Multi-source engineering process target index intelligent control method based on deep learning

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a multi-source engineering process target index intelligent control method based on deep learning. The method comprises the following steps: step 1, collecting multi-source sensing data of a target engineering process execution device, and constructing a process time sequence sample set containing a multi-source feature vector and a corresponding target index value; 2, forming a target index dynamic prediction model by utilizing the process time sequence sample set and combining a process state coding module and a target index prediction decoding module; 3, dynamically predicting the model based on a target index; and 4, executing the generated process control sequence, collecting real-time data, obtaining a real-time target index value, and performing incremental updating on the target index dynamic prediction model by using the formed online correction sample. According to the method, energy consumption can be reduced, dependence on human experience is reduced, and the intelligence and automation level of the engineering process is remarkably improved on the whole.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent control method for target indicators of multi-source engineering processes based on deep learning. Background Technology

[0002] In industrial settings such as engineering construction, equipment operation, and material processing, process quality is typically influenced by multiple factors, including equipment operating conditions, raw material status, environmental conditions, and operational strategies, making it a typical dynamic, multivariate coupled system. Traditional process quality control models mostly rely on manual experience to set operating parameters and determine quality compliance through periodic sampling or offline measurement. These methods exhibit significant lag and locality, making it difficult to reflect the evolutionary characteristics of the process over continuous time. Understanding the internal state of the system is also highly dependent on experience, failing to meet the requirements of modern engineering for real-time performance, stability, and automation. With the development of sensor and communication technologies, construction machinery and industrial equipment can collect real-time data on equipment operation and material response using various sensors such as vibration, load, speed, position, and temperature and humidity conditions, making it possible to use data-driven models for process quality prediction. However, existing technologies generally employ statistical regression models or simple machine learning methods. These models can only establish static mapping relationships and are difficult to effectively describe the nonlinear evolution of the system state over continuous time as operational strategies change.

[0003] In recent years, deep learning methods have been applied to industrial process monitoring and quality prediction, such as using convolutional networks to extract features or recurrent networks to analyze time series. However, these methods have certain limitations in engineering practice, mainly manifested in severe structural black-boxing, a lack of interpretability corresponding to actual physical processes, and difficulty in accurately characterizing the subtle changes in the system caused by operating parameters over continuous time. Furthermore, traditional deep learning models often rely on discrete time-step modeling, making it difficult to construct a unified and effective dynamic process model when faced with multi-source sensor data of varying frequencies and dimensionalities. In addition, existing technologies typically employ simple rule-based control or greedy optimization strategies based on single-step predictions for control decisions, failing to achieve overall optimization of operational behavior across multiple future time steps. Since engineering processes generally exhibit cumulative effects, local optimum strategies often lead to a decline in global performance. Summary of the Invention

[0004] The main objective of this invention is to provide an intelligent control method for multi-source engineering process target indicators based on deep learning. By constructing an overall framework that includes process state encoding, a twin model of NeuralODE and Koopman operators, and a target indicator dynamic prediction module, it achieves fine representation and forward-looking prediction of complex engineering process states at multiple time steps. Based on the prediction results, it generates optimized process control sequences. Combined with online correction and incremental updates, the model can continuously adapt to changes in material properties and construction environment. This ensures that the target indicators stably meet design requirements while improving construction efficiency, reducing energy consumption, and reducing reliance on human experience, thus significantly improving the overall intelligence and automation level of the engineering process.

[0005] To address the aforementioned technical problems, this invention provides an intelligent control method for target indicators of multi-source engineering processes based on deep learning, comprising: A deep learning-based intelligent control method for multi-source engineering process target indicators, comprising the following steps: Step 1: Collect multi-source sensor data from the target engineering process execution equipment, obtain target index values, and construct a process time series sample set containing multi-source feature vectors and corresponding target index values; Step 2: Using the process time series sample set, jointly train a NeuralODE process sub-model and a Koopman operator process sub-model, and combine the process state encoding module and the target index prediction decoding module to form a target index dynamic prediction model. Step 3: Based on the dynamic prediction model of the target indicators, generate a process control sequence for the execution of the target engineering process through iterative scoring; Step four: Execute the generated process control sequence, collect real-time data and obtain real-time target indicator values, compare the obtained real-time target indicator values ​​with the predicted target indicator values ​​to form online calibration samples, and use the formed online calibration samples to incrementally update the target indicator dynamic prediction model.

[0006] Further, step one includes: installing a vibration acceleration sensor, wheel load sensor, position acquisition module, velocity acquisition module, and temperature and moisture content acquisition module on a target engineering process execution device, and acquiring vertical vibration acceleration data, wheel load data, position data, velocity data, temperature data, and moisture content data at sampling frequencies of 100Hz, 10Hz, 1Hz, 1Hz, 1Hz, and 1Hz, respectively; measuring the target index values ​​using a nuclear density meter or a load-bearing plate testing device; and dividing the acquired vibration acceleration data into 1-second segments within a process time window. The mean, standard deviation, and extreme values ​​are calculated. The mean of the collected wheel load data and the collected speed data are calculated within the same process time window. The last sample value of the collected temperature and water content data is taken within the same process time window to form a multi-source feature vector. The multi-source feature vector of the same spatial unit is combined with the corresponding target index value in order of pass / fail to form a time series sample. The time series samples are summarized and processed by missing value filling, outlier removal, and numerical scaling to construct a process time series sample set, which is divided into training set, validation set and test set according to the sample size ratio of 7:2:1.

[0007] Further, step two includes: constructing a process state encoding module and a target indicator prediction and decoding module, inputting the multi-source feature vector of each time step into the process state encoding module; the process state encoding module sequentially includes four fully connected layers, the first fully connected layer has the same input dimension as the length of the multi-source feature vector and an output dimension of 128, the second fully connected layer has an input dimension of 128 and an output dimension of 128, the third fully connected layer has an input dimension of 128 and an output dimension of 64, and the fourth fully connected layer has an input dimension of 64 and an output dimension of 1; the activation function of the first to third fully connected layers is a rectified linear unit, and the fourth fully connected layer does not use an activation function; the 64-dimensional output of the third fully connected layer is defined as the process state vector, and the 1-dimensional output of the fourth fully connected layer is defined as the static target indicator prediction value; the first to third fully connected layers constitute the process state encoding module, and the fourth fully connected layer constitutes the target indicator prediction and decoding module.

[0008] Furthermore, step two also includes constructing a NeuralODE process sub-model, which consists of a state input splicing module and a continuous-time evolution network. The state input splicing module splices the process state vector of the current time step with the 16-dimensional control vector in the feature dimension to form a state control vector of length 80. The 16-dimensional control vector is obtained by one-hot encoding of the speed level, vibration mode level, vibration intensity level, and sprinkler switch state of the target engineering process execution equipment and then transforming it through two fully connected layers with output dimensions of 32 and 16 respectively. The continuous-time evolution network includes three fully connected layers, with input dimension 80 and output dimension 128, input dimension 128 and output dimension 128, and input dimension 128 and output dimension 64 respectively. The first and second fully connected layers are activated using rectified linear units, while the third fully connected layer does not use an activation function. The continuous-time evolution network receives the state control vector and outputs a 64-dimensional derivative estimation vector.

[0009] Furthermore, the NeuralODE process sub-model performs numerical integration as follows: each process operation is divided into one time interval, and this time interval is evenly divided into four sub-time intervals; within each sub-time interval, a four-fold function evaluation is performed: (i) the first derivative estimate is obtained using the current process state vector and the current control vector through a continuous-time evolution network; (ii) based on the first derivative estimate, the first intermediate state is obtained by shifting forward half the length of the sub-time interval from the original state, and the second derivative estimate is obtained using the first intermediate state and the current control vector through a continuous-time evolution network; (iii) based on the second derivative estimate, the second derivative estimate is obtained by shifting forward half the length of the sub-time interval from the original state. (iv) The state is shifted forward by half a sub-time interval to obtain the second intermediate state. The second intermediate state and the current control vector are used to obtain the third derivative estimate through a continuous-time evolution network. The second intermediate state is shifted forward by one sub-time interval to obtain the sub-time interval endpoint state. The sub-time interval endpoint state and the current control vector are used to obtain the fourth derivative estimate through a continuous-time evolution network. The first, second, third and fourth derivative estimates are weighted and summed in a ratio of 1:2:2:1 and multiplied by the sub-time interval length to obtain the process state increment. The obtained process state increment is added to the original process state vector to obtain the process state vector at the end of the time interval.

[0010] Furthermore, step two also includes constructing a Koopman operator process sub-model, which includes: an upscaling mapping network, consisting of two fully connected layers, with an input dimension of 64 and an output dimension of 128, and an input dimension of 128 and an output dimension of 128, both activated by rectified linear units, which inputs the process state vector to obtain a 128-dimensional high-dimensional state vector; a linear evolution matrix, which is a 128-row, 128-column trainable weight matrix, used to perform matrix multiplication on the 128-dimensional high-dimensional state vector to obtain the high-dimensional state prediction vector for the next time step; and a downscaling decoding network, consisting of two fully connected layers, with an input dimension of 128 and an output dimension of 128, and an input dimension of 128 and an output dimension of 64, where the first fully connected layer is activated by rectified linear units and the second fully connected layer does not use an activation function, which inputs the high-dimensional state prediction vector to obtain the process state prediction vector for the corresponding time step.

[0011] Furthermore, the joint training in step two includes: calculating the temporal prediction error of the NeuralODE process sub-model and the temporal prediction error of the Koopman operator process sub-model; inputting the process state prediction vectors obtained from the two sub-models into the target index prediction decoding module to obtain the NeuralODE target index prediction value and the Koopman target index prediction value, and calculating the mean square error between the two target index prediction values ​​and the target index value; multiplying the temporal prediction errors of the two sub-models by 0.5 and adding them together, and then adding them to the calculated mean square error multiplied by 0.5 to form a comprehensive error that includes the process state error and the target index error; using the constructed comprehensive error, jointly updating the weight parameters of the first to third fully connected layers of the process state encoding module, the continuous time evolution network, the upscaling mapping network, the linear evolution matrix, the downscaling decoding network, and the target index prediction decoding module.

[0012] Further, step three includes: dividing the prediction time range of the control sequence to be generated into 10 consecutive time steps; limiting the candidate control actions of each time step to the Cartesian product of the speed level set, vibration mode set, vibration intensity set, and sprinkler switch state set; for the first time step, for all candidate control actions, using the initial multi-source feature vector and the current candidate control action as input to the target index dynamic prediction model to obtain the target index prediction value; calculating the average absolute difference between the obtained target index prediction value and the design target index value as the target index deviation index; calculating the weighted sum of the speed level and vibration intensity level of the current candidate control action as the energy consumption index; calculating the weighted sum of the target index deviation index and the energy consumption index as the total score; selecting the control action with the smallest total score as the target control action of the first time step; iterating sequentially to the 10th time step, and at each time step, predicting and scoring based on the process state prediction vector of the previous time step and the candidate control action of the current time step through the target index dynamic prediction model to construct a process control sequence of length 10.

[0013] Furthermore, the incremental update in step four includes: after passing through every 10 spatial units, measuring the real-time target index value of one spatial unit, comparing the measured real-time target index value with the target index prediction value generated by the target index dynamic prediction model to obtain the target index error, and combining the target index error with the multi-source feature vector of the corresponding time step to form an online correction sample; when the number of online correction samples reaches 50, inputting the 50 online correction samples into the target index dynamic prediction model; and performing incremental updates for 5 training cycles with a learning rate of 0.0001 on the third fully connected layer of the process state encoding module, the last fully connected layer of the continuous-time evolution network of the NeuralODE process sub-model, the linear evolution matrix of the Koopman operator process sub-model, and the fully connected layer of the target index prediction decoding module.

[0014] Furthermore, step four also includes: recalculating the process control sequence for the next 10 time steps using the latest parameters after each training cycle of incremental updates; replacing the original control sequence in the process control terminal; and enabling the target engineering process execution equipment to run according to the process control sequence generated by the online corrected target index dynamic prediction model.

[0015] The intelligent control method for multi-source engineering process target indicators based on deep learning proposed in this invention has the following beneficial effects: By combining multi-source sensor data, deep learning modeling methods, and continuous-time dynamic description, this invention achieves dynamic, high-precision prediction and intelligent control of engineering process target indicators, which has significant advantages over traditional quality control methods that rely on manual experience and offline detection. This invention employs a process state encoding mechanism based on time-series samples, enabling the model to extract key state features reflecting material evolution and equipment operating condition changes from high-dimensional heterogeneous data, thus overcoming the shortcomings of traditional static feature methods that cannot reflect process dynamics. Simultaneously, this invention jointly introduces the NeuralODE process sub-model and the Koopman operator process sub-model, enabling the system to simultaneously obtain nonlinear continuous-time modeling capabilities and linear analytical state evolution capabilities. The complementary use of these two models improves the stability of state prediction for complex engineering processes. Furthermore, the dynamic prediction model for target indicators proposed in this invention can perform forward-looking extrapolation over multiple future time steps, allowing control strategy generation to move beyond single-step feedback and perform comprehensive scoring and optimization over the time series dimension. This ensures that the generated process control sequence guarantees the quality of target indicators while also considering construction efficiency and energy consumption performance. This invention also incorporates an online correction and incremental update mechanism, enabling the model to automatically adapt to changes in material properties and environmental conditions based on real-time feedback during construction, continuously maintaining prediction accuracy and control performance, and solving the problem of easy degradation in existing models. Based on the above design, this invention achieves a complete closed loop from data acquisition, state characterization, dynamic prediction to control decision-making, enabling intelligent, refined, and adaptive process control in various engineering construction scenarios. It significantly improves the stability of target indicators, reduces the risk of over-operation, and enhances construction efficiency, providing an effective solution for intelligent management and control of engineering quality. Attached Figure Description

[0016] Figure 1 A schematic diagram of the fourth-order Runge-Kutta numerical integration process of NeuralODE provided in an embodiment of the present invention; Figure 2 A schematic diagram of the three-dimensional evolution trajectory of the process state vector in continuous time steps, provided for an embodiment of the present invention; Figure 3 This is a grayscale image showing the spatial unit compaction distribution provided in an embodiment of the present invention. Detailed Implementation

[0017] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0018] A deep learning-based intelligent control method for multi-source engineering process target indicators, comprising the following steps: Step 1: Collect multi-source sensor data from the target engineering process execution equipment, obtain target index values, and construct a process time series sample set containing multi-source feature vectors and corresponding target index values.

[0019] In one embodiment of this method, step one involves systematically collecting and processing multi-source sensor data from the target engineering process execution device, and combining this data with the target index value to construct a process time series sample set containing multi-source feature vectors and corresponding target index values. This enables the subsequent dynamic prediction model of the target index to not only reflect the instantaneous state of a single process operation, but also to reflect the evolution of the same spatial unit under the action of multiple consecutive process operations.

[0020] In one specific implementation, taking the intelligent control of compaction degree of fly ash roadbed composite structure as an example, the target engineering process execution equipment is the compaction construction machinery, which can be a road roller. First, a vibration acceleration sensor, a wheel load sensor, a position acquisition module, a speed acquisition module, and a temperature and moisture content acquisition module are installed on one of the target engineering process execution devices. The vibration acceleration sensor is fixed to the vibrating wheel frame of the road roller to collect vertical vibration acceleration data; the wheel load sensor is arranged on the axle or wheel structure to reflect the change in contact load between the wheel and the fly ash roadbed composite structure; the position acquisition module can use a satellite positioning device combined with an inertial measurement device to obtain the spatial position of the target engineering process execution device on the construction section; the speed acquisition module can calculate the forward speed using wheel speed signals or hydraulic system flow signals; the temperature and moisture content acquisition module can include an infrared temperature sensor and a contact-type moisture content sensor to characterize the temperature and moisture content of the fly ash mixture surface.

[0021] To ensure the collected data fully reflects the dynamic characteristics of the compaction process while avoiding the storage and processing burden of redundant data, vibration acceleration data is acquired at a high sampling frequency, such as 100 samples per second. This sampling frequency is sufficient to cover the main vibration frequency and harmonic components of the roller's vibration system, preserving the peak values, periodicity, and impact characteristics of the vibration acceleration signal. Wheel load data is acquired at a sampling frequency of 10 samples per second. This captures the slow fluctuations in load caused by changes in subgrade stiffness during compaction without generating excessive redundant data. Location, velocity, temperature, and moisture content data can be acquired at a sampling frequency of 1 sample per second. This generates one record per second, which is sufficient to accurately describe the trajectory, speed, and environmental and material conditions, while also facilitating subsequent data alignment and segmentation by second.

[0022] At the construction site, the road section is spatially divided along the roller's direction of travel. This can be done by dividing the section into multiple adjacent spatial units, for example, assigning a spatial unit number every 1 meter. Additionally, the width and lateral position within each spatial unit can be recorded as additional attributes to facilitate later analysis of lateral differences. When the equipment passes through a spatial unit, sensors continuously collect multi-source data within the corresponding time range. To ensure consistency with the actual construction process, the time range from the roller's entry into the spatial unit to its exit can be defined as a process time window. The actual length of this window depends on the roller's speed; for example, at a speed of 2 meters per second, the process time window for a 1-meter spatial unit is approximately 0.5 seconds. When the roller performs multiple operations within the same spatial unit, each operation corresponds to a separate process time window.

[0023] To obtain the target index values, within each spatial unit, after completing the preset minimum number of process operations, the compaction degree can be measured at a representative location of that spatial unit using a nuclear density meter, or indirectly calculated using a bearing plate test device through the relationship between deformation and load. The advantages of a nuclear density meter are its fast measurement speed, obtaining density and moisture content at a location within minutes, and its penetration depth is suitable for the typical thickness of fly ash subgrade composite structures. While the bearing plate test device takes slightly longer, it provides indicators related to bearing capacity, helping to supplement and verify the nuclear density meter measurement results. In practical implementation, the nuclear density meter can be used preferentially for measurements in most spatial units, while the bearing plate test device can be used for verification at key locations or locations exhibiting significant anomalies. This approach ensures both measurement accuracy and construction efficiency.

[0024] To facilitate the subsequent construction of multi-source feature vectors, the collected vibration acceleration data needs to be segmented and calculated within each process time window. Specifically, the process time window can be segmented in 1-second increments. When the process time window length is less than 1 second, it can still be divided into 1-second time scales, with periods less than 1 second considered as one segment. When the process time window length is greater than 1 second, it is divided into multiple segments of 1 second each, and statistics are calculated for each segment. For the vibration acceleration data within each 1-second segment, its arithmetic mean, standard deviation, and extreme values ​​are calculated. The arithmetic mean reflects the overall level of vibration intensity within this time period, the standard deviation reflects the degree of fluctuation in vibration intensity within this time period, and the extreme values ​​reflect the maximum or minimum impact response that occurs during compaction. The advantage of this approach is that it compresses the originally numerous raw vibration acceleration sampling points into a small number of statistical features, preserving important amplitude and fluctuation information about vibration behavior while significantly reducing the dimensionality of the input data, making subsequent deep learning training more stable.

[0025] The aforementioned vibration acceleration statistical features, wheel load data average, velocity data average, temperature representative value, and water content representative value are concatenated in a fixed order to form a multi-source feature vector. To ensure the stability of subsequent deep learning training, the meaning and order of each dimension in the multi-source feature vector need to remain completely consistent. For example, it can be agreed that the first three dimensions correspond to the arithmetic mean, standard deviation, and extreme values ​​of vibration acceleration; the next dimension corresponds to the wheel load data average; the following dimension corresponds to the velocity data average; and the last two dimensions correspond to the temperature representative value and the water content representative value. This method, through this fixed-order definition, ensures that the same dimension has a consistent physical meaning across different spatial units and different process operation cycles, avoiding the model's inability to learn effective patterns due to inconsistent data arrangement.

[0026] For the same spatial unit, when the target engineering process execution equipment performs multiple process operations on that spatial unit, a multi-source feature vector can be constructed for each process operation and arranged in order of the number of operations. In this way, the multi-source feature vectors of the same spatial unit at the 1st, 2nd, and subsequent processes will form a sequence arranged in chronological order. The corresponding target index values ​​are then appended to this sequence to form a time-series sample. For cases where the target index values ​​are measured using a nuclear density meter, these values ​​can be used as reference results for that spatial unit at that construction stage. Generally, the target index values ​​can be measured once after a preset number of operations, or after a specific number of operations, such as once after the 3rd and 5th operations. In this case, multiple time-series samples can be constructed, each using the target index values ​​at different measurement times as the output label of the sequence. This method, through this approach, allows the process time-series sample set to include both compaction labels at a single target time and multi-target labels from different stages, thereby increasing the diversity of the training samples.

[0027] After missing value imputation and outlier removal, the multi-source feature vectors need to be numerically scaled. Since the numerical ranges and dimensions of vibration acceleration statistics, wheel load data, velocity data, temperature data, and water content data differ, directly inputting them into a deep learning model can easily lead to features with larger numerical ranges having an excessive impact on model training, while features with smaller numerical values ​​are weakened. To address this issue, one approach is to linearly scale each feature dimension to a numerical range between 0 and 1 based on its minimum and maximum values ​​in the training set, or to standardize it by subtracting the mean and dividing by the standard deviation. Regardless of the method used, applying the same scaling rule to the same feature dimension across all samples in the entire dataset ensures that features of different dimensions are numerically similar. This allows the deep learning model to focus on multiple features simultaneously during training, without one feature's excessively large numerical range masking the role of other features.

[0028] After processing, the time-series samples of all spatial units are aggregated to form a process time-series sample set. To ensure the consistency of statistical distribution during training, all time-series samples can be randomly shuffled to ensure a uniform distribution of samples from different spatial units and construction stages within the set. Then, the process time-series sample set is divided into a training set, a validation set, and a test set according to a 7:2:1 ratio of sample size. The training set is used to fit the parameters of the dynamic prediction model for the target index, the validation set is used to monitor model performance and prevent overfitting during training, and the test set is used to independently evaluate the model's generalization ability after training. This ratio balances the amount of training data and evaluation data, conforms to common practices in deep learning, and facilitates stable and reliable model training and evaluation in subsequent steps.

[0029] Step 2: Using the process time series sample set, jointly train a NeuralODE process sub-model and a Koopman operator process sub-model, and combine them with the process state encoding module and the target index prediction decoding module to form a target index dynamic prediction model.

[0030] In one embodiment of this method, after the process time series sample set is constructed, a NeuralODE process sub-model and a Koopman operator process sub-model are jointly trained using the process time series sample set. Combined with the process state encoding module and the target index prediction decoding module, a target index dynamic prediction model is formed to describe the change law of compaction state of fly ash roadbed composite structure with time under continuous process operation and the corresponding compaction degree output.

[0031] First, the process state encoding module and the target index prediction and decoding module are constructed and pre-trained. The process state encoding module receives multi-source feature vectors from each time step in the process time series sample set, and inputs these multi-source feature vectors sequentially into four fully connected layers. The first fully connected layer has the same input dimension as the length of the multi-source feature vector and an output dimension of 128; the second fully connected layer has an input dimension of 128 and an output dimension of 128; the third fully connected layer has an input dimension of 128 and an output dimension of 64; and the fourth fully connected layer has an input dimension of 64 and an output dimension of 1. The first to third fully connected layers use rectified linear units as activation functions, while the fourth fully connected layer does not use an activation function. The 64-dimensional output of the third fully connected layer serves as the process state vector generated by the process state encoding module, and the 1-dimensional output of the fourth fully connected layer serves as the static target index prediction value generated by the target index prediction and decoding module.

[0032] In the pre-training phase, each multi-source feature vector in the process time-series sample set is paired with its corresponding target index value to form a static mapping training sample. The multi-source feature vector is input into the process state encoding module and the target index prediction decoding module to obtain a static target index prediction value. This prediction value is then compared with the target index value, and the mean squared error (MSE) between the static MSE prediction value and the target index value is calculated. The average of the MSEs of all training samples is used as the loss metric. An optimization strategy with a batch size of 64 and a learning rate of 0.001 is employed. The weights and bias parameters of the four fully connected layers are iteratively updated in the deep learning training environment until the average MSE on the validation set decreases by less than a preset threshold over 20 consecutive training epochs, at which point training stops. Through this pre-training process, the process state encoding module learns a way to compress the multi-source feature vector into a 64-dimensional process state vector, which retains key information that interprets the target index value. Simultaneously, the target index prediction decoding module learns the mapping relationship to restore the process state vector to the compaction estimate. The advantage of this approach is that the subsequent NeuralODE process sub-model and Koopman operator process sub-model only need to focus on the change of the process state vector over time, without having to directly process the original multi-source feature vectors. This significantly reduces the dimensionality that needs to be processed and allows the time evolution to take place in a pre-trained compacted state space, making it easier to obtain smooth and stable dynamic features.

[0033] Based on this, a NeuralODE process sub-model is constructed to characterize the nonlinear continuous evolution of the process state vector under short time scales and local operating conditions. The NeuralODE process sub-model consists of state input concatenation and a continuous-time evolution network. For each time step in the process time series sample set, the 64-dimensional process state vector of that time step is first concatenated with the 16-dimensional control vector representing the current control command in the feature dimension to form a state control vector of length 80. The 16-dimensional control vector is obtained as follows: the speed level, vibration mode level, vibration intensity level, and sprinkler switch state of the target engineering process execution equipment are individually encoded, where each discrete level is represented by a fixed-length Boolean vector. Then, these Boolean vectors are concatenated to form an intermediate control vector. To avoid an overly sparse control space due to direct use of Boolean vectors, the intermediate control vector can be sequentially input into two fully connected layers. The first fully connected layer outputs a dimension of 32, and the second fully connected layer outputs a dimension of 16. Nonlinearity is introduced through the rectified linear unit activation function, ensuring that the 16-dimensional control vector not only reflects the gear position itself but also implicitly represents the similarity between different gear combinations. Encoding control commands in this way makes similar speed or vibration gears closer in the control vector space, thus making it easier for the NeuralODE process sub-model to learn the impact of control changes on the compaction state evolution during training.

[0034] A continuous-time evolution network receives a state control vector of length 80, passes it sequentially through three fully connected layers, and outputs a 64-dimensional derivative estimation vector. The structure of the three fully connected layers is as follows: the first fully connected layer has an input dimension of 80 and an output dimension of 128; the second fully connected layer has an input dimension of 128 and an output dimension of 128; and the third fully connected layer has an input dimension of 128 and an output dimension of 64. The first and second fully connected layers use rectified linear unit activation functions to introduce nonlinearity, while the third fully connected layer does not use an activation function, so that the output 64-dimensional derivative estimation vector can vary freely in both positive and negative directions. The derivative estimation vector can be understood as the trend of the process state vector changing per unit time under the combined effects of the current compaction state and the current control command. The advantage of using this continuous-time evolution network is that it can perform numerical integration over time intervals of any length, without being limited to a fixed time step, allowing the same model to adapt to construction processes at different speeds.

[0035] To transform the derivative estimate vector into the actual process state vector change, the NeuralODE process sub-model performs numerical integration within the time interval corresponding to each process operation. Each process operation is divided into one time interval, which is then evenly divided into four sub-time intervals. Within each sub-time interval, the integration process is approximated by a fourth-order function evaluation. Specifically, at the beginning of the sub-time interval, the current process state vector and the current control vector are input into the continuous-time evolution network to obtain the first derivative estimate. Then, the process state vector is shifted forward by half a sub-time interval length along the derivative estimation direction to obtain the first intermediate state. This intermediate state and the current control vector are then input into the continuous-time evolution network to obtain the second derivative estimate. Next, in the same manner, the process state vector is shifted forward by half a sub-time interval length along the second derivative estimation direction to obtain the second intermediate state. This intermediate state and the current control vector are then input into the continuous-time evolution network to obtain the third derivative estimate. Finally, using the second intermediate state as the approximate endpoint state, the process is shifted forward by one sub-time interval length along the direction of the third derivative estimation to obtain the endpoint state of the sub-time interval. This endpoint state is then input into the continuous-time evolution network along with the current control vector to obtain the fourth derivative estimate. These four derivative estimates are weighted and summed in a 1:2:2:1 ratio, and then multiplied by the sub-time interval length to obtain the process state increment within that sub-time interval. The use of four evaluations in a 1:2:2:1 ratio is intended to balance the derivative information at the start, intermediate, and endpoint positions of the sub-time interval while ensuring a moderate computational load. Repeated evaluation of intermediate states reduces local errors caused by using only the start or endpoint derivatives, thus achieving higher overall numerical integration accuracy. This process is executed sequentially for each of the four sub-time intervals, and the process state increment of each sub-time interval is accumulated onto the initial process state vector to obtain the process state vector at the end of the entire process operation. For a time series sample of length 20, the process state vector and corresponding control vector of the first time step can be selected as the initial state and initial control command. The numerical integration process of the above process operation is executed multiple times consecutively to obtain the process state prediction vectors for time steps 2 to 20. These process state prediction vectors are then compared step-by-step with the process state vectors output by the process state encoding module based on the real multi-source feature vectors. The mean square error is calculated and averaged to obtain the time series prediction error of the NeuralODE process sub-model.

[0036] refer to Figure 1This figure illustrates the state evolution of the NeuralODE process submodel during the execution of a fourth-order Runge-Kutta numerical integration. The horizontal axis represents time (t) in seconds (s), with the vertical scale ranging from 0 to 1.0 seconds, incremented every 0.1 seconds. The vertical axis represents the compaction state component (s), ranging from 0 to 2.0, with increments every 0.2 seconds. A thin dashed curve represents the precise evolution trajectory of a certain dimension of the process state vector, serving as a theoretical reference solution. This curve exhibits exponential reaching characteristics, starting from the initial state s = 0.3, gradually increasing with time and approaching a stable value. Based on this precise solution curve, the figure also marks the four-fold function evaluation process performed by the fourth-order Runge-Kutta method within a single time interval. Each time interval is set to 0.25 seconds. Within the first time interval, starting from time t=0: First, a solid black dot is marked at the starting position, corresponding to the k1 evaluation position. This indicates that the first function call is performed on the continuous-time evolution network at the beginning of the sub-interval, obtaining the derivative estimate k1 at this position. An arrow is drawn from this point along the derivative direction. The length and direction of the arrow reflect the influence of the derivative estimate on the state evolution. The arrow points to the upper right, indicating that the compaction state component has a positive growth trend at this position. Then, at time t=0.125, the midpoint of the sub-interval, a hollow dot is marked, corresponding to the k2 evaluation position. The state value at this position is the first intermediate state obtained by shifting the starting state forward by half the sub-interval length along the derivative direction of k1. An arrow is also drawn from this hollow dot, indicating the derivative estimate k2 obtained by performing a second function evaluation in this intermediate state. Immediately afterwards, near the same time t=0.125, a hollow box is marked, corresponding to the k3 evaluation position. The state value at this position is the second intermediate state obtained by shifting forward half a sub-time interval length along the direction of the derivative of k2 from the starting state. An arrow drawn from this hollow box represents the derivative estimate k3 obtained from the third function evaluation. Finally, at time t=0.25 seconds, which is the end of the sub-time interval, a solid triangle is marked, corresponding to the k4 evaluation position. The state value at this position is the end state obtained by shifting forward one sub-time interval length along the direction of the derivative of k3 from the second intermediate state. The fourth function evaluation is performed at this position. The figure clearly labels "k1 starting point," "k2 midpoint 1," "k3 midpoint 2," and "k4 end point" within the first sub-time interval to facilitate understanding the positional relationship of the four evaluations. These four derivative estimates are weighted and combined in a 1:2:2:1 ratio when calculating the final state increment. The subsequent three sub-time intervals (from 0.25 seconds to 1.0 seconds) also follow the above four evaluation process, repeating the calculation of k1, k2, k3, and k4 within each sub-time interval.In this way, the time interval corresponding to the entire process operation is evenly divided into four sub-time intervals. Each sub-time interval is approximated by a fourth-order function evaluation. This approach ensures a moderate computational load while taking into account the derivative information at the start, middle, and end points of the sub-time intervals, thus achieving high overall numerical integration accuracy. The thin dashed line in the upper right corner of the figure is labeled "exact solution," used to compare and illustrate the closeness between the numerical integration results of the RK4 method and the theoretical exact solution.

[0037] Unlike the NeuralODE process sub-model, which focuses on local nonlinear evolution, the Koopman operator process sub-model approximates the overall evolution of compaction states in a high-dimensional observation space, thereby enhancing the stability of long-term predictions. The Koopman operator process sub-model comprises an upscaling mapping network, a linear evolution matrix, and a downscaling decoding network. The upscaling mapping network consists of two fully connected layers: the first fully connected layer has an input dimension of 64 and an output dimension of 128, and the second fully connected layer also has an input dimension of 128 and an output dimension of 128. Both fully connected layers use rectified linear unit activation functions. The process state vector, after being input into the upscaling mapping network, is mapped to a 128-dimensional high-dimensional state vector. The purpose of mapping to a high-dimensional space is to increase the dimension so that the evolution trajectory, which originally exhibits a strongly nonlinear relationship in the 64-dimensional compaction state space, can be approximated as a linear change in the 128-dimensional space. This allows a linear evolution matrix to be used to update the high-dimensional state vector, mitigating the problem of accumulated nonlinear errors in long-term predictions.

[0038] refer to Figure 2This diagram illustrates the projection of the three-dimensional evolution trajectory of the process state vector across continuous time steps. The diagram uses a three-dimensional coordinate system, where the horizontal axis represents the first principal component s1 of the process state vector, the vertical axis represents the second principal component s2, and the vertical axis represents time step n. Since the process state vector is a 64-dimensional vector, to visualize it on a two-dimensional plane, the diagram projects the 64-dimensional process state vector onto the plane formed by the first two principal components s1 and s2 using principal component analysis. Simultaneously, time step n is introduced as a third dimension, forming the evolution trajectory in three-dimensional space. Arrows are labeled on all three axes to indicate the positive direction. The values ​​of the s1 and s2 axes range approximately from -1.2 to 1.2, while the value of the time step n axis ranges from -1 to 1, corresponding to the normalized time step sequence. The axis labels "State Component s1," "State Component s2," and "Time Step n" are labeled in appropriate positions on the coordinate system using bold 22-point font for clear visibility. The figure shows a thick solid line curve representing the complete evolution trajectory of the process state vector from the initial state to the target state. This trajectory exhibits a spiraling upward shape, indicating that the process state vector moves along a spiral path in the s1-s2 plane, while simultaneously rising vertically with increasing time step n. The trajectory begins at time step n=-1, corresponding to the initial state of the compaction process, marked with a solid black dot and labeled "Initial State" near this point. The trajectory terminates at time step n=1, corresponding to the state when the compaction process reaches the target compaction degree, marked with a hollow triangle and labeled "Target State". Hollow dots are marked at several key time step positions along this spiraling upward trajectory; these dots represent state snapshots sampled every four time steps. These sampling points show that as the time step progresses, the spiral radius gradually decreases, indicating that the oscillation amplitude of the process state vector in the s1-s2 plane gradually decays, and the compaction process gradually converges to a stable state. This spiral decay characteristic conforms to the physical law that the compaction effect of fly ash roadbed composite structures gradually accumulates and the increment gradually decreases under continuous multiple process operations. To more clearly illustrate the spatial relationship of the trajectory, the projection of this three-dimensional trajectory onto the bottom plane (the plane of time step n=-1) is also drawn in the figure. This projection is represented by a thin dashed line and also presents a spiral shape, helping readers understand the motion trajectory of the process state vector in the two-dimensional state space. The outer circle of the projected spiral corresponds to the initial time step, and the inner circle corresponds to the later time steps, intuitively reflecting the convergence characteristics of the state evolution. The upper left corner of the figure is labeled with explanatory text "Principal component projection trajectory of 64-dimensional process state vector" and "Showing the dynamic evolution of 20 consecutive time steps," indicating that the figure shows the evolution trajectory of the 64-dimensional process state vector corresponding to each time step after the 20-time series sample has been processed by the process state encoding module, obtained by dimensionality reduction projection.This three-dimensional visualization method allows for an intuitive observation of the dynamic behavior of the NeuralODE process sub-model and the Koopman operator process sub-model when predicting the temporal evolution of the process state vector, thus verifying the model's ability to characterize the continuity, smoothness, and physical rationality of the compaction process.

[0039] The linear evolution matrix is ​​a 128x128 trainable weight matrix used to perform matrix multiplication on the high-dimensional state vector to generate the high-dimensional state prediction vector for the next time step. During training, each element of the linear evolution matrix is ​​optimized so that after multiple consecutive linear transformations of the high-dimensional state vector of the training samples, it can approximate as closely as possible to the high-dimensional representation obtained from the real process state vector through the dimension-upgrading mapping network. The dimension-reduction decoding network consists of two fully connected layers. The first fully connected layer has an input dimension of 128 and an output dimension of 128, using a rectified linear unit activation function. The second fully connected layer has an input dimension of 128 and an output dimension of 64, without using an activation function. The dimension-reduction decoding network restores the high-dimensional state prediction vector to a 64-dimensional process state prediction vector. Overall, the dimension-upgrading mapping network and the dimension-reduction decoding network form a round-trip path between the high-dimensional observation space and the compacted state space, with the linear evolution matrix performing time advancement in the high-dimensional observation space. For a time series sample of length 20, the process state encoding module first converts the multi-source feature vectors of the 20 time steps into 20 process state vectors. Then, the process state vector of the first time step is input into a dimension-upgrading mapping network to obtain the first high-dimensional state vector. Based on this, a linear evolution matrix is ​​applied multiple times to perform 19 consecutive linear transformations on the high-dimensional state vector, successively obtaining the high-dimensional state prediction vectors for the 2nd to 20th time steps. These high-dimensional state prediction vectors are then restored to process state prediction vectors using a dimension-reduction decoding network. The process state prediction vectors for the 2nd to 20th time steps are compared step-by-step with the corresponding time-step process state vectors output by the process state encoding module. The mean square error is calculated and averaged to obtain the time series prediction error of the Koopman operator process sub-model. This method of dimension-upgrading plus linear evolution allows for the approximation of the long-term evolution of complex compaction processes in high-dimensional space using relatively simple linear dynamics, resulting in more stable predictions over long time periods.

[0040] To enable the NeuralODE process sub-model and the Koopman operator process sub-model to work collaboratively in the same compaction state space, they need to be jointly trained. During joint training, for each time series sample, on the one hand, a compaction state prediction sequence is obtained using the NeuralODE process sub-model, and the temporal prediction error of the NeuralODE process sub-model is calculated; on the other hand, another compaction state prediction sequence is obtained using the Koopman operator process sub-model, and the temporal prediction error of the Koopman operator process sub-model is calculated. These two temporal prediction errors are multiplied by 0.5 respectively and then added together to ensure that the two dynamic modeling methods have equal importance in the overall error. To include compaction output in the constraints, at each time step, the process state prediction vectors generated by the NeuralODE process sub-model and the Koopman operator process sub-model are input into the target index prediction decoding module, respectively. This yields the NeuralODE target index prediction value and the Koopman target index prediction value, which are then compared with the target index value at the corresponding time step. The mean square error is calculated, and the average of the mean square errors of the two predictions is taken. This average is multiplied by 0.5 and added to the aforementioned error, forming a comprehensive error that includes both process state error and target index error. This setup is because the process state vector itself is an intermediate representation; it must maintain continuity and consistency over time while ensuring accurate recovery of compaction after passing through the target index prediction decoding module. Therefore, the comprehensive error simultaneously constrains both state-level and output-level predictions, enabling the process state encoding module, the NeuralODE process sub-model, and the Koopman operator process sub-model to converge collaboratively under a unified objective. This facilitates obtaining a physically sound and accurately outputting dynamic prediction model for the target index.

[0041] In the specific training process, an optimization strategy with a batch size of 32 and a learning rate of 0.0005 can be adopted. The first to third fully connected layers of the process state encoding module, the continuous-time evolution network of the NeuralODE process sub-model, the dimensionality-upgrading mapping network of the Koopman operator process sub-model, the linear evolution matrix and dimensionality-reduction decoding network, and the fully connected layers of the target index prediction decoding module are all set as updatable parameters. For each batch of time-series samples, the comprehensive error is calculated as described above, and the gradient of the comprehensive error with respect to all the above parameters is backpropagated using the automatic differentiation function of the deep learning training framework to update the parameters of each part. The training process can be set to a maximum of 200 training epochs, with each training epoch traversing the entire training set. After each training epoch, the comprehensive error is calculated on the validation set. If the decrease in the comprehensive error of the validation set is less than a preset threshold within 10 consecutive training epochs, training is stopped early to prevent overfitting. After training, the process state encoding module, NeuralODE process sub-model, Koopman operator process sub-model and target index prediction decoding module are combined and used together. During runtime, the multi-source feature vector and control vector are used as inputs, and the output is the process state vector that evolves over time and the corresponding target index prediction value. This combination constitutes the target index dynamic prediction model.

[0042] Step 3: Based on the dynamic prediction model of the target indicators, generate a process control sequence for the execution of the target engineering process through iterative scoring; Step four: Execute the generated process control sequence, collect real-time data and obtain real-time target indicator values, compare the obtained real-time target indicator values ​​with the predicted target indicator values ​​to form online calibration samples, and use the formed online calibration samples to incrementally update the target indicator dynamic prediction model.

[0043] In one embodiment of this method, after obtaining the dynamic prediction model of the target index, the process of generating a process control sequence for the execution of the target project process through iterative scoring is executed sequentially according to the predetermined construction target and on-site working conditions. The process of carrying out on-site construction based on the process control sequence and incrementally updating the dynamic prediction model of the target index is also executed, so that the compaction process of the fly ash roadbed composite structure transitions from offline modeling to online closed-loop control.

[0044] In one specific implementation, before the process begins, for a specific construction section of the current fly ash roadbed composite structure, the spatial unit division results, thickness information of each structural layer, fly ash mix proportion information, and design target index values ​​for that construction section are first obtained. To balance process operation quality and control flexibility within a limited computational load, the prediction time range of the control sequence to be generated is divided into 10 consecutive time steps. Each time step can correspond to the target engineering process execution equipment advancing a fixed distance along the construction direction, such as 1 meter, or it can correspond to a fixed time length, such as 10 seconds. Choosing 10 time steps as the prediction range serves two purposes: firstly, it covers a representative construction length, ensuring spatial continuity of the control sequence; secondly, multiple candidate control actions need to be scored at each time step. If the number of time steps is too large, the computational complexity will increase rapidly, hindering the real-time generation of control sequences on the construction site. Therefore, 10 time steps represent a trade-off between prediction accuracy and computational cost.

[0045] At each time step, a predefined set of candidate control actions is established. These candidate actions are the Cartesian product of a set of speed levels, a set of vibration mode levels, a set of vibration intensity levels, and a set of water spray switch states. For example, the speed level set may include three travel speeds: 2, 4, and 6, corresponding to low-speed, medium-speed, and high-speed construction conditions, respectively. The vibration mode set may include two modes: static pressure mode and vibration mode. The static pressure mode is suitable for avoiding excessive vibration when the structural layer is thin or close to the target compaction degree, while the vibration mode is suitable for rapidly increasing compaction degree in the initial or middle stages. The vibration intensity level set may include three preset intensity levels, each corresponding to a different excitation force output. The water spray switch state set includes two states: on and off, used to control the surface moisture content of the fly ash mixture. Thus, at each time step, there are a total of 3 x 2 x 3 x 2 = 36 candidate control actions. Discretizing the control variables into a finite number of practically selectable gear combinations allows direct correspondence to the operation panel settings of the target engineering process execution equipment, enabling the generated process control sequence to be seamlessly executed by the mechanical equipment. At the same time, by enumerating a finite candidate set, the uncertainties brought about by complex continuous optimization algorithms during real-time operation on site can be avoided.

[0046] In the first time step, to evaluate the merits of each candidate control action, the initial working condition needs to be input into the target index dynamic prediction model. The initial working condition can be described by the multi-source feature vectors of the current construction section and the initial control command. Specifically, before construction begins or after the completion of the previous stage, the latest multi-source feature vectors of all spatial units in the current construction section can be collected, and the process state encoding module can be used to calculate the process state vectors corresponding to these multi-source feature vectors, which serve as the initial values ​​of the compaction state in the target index dynamic prediction model. Simultaneously, each candidate control action is converted into a 16-dimensional control vector, combined with the corresponding process state vector, and then the NeuralODE process sub-model and the Koopman operator process sub-model predict the process state prediction vector after one time step under the action of the candidate control action. Finally, the target index prediction decoding module obtains the target index prediction values ​​for each spatial unit and each structural layer at the end of the time step. In this way, the impact of each candidate control action on the compaction degree of the fly ash subgrade composite structure can be simulated within the model without actually executing the construction actions. This is equivalent to testing all possible operation options in advance and then selecting the most advantageous one.

[0047] To select the most suitable control action from multiple candidate control actions, a scoring mechanism is needed to evaluate the comprehensive impact of each candidate control action on compaction quality and energy consumption. In this implementation, the scoring is divided into two parts: a target index deviation index and an energy consumption index. The target index deviation index measures the average deviation between the predicted target index value and the design target index value when a certain candidate control action is executed. Specifically, for each spatial unit and each structural layer, the absolute difference between the predicted target index value and the design target index value for that spatial unit and structural layer is calculated, and then averaged over all spatial units and all structural layers in the entire construction section. Using the absolute difference instead of only considering the direction of deviation is because in actual construction, both low and high compaction can cause problems; low compaction may lead to insufficient bearing capacity, while high compaction may lead to increased structural brittleness and material waste. By taking the absolute value and then averaging, both over-compaction and under-compaction can be penalized simultaneously, and by averaging in the spatial and structural layer directions, this index reflects the compaction uniformity of the entire construction area. If a candidate control action causes a large deviation only in a local area while meeting the requirements in other areas, the mean absolute difference will increase moderately to remind the control strategy to avoid producing extreme effects in local areas.

[0048] Energy consumption indicators measure the energy consumed by the equipment executing a target engineering process when performing a candidate control action. Speed ​​and vibration intensity levels directly affect the machine's fuel consumption and the workload of its components. Lower speed levels mean the roller stays on a unit length of road, increasing the number of vibrations and compaction cycles over the same distance, thus increasing energy consumption. Higher vibration intensity levels result in greater excitation force from the excitation system, higher loads on the hydraulic and drive systems, and consequently, increased energy consumption per unit time. Therefore, a speed weight can be preset for each speed level (e.g., 1.0 for speed level 2, 0.7 for speed level 4, and 0.5 for speed level 6); and an intensity weight can be preset for each vibration intensity level (e.g., 0.5 for low intensity, 1.0 for medium intensity, and 1.5 for high intensity). Energy consumption indicators can be obtained by multiplying or weighting the speed and intensity weights, and are used to estimate the relative energy consumption level of the candidate control action per unit road segment. This approach eliminates the need for precise calculation of absolute fuel or electricity consumption. Instead, it uses relative weights to characterize the energy consumption differences of different control actions, making it easier to introduce energy consumption constraints during the scoring process.

[0049] After obtaining the target deviation index and energy consumption index, a total score is calculated for each candidate control action. The total score can be based primarily on the target deviation index and secondarily on the energy consumption index. For example, the target deviation index can be multiplied by 0.7, the energy consumption index by 0.3, and then the two can be added together to obtain the total score. The intention of this setting is to ensure that compaction quality is given higher priority, while energy consumption is treated as a secondary indicator, guiding the control strategy to prioritize the lower energy consumption among multiple candidate control actions with similar quality. If the target deviation index of a candidate control action is slightly lower than that of another candidate control action, but the energy consumption index is significantly higher than the latter, the total score may not be the lowest, thus avoiding excessive energy consumption for a small improvement in quality. By establishing this explicit trade-off between quality and energy consumption, compaction control can better conform to the principle of "quality first, with consideration for economy" in engineering practice.

[0050] After calculating the total score for all candidate control actions in the first time step, the control action with the smallest total score is selected as the target control action for the first time step. Next, the process proceeds to the second time step. In the second time step, the process state prediction vector after the target control action in the first time step is used as the new initial compaction state. The spatial unit division and design target index values ​​of the current construction section remain unchanged. All candidate control actions for the second time step are enumerated again. For each candidate control action, it is converted into a control vector, which, along with the new process state prediction vector, is input into the target index dynamic prediction model to obtain the predicted target index value after executing the candidate control action. The target index deviation index and energy consumption index are calculated in the same way, and the total score is obtained. The control action with the smallest total score is selected as the target control action for the second time step. Following this method, the process is iterated from the first time step to the tenth time step. At each time step, prediction and scoring are performed based on the process state prediction vector from the previous time step and the candidate control actions of the current time step. This process continuously supplements and expands the target control action sequence, ultimately constructing a process control sequence of length 10.

[0051] This iterative scoring method generates the process control sequence. The decision at each time step fully considers the response of the dynamic prediction model of the target index to the current working conditions and control actions. Furthermore, since the initial state of each step originates from the prediction result of the previous step, the entire control sequence remains continuous in time. This allows the compaction degree of the fly ash subgrade composite structure to gradually approach the design target value under the superimposed effects of multiple time steps. Compared to optimizing multiple time steps simultaneously, iterative scoring significantly reduces computational complexity, making it suitable for real-time operation on the control terminal at the construction site. It also retains indirect consideration of the impact on future steps, avoiding short-sighted behavior caused by purely greedy control.

[0052] In another alternative implementation, the prediction time range of the process control sequence can be set to 5 time steps for scenarios with shorter construction sections or requiring more frequent manual intervention, or it can be set to 20 time steps for scenarios with longer construction sections and where a longer forward-looking control sequence is desired. Regarding the scoring strategy, the weight of the deviation indicator from the target indicator can be increased to 0.8, while the weight of the energy consumption indicator can be decreased to 0.2, to suit projects with extremely high compaction requirements and low energy sensitivity. If there is high-noise energy consumption measurement data at the construction site, the energy consumption indicator can be simplified by setting the energy consumption level only based on the speed level, without introducing the weight of the vibration intensity level, thus enabling iterative scoring even with incomplete information.

[0053] refer to Figure 3This image displays a grayscale map showing the spatial distribution of compaction degree in various spatial units of a construction section after the 8th pass of the compaction process. The horizontal axis represents the longitudinal position (x) of the construction section along the roller's direction of travel, in meters (m), ranging from 0 to 100 meters, with a graduation mark every 10 meters. The vertical axis represents the lateral position (y) perpendicular to the direction of travel, in meters (m), ranging from 0 to 50 meters, with a graduation mark every 5 meters. The entire drawing area is divided into a fine grid, with each cell corresponding to a spatial unit of the construction site. Grayscale values ​​are used to represent the compaction degree of each spatial unit. Areas with higher compaction degree are represented by darker gray, and areas with lower compaction degree are represented by lighter gray. Specifically, areas with 97% compaction degree are displayed as dark gray, areas with 88% compaction degree are displayed as light gray, and intermediate compaction degree values ​​show a continuous transition between dark and light gray. This grayscale coding method allows for a clear observation of the spatial distribution and uniformity of compaction throughout the construction section. The graph shows a distinct spatial variation in compaction distribution. In the central area of ​​the construction section, approximately at x=30 meters longitudinally and y=25 meters laterally, a darker gray area appears, corresponding to a high-compaction zone. A hollow circle marks a feature point within this zone, with a leader line indicating "High-compaction zone K=96%", signifying that the compaction at this location reaches 96%, meeting the design target compaction requirement. The formation of this high-compaction zone is related to the extended working time of the roller in this area and sufficient vibration energy input. At the edges of the construction section, particularly in the lower left corner, approximately at x=10 meters longitudinally and y=10 meters laterally, a lighter gray area appears. This area is marked with a hollow circle and labeled "Edge under-compaction K=90%", indicating that the compaction at this location is only 90%, lower than the design target value. This under-compaction phenomenon at the edges is caused by insufficient compaction energy input when the roller turns or turns around at the edge of the construction section, resulting in unstable contact between the vibratory drum and the roadbed. It is a common weak point in actual construction. In the upper right area of ​​the construction section, roughly at x=80 meters longitudinally and y=40 meters laterally, a particularly dark gray area appears, marked "Local Over-compaction K=97%", indicating that the compaction degree at this location reaches 97%, exceeding the design target value. Local over-compaction may be due to excessive repeated compaction by the roller at this location, or an excessively high vibration intensity setting, leading to excessive concentration of compaction energy. Although the compaction degree value in this area is high, over-compaction may cause increased brittleness and microcracks in the fly ash roadbed composite structure, while also wasting energy. Overall observation of the grayscale map reveals that the compaction degree distribution is not completely uniform but exhibits certain fluctuations. In the central area, the compaction degree distribution is relatively uniform and close to the target value, while in the edge areas and some local locations, the compaction degree deviates significantly from the target value.This spatial distribution characteristic reflects the combined influence of various factors during the process, including roller travel speed, vibration intensity, number of passes, initial material condition, and boundary effects. A vertical color bar is drawn on the right side of the graph to indicate the correspondence between grayscale values ​​and compaction values. The color bar transitions from dark gray to light gray from top to bottom. The compaction scale values ​​are marked on the left side of the bar, from top to bottom: 97%, 95%, 93%, 91%, 89%, and 88%, each corresponding to a position on the bar. The axis label "Compaction K (%)" is marked on the right side of the bar, written vertically. By referring to this color bar, the compaction value at any position in the grayscale graph can be accurately read. The upper left corner of the graph is labeled with the explanatory text "Spatial Distribution of Compaction After the 8th Pass of Processing," indicating that this grayscale graph reflects the spatial distribution of the measured compaction values ​​obtained through a nuclear density meter or other testing methods, or the predicted values ​​of the target index calculated by the target index dynamic prediction model, after completing 8 passes of the process. This grayscale image provides an intuitive visualization basis for process operation quality assessment and control strategy optimization. By identifying the spatial locations of high-compaction zones, under-compaction zones, and over-compaction zones, construction managers can adjust the roller travel path, speed setting, and vibration intensity setting for subsequent compaction passes to ensure that under-compaction areas at the edges are compacted, preventing further over-compaction in local areas, and ultimately achieving uniformity and consistency in the compaction distribution of the entire construction section, meeting the engineering quality requirements of fly ash subgrade composite structures.

[0054] After generating the process control sequence, it is distributed to the target engineering process execution equipment to implement the control strategy at the construction site. The target engineering process execution equipment is equipped with a process control terminal, which controls speed, vibration mode, vibration intensity, and water spraying status according to the target control actions in the process control sequence. The process control terminal can execute control cycles at fixed time intervals, such as once every second. At the beginning of each control cycle, the process control terminal reads the target control action corresponding to the current time step, parses the target speed level, target vibration mode, target vibration intensity, and target water spraying status from the process control sequence, and converts these target control actions into drive commands, which are then sent to the walking system, vibration system, and water spraying system, enabling the target engineering process execution equipment to operate according to the target control actions within the current time step. During the continuous execution of the control cycle, multi-source sensing devices continuously collect vibration acceleration data, wheel load data, position data, speed data, temperature data, and water content data at the aforementioned sampling frequency, and generate a new multi-source feature vector at the end of each process time window.

[0055] To ensure the dynamic prediction model of the target index continuously reflects the actual site conditions, real-time compaction degree detection is required simultaneously with the execution of the process control sequence, forming an online calibration sample. In this embodiment, after the target engineering process execution equipment passes through 10 spatial units along the construction direction, one of these 10 spatial units is selected as the detection object. Spatial units located in the middle position can be preferentially selected to reduce the interference of boundary effects at the start and end points of construction on the detection results. Alternatively, in another optional embodiment, the spatial unit whose compaction degree deviates most from the design target index value according to the current dynamic prediction model can be selected for detection, thus prioritizing the verification of locations with higher risks. For the selected spatial unit, real-time compaction degree detection is performed using a nuclear density meter to obtain the real-time target index value for that spatial unit. Simultaneously, the predicted target index value for that spatial unit at the same time step is read from the dynamic prediction model. The real-time target index value is compared with the predicted target index value, and the difference between the two is calculated. This difference reflects the deviation between the model's prediction result and the actual compaction effect under current environmental conditions, material state, and mechanical state.

[0056] During incremental updates, these 50 online calibration samples are input into the target indicator dynamic prediction model. For each online calibration sample, the multi-source feature vector is first converted into a process state vector using the process state encoding module. Then, combined with the corresponding control vector, the process state prediction vector and the target indicator prediction value are calculated using the NeuralODE process sub-model and the Koopman operator process sub-model. The target indicator prediction value is then compared with the real-time target indicator value in that online calibration sample, and the squared difference between the two is calculated. This difference is averaged over all 50 online calibration samples to obtain the target error during the incremental update process. Based on the target error, the weights of the third fully connected layer of the process state encoding module, the last fully connected layer of the continuous-time evolution network of the NeuralODE process sub-model, the linear evolution matrix of the Koopman operator process sub-model, and the fully connected layer of the target indicator prediction decoding module are updated. By updating only the parameters of these specific layers while keeping the parameters of other layers unchanged, the magnitude of changes in the model structure can be effectively controlled. While retaining the global compaction rules learned in the original offline training phase, the local behavior of the model adapts to the specific environment of the current work site. For example, the third fully connected layer directly determines the final expression of the process state vector. Adjusting this layer can make the compaction state representation more adaptable to the current data while retaining the main structure. The last fully connected layer of the continuous-time evolution network directly determines the output shape of the derivative estimate. Fine-tuning it can refine the local dynamic response of the NeuralODE process sub-model. The linear evolution matrix controls the overall evolution direction of the Koopman operator process sub-model in high-dimensional space. Small adjustments can improve the stability of long-term predictions. The fully connected layer of the target index prediction decoding module directly affects the mapping from the process state vector to the target index prediction value. Incremental updates can reduce the systematic deviation between the prediction output and real-time measurement.

[0057] To avoid excessive perturbation to the model in a single incremental update, the incremental update process typically uses a small learning rate and a limited number of training epochs. In this implementation, the learning rate can be set to 0.0001, and the number of training epochs can be set to 5. In each training epoch, the target error is calculated using all 50 online correction samples, and a parameter update is performed. This allows the model to gradually approach the new data distribution in multiple small steps, rather than undergoing a drastic change all at once. By setting a small learning rate, the magnitude of each parameter update can be limited, allowing the model to absorb new information while retaining existing effective features as much as possible, thus achieving smooth model evolution. As construction conditions change slowly, such as a gradual decrease in water content or changes in ambient temperature, the model can gradually adjust itself through multiple incremental updates, ensuring that the prediction results continuously closely match on-site observations.

[0058] In another optional implementation, the number of online calibration samples triggering incremental updates can be adjusted from 50 to 30 or 80. If the sensors at the construction site are highly stable and the data quality is good, the number of online calibration samples per batch can be reduced, for example, to 30, so that the model can respond more quickly to changes in construction conditions. If the construction environment fluctuates greatly or the detection frequency is low, the number of samples per batch can be increased, for example, to 80, so that the impact of individual abnormal data on the update direction can be balanced by a larger sample set. In addition, an average target index error threshold can be set. Incremental updates and control sequence recalculation are only initiated when the average error of the most recent few online calibration samples exceeds this threshold, thereby reducing invalid updates when the model prediction is already sufficiently accurate, further reducing the computational burden and improving construction efficiency. Through these optional implementations, the frequency and intensity of incremental updates can be flexibly configured according to the characteristics of different engineering projects and the computing power of the control system.

[0059] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for intelligent control of target indicators in multi-source engineering processes based on deep learning, characterized in that, The method includes the following steps: Step 1: Collect multi-source sensor data from the target engineering process execution equipment, obtain target index values, and construct a process time series sample set containing multi-source feature vectors and corresponding target index values; Step 2: Using the process time series sample set, jointly train a NeuralODE process sub-model and a Koopman operator process sub-model, and combine the process state encoding module and the target index prediction decoding module to form a target index dynamic prediction model. Step 3: Based on the dynamic prediction model of the target indicators, generate a process control sequence for the execution of the target engineering process through iterative scoring; Step four: Execute the generated process control sequence, collect real-time data and obtain real-time target indicator values, compare the obtained real-time target indicator values ​​with the predicted target indicator values ​​to form online calibration samples, and use the formed online calibration samples to incrementally update the target indicator dynamic prediction model.

2. The method according to claim 1, characterized in that, Step one includes: installing a vibration acceleration sensor, wheel load sensor, position acquisition module, velocity acquisition module, and temperature and moisture content acquisition module on a target engineering process execution device, and acquiring vertical vibration acceleration data, wheel load data, position data, velocity data, temperature data, and moisture content data at sampling frequencies of 100Hz, 10Hz, 1Hz, 1Hz, 1Hz, and 1Hz respectively; measuring the target index values ​​using a nuclear density meter or a load-bearing plate testing device; and calculating the average value of the acquired vibration acceleration data in 1-second segments within a process time window. Values, standard deviations, and extreme values ​​are calculated. The mean of the collected wheel load data and the collected speed data are calculated within the same process time window. The last sampled value of the collected temperature and water content data is taken within the same process time window to form a multi-source feature vector. The multi-source feature vector of the same spatial unit is combined with the corresponding target index value in order of pass / fail to form a time series sample. The time series samples are summarized and processed by missing value filling, outlier removal, and numerical scaling to construct a process time series sample set, which is divided into training set, validation set and test set according to the sample size ratio of 7:2:

1.

3. The method according to claim 1, characterized in that, Step two includes: constructing a process state encoding module and a target indicator prediction and decoding module. The multi-source feature vector at each time step is input into the process state encoding module. The process state encoding module consists of four fully connected layers. The first fully connected layer has the same input dimension as the multi-source feature vector and an output dimension of 128. The second fully connected layer has an input dimension of 128 and an output dimension of 128. The third fully connected layer has an input dimension of 128 and an output dimension of 64. The fourth fully connected layer has an input dimension of 64 and an output dimension of 1. The activation function for the first to third fully connected layers is a rectified linear unit (RCU). The fourth fully connected layer does not use an activation function. The 64-dimensional output of the third fully connected layer is defined as the process state vector, and the 1-dimensional output of the fourth fully connected layer is defined as the static target indicator prediction value. The first to third fully connected layers constitute the process state encoding module, and the fourth fully connected layer constitutes the target indicator prediction and decoding module.

4. The method according to claim 3, characterized in that, Step two also includes constructing a NeuralODE process sub-model, which consists of a state input splicing module and a continuous-time evolution network. The state input splicing module splices the process state vector of the current time step with the 16-dimensional control vector in the feature dimension to form a state control vector of length 80. The 16-dimensional control vector is obtained by one-hot encoding of the speed level, vibration mode level, vibration intensity level, and sprinkler switch state of the target engineering process execution equipment and then transforming it through two fully connected layers with output dimensions of 32 and 16 respectively. The continuous-time evolution network includes three fully connected layers, with an input dimension of 80 and an output dimension of 128, an input dimension of 128 and an output dimension of 128, and an input dimension of 128 and an output dimension of 64 respectively. The first and second fully connected layers are activated using rectified linear units, while the third fully connected layer does not use an activation function. The continuous-time evolution network receives the state control vector and outputs a 64-dimensional derivative estimation vector.

5. The method according to claim 4, characterized in that, The NeuralODE process sub-model performs numerical integration as follows: each process operation is divided into one time interval, and this time interval is evenly divided into four sub-time intervals; within each sub-time interval, a four-fold function evaluation is performed: (i) the first derivative estimate is obtained by passing the current process state vector and the current control vector through a continuous-time evolution network; (ii) based on the first derivative estimate, the first intermediate state is obtained by shifting forward half the length of the sub-time interval from the original state, and the second derivative estimate is obtained by passing the first intermediate state and the current control vector through a continuous-time evolution network; (iii) based on the second derivative estimate, the second derivative estimate is obtained by shifting forward half the length of the sub-time interval from the original state. (iv) Shift the second intermediate state by half a sub-time interval to obtain the second intermediate state. Use the second intermediate state and the current control vector through a continuous-time evolution network to obtain the third derivative estimate. Shift the second intermediate state forward by one sub-time interval to obtain the sub-time interval endpoint state. Use the sub-time interval endpoint state and the current control vector through a continuous-time evolution network to obtain the fourth derivative estimate. Weight the first, second, third, and fourth derivative estimates in a ratio of 1:2:2:1 and multiply by the sub-time interval length to obtain the process state increment. Add the obtained process state increment to the original process state vector to obtain the process state vector at the end of the time interval.

6. The method according to claim 3, characterized in that, Step two also includes constructing the Koopman operator process sub-model, which includes: an upscaling mapping network, consisting of two fully connected layers, with an input dimension of 64 and an output dimension of 128, and an input dimension of 128 and an output dimension of 128, both activated by rectified linear units (CLUs), which takes the process state vector as input to obtain a 128-dimensional high-dimensional state vector; a linear evolution matrix, which is a 128-row, 128-column trainable weight matrix, used to perform matrix multiplication on the 128-dimensional high-dimensional state vector to obtain the high-dimensional state prediction vector for the next time step; and a downscaling decoding network, consisting of two fully connected layers, with an input dimension of 128 and an output dimension of 128, and an input dimension of 128 and an output dimension of 64, where the first fully connected layer is activated by CLUs and the second fully connected layer does not use an activation function, taking the high-dimensional state prediction vector as input to obtain the process state prediction vector for the corresponding time step.

7. The method according to any one of claims 3, 4, and 6, characterized in that, The joint training in step two includes: calculating the temporal prediction error of the NeuralODE process sub-model and the Koopman operator process sub-model; inputting the process state prediction vectors obtained from the two sub-models into the target index prediction decoding module to obtain the NeuralODE target index prediction value and the Koopman target index prediction value, and calculating the mean square error between the two target index prediction values ​​and the target index value; multiplying the temporal prediction errors of the two sub-models by 0.5 and adding them together, and then adding them to the calculated mean square error multiplied by 0.5 to form a comprehensive error that includes the process state error and the target index error; using the comprehensive error, jointly updating the weight parameters of the first to third fully connected layers of the process state encoding module, the continuous time evolution network, the up-dimensional mapping network, the linear evolution matrix, the down-dimensional decoding network, and the target index prediction decoding module.

8. The method according to claim 1, characterized in that, Step 3 includes: dividing the prediction time range of the control sequence to be generated into 10 consecutive time steps; limiting the candidate control actions of each time step to the Cartesian product of the speed level set, vibration mode set, vibration intensity set, and sprinkler switch state set; for the first time step, for all candidate control actions, using the initial multi-source feature vector and the current candidate control action as input to the target index dynamic prediction model to obtain the target index prediction value; calculating the average absolute difference between the obtained target index prediction value and the design target index value as the target index deviation index; calculating the weighted sum of the speed level and vibration intensity level of the current candidate control action as the energy consumption index; calculating the weighted sum of the target index deviation index and the energy consumption index as the total score; selecting the control action with the smallest total score as the target control action of the first time step; iterating sequentially to the 10th time step, and at each time step, predicting and scoring based on the process state prediction vector of the previous time step and the candidate control action of the current time step through the target index dynamic prediction model to construct a process control sequence of length 10.

9. The method according to claim 1 or 3, characterized in that, The incremental update in step four includes: after passing through every 10 spatial units, the real-time target index value of one spatial unit is measured, and the measured real-time target index value is compared with the target index prediction value generated by the target index dynamic prediction model to obtain the target index error. The target index error and the multi-source feature vector of the corresponding time step are used to form an online correction sample. When the number of online correction samples reaches 50, the 50 online correction samples are input into the target index dynamic prediction model. Incremental updates are performed for 5 training cycles with a learning rate of 0.0001 on the third fully connected layer of the process state encoding module, the last fully connected layer of the continuous time evolution network of the NeuralODE process sub-model, the linear evolution matrix of the Koopman operator process sub-model, and the fully connected layer of the target index prediction decoding module.

10. The method according to claim 9, characterized in that, Step four also includes: recalculating the process control sequence for the next 10 time steps using the latest parameters after each training cycle of incremental updates; replacing the original control sequence in the process control terminal; and enabling the target engineering process execution equipment to run according to the process control sequence generated by the online calibrated target index dynamic prediction model.

Images