainable parameters for the GRU and its fully connected output layer, according to norm Weighted, This represents the weight decay coefficient, which is dimensionless. Indicates the trigger sequence number Measured data distribution drift score (after...) (Normalized), dimensionless, The drift penalty weight is dimensionless. Indicates the trigger sequence number Predictive uncertainty score (by (Normalized), dimensionless, The uncertainty penalty weight is dimensionless. This represents the set of parameters after one gradient step. Indicates the trigger sequence number The learning rate used, Indicates the upper limit of the learning rate. Indicates the basic learning rate. This represents the amplification factor of drift with respect to the learning rate; it is dimensionless. This represents the inhibition coefficient of the error with respect to the learning rate; it is dimensionless. This represents the dimensionless mean of the most recently predicted residuals calculated using the exponential sliding method. Represents the gradient with respect to the set of parameters.

[0087] i) Employ error-adaptive learning rate and cover Upper limit; the number of gradient steps within a single trigger does not exceed , This indicates the maximum number of gradient steps within a single trigger.

[0088] j) Calculate the validation loss using an independent sliding validation subset. If the improvement relative to the baseline before triggering is insufficient to reach the threshold, roll back the parameters; if the improvement reaches the threshold, solidify the data. Specifically, solidification means marking the current parameter set as the valid version and recording the version number and timestamp for subsequent rollback or comparison. The default validation set uses the sliding subset of the most recent one to three windows, and the validation cycle is triggered synchronously with the incremental update. To ensure the independence of online validation, the validation subset does not overlap with the incremental samples used for updates. If overlap is unavoidable, overlapping samples are removed first, and subsequent samples are added. If a slight improvement occurs but does not reach the threshold, maintain the existing parameters and extend the observation period by no more than two minutes; if multiple consecutive triggers fail to improve the data, automatically reduce the learning rate and increase the trigger threshold to reduce invalid updates.

[0089] The implementation steps for this step are as follows:

[0090] Trigger threshold and order: Setting ; These are the drift / uncertainty thresholds for entering incremental updates. or Then proceed to step h);

[0091] Regular expression object: It also includes the weights and biases of each gate in the GRU, as well as the weights and biases of the output layer, and applies them uniformly. Penalties are imposed to prevent imbalance caused by suppressing only a single sub-block;

[0092] Update count limit: The number of gradient steps within a single trigger shall not exceed [number missing]. If the validation set improvement has reached the threshold before the number of steps is reached. Stop early This indicates the minimum improvement required to stop the process early;

[0093] Rollback strategy: Save For snapshot purposes, if the online validation loss does not decrease sufficiently relative to the baseline... If an increase occurs, roll back to the snapshot. This indicates the minimum improvement required for rollback criteria; consecutive rollbacks exceeding [a certain threshold]... At that time, a temporary upgrade was implemented. and reduce Entering a period of steady observation, Indicates the upper limit of the count for consecutive rollbacks;

[0094] Temperature correlation: When the ambient temperature changes significantly over a short period of time, it can be detected through... The estimate incorporates the temperature uncertainty component and is updated by step f). ;

[0095] Specifically, this incremental learning design closes the triggering, objective, stepping, and safeguard processes within a single triggering cycle. The triggering side uses drift and uncertainty as dual thresholds, allowing it to enter the adaptive phase when distribution changes or model confidence decreases. The objective function, in addition to the baseline error, incorporates three penalties: weight decay, drift, and uncertainty, ensuring updates closely reflect real-world data while avoiding overfitting and blindly following noise. The learning rate employs an adaptive approach that combines error suppression and drift amplification, further enhanced by a global upper limit, enabling faster convergence during significant changes and robustness during periods of rising error. An upper limit on the number of triggers and early stopping together limit the energy consumption and latency of a single trigger. Online validation and rollback act as safety valves, preventing accidental noise batches from guiding the model in the wrong direction for fine-tuning. The observation period for continuous rollback triggers further reduces over-response to unstable data segments. For example, during the observation period, only monitoring and recording are performed, without changing model parameters. Monitoring indicators include the moving average and variance of the prediction residuals, uncertainty estimation, and the time series of resource utilization. The default length of the observation period is between 30 and 120 seconds, which is adaptively adjusted according to the construction schedule and data stability. If the data stabilizes and the verification loss falls back to near the baseline during the observation period, the rollback count is cleared and regular triggering is resumed; if the situation continues to deteriorate, a degradation prompt is output and manual verification of the sample or sensor status is recommended.

[0096] In one embodiment, the demulsification threshold is set by reference test results or field calibration results of historical samples and can be updated according to the batch of construction materials.

[0097] Optionally, when updating a batch, the batch identifier and corresponding threshold are recorded. The threshold source includes indoor reference results or rapid on-site comparison tests, prioritizing data consistent with or similar to the current environmental conditions as reference. The default review period is at the beginning of each new batch and before the first operation after significant changes in environmental conditions. New thresholds can only be enabled after the review is passed. When a threshold cannot be reliably determined, the system retains the previous valid threshold and provides a trend reference only based on rate estimation.

[0098] In one embodiment, preprocessing further includes baseline estimation and drift detection for the initial conductivity segment. When long-term drift is detected, baseline reestimation is performed, and online updates are paused during the reestimation period. Further, the default triggering conditions for baseline reestimation include the establishment of a long-term drift determination or a significant shift in the statistical characteristics of the initial segment. The reestimation window covers several recent stable segments to improve robustness. During the reestimation period, model parameters and window configuration are frozen, with a minimum freeze time of 10 seconds and no more than 60 seconds, determined by a combination of data stability and resource consumption. If instability is still determined after the freeze period ends, the freeze is extended by one iteration, and a prompt is displayed to check sensor contact and sample continuity.

[0099] In one embodiment, the lightweight time series prediction model is an edge model consisting of one or two layers of gated recurrent units and a fully connected output layer, and the number of model parameters and inference latency are limited to not exceeding a preset resource threshold.

[0100] In one embodiment, the lightweight time series prediction model is pre-trained by transfer learning before deployment: the model weights are initialized with historical demulsification data under multiple environmental conditions and fine-tuned on a small amount of data collected from the target construction scenario.

[0101] In one embodiment, the embedded display interface displays: an estimated demulsification rate, a demulsification status indicator, a data quality indicator, and alarm prompts, and supports providing prompts from the paving / compacting machine based on the estimated values;

[0102] Similarly, timing prompts are based on rule mappings of estimated time or rate thresholds, and by default use a combination of three-color status indicators and text prompts. Prompt refresh is synchronized with estimate refresh, and the minimum prompt duration is no less than one second to ensure readability. If data quality or model confidence is low, the prompt automatically downgrades to a "please review" status, without providing specific operational suggestions. When the system is in resource-constrained or rate-limited mode, only status indicators and critical alarms are retained, while auxiliary information is paused to reduce load.

[0103] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for determining the demulsification rate of emulsified asphalt as described in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0105] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for determining the demulsification rate of emulsified asphalt, characterized in that, include: Step S1: Collect time-series physical parameter data of emulsified asphalt samples using a portable sensing unit; The time-series physical parameter data includes at least electrical conductivity; Step S2: Perform preprocessing on the time-series physical parameter data, including temperature compensation and baseline normalization; Step S3: The preprocessed sequence is segmented using a sliding window and used as input to a lightweight time series prediction model. The lightweight time series prediction model adopts a gated recurrent unit structure and runs on an edge processing unit. Step S4: Output the demulsification rate estimate, which is defined as: a quantitative characterization index of the sample demulsification process, including at least one of the following: the estimated time required for the sample to reach the preset demulsification threshold, or the estimated demulsification rate of the sample at the measurement time. Step S5: Update the parameters of the lightweight time series prediction model incrementally based on the online learning mechanism, and display the estimated value of the demulsification rate in real time on the embedded display interface; The sliding window has a configured window length and step size. The window length is used to cover the key stages of the demulsification process, and the step size is used to control the inference frequency and is adaptively adjusted based on data stability. The adaptive adjustment steps for the window length and step size include: a) at time For the preprocessed conductivity and temperature compensation sequences, the short window variance, normalized change point score and normalized value of short-term prediction residual are calculated to form the input metric set; b) Based on the comprehensive stability index, the window length and step size are adjusted in conjunction, so that the window is lengthened and the step size is increased during the stable phase, and the window is shortened and the step size is reduced during the fluctuating phase. c) When the relative change of the stability index exceeds the hysteresis threshold, or when the short-term prediction error shows a significant decrease in the candidate combination, perform parameter updates; otherwise, remain inert. d) Apply upper and lower bounds and inference time constraints to the candidate window length and step size. If they are close to the resource limit, prioritize the combination with the larger step size. e) Use the updated window length and step size for sliding window slicing, enter edge model inference, and feed the latest residuals back into step a) for the next round of evaluation; The online learning mechanism is triggered when data distribution drift is detected or the model confidence is lower than the threshold. It adopts incremental gradient update with regularization constraints and sets an upper limit on the learning rate and an upper limit on the number of updates. The step of using incremental gradient update with regularization constraints includes: f) Calculate the two types of signals after inference: drift score With uncertainty ;when Above the threshold or If the value is above the threshold, the incremental update process begins; if the value is below the threshold at the same time, the update is skipped, and only the monitored value is recorded. g) Sample size from the nearest window segment. The samples are used to form incremental batches. Indicates the incremental batch size; h) Perform one or more gradient steps with an objective function that includes a penalty term, and apply the regularization term to the joint parameter set of the GRU weights and biases. i) Employ error-adaptive learning rate and cover Upper limit; the number of gradient steps within a single trigger does not exceed , This indicates the maximum number of gradient steps within a single trigger. j) Calculate the validation loss using an independent sliding validation subset. If the improvement relative to the baseline before triggering is less than the threshold, roll back the parameters; if the improvement reaches the threshold, solidify the parameters.

2. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The temperature compensation includes temperature normalization of the conductivity sequence based on the temperature measured by the temperature sensor, and enabling compensation updates when the ambient temperature changes rapidly.

3. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, In the adaptive adjustment steps of the window length and step size: Based on the comprehensive stability index, the window length and step size are adjusted in tandem, so that the window is lengthened and the step size is increased during the stable phase, and the window is shortened and the step size is decreased during the fluctuating phase. ， ， in, Indicates time The window length, in units of sample size. Indicates time The step size, in units of sample size. This indicates the lower bound of the window length, expressed in units of the number of samples. This indicates the upper bound of the window length, expressed in units of the number of samples. Represents the exponential compression factor, a dimensionless constant. Indicates time The stability index is dimensionless. This indicates the lower bound of the step size, expressed in units of sample size. The ratio of step size to window length is a dimensionless constant. This represents the floor operator; In the formula: ， in, Indicates time The stability index is dimensionless. Indicates time The short-window variance calculated within the most recent fixed sample segment, in units of , This represents the reference value for stable baseline variance, in units of [unit name missing]. , Indicates time The variable point score, normalized to Dimensionless Indicates time The normalized value of the one-step prediction residual, dimensionless. , , Let each of the three weights be a non-negative dimensionless constant and satisfy the following conditions: ; The execution details and criteria of the adaptive adjustment step are as follows: Trigger: If relative change Update if hysteresis exceeds the set upper threshold. If it falls below the lower threshold, it remains unchanged; Candidate search: Compare the rolling mean of short-term prediction errors across a discrete set; if the improvement is less than the minimum improvement, do not switch combinations. constraint: , When runtime latency is close to the limit, prioritize increasing it. .

4. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, In the step of using incremental gradient update with regularization constraints: With one or more gradient steps performed using an objective function that includes a penalty term, the regularization term is applied to the joint parameter set of the GRU weights and biases in the following way: ， ， ， in, Indicates the trigger sequence number The incremental objective function is dimensionless. Indicates predicted value Compared with the labeled value The baseline loss between them is dimensionless. Indicates the trigger sequence number The training parameter set of the GRU and its fully connected output layer, according to norm Weighted, This represents the weight decay coefficient, which is dimensionless. Indicates the trigger sequence number The measured data distribution drift score is dimensionless. The drift penalty weight is dimensionless. Indicates the trigger sequence number The prediction uncertainty score, dimensionless. The uncertainty penalty weight is dimensionless. This represents the set of parameters after one gradient step. Indicates the trigger sequence number The learning rate used, Indicates the upper limit of the learning rate. Indicates the basic learning rate. This represents the amplification factor of drift with respect to the learning rate; it is dimensionless. This represents the inhibition coefficient of the error with respect to the learning rate; it is dimensionless. This represents the dimensionless mean of the most recently predicted residuals calculated using the exponential sliding method. This represents the gradient with respect to the set of parameters.

5. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The demulsification threshold is set based on reference test results or on-site calibration results of historical samples and is updated according to the batch of construction materials.

6. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The preprocessing also includes baseline estimation and drift detection for the initial conductivity segment; when long-term drift is detected, baseline re-estimation is performed and online updates are paused during the re-estimation.

7. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The lightweight time series prediction model is an edge model consisting of one or two layers of gated loop units and a fully connected output layer, and the number of model parameters and inference latency are limited to not exceeding a preset resource threshold.

8. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The lightweight time-series prediction model is pre-trained through transfer learning before deployment: the model weights are initialized with historical demulsification data under multiple environmental conditions, and fine-tuned on a small amount of data collected from the target construction scenario.

9. The method for determining the demulsification rate of emulsified asphalt as described in claim 1, characterized in that, The embedded display interface includes: estimated demulsification rate, demulsification status indication, data quality indication and alarm prompts, and supports providing prompts from the paving / compacting machine based on the estimated values.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for determining the demulsification rate of emulsified asphalt as described in any one of claims 1 to 9.

Images