A double-wheel milling construction data management and remote monitoring system and method

Abstract

The application discloses a double-wheel milling construction data management and remote monitoring system and method, and belongs to the technical field of double-wheel milling. In view of the problem that geological identification, slot quality prediction and construction parameter control links are mutually disconnected in the prior art, construction precision is low under complex strata, and parameter adjustment is lagged, a full-closed-loop automatic control system is constructed: multi-frequency data accurate time alignment is realized through a multi-source heterogeneous data acquisition module; six types of complex geology are identified in real time by adopting multi-dimensional spectrum coupling feature extraction and fuzzy soft identification technology combined with a lightweight CNN; slot quality per meter per pass is predicted by using a cross-attention fusion Transformer-LSTM hybrid model, and different geologies are adapted through dynamic condition modulation; a model predictive control algorithm with Smith predictor delay compensation is used to automatically optimize construction parameters; and an edge-cloud collaborative architecture and an online incremental learning mechanism are designed, so that slot quality and construction efficiency are significantly improved, and equipment failure rate and geological risk are reduced.

Description

Technical Field

[0001] This invention relates to the field of twin-wheel milling technology, and specifically to a twin-wheel milling construction data management and remote monitoring system and method. Background Technology

[0002] Twin-wheel trenching machines are construction equipment for diaphragm walls and are widely used in major infrastructure projects such as deep foundation pit support and water conservancy and hydropower projects. As underground engineering develops towards ultra-deep and ultra-complex geological conditions, twin-wheel trenching construction frequently encounters complex working conditions such as isolated boulders, soft and hard interlayers, and strata junctions, which places extremely high demands on the accuracy, efficiency, and safety of trenching.

[0003] Currently, twin-wheel milling operations still largely rely on manual experience or simple automated control based on single feedback. While existing technologies have incorporated sensor monitoring and geological identification to some extent, they suffer from fundamental deficiencies in achieving closed-loop control: the three components of geological identification, trenching quality prediction, and construction parameter control in existing systems are isolated from each other, failing to form an effective collaborative closed loop.

[0004] Specifically, existing geological identification methods largely rely on the temporal characteristics of single signals for coarse classification, making it difficult to accurately identify complex geological formations such as isolated boulders and interlayers that significantly impact construction. Trenching quality prediction often employs a single time-series model, failing to effectively integrate multi-source heterogeneous information such as real-time operating parameters, geological data, and historical data, resulting in low prediction accuracy and lag. Furthermore, construction parameter control often relies solely on passive adjustments based on single variables such as torque exceeding limits, failing to incorporate real-time geological identification results and future trenching quality predictions as feedforward inputs into control decisions. This disconnect leads to severely delayed adjustments to construction parameters when encountering complex geological formations, and the system is highly susceptible to oscillations due to a lack of predictability in control actions, hindering smooth and accurate automatic optimization. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this application provides a dual-wheel milling construction data management and remote monitoring system.

[0006] To achieve the above objectives, the technical solution adopted in this application is: a dual-wheel milling construction data management and remote monitoring system, comprising:

[0007] The multi-source heterogeneous data acquisition module is used to acquire multi-source heterogeneous data in real time during the construction process of the dual-wheel milling machine. The multi-source heterogeneous data includes at least real-time construction operation parameters, on-site geological stratification data, and historical construction data under the same working conditions.

[0008] The real-time prediction and risk warning module for trenching quality is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to predict the trenching quality meter by meter and cut by cut in real time based on the multi-source heterogeneous data using a pre-trained Transformer-LSTM hybrid neural network model, and output a corresponding level of risk warning based on the prediction results. The Transformer component is used to extract global correlation features between multi-source heterogeneous data, and the LSTM component is used to model the long-term dependency relationship of time-series construction data.

[0009] The real-time geological type identification module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to extract the multi-dimensional spectral features of the milling wheel torque signal and the milling wheel vibration signal, and input them into a pre-trained lightweight convolutional neural network classifier to identify the geological type of the current construction stratum in real time. The geological type includes at least soft soil, sandy soil, rock, boulders, soft and hard interlayers and strata junctions.

[0010] The automatic optimization control module for construction parameters is connected to the real-time prediction and risk warning module for trenching quality and the real-time geological type identification module, respectively. It is used to quickly and smoothly optimize and adjust the construction parameters of the twin-wheel milling machine based on the real-time identified geological type and trenching quality prediction results, and to build a fully closed-loop automatic control system for geological identification, quality prediction and parameter optimization.

[0011] Furthermore, the multi-source heterogeneous data acquisition module includes a time alignment unit. The time alignment unit is used to establish a data frame structure with timestamps for various types of data, using the high-frequency sampling clock of the dual-wheel milling main control system as the reference time source. For data sources with different sampling frequencies, adaptive resampling based on spline interpolation and sliding window downsampling based on energy conservation are adopted respectively. For non-periodic data sources, the nearest neighbor time association strategy is used to map to the reference time axis and add deviation marks. The aligned multi-source data is spliced ​​into a multi-channel synchronous data frame with a unified timestamp as the input for subsequent models.

[0012] Furthermore, it also includes an edge-cloud collaborative transmission and monitoring architecture, which includes:

[0013] Edge computing nodes deployed at the construction site are used to perform lightweight preprocessing and feature extraction on the raw data before compression and uploading.

[0014] The cloud server is communicatively connected to the edge computing node;

[0015] Among them, the edge computing nodes and the cloud server adopt an adaptive compression transmission strategy and combine it with sequence number confirmation and breakpoint resume protocol to build a two-layer collaborative architecture of edge inference and cloud inference.

[0016] Furthermore, the Transformer-LSTM hybrid neural network model in the real-time prediction and risk warning module for trenching quality adopts a cross-attention fusion architecture: the Transformer component extracts global correlation features through a multi-head self-attention mechanism, the LSTM component models long-term dependence of time-series data, and a cross-attention fusion layer is set between the two to use the global correlation features as the LSTM context vector and calculate the influence weights. After fusion, the meter-by-meter and cutter-by-cut prediction values ​​of trench wall verticality deviation, trench width deviation and sediment thickness are output.

[0017] Furthermore, the real-time prediction and risk warning module for trenching quality also includes an online incremental learning unit. The online incremental learning unit is used to extract incremental training samples at fixed time intervals, fine-tune them using an elastic weight consolidation strategy and apply regularization constraints, combine historical data playback samples for mixed training, set a model performance monitor, and automatically roll back and trigger full retraining when the accuracy is lower than a preset lower limit.

[0018] Furthermore, the real-time geological type identification module includes a multi-dimensional spectral feature extraction unit. The multi-dimensional spectral feature extraction unit is used to perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signals respectively, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient of wavelet coefficients at the same scale of the two signals and the spectral coherence function to construct cross-signal spectral coupling features, and splice the statistical features and coupling features to form a multi-dimensional spectral feature vector as the input of the classifier.

[0019] Furthermore, the real-time geological type identification module also includes a stratigraphic transition zone fuzzy soft identification unit. The stratigraphic transition zone fuzzy soft identification unit is used to receive the probability distribution vectors of various geological types output by the classifier, convert them into membership degrees through a fuzzy membership function, and determine the stratigraphic transition zone when the difference between the highest and second highest membership degrees is less than a preset distinction threshold and estimate the width. The membership degree vector and the transition zone label are then output to the fully closed-loop construction parameter automatic optimization control module.

[0020] Furthermore, the model prediction control algorithm in the fully closed-loop construction parameter automatic optimization control module adopts a multi-objective weighted function as the optimization objective function. The multi-objective weighted function includes three cost terms: trenching quality, construction efficiency, and parameter change. The weight of each cost term is dynamically adjusted according to the real-time geological type. At the same time, hard constraints are set for milling wheel torque, rotation speed, mud pressure, and parameter change rate.

[0021] Furthermore, the fully closed-loop construction parameter automatic optimization control module also includes a delay compensation unit based on the Smith predictor. The delay compensation unit is used to establish a prediction model for the positive and feedback channels of the construction process, introduce the delayed trenching quality state predicted by the Smith predictor to replace the delayed measurement value for optimization, dynamically estimate the delay time through step test and cross-correlation analysis, and adaptively adjust the control frequency and prediction time domain under large delay conditions.

[0022] It also includes a dynamic condition modulation module, which is connected between the real-time geological type identification module and the trenching quality real-time prediction and risk warning module. The dynamic condition modulation module is used to take the geological identification results and fuzzy membership degree as condition variables, generate scaling and offset coefficients through the feature linear modulation mechanism to perform affine transformation on the model fusion feature vector, and enhance the corresponding feature channel weights according to different geological types, so that the single model can adapt to different geological conditions.

[0023] This invention also provides a method for managing and remotely monitoring construction data of a dual-wheel milling machine, comprising the following steps:

[0024] S1: Real-time acquisition of multi-source heterogeneous data from dual-wheel milling operations, including operating parameters, geological stratification, and historical data under the same working conditions; using the high-frequency clock of the main control system as a reference, data frames with timestamps are established for various types of data; spline interpolation resampling and energy conservation downsampling are used for data with different sampling frequencies; non-periodic data are correlated with nearest neighbor time and marked with deviations, and spliced ​​into multi-channel synchronous data frames; after lightweight preprocessing and feature extraction at the field edge nodes, the data is uploaded to the cloud through adaptive compression and breakpoint resume protocol, constructing an edge-cloud dual-layer collaborative inference architecture;

[0025] S2: Perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signal, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient and spectral coherence function at the same scale to construct cross-signal coupling features, and splice them to form a multi-dimensional spectral feature vector; input the lightweight CNN classifier to obtain the probability distribution of geological types, and after fuzzy membership function transformation, when the difference between the highest and second highest membership degrees is less than the threshold, it is determined to be a stratigraphic transition zone and its width is estimated, and strata such as soft soil, sand, rock, boulders, soft and hard interlayers and junctions are identified in real time;

[0026] S3: Using geological identification results and membership degrees as conditional variables, scaling and offset coefficients are generated through feature linear modulation. Affine transformation is performed on the fusion feature vector of the trenching quality prediction model, and the corresponding feature channel weights are dynamically enhanced according to geological type.

[0027] S4: Based on multi-channel synchronous data frames, a Transformer-LSTM hybrid model with cross-attention fusion is used to predict the trenching quality meter by meter and blade by blade: Transformer extracts global correlation features from multi-source data, LSTM models long-term temporal dependencies, and the cross-attention layer uses global features as LSTM context vectors to calculate influence weights, outputting predicted values ​​for trench wall verticality, trench width, and sludge thickness, and providing graded early warnings; incremental samples are extracted at fixed intervals, and an elastic weight consolidation strategy combined with historical playback is used to fine-tune the model. A performance monitor is set up, and when the accuracy is lower than the threshold, automatic rollback and full retraining are triggered.

[0028] S5: Based on geological identification and quality prediction results, a model predictive control algorithm is used to automatically optimize construction parameters: a multi-objective weighted function is constructed with trenching quality, construction efficiency, and parameter changes as cost terms, and the weights are dynamically adjusted according to the geological type. At the same time, hard constraints are set for torque, rotational speed, mud pressure, and parameter change rate. A construction forward and reverse prediction model is established, and a Smith predictor is introduced to compensate for measurement delay. The delay time is dynamically estimated through step test and cross-correlation analysis. Under the condition of large delay, the control frequency and prediction time domain are adaptively adjusted to construct a closed-loop control system of geological identification-quality prediction-parameter optimization.

[0029] Beneficial effects:

[0030] 1. This invention, through a technical solution combining multidimensional spectral coupling feature extraction and fuzzy soft recognition, overcomes the fundamental limitations of existing technologies that rely solely on the temporal features of a single signal. This significantly improves recognition accuracy, greatly enhancing the comprehensive recognition accuracy for six geological types: soft soil, sand, rock, boulders, soft-hard interlayers, and stratigraphic transition zones. It also significantly improves the accuracy of identifying stratigraphic transition zones and can accurately estimate the width of these zones, providing a reliable basis for refined construction. By analyzing the spectral characteristics of milling wheel torque and vibration signals, geological changes can be identified as early as 0.1-0.2 meters into abnormal strata, providing early warning of geological risks. It can also detect complex geological features such as boulders and karst caves missed in previous explorations, avoiding major accidents such as cutter tooth breakage, milling wheel jamming, and trench wall collapse, significantly reducing equipment failure rates and greatly minimizing downtime caused by geological anomalies.

[0031] 2. This invention employs a Transformer-LSTM cross-attention fusion architecture combined with a dynamic condition modulation mechanism to achieve real-time prediction of trenching quality meter by meter and cut by cut. It can accurately reflect the impact of each milling parameter on trenching quality, providing precise guidance for real-time parameter adjustment. Construction parameters can be adjusted in a timely manner when trenching quality shows a deviation trend, greatly reducing the trenching quality non-conformity rate and significantly reducing the risk of project delay. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of a twin-wheel milling construction data management and remote monitoring method according to an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0035] Example 1

[0036] This embodiment provides a dual-wheel milling machine construction data management and remote monitoring system, including:

[0037] A multi-source heterogeneous data acquisition module is used to acquire multi-source heterogeneous data in real time during the twin-wheel milling construction process. This multi-source heterogeneous data includes at least real-time construction operation parameters, on-site geological stratification data, and historical construction data under the same working conditions. The module acquires these data through the twin-wheel milling main control system interface, vibration sensors, a geological exploration database, and a historical construction database. Specifically, the real-time construction operation parameters include milling wheel torque, rotational speed, feed rate, mud pressure, and vibration acceleration. The on-site geological stratification data consists of data from previous exploration boreholes and geological records supplemented during construction. The historical construction data under the same working conditions includes construction parameters and trenching quality records under the same geological conditions. Due to the significant differences in sampling frequencies and inconsistent time bases among the different data sources, precise time alignment is necessary to achieve multi-source data fusion. This unit uses the high-frequency sampling clock of the twin-wheel milling main control system (1000Hz) as the sole reference time source and employs a differentiated alignment strategy for the three types of data.

[0038] Adaptive resampling based on cubic spline interpolation:

[0039] Suitable for data sources with sampling frequencies lower than the reference frequency, it generates equally spaced data points aligned with the reference time axis through natural cubic spline interpolation, ensuring the smoothness and continuity of the signal.

[0040] ,

[0041] in, The signal value at time t after interpolation;

[0042] These are two adjacent sampling times of the original data;

[0043] The cubic spline coefficients for the k-th interval are determined by the boundary conditions. , Solve for the natural boundary conditions where the second derivative at the endpoints is 0.

[0044] Energy-conserving sliding window sampling:

[0045] It is suitable for high-frequency data sources with sampling frequencies higher than the reference frequency. It adopts sliding window averaging downsampling while preserving signal energy and variance characteristics, thus avoiding the loss of high-frequency information.

[0046] ;

[0047] ;

[0048] 2 .

[0049] in: is the average value of the i-th downsampling window; N is the downsampling factor (original sampling frequency / reference sampling frequency); Let be the signal energy of the i-th window, used to characterize the signal strength; Let be the sample variance of the i-th window, used to characterize high-frequency fluctuation characteristics.

[0050] Nearest neighbor time correlation for non-periodic data:

[0051] It is suitable for non-periodic discrete data such as geological exploration data and quality inspection data. It adopts the nearest neighbor time mapping strategy to associate with the reference time axis and adds time deviation mark to evaluate the reliability of the data.

[0052] ; ;

[0053] in, The original timestamps are for non-periodic data; T is the set of all time points on the baseline time axis. The base timestamp after mapping; For time deviation marking, when The time stamps are marked as low-confidence data and are used only for model training; they are not involved in real-time control.

[0054] Aligned multi-source data are concatenated according to a unified timestamp into a dimension. Multi-channel synchronous data frames serve as input for all subsequent models.

[0055] The real-time prediction and risk warning module for trenching quality is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to predict the trenching quality meter by meter and cut by cut in real time based on the multi-source heterogeneous data using a pre-trained Transformer-LSTM hybrid neural network model, and output a corresponding level of risk warning based on the prediction results. The Transformer component is used to extract global correlation features between multi-source heterogeneous data, and the LSTM component is used to model the long-term dependency relationship of time-series construction data.

[0056] This architecture combines the Transformer's ability to capture globally correlated features with the LSTM's ability to model long-term temporal dependencies. It achieves deep fusion of the two through a cross-attention fusion layer, solving the problem that a single model cannot simultaneously handle multi-source heterogeneous data and time-series data.

[0057] Firstly, regarding the Transformer's multi-head self-attention mechanism:

[0058] It is used to input multi-channel synchronous data frames into the Transformer encoder and extract global correlation features between different data sources through a multi-head self-attention mechanism.

[0059] ;

[0060] ;

[0061] ;

[0062] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, which are obtained from the input features through linear transformation; is the dimension of the key matrix, used to scale the dot product result to avoid gradient vanishing; h is the number of attention heads, which is set to h=8 in this invention; , , , is a trainable weight matrix.

[0063] The output dimension of the Transformer encoder is Globally related feature sequences Each moment's features contain information from all historical moments and all data sources.

[0064] For the cross-attention fusion layer:

[0065] The global associated feature sequence used for the Transformer output is fused with the hidden state of the LSTM through cross-attention, so as to realize the dynamic weighting of global features at each time step.

[0066] ;

[0067] in, This represents the hidden state of the LSTM at the previous time step; , For a trainable weight matrix, For bias terms; Let T be the attention weight distribution of length T, representing the importance of each time step in the global feature sequence for the prediction at the current time step; This is the weighted context vector; To determine the final hidden state after fusion, the meter-by-meter and cut-by-cut prediction values ​​of the verticality deviation of the tank wall, the width deviation of the tank, and the thickness of the sediment are output through the fully connected layer.

[0068] The working principle of the online incremental learning unit is as follows: an elastic weight consolidation strategy is used to realize the online incremental learning of the model, so that the model can continuously optimize its performance using new construction data, while retaining the knowledge learned previously.

[0069] 2 ;

[0070] in, This is the total loss function; The mean squared error loss for the new data; The regularization intensity coefficient is taken in this invention. ; The diagonal elements of the Fisher information matrix are used to measure parameters. The importance of the previous task was calculated using 1000 representative samples from the previous task via the Monte Carlo method. These are the parameter values ​​after the model has been trained on the old task.

[0071] This unit extracts incremental training samples at fixed 24-hour intervals, fine-tunes them using the aforementioned loss function, and mixes in 20% historical data replay samples. A model performance monitor is set up so that when the prediction accuracy of 100 consecutive samples falls below a preset lower limit, it automatically rolls back to the previous stable version and triggers a full retraining.

[0072] The real-time geological type identification module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to extract the multi-dimensional spectral features of the milling wheel torque signal and the milling wheel vibration signal, and input them into a pre-trained lightweight convolutional neural network classifier to identify the geological type of the current construction stratum in real time. The geological type includes at least soft soil, sandy soil, rock, boulders, soft and hard interlayers and strata junctions.

[0073] The spectral characteristics of milling wheel torque and vibration signals are closely related to the geological properties. This unit extracts the time-frequency domain features of the signals through multi-resolution continuous wavelet transform and constructs cross-signal spectral coupling features, which solves the problem that single time-domain features cannot distinguish complex geological conditions.

[0074] Milling wheel torque signal The vibration signal V(t) is subjected to multi-resolution continuous wavelet transform to convert the one-dimensional time-domain signal into a two-dimensional time-frequency domain representation.

[0075] ;

[0076] in, For signal Wavelet coefficients at scale a and translation b; a is the scale factor, which controls the scaling of the wavelet; b is the translation factor, which controls the translation of the wavelet. For the inner product operation to be performed, the second function must be the complex conjugate;

[0077] For the Morlet mother wavelet, the expression is: Center frequency =5.

[0078] Wavelet coefficient statistical features are used to extract five statistical features of wavelet coefficients at each scale, comprehensively characterizing the amplitude, energy, and distribution characteristics of the signal.

[0079] The automatic optimization control module for construction parameters is connected to the real-time prediction and risk warning module for trenching quality and the real-time geological type identification module, respectively. It is used to quickly and smoothly optimize and adjust the construction parameters of the twin-wheel milling machine based on the real-time identified geological type and trenching quality prediction results, and to build a fully closed-loop automatic control system for geological identification, quality prediction and parameter optimization.

[0080] This module receives geological type, fuzzy membership degree, and transition zone width information from the real-time geological type identification module, as well as meter-by-meter and cutter-by-cut trenching quality prediction values ​​from the real-time trenching quality prediction and risk warning module; it directly connects to the dual-wheel milling main control system to output optimized instructions for key construction parameters such as milling wheel speed, feed rate, and mud pressure.

[0081] Model predictive control (MMC) is an advanced model-based control algorithm that employs rolling time-domain optimization. At each sampling time, based on the current system state, it solves for the optimal control sequence over a finite future time domain, executing only the first control variable of the sequence, and repeating the process at the next time step. This rolling optimization mechanism enables the system to continuously adjust its control strategy based on the latest information, exhibiting strong robustness and adaptability.

[0082] Furthermore, the multi-source heterogeneous data acquisition module includes a time alignment unit. The time alignment unit is used to establish a data frame structure with timestamps for various types of data, using the high-frequency sampling clock of the dual-wheel milling main control system as the reference time source. For data sources with different sampling frequencies, adaptive resampling based on spline interpolation and sliding window downsampling based on energy conservation are adopted respectively. For non-periodic data sources, the nearest neighbor time association strategy is used to map to the reference time axis and add deviation marks. The aligned multi-source data is spliced ​​into a multi-channel synchronous data frame with a unified timestamp as the input for subsequent models.

[0083] This unit is the core processing unit of the multi-source heterogeneous data acquisition module. It receives raw data from various sensors, the main control system, geological databases, and historical databases, and outputs multi-channel synchronous data frames with perfect time alignment. This is a prerequisite for the normal operation of all subsequent intelligent algorithms.

[0084] Using the 1000Hz high-frequency sampling clock of the dual-wheel milling main control system as the globally unique reference time source, a global timestamp with microsecond-level precision is generated. All raw data entering the system is immediately stamped with the current global timestamp, overwriting the time stamp of the data source itself. Based on the sampling frequency characteristics of the data source, adaptive resampling, sliding window downsampling, or nearest neighbor time association strategies are used for alignment. The credibility of the aligned data is evaluated, and low-credibility marks are added to data with large deviations. All aligned data are spliced ​​into a multi-channel synchronous data frame with a unified timestamp and output to subsequent modules.

[0085] This unit selects the high-frequency sampling clock of the dual-wheel milling main control system as the sole reference time source, rather than using network time protocols or GPS clocks, for the following reasons: The main control system clock is the control center of the entire device, and all execution commands are issued based on this clock. Using it as a reference ensures accurate synchronization between control commands and feedback data; the main control system clock uses an industrial-grade crystal oscillator with a daily error of less than 1 millisecond, which is far higher than the accuracy of NTP and GPS clocks on the construction site; it is not affected by network conditions and can still work normally even in the event of a network interruption; the global timestamp generated by this unit uses the Unix timestamp format, accurate to microseconds, ensuring that the time synchronization accuracy of all data is less than 1 millisecond.

[0086] Furthermore, it also includes an edge-cloud collaborative transmission and monitoring architecture, which includes:

[0087] This architecture resolves the contradiction between the poor real-time performance of pure cloud architectures and the insufficient computing power of pure edge architectures, and constructs a two-layer collaborative mode of real-time edge inference and offline cloud training:

[0088] For the surface-source computing nodes, an industrial-grade edge gateway deployed at the construction site is responsible, specifically the NVIDIA Jetson Xavier NX model.

[0089] Lightweight preprocessing and feature extraction for raw data; geological identification and parameter optimization tasks with high real-time requirements; data compression: using the LZ4 adaptive compression algorithm, the compression ratio can reach 10:1.

[0090] Cloud servers: Deployed in enterprise data centers or public clouds, responsible for: complex model training and incremental update tasks; big data analysis and historical data mining; centralized monitoring and remote management of multiple devices.

[0091] Transmission protocol: The TCP / IP protocol combined with sequence number confirmation and breakpoint resume mechanism is used between the edge and the cloud. In construction sites with large fluctuations in network bandwidth, the data loss rate is less than 0.1% and the transmission delay is less than 1 second.

[0092] Edge computing nodes deployed at the construction site are used to perform lightweight preprocessing and feature extraction on the raw data before compression and uploading.

[0093] The cloud server is communicatively connected to the edge computing node;

[0094] Among them, the edge computing nodes and the cloud server adopt an adaptive compression transmission strategy and combine it with sequence number confirmation and breakpoint resume protocol to build a two-layer collaborative architecture of edge inference and cloud inference.

[0095] Furthermore, the Transformer-LSTM hybrid neural network model in the real-time prediction and risk warning module for trenching quality adopts a cross-attention fusion architecture: the Transformer component extracts global correlation features through a multi-head self-attention mechanism, the LSTM component models long-term dependence of time-series data, and a cross-attention fusion layer is set between the two to use the global correlation features as the LSTM context vector and calculate the influence weights. After fusion, the meter-by-meter and cutter-by-cut prediction values ​​of trench wall verticality deviation, trench width deviation and sediment thickness are output.

[0096] The above-mentioned structure receives the multi-channel synchronous data frames output by the time alignment unit, outputs the meter-by-meter and cutter-by-cut prediction values ​​of three key tank forming quality indicators: verticality deviation of the tank wall, width deviation of the tank, and thickness of the sediment, and generates corresponding risk warnings based on the prediction values.

[0097] First, the multi-channel synchronous data frames are converted into the model input format and standardized. Then, the global correlation features between the multi-source data are extracted through the multi-head self-attention mechanism to generate a global feature sequence. The input data is then used for time-series modeling to capture the cumulative impact of changes in construction parameters on the trenching quality. Using the hidden state of the LSTM at the previous time step as the query, the influence weight of the global feature sequence on the prediction at the current time step is calculated, and a dynamic context vector is generated and fused with the LSTM hidden state.

[0098] Specifically, the grooving quality achieved by twin-wheel milling is the result of multiple factors working together, with complex nonlinear relationships between parameters from different data sources and at different time points. The Transformer's multi-head self-attention mechanism can compute the correlation strength between all features at all time points in parallel, capturing all global information that would otherwise require LSTM to process step by step.

[0099] Trench formation quality exhibits a significant cumulative effect; the current trench formation quality depends not only on the current construction parameters but also on the combined effect of all construction parameters over a past period. LSTM, through its gating mechanism, effectively captures this long-term temporal dependency, thus solving the gradient vanishing problem of traditional RNNs.

[0100] Instead of concatenating or adding global features and temporal features with fixed weights, the cross-attention fusion layer dynamically calculates the importance weight of each time step in the global feature sequence to the current prediction based on the current temporal state of the LSTM through the cross-attention mechanism, thereby achieving deep adaptive fusion of the two.

[0101] Furthermore, the real-time prediction and risk warning module for trenching quality also includes an online incremental learning unit. The online incremental learning unit is used to extract incremental training samples at fixed time intervals, fine-tune them using an elastic weight consolidation strategy and apply regularization constraints, combine historical data playback samples for mixed training, set a model performance monitor, and automatically roll back and trigger full retraining when the accuracy is lower than a preset lower limit.

[0102] This unit runs on a cloud server and regularly collects verified construction data and actual trenching quality test results from edge nodes. It then performs online fine-tuning and updates to the trenching quality prediction model and synchronously distributes the updated model to all edge computing nodes, achieving global unified evolution of the model.

[0103] Its specific working principle is as follows: Every day at midnight, it automatically extracts all valid construction data uploaded by edge nodes in the past 24 hours; the extracted samples are cleaned, filtered, and labeled, and divided into training and validation sets; the incremental training samples are mixed with historical data playback samples in a certain proportion to construct the final training set; based on the mixed training set, the model is fine-tuned using an elastic weight consolidation strategy to protect parameters important to the old tasks; the accuracy of the new model is tested on a comprehensive validation set containing both new and old data; if the accuracy meets the standard, the new model is synchronously distributed to all edge nodes; if it does not meet the standard, it is automatically rolled back to the previous stable version; if the accuracy fails to meet the standard for three consecutive incremental training iterations, a full retraining process is automatically triggered.

[0104] This unit only extracts valid samples with genuine quality labels for training to avoid contaminating the model with unlabeled or incorrectly labeled data. The criteria for valid samples are: containing a complete 10-second construction parameter sequence and corresponding trenching quality inspection results; continuous construction process without interruption or sensor malfunction; and trenching quality inspection results that have been manually verified and are highly reliable. For the extracted samples, this unit automatically performs standardization processing, unifying the data format and numerical range, while removing outliers and duplicate data. Finally, the samples are divided into training and validation sets.

[0105] Furthermore, the real-time geological type identification module includes a multi-dimensional spectral feature extraction unit. The multi-dimensional spectral feature extraction unit is used to perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signals respectively, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient of wavelet coefficients at the same scale of the two signals and the spectral coherence function to construct cross-signal spectral coupling features, and splice the statistical features and coupling features to form a multi-dimensional spectral feature vector as the input of the classifier.

[0106] This unit receives time-aligned milling wheel torque signals and triaxial vibration acceleration signals from the multi-source heterogeneous data acquisition module, and outputs a multi-dimensional spectral feature vector containing single-signal time-frequency statistical characteristics and cross-signal coupling characteristics, which is the premise and foundation for achieving high-precision geological identification.

[0107] The original torque and vibration signals are bandpass filtered, detrended, and outlier removed to eliminate the effects of environmental noise and sensor drift. The preprocessed one-dimensional time-domain signal is converted into a two-dimensional time-frequency domain representation to obtain wavelet coefficient sequences at different scales. Multiple statistics are calculated for the wavelet coefficients at each scale to comprehensively characterize the amplitude, energy, and distribution characteristics of the signal in different frequency bands. The cross-correlation coefficient and spectral coherence function of the torque and vibration signals at the same scale are calculated to capture the coupling strength and frequency correlation between the two signals. All single-signal statistical features and cross-signal coupling features are concatenated in scale order to form a unified multi-dimensional spectral feature vector, which is then input into the subsequent classifier.

[0108] The signal preprocessing involves addressing the significant noise and interference present in the raw acquired torque and vibration signals, including equipment vibration, hydraulic system fluctuations, and sensor electromagnetic interference. This unit first employs a Butterworth bandpass filter to remove the DC component and high-frequency noise, retaining only the effective frequency components relevant to formation cutting. Then, polynomial fitting is used to remove the slow trend term and eliminate the influence of sensor drift. Finally, the 3σ criterion is used to eliminate outliers in the signal, ensuring the accuracy of subsequent feature extraction.

[0109] Multi-scale continuous wavelet transform, using traditional Fourier transform, can only obtain the global frequency distribution of a signal and cannot reflect the changes in frequency components over time, making it unsuitable for analyzing non-stationary signals such as those from twin-wheel milling operations. This unit employs multi-scale continuous wavelet transform to decompose a one-dimensional time-domain signal into wavelet coefficient sequences at multiple different scales. Each scale corresponds to a specific frequency range: smaller scales correspond to high-frequency components, reflecting rapid changes in the signal; larger scales correspond to low-frequency components, reflecting slow changes in the signal.

[0110] Through this multi-scale decomposition, this unit can simultaneously obtain localized information of the signal in both time and frequency dimensions, clearly capturing the differences in the time-frequency characteristics of the signal under different geological conditions. For example, when the milling wheel cuts a boulder, the signal will show obvious energy spikes in the high-frequency band; while when cutting soft soil, the signal energy is mainly concentrated in the low-frequency band.

[0111] The construction of cross-signal spectrum coupling features involves the fact that when a milling wheel cuts through a formation, the torque signal reflects the magnitude of the cutting resistance, while the vibration signal reflects the impact characteristics during the cutting process. There is a close mechanical coupling relationship between the two, and this coupling relationship varies significantly with different geological types.

[0112] In soft soil layers, the cutting process is smooth, and the coupling between torque and vibration signals is weak.

[0113] In rock formations, the cutting process generates intense impacts, significantly enhancing the coupling between torque and vibration signals.

[0114] In the transition zone between soft and hard interlayers and strata, the coupling relationship undergoes continuous and significant changes.

[0115] This coupling relationship is quantitatively characterized by calculating the cross-correlation coefficient and spectral coherence function of the torque signal and vibration signal at the same scale:

[0116] Cross-correlation coefficient: reflects the degree of linear correlation between two signals in the time domain, and its value ranges from -1 to 1. The larger the absolute value, the stronger the correlation.

[0117] Spectral coherence function: reflects the degree of correlation between two signals in the frequency domain. The value ranges from 0 to 1. The larger the value, the stronger the coupling between the two signals at that frequency.

[0118] These cross-signal coupling features can effectively distinguish different geological types with similar single signal features, especially for soft and hard interlayers and stratigraphic transition zones, where their identification ability is far higher than that of single signal features.

[0119] The multidimensional spectral feature vector concatenation method sequentially concatenates five single-signal statistical features and two cross-signal coupling features at each scale in scale order, forming a multidimensional spectral feature vector with a dimension of "scale number × 7". This invention uses eight logarithmically spaced scales for decomposition, ultimately generating a 56-dimensional feature vector. This feature vector contains all the key information of the signal in different frequency bands, possessing extremely high discriminative power, enabling subsequent classifiers to quickly and accurately identify the geological type of the current construction.

[0120] Furthermore, the real-time geological type identification module also includes a stratigraphic transition zone fuzzy soft identification unit. The stratigraphic transition zone fuzzy soft identification unit is used to receive the probability distribution vectors of various geological types output by the classifier, convert them into membership degrees through a fuzzy membership function, and determine the stratigraphic transition zone when the difference between the highest and second highest membership degrees is less than a preset distinction threshold and estimate the width. The membership degree vector and the transition zone label are then output to the fully closed-loop construction parameter automatic optimization control module.

[0121] This unit receives the probability distribution vectors of six geological types from a lightweight convolutional neural network classifier, and outputs the fuzzy membership vectors of each geological type, stratigraphic transition zone markers, and estimated transition zone widths. It is a key link connecting geological identification and construction parameter optimization, and directly determines the smoothness and accuracy of parameter adjustment under complex stratigraphic conditions.

[0122] Specifically, the probability distribution vectors of the six geological types output by the convolutional neural network classifier are obtained; the hard probability distribution is converted into a fuzzy membership vector through a fuzzy membership function, amplifying the differences of high-probability categories and suppressing low-probability noise; the difference between the highest and second-highest membership is calculated and compared with a preset discrimination threshold to determine whether the current stratum is a transition zone; for areas determined to be transition zones, the actual width of the transition zone is quantitatively estimated by monitoring the rate of change of membership over time and combining it with the real-time feed speed of the dual-wheel milling machine; the fuzzy membership vector, transition zone label, and transition zone width are simultaneously output to the automatic optimization control module for construction parameters and the dynamic condition modulation module.

[0123] The probability distribution output by a convolutional neural network classifier is a hard probability estimate based on the training data. For mixed types such as geological transition zones, the probability distribution usually shows two high probability values ​​coexisting. Directly using hard classification will lose a lot of gradual information.

[0124] This unit performs a nonlinear transformation on the original probability distribution using a fuzzy membership function. Its core functions are: firstly, to amplify the differences between high-probability categories, improving the distinguishability of major geological types; and secondly, to suppress the influence of low-probability noise, enhancing the stability of the identification results. The transformed fuzzy membership vector can more accurately reflect the actual proportion of different geological types in the current strata.

[0125] The essential characteristic of stratigraphic transition zones is the continuous variation in the compositional ratio of the two geological types. This is reflected in the fuzzy membership vector as a significantly smaller difference between the highest and second-highest membership degree compared to a single geological type. For a single geological type, the highest membership degree is typically greater than 0.9, and the difference between it and the second-highest membership degree is greater than 0.8; while for transition zones, the highest membership degree is typically between 0.5 and 0.8, and the difference between it and the second-highest membership degree is less than 0.3.

[0126] This unit determines the transition zone by calculating the difference between the highest and second-highest membership degrees and comparing it with a preset distinction threshold. This invention has undergone extensive engineering verification, setting the preset distinction threshold to 0.3: when the difference is greater than or equal to 0.3, it is determined to be a single geological type; when the difference is less than 0.3, it is determined to be a transition zone between two geological types. This determination method can accurately distinguish between a single geological type and a transition zone, with a false positive rate of less than 5%.

[0127] This unit estimates the width of the transition zone by monitoring the rate of change of the highest membership degree over time. As the strata transition from one geological type to another, the original highest membership degree gradually decreases, while a new highest membership degree gradually increases. By calculating the rate of change of the highest membership degree and combining it with the real-time feed rate of the twin-wheel milling machine, the actual width of the transition zone can be accurately estimated. For example, if the highest membership degree decreases from 0.9 to 0.2 in 30 seconds, and the twin-wheel milling machine feed rate is 2 meters per hour, then the width of the transition zone is approximately 0.017 meters.

[0128] The estimation method can dynamically update the width of the transition zone in real time, with an estimation error of less than 0.1 meters, which fully meets the accuracy requirements for adjusting construction parameters.

[0129] This unit not only outputs the transition zone markers and widths, but more importantly, it outputs the complete fuzzy membership vector, rather than a single geological type label. This is because in the transition zone area, construction parameters should not abruptly switch from one geological model to another, but should be smoothly adjusted according to the membership ratio of the two geological types.

[0130] The automatic optimization control module for construction parameters calculates a weighted average of the optimal parameters corresponding to the two geological types based on the membership vector, generating the optimal construction parameters for the current transition zone. For example, in the transition zone from soft soil to rock, when the membership degree of soft soil is 0.6 and that of rock is 0.35, the construction parameters will be a weighted combination of 60% of the soft soil parameters and 35% of the rock parameters, achieving a continuous and smooth transition of parameters and avoiding equipment shocks and fluctuations in trenching quality caused by abrupt parameter changes.

[0131] Furthermore, the model prediction control algorithm in the fully closed-loop construction parameter automatic optimization control module adopts a multi-objective weighted function as the optimization objective function. The multi-objective weighted function includes three cost terms: trenching quality, construction efficiency, and parameter change. The weight of each cost term is dynamically adjusted according to the real-time geological type. At the same time, hard constraints are set for milling wheel torque, rotation speed, mud pressure, and parameter change rate.

[0132] The system receives geological type, fuzzy membership degree, and transition zone width information from the real-time geological type identification module, as well as meter-by-meter and cutter-by-cut trenching quality prediction values ​​from the real-time trenching quality prediction and risk warning module. It then generates optimal control commands for key construction parameters such as milling wheel speed, feed rate, and mud pressure, directly driving the dual-wheel milling main control system. The system acquires the current geological identification results, trenching quality prediction values, actual equipment operating status, and target trenching quality values. Based on the real-time geological type and fuzzy membership degree, it automatically calculates the weight coefficients of three cost terms. Based on the prediction model, it constructs a quadratic programming problem containing multi-objective weighted functions and hard constraints. Using an efficient numerical solver, it solves the optimal construction parameter control sequence within the finite time domain. Only the first parameter command of the control sequence is executed; the process is repeated at the next sampling time based on the latest status information.

[0133] Furthermore, the fully closed-loop automatic optimization control module for construction parameters also includes a delay compensation unit based on a Smith predictor. This delay compensation unit is used to establish a prediction model for the forward and feedback channels of the construction process. It introduces the delayed trenching quality state predicted by the Smith predictor to replace the delayed measurement value in the optimization process. Through step testing and cross-correlation analysis, it dynamically estimates the delay time and adaptively adjusts the control frequency and prediction time domain under large delay conditions. Running within the automatic optimization control module, it receives control commands from the model predictive control algorithm and actual trenching quality feedback data, and outputs the equivalent delay-free system state after delay compensation for optimization by the model predictive control algorithm. This is a crucial step in solving the control challenges of large delay systems.

[0134] Specifically, a forward transfer function model and feedback channel model for the twin-wheel milling construction process are established based on historical construction data using a system identification method. Step tests and cross-correlation analysis are used to monitor the time difference between control commands and trenching quality feedback in real time, dynamically estimating the system's pure delay time. The Smith predictor calculates the delayed trenching quality state in advance based on historical control commands and the prediction model, generating an equivalent zero-delay system state. The estimated zero-delay state replaces the actual delay measurement value and is input into the model predictive control algorithm for optimization. When the estimated delay time exceeds a preset threshold, the control frequency, prediction time domain, and control time domain are automatically adjusted to ensure system stability under extreme large-delay conditions.

[0135] Specifically, the Smith predictor runs two models in parallel: a zero-delay predictive model to simulate the immediate effect of control commands, and a delayed predictive model to simulate the delay characteristics of the actual system. By combining the outputs of the two models with the feedback from the actual system, an equivalent zero-delay system state is generated. When the predictive model matches the characteristics of the actual system, the controller sees a system with absolutely no delay, thus completely eliminating the impact of large delays on the stability of the control system and enabling the system's response speed and stability to reach the same level as a zero-delay system.

[0136] The biggest drawback of traditional Smith predictors is their use of a fixed delay time, while the delay time in a twin-wheel milling process is dynamically changing. This unit employs a method combining step testing and cross-correlation analysis to dynamically estimate the pure delay time of the system in real time.

[0137] Step test: When the system starts up or undergoes significant changes in operating conditions, a small-amplitude step control command is automatically applied to measure the response time of the tank formation quality and obtain an initial estimate of the delay time.

[0138] Cross-correlation analysis: During normal construction, the correlation between the control command sequence and the trenching quality feedback sequence is continuously analyzed. The time difference corresponding to the correlation peak is the current pure delay time.

[0139] Delay time update: The estimated delay time is updated every 10 minutes, and a smoothing filter mechanism is set to avoid frequent changes in delay time from affecting system stability.

[0140] When the estimated delay exceeds 20 seconds, the system automatically enters a large delay adaptive mode, adjusting key parameters of the control algorithm to ensure system stability.

[0141] The prediction time domain was increased from the default 20 sampling periods to 30 sampling periods to ensure that the prediction range could cover the entire delay time.

[0142] Reducing the control time domain from the default 5 sampling periods to 3 sampling periods decreases the complexity of the optimization problem and increases the solution speed.

[0143] The control frequency was reduced from the default 10Hz to 5Hz to avoid system oscillation caused by overly frequent control adjustments.

[0144] By appropriately tightening the constraints on the rate of change of parameters, the smoothness of control commands can be further guaranteed.

[0145] Through the above adaptive adjustments, the system can maintain stable operation even under extreme conditions with a delay of up to 30 seconds, and control the overshoot to be less than 5%.

[0146] Furthermore, it also includes a dynamic condition modulation module, which is connected between the real-time geological type identification module and the trenching quality real-time prediction and risk warning module. It is used to take the geological identification results and fuzzy membership degree as condition variables, generate scaling and offset coefficients through a feature linear modulation mechanism to perform affine transformation on the model fusion feature vector, and enhance the corresponding feature channel weights according to different geological types, so that the single model can adapt to different geological conditions.

[0147] It is used to receive the fuzzy membership vectors of 6 geological types output by the real-time geological type identification module, generate scaling coefficients and offset coefficients consistent with the fusion feature dimensions of the trenching quality prediction model, and output the intermediate fusion features of the model to the prediction output layer after performing affine transformation, so as to realize the dynamic guidance of geological conditions on the quality prediction process.

[0148] Specifically, the system obtains fuzzy membership vectors of six geological types output by the real-time geological type identification module. These vectors reflect the actual proportion of each geological type in the current strata. Two independent lightweight fully connected networks are used to map the geological membership vectors to scaling and offset coefficient vectors, respectively, with the same dimensions as the fused feature vectors. The generated scaling and offset coefficients are then used to perform an element-wise affine transformation on the fused feature vectors output by the cross-attention fusion layer in the trenching quality prediction model. The modulated feature vectors are then input into the final output layer of the trenching quality prediction model to generate trenching quality prediction values ​​adapted to the current geological conditions.

[0149] Affine transformations are used to dynamically weight feature channels: for important feature channels under current geological conditions, their scaling factor is increased to enhance their feature response; for less important feature channels, their scaling factor is decreased to suppress interference from irrelevant information. Simultaneously, the overall distribution of features is adjusted using offset factors to better align with the trenching quality mapping patterns under current geological conditions. This dynamic adjustment mechanism allows the model to automatically focus on the most critical features of the current geological conditions, thereby significantly improving prediction accuracy.

[0150] Hard geological labels can only indicate whether a stratum is or is not a certain geological type, and cannot reflect the gradual changes in strata. In contrast, fuzzy membership vectors can accurately represent the mixing ratio of various geological types in the current strata. The modulation coefficients generated based on fuzzy membership will change continuously and smoothly with the changes in geological composition. In geological transition zones, the model's prediction results will also transition continuously without any abrupt changes.

[0151] All parameters are jointly trained end-to-end with the parameters of the trenching quality prediction model. During training, the model automatically learns the optimal modulation coefficients corresponding to different geological types, that is, it automatically learns which features are more important under which geological conditions. This joint training method requires no manual intervention and can fully explore the intrinsic correlation between geological conditions and trenching quality characteristics, so as to achieve the optimal modulation effect.

[0152] Example 2

[0153] This embodiment also provides a method for managing and remotely monitoring construction data of dual-wheel milling machines, including the following steps:

[0154] S1: Real-time acquisition of multi-source heterogeneous data from dual-wheel milling operations, including operating parameters, geological stratification, and historical data under the same working conditions; using the high-frequency clock of the main control system as a reference, data frames with timestamps are established for various types of data; spline interpolation resampling and energy conservation downsampling are used for data with different sampling frequencies; non-periodic data are correlated with nearest neighbor time and marked with deviations, and spliced ​​into multi-channel synchronous data frames; after lightweight preprocessing and feature extraction at the field edge nodes, the data is uploaded to the cloud through adaptive compression and breakpoint resume protocol, constructing an edge-cloud dual-layer collaborative inference architecture;

[0155] S2: Perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signal, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient and spectral coherence function at the same scale to construct cross-signal coupling features, and splice them to form a multi-dimensional spectral feature vector; input the lightweight CNN classifier to obtain the probability distribution of geological types, and after fuzzy membership function transformation, when the difference between the highest and second highest membership degrees is less than the threshold, it is determined to be a stratigraphic transition zone and its width is estimated, and strata such as soft soil, sand, rock, boulders, soft and hard interlayers and junctions are identified in real time.

[0156] S3: Using geological identification results and membership degrees as conditional variables, scaling and offset coefficients are generated through feature linear modulation. Affine transformation is performed on the fusion feature vector of the trenching quality prediction model, and the weights of corresponding feature channels are dynamically enhanced according to geological type.

[0157] S4: Based on multi-channel synchronous data frames, a Transformer-LSTM hybrid model with cross-attention fusion is used to predict the trenching quality meter by meter and blade by blade: Transformer extracts global correlation features from multi-source data, LSTM models long-term temporal dependencies, and the cross-attention layer uses global features as LSTM context vectors to calculate influence weights, outputting predicted values ​​for trench wall verticality, trench width, and sludge thickness, and providing graded early warnings; incremental samples are extracted at fixed intervals, and an elastic weight consolidation strategy combined with historical playback is used to fine-tune the model. A performance monitor is set up, and when the accuracy is lower than the threshold, automatic rollback and full retraining are triggered.

[0158] S5: Based on geological identification and quality prediction results, a model predictive control algorithm is used to automatically optimize construction parameters: a multi-objective weighted function is constructed with trenching quality, construction efficiency, and parameter changes as cost terms, and the weights are dynamically adjusted according to the geological type. At the same time, hard constraints are set for torque, rotational speed, mud pressure, and parameter change rate. A construction forward and reverse prediction model is established, and a Smith predictor is introduced to compensate for measurement delay. The delay time is dynamically estimated through step test and cross-correlation analysis. Under the condition of large delay, the control frequency and prediction time domain are adaptively adjusted to construct a closed-loop control system of geological identification-quality prediction-parameter optimization.

[0159] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A dual-wheel milling machine construction data management and remote monitoring system, characterized in that, include: The multi-source heterogeneous data acquisition module is used to acquire multi-source heterogeneous data in real time during the construction process of the dual-wheel milling machine. The multi-source heterogeneous data includes at least real-time construction operation parameters, on-site geological stratification data, and historical construction data under the same working conditions. The real-time prediction and risk warning module for trenching quality is communicatively connected to the multi-source heterogeneous data acquisition module. Based on the multi-source heterogeneous data, it uses a pre-trained Transformer-LSTM hybrid neural network model to predict the trenching quality meter by meter and cutter by cutter in real time, and outputs a corresponding level of risk warning based on the prediction results. The Transformer component is used to extract global correlation features between multi-source heterogeneous data, and the LSTM component is used to model the long-term dependencies of time-series construction data. The real-time geological type identification module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to extract the multi-dimensional spectral features of the milling wheel torque signal and the milling wheel vibration signal, and input them into a pre-trained lightweight convolutional neural network classifier to identify the geological type of the current construction stratum in real time. The geological type includes at least soft soil, sandy soil, rock, boulders, soft and hard interlayers and strata junctions. The construction parameter optimization control module is connected to the trenching quality real-time prediction and risk warning module and the real-time geological type identification module, respectively. It is used to quickly and smoothly optimize and adjust the construction parameters of the twin-wheel milling machine based on the real-time identified geological type and trenching quality prediction results, and to build a fully closed-loop automatic control system for geological identification, quality prediction and parameter optimization.

2. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module includes a time alignment unit. The time alignment unit is used to establish a data frame structure with timestamps for various types of data, using the high-frequency sampling clock of the dual-wheel milling main control system as the reference time source. For data sources with different sampling frequencies, adaptive resampling based on spline interpolation and sliding window downsampling based on energy conservation are adopted respectively. For non-periodic data sources, the nearest neighbor time association strategy is used to map to the reference time axis and add deviation marks. The aligned multi-source data is spliced ​​into a multi-channel synchronous data frame with a unified timestamp as the input for subsequent models.

3. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, It also includes an edge-cloud collaborative transmission and monitoring architecture, which includes: Edge computing nodes deployed at the construction site are used to perform lightweight preprocessing and feature extraction on the raw data before compression and uploading. The cloud server is communicatively connected to the edge computing node; Among them, the edge computing nodes and the cloud server adopt an adaptive compression transmission strategy and combine it with sequence number confirmation and breakpoint resume protocol to build a two-layer collaborative architecture of edge inference and cloud inference.

4. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The Transformer-LSTM hybrid neural network model in the real-time prediction and risk warning module for trenching quality adopts a cross-attention fusion architecture. The Transformer component extracts global correlation features through a multi-head self-attention mechanism, while the LSTM component models long-term dependencies in time-series data. A cross-attention fusion layer is set between the two to use the global correlation features as the LSTM context vector and calculate the influence weights. After fusion, the meter-by-meter and cutter-by-cut prediction values ​​of the verticality deviation of the tank wall, the width deviation of the tank, and the thickness of the sediment are output.

5. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The real-time prediction and risk warning module for trenching quality also includes an online incremental learning unit. The online incremental learning unit is used to extract incremental training samples at fixed time intervals, fine-tune them using an elastic weight consolidation strategy and apply regularization constraints, combine historical data playback samples for mixed training, set a model performance monitor, and automatically roll back and trigger full retraining when the accuracy is lower than the preset lower limit.

6. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The real-time geological type identification module includes a multi-dimensional spectral feature extraction unit. The multi-dimensional spectral feature extraction unit is used to perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signals respectively, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient of wavelet coefficients at the same scale of the two signals and the spectral coherence function to construct cross-signal spectral coupling features, and splice the statistical features and coupling features to form a multi-dimensional spectral feature vector as the input of the classifier.

7. The dual-wheel milling construction data management and remote monitoring system according to claim 5, characterized in that, The real-time geological type identification module also includes a stratigraphic transition zone fuzzy soft identification unit. The stratigraphic transition zone fuzzy soft identification unit is used to receive the probability distribution vectors of various geological types output by the classifier, convert them into membership degrees through a fuzzy membership function, and determine the stratigraphic transition zone when the difference between the highest and second highest membership degrees is less than a preset distinction threshold and estimate the width. The membership degree vector and the transition zone label are then output to the fully closed-loop construction parameter automatic optimization control module.

8. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The model prediction control algorithm in the fully closed-loop construction parameter automatic optimization control module adopts a multi-objective weighted function as the optimization objective function. The multi-objective weighted function includes three cost terms: trenching quality, construction efficiency, and parameter change. The weight of each cost term is dynamically adjusted according to the real-time geological type. At the same time, hard constraints are set for milling wheel torque, rotation speed, mud pressure, and parameter change rate.

9. The dual-wheel milling construction data management and remote monitoring system according to claim 1, characterized in that, The fully closed-loop construction parameter automatic optimization control module also includes a delay compensation unit based on the Smith predictor. The delay compensation unit is used to establish a prediction model of the positive and feedback channels of the construction process, introduce the delayed trenching quality state predicted by the Smith predictor to replace the delayed measurement value for optimization, dynamically estimate the delay time through step test and cross-correlation analysis, and adaptively adjust the control frequency and prediction time domain under large delay conditions. It also includes a dynamic condition modulation module, which is connected between the real-time geological type identification module and the trenching quality real-time prediction and risk warning module. The dynamic condition modulation module is used to take the geological identification results and fuzzy membership degree as condition variables, generate scaling and offset coefficients through the feature linear modulation mechanism to perform affine transformation on the model fusion feature vector, and enhance the corresponding feature channel weights according to different geological types, so that the single model can adapt to different geological conditions.

10. A method for managing and remotely monitoring construction data of a dual-wheel milling machine, characterized in that, Includes the following steps: S1: Real-time acquisition of multi-source heterogeneous data from dual-wheel milling operations, including operating parameters, geological stratification, and historical data under the same working conditions; using the high-frequency clock of the main control system as a reference, establish data frames with timestamps for various types of data; use spline interpolation resampling and energy conservation downsampling for data with different sampling frequencies; use nearest neighbor time correlation for non-periodic data and mark the deviation, and splice them into multi-channel synchronous data frames. After lightweight preprocessing and feature extraction at the on-site edge nodes, the data is uploaded to the cloud via adaptive compression and breakpoint resume protocol to build a two-layer collaborative inference architecture between the edge and the cloud. S2: Perform multi-resolution continuous wavelet transform on the milling wheel torque and vibration signal, extract the statistical features of wavelet coefficients at each scale, calculate the cross-correlation coefficient and spectral coherence function at the same scale to construct cross-signal coupling features, and splice them to form a multi-dimensional spectral feature vector; input the lightweight CNN classifier to obtain the probability distribution of geological types, and after fuzzy membership function transformation, when the difference between the highest and second highest membership degrees is less than the threshold, it is determined to be a stratigraphic transition zone and its width is estimated, and strata such as soft soil, sand, rock, boulders, soft and hard interlayers and junctions are identified in real time; S3: Using geological identification results and membership degrees as conditional variables, scaling and offset coefficients are generated through feature linear modulation. Affine transformation is performed on the fusion feature vector of the trenching quality prediction model, and the corresponding feature channel weights are dynamically enhanced according to geological type. S4: Based on multi-channel synchronous data frames, a Transformer-LSTM hybrid model with cross-attention fusion is used to predict the trenching quality meter by meter and blade by blade: Transformer extracts global correlation features from multi-source data, LSTM models long-term temporal dependencies, and the cross-attention layer uses global features as LSTM context vectors to calculate influence weights, outputting predicted values ​​for trench wall verticality, trench width, and sludge thickness, and providing graded early warnings; incremental samples are extracted at fixed intervals, and an elastic weight consolidation strategy combined with historical playback is used to fine-tune the model. A performance monitor is set up, and when the accuracy is lower than the threshold, automatic rollback and full retraining are triggered. S5: Based on geological identification and quality prediction results, a model predictive control algorithm is used to automatically optimize construction parameters: a multi-objective weighted function is constructed with trenching quality, construction efficiency, and parameter changes as cost terms, and the weights are dynamically adjusted according to the geological type. At the same time, hard constraints are set for torque, rotational speed, mud pressure, and parameter change rate. A construction forward and reverse prediction model is established, and a Smith predictor is introduced to compensate for measurement delay. The delay time is dynamically estimated through step test and cross-correlation analysis. Under the condition of large delay, the control frequency and prediction time domain are adaptively adjusted to construct a closed-loop control system of geological identification-quality prediction-parameter optimization.

Images