Tunnel lining construction intelligent trolley and control system thereof

Abstract

The application relates to the technical field of tunnel construction intelligent control, and discloses a tunnel lining construction intelligent trolley and a control system thereof. The system comprises a state change time point determined through event listening, according to which multi-source sensor data are intercepted and features are extracted. A mode matching module generates a quality difference measure by comparing a standard with an abnormal mode. A causal analysis framework with an operation frequency as an adjusting variable is innovatively introduced, the influence weight of each sensor data stream on quality is dynamically evaluated, and key quality factors are accurately identified. An optimized feature set is partitioned and merged based on a construction waveform time sequence relationship, and a control instruction is generated by adopting an attention mechanism. The scheme realizes the transformation from static threshold judgment to dynamic causal inference, adaptive control is realized through time sequence context understanding, and the quality diagnosis accuracy and the adaptive capacity of the system to complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for tunnel construction, specifically to an intelligent trolley for tunnel lining construction and its control system. Background Technology

[0002] In tunnel lining construction, existing technologies primarily rely on single-sensor monitoring based on fixed thresholds or simple multi-sensor data fusion methods. Current systems typically process various sensor signals independently and set static safety thresholds for each parameter. When a parameter exceeds a threshold, the system triggers an alarm or executes a preset shutdown command. For multi-source data, a common practice is to use empirical weighted averaging or logical combinations based on fixed rules for status judgment; these data fusion models lack adaptability. Operators must set these thresholds and rules based on personal experience; the system itself cannot autonomously learn from historical data or real-time operations to identify key parameters that decisively influence construction quality. This approach treats all sensor data as equally or statically important, ignoring the dynamic changes in the impact of each parameter on the final quality under different operating modes during construction.

[0003] In terms of control command generation, existing technologies generally employ pre-programmed sequential control logic or conditional judgments based on instantaneous states. The control system selects an action from a predefined command library based on whether the sensor readings at the current moment meet specific conditions. This method cannot analyze dynamic interaction characteristics such as the alignment and phase relationship between pressure change curves and vibration waveforms on the time axis. The control strategy is reactive and fragmented, failing to understand a continuous construction process as a complete context. The system cannot simulate the operational wisdom of skilled workers, i.e., dynamically adjusting the control focus according to the phased progress of the process, resulting in poor adaptability to complex working conditions. The generation of control commands is difficult to achieve precision and smooth transition, hindering further improvement in the automation and intelligence level of lining construction.

[0004] The drawback of existing technologies lies in their static and relational analytical frameworks, and their instantaneous and isolated control strategies. This makes it difficult for the system to accurately identify the root causes affecting quality, and the generated control instructions lack adaptability to the dynamic evolution of the entire construction process. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent trolley for tunnel lining construction and its control system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a control system for an intelligent trolley used for tunnel lining construction, the system comprising:

[0007] The event monitoring module is used to monitor the occurrence signals of trolley operation events and mark the reference time points of construction status changes in the continuously acquired sensor data stream;

[0008] The data acquisition and processing module is used to extract multi-source sensor data within a fixed time interval before and after a reference time point, and to transform and process the multi-source sensor data to derive a characteristic representation of the construction status.

[0009] The pattern matching module is used to feed the feature representation into the pattern matching unit, evaluate the pattern consistency between standard construction patterns and abnormal construction patterns, and generate a quality difference metric.

[0010] The causal analysis module is used to use the trolley operation frequency as a moderating variable, process quality difference measurement through the causal analysis framework, and estimate the impact weight of each sensor data stream on construction quality.

[0011] The feature optimization module is used to integrate the dynamic indicators of construction resistance and the trolley movement trajectory indicators, optimize the selection conditions of influence weights, screen out the feature representations corresponding to high influence weights, and form an optimized feature set.

[0012] The instruction generation module is used to partition and merge the optimized feature set according to the timing alignment relationship of the construction waveform, and generate driving instructions using an attention allocation mechanism.

[0013] Preferably, the signal for monitoring the occurrence of the trolley operation event includes: continuously receiving control pulse signals from the trolley actuator, detecting edge changes of the control pulse signals as event trigger markers; recording the precise timestamp of the event trigger marker, and comparing the time difference between the timestamp and the displacement mutation point in the displacement sensor data sequence; when the time difference is less than a set tolerance, confirming the timestamp as a reference time point for the change in construction status.

[0014] Preferably, the step of extracting multi-source sensor data within a fixed time interval before and after a reference time point includes: taking the reference time point as the center, extending forward and backward by a preset time duration to obtain the raw readings of the accelerometer, pressure sensor, and displacement sensor within the time window; performing a moving average filter on the raw readings to suppress noise, and then calculating the data variance and mean of each sensor channel as preliminary statistics; standardizing the preliminary statistics to unify the data scale of each channel, and then using principal component analysis to reduce dimensionality and generate a feature representation of the construction status.

[0015] Preferably, the step of feeding the feature representations into the pattern matching unit to evaluate the pattern consistency between the standard construction pattern and the abnormal construction pattern, and generating a quality difference metric, includes: constructing a dual-branch structure for the pattern matching unit, each branch containing a convolutional layer and a fully connected layer to process the feature representations of the standard construction pattern and the abnormal construction pattern, respectively; calculating the Euclidean distance between the output vectors of the two branches, and converting the Euclidean distance into a similarity score using the sigmoid function; adjusting the weights of the feature representations according to the similarity score to maximize the similarity score in the standard pattern and minimize the similarity score in the abnormal pattern, thereby outputting a quality difference metric.

[0016] Preferably, the step of using the trolley operation frequency as a moderating variable and processing the quality difference measure through a causal analysis framework to estimate the impact weight of each sensor data stream on construction quality includes: counting the number of trolley operation events occurring per unit time and dividing the operation frequency into three intervals: low, medium, and high; establishing a structural equation model of the causal analysis framework, with the quality difference measure as the dependent variable, the construction status as the processing variable, and the operation frequency as the moderating variable; calculating the marginal effect of each sensor data stream when the processing variable changes through least squares estimation, and aggregating the average marginal effect of all samples to obtain the impact weight of each sensor data stream.

[0017] Preferably, the dynamic index of construction resistance is obtained by: extracting low-frequency components related to construction resistance from the feature representation, performing spectral analysis on the low-frequency components, calculating the distribution ratio of spectral energy within a predetermined frequency band; and combining the displacement change rate in the trolley motion trajectory index to perform a weighted summation of the distribution ratio to generate the dynamic index of construction resistance.

[0018] Preferably, the trolley motion trajectory index is obtained by: separating the trolley displacement trajectory data from the quality difference measurement, calculating the first and second derivatives of the displacement trajectory as motion features; performing time warping on the motion features, aligning trajectory segments of different construction stages, and then calculating the eigenvalues ​​of the covariance matrix of the trajectory segments to derive the trolley motion trajectory index.

[0019] Preferably, the selection conditions for the optimized influence weights include: inputting the dynamic index of construction resistance and the trolley motion trajectory index into a pre-trained support vector machine model, wherein the decision function output of the support vector machine model is used to adjust the threshold of the influence weights; and comparing the influence weights of each sensor data stream according to the adjusted threshold, retaining the feature representations corresponding to the sensor data streams with influence weights higher than the threshold, thereby forming an optimized feature set.

[0020] Preferably, partitioning the optimized feature set according to the time alignment relationship of the construction waveform includes: extracting zero-crossing points and extreme points from the construction waveform signal, dividing the waveform into multiple phase intervals; grouping the optimized feature set according to the time boundaries of the phase intervals, and performing time synchronization correction on the feature representations within each group;

[0021] The method of generating driving instructions using an attention allocation mechanism includes: calculating a self-attention score for the feature representation of each phase interval; performing a weighted average of the feature representation based on the self-attention score to obtain the interval representation; inputting all interval representations into a long short-term memory network for sequence modeling; and mapping the final hidden state of the network to driving instructions through a fully connected layer.

[0022] Preferably, the present invention also includes an intelligent trolley for tunnel lining construction, the trolley including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the functions of the control system of the intelligent trolley for tunnel lining construction as described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] By employing a causal inference model with the trolley operation frequency as the core moderating variable, this system can dynamically analyze the intrinsic causal relationship between multi-source sensor data and construction quality difference measurements, rather than superficial statistical correlations. Introducing the dynamic working condition of construction rhythm as a moderating variable into the analytical framework allows the system to identify the true driving effect of each sensor parameter on the final lining quality under different operational intensities and quantify it as a specific influence weight. This achieves precise isolation of key quality influencing factors, and system resources can automatically focus on sensor data streams that consistently exhibit high causal effects under different construction rhythms. This improves the targeting and anti-interference capability of diagnosing the root causes of anomalies, avoiding false alarms and missed alarms caused by non-causal correlations.

[0025] Based on refined time-series alignment analysis of waveforms throughout the entire construction process, the optimized feature set is dynamically partitioned and merged within a specific context, and then an attention mechanism is employed to generate control commands. This technology enables the system to move beyond isolated responses to instantaneous states and instead understand the construction process from the perspective of macroscopic process segments. By analyzing the alignment, phase, and shape relationships of different sensor signal waveforms along the time axis, the system can divide continuous operations into stages with clear engineering significance. On this basis, the attention mechanism can dynamically adjust the degree of attention to different control objectives according to the current stage and real-time characteristics. This achieves a fundamental shift in control strategy from a fixed procedure to context-adaptive approaches. The command generation process can simulate expert experience, smoothly shifting the control focus at different stages of the process. Therefore, when facing complex, nonlinear, and time-varying construction processes, it exhibits higher control flexibility and decision-making intelligence, making the trolley movements more aligned with actual process requirements. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the working principle of the control system of the intelligent trolley for tunnel lining construction described in this invention.

[0027] Figure 2 A flowchart for monitoring the occurrence of trolley operation events;

[0028] Figure 3 A flowchart generated for pattern matching and quality difference measurement;

[0029] Figure 4 A graph showing the variation of eigenvalues ​​in the covariance matrix of the intelligent trolley's motion trajectory during tunnel lining construction.

[0030] Figure 5 A phase interval attention score distribution diagram for the intelligent trolley instruction generation module in tunnel lining construction. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figure 1This invention provides a control system for an intelligent trolley used in tunnel lining construction. The system includes: an event monitoring module responsible for real-time monitoring of trolley operation events, such as identifying reference time points for changes in construction status by detecting edge changes in control pulse signals; this reference time point serves as the benchmark for data processing. A data acquisition and processing module extracts multi-source sensor data within a fixed time interval based on the reference time point, including readings from acceleration, pressure, and displacement sensors, and performs filtering, standardization, and dimensionality reduction on this data to generate a feature representation of the construction status. A pattern matching module uses a dual-branch structure to evaluate the consistency of the feature representation with standard or abnormal construction modes, outputting a quality difference metric. A causal analysis module uses the trolley operation frequency as a moderating variable, analyzes the quality difference metric through structural equation modeling, and estimates the influence weight of each sensor data stream on construction quality. A feature optimization module integrates dynamic indicators of construction resistance and trolley motion trajectory indicators, optimizes the selection conditions for influence weights, and selects features with high influence weights to form an optimized feature set. An instruction generation module partitions and merges the optimized feature set according to the temporal alignment relationship of the construction waveform, and applies an attention allocation mechanism to generate drive instructions to control the movement of the trolley and the operation of the actuators.

[0033] Example 1: See Figure 2 In practical implementation, the event monitoring module is used to monitor the occurrence signals of trolley operation events. The event monitoring module continuously receives control pulse signals from the trolley actuator and detects edge changes of the control pulse signals as event trigger flags. The event monitoring module records the precise timestamp of the event trigger flag and compares the time difference between the timestamp and the displacement change point in the displacement sensor data sequence. When the time difference is less than the set tolerance, the event monitoring module confirms that the timestamp is the reference time point for the change in construction status. The implementation of the event monitoring module relies on a high-speed digital signal processing circuit. The high-speed digital signal processing circuit can capture the rising or falling edge changes of the control pulse signal in real time and store the timestamp of the event trigger flag in a buffer. The displacement sensor data sequence is acquired through high sampling rate. The system uses an interpolation algorithm to align the timestamp and displacement change point. The set tolerance value is determined based on the sensor response time and communication delay. For example, the tolerance can be set to the millisecond level. The precise time synchronization of the event monitoring module is the basis for construction status monitoring. In some embodiments, the event monitoring module may further include a redundancy detection mechanism, which improves reliability through multiple event trigger flag verification. The buffer of the event monitoring module adopts a circular storage structure to avoid data overflow. The edge detection algorithm of the event monitoring module uses a digital filter to preprocess the control pulse signal to eliminate jitter interference. The timestamp comparison process of the event monitoring module adopts the least squares fitting method to optimize the time difference calculation accuracy. The set tolerance value of the event monitoring module can be dynamically adjusted according to the construction environment. For example, the tolerance value is reduced under high-speed operation to improve sensitivity.

[0034] The data acquisition and processing module extracts multi-source sensor data within a fixed time interval before and after a reference time point. Centered on the reference time point, the module extends forward and backward by preset durations to acquire raw readings from the accelerometer, pressure sensor, and displacement sensor within this time window. The module applies a moving average filter to the raw readings to suppress noise, then calculates the variance and mean of each sensor channel as preliminary statistics. These preliminary statistics are standardized to ensure uniform data scale across channels. Principal component analysis (PCA) is then used for dimensionality reduction to generate a feature representation of the construction status. The preset duration of the data acquisition and processing module is dynamically adjusted according to the construction cycle to ensure the time window covers all operational events. The window size for the moving average filter is configured based on the sensor noise characteristics. The variance and mean are calculated using a moving window statistical method for real-time updates. Standardization uses the Z-score method to convert the data into a distribution with a mean of zero and a variance of one. Principal component analysis is achieved through eigenvalue decomposition, selecting the principal component with the highest contribution rate as the feature representation. The dimensionality of the feature representation is adaptively determined based on the complexity of the construction status. In practical implementation, the multi-source sensor data of the data acquisition and processing module includes vibration data measured by accelerometers, hydraulic data measured by pressure sensors, and position data measured by displacement sensors. The moving average filtering algorithm of the data acquisition and processing module adopts a weighted average method to emphasize the importance of recent data. The preliminary statistical calculation of the data acquisition and processing module includes peak-to-peak value and standard deviation as supplementary indicators to enhance the richness of feature representation. The standardization processing of the data acquisition and processing module also includes an outlier removal step, which is based on the three-standard-deviation principle to eliminate the influence of outliers. The principal component analysis process of the data acquisition and processing module uses the covariance matrix to calculate the eigenvector. The selection of eigenvectors is based on the cumulative contribution rate threshold, for example, the threshold is set to 95%. The feature representation output of the data acquisition and processing module is a low-dimensional vector, which is convenient for subsequent processing by the pattern matching module.

[0035] The event listening module and the data acquisition and processing module work together seamlessly by sharing a reference time point. The reference time point output by the event listening module serves as the trigger signal for the data acquisition and processing module. Upon receiving the reference time point, the data acquisition and processing module immediately initiates the data capture process. Communication between the event listening module and the data acquisition and processing module uses a high-speed bus protocol to ensure low-latency data transmission. The buffer of the event listening module is directly mapped to the storage area of ​​the data acquisition and processing module to avoid data copying overhead. The tolerance value setting of the event listening module is coordinated and optimized with the preset duration of the data acquisition and processing module. For example, the tolerance value is less than one-tenth of the preset duration to maintain the integrity of the time window. In some embodiments, the event monitoring module may also integrate a watchdog timer to monitor the operating status of the event monitoring module and prevent deadlock. The data acquisition and processing module may also include a data compression unit that uses a lossless compression algorithm to reduce storage space usage. The software implementation of the event monitoring module and the data acquisition and processing module is based on a real-time operating system, which provides task scheduling and interrupt handling functions. The edge detection hardware of the event monitoring module is implemented using an FPGA to ensure high response speed. The filtering and statistical calculations of the data acquisition and processing module are accelerated using a DSP chip to improve processing efficiency. Optionally, the event monitoring module can support the detection of multiple event types, such as temperature sensor events or humidity sensor events, to expand the monitoring range. The data acquisition and processing module can adapt to different sensor interface protocols, such as I2C or SPI interfaces, to enhance compatibility. Optionally, the calibration process of the event monitoring module and the data acquisition and processing module is performed periodically. The calibration process uses a standard signal source to verify the accuracy of the timestamps and data. The reference time point output format of the event monitoring module adopts the international standard time code, which facilitates cross-system synchronization. The feature representation of the data acquisition and processing module is stored in non-volatile memory for use in historical data analysis.

[0036] The specific implementation details of the event monitoring module include: the acquisition method of the control pulse signal, the control pulse signal coming from the drive circuit of the trolley actuator, the use of an opto-isolator to isolate the control pulse signal to prevent electrical interference, the use of a Schmitt trigger to shape the signal in the edge detection circuit to improve edge recognition accuracy, the use of a high-precision crystal oscillator as the clock source for the timestamp counter, with a clock frequency reaching the megahertz level, the direct transfer of displacement sensor data sequences to memory via DMA to reduce CPU load, the use of binary search to optimize search efficiency in the time difference comparison algorithm, and the storage of the set tolerance values ​​in the configuration file, which supports online updates to adapt to changes in operating conditions. The specific implementation details of the data acquisition and processing module include a synchronous acquisition mechanism for multi-source sensor data, which uses hardware triggers to simultaneously start sampling from all sensors; the moving average filtering of the data acquisition and processing module is implemented as a circular buffer operation, with the size of the circular buffer matching the filtering window; the variance and mean calculation of the data acquisition and processing module uses a recursive algorithm, which avoids repeated traversal of data; the standardized processing parameters of the data acquisition and processing module are pre-calculated based on a training dataset that covers typical construction scenarios; the principal component analysis model of the data acquisition and processing module is obtained through offline learning, which uses a large amount of historical data to train the feature transformation matrix; and the feature representation output format of the data acquisition and processing module is a fixed-length array, with array elements being floating-point numbers. In practical implementation, the integration test of the event monitoring module and the data acquisition and processing module is conducted by simulating trolley operation events. The simulated trolley operation events generate standard pulse signals and sensor data sequences. The performance indicators of the event monitoring module and the data acquisition and processing module include response time and data throughput. The response time is required to be less than milliseconds, and the data throughput meets the requirements of multi-channel high-speed sampling. The power management of the event monitoring module and the data acquisition and processing module adopts a low-power design. The low-power design enters sleep mode when idle to save energy. The fault diagnosis function of the event monitoring module and the data acquisition and processing module records the operation log, which is used for later maintenance analysis.

[0037] The software implementation of the event monitoring module is an interrupt service routine, triggered when the control pulse signal edge changes. The hardware includes an ADC converter and a digital comparator. The ADC converter digitizes the analog pulse signal, and the digital comparator detects edge events. The event monitoring module's buffer management uses a mutex mechanism to prevent multi-threaded access conflicts. Displacement sensor data alignment uses timestamp interpolation to compensate for sampling jitter. The software implementation of the data acquisition and processing module is a multi-threaded task, processing different sensor channels in parallel. The hardware includes a multiplexer and signal conditioning circuit. The multiplexer switches sensor inputs, and the signal conditioning circuit amplifies and filters the raw signal. The moving average filtering algorithm is optimized for vectorized operations, utilizing SIMD instructions for speed improvement. The standardized processing in the data acquisition and processing module adjusts parameters in real time based on the data distribution within a sliding window. Principal component analysis in the data acquisition and processing module uses incremental learning technology, which adapts to slow changes in construction conditions. It is understandable that the robust design of the event monitoring module and the data acquisition and processing module is intended to withstand harsh industrial environments. This robust design includes dustproof and shockproof measures. The communication protocol of the event monitoring module and the data acquisition and processing module adopts the industrial Ethernet standard, which ensures highly reliable data transmission. Version control of the event monitoring module and the data acquisition and processing module supports remote upgrades, which are performed through a secure encrypted channel. In some embodiments, the event monitoring module can also integrate external synchronization signals from GPS or a network time protocol to improve global timestamp consistency. The data acquisition and processing module can also include a data quality assessment unit that calculates the signal-to-noise ratio and integrity indicators as auxiliary information for feature representation.

[0038] Example 2: See Figure 3In its implementation, the pattern matching module feeds the feature representations into the pattern matching unit, evaluates the pattern consistency between the standard construction pattern and the abnormal construction pattern, and generates a quality difference metric. The pattern matching unit adopts a dual-branch structure, with each branch containing convolutional layers and fully connected layers to process the feature representations of the standard and abnormal construction patterns, respectively. The convolutional layers are used to extract local feature patterns, while the fully connected layers perform global feature integration. The output vectors of the two branches are used to calculate the similarity using Euclidean distance, and the Euclidean distance is converted into a similarity score using the sigmoid function. The similarity score is used to adjust the weights of the feature representations, making the similarity score under the standard pattern more consistent. The pattern matching unit (BMU) maximizes the similarity score under anomalous patterns, thereby outputting a quality difference metric. Its dual-branch structure optimizes the parameters of convolutional and fully connected layers using labeled data during training. The filter size and stride of the convolutional layers are set according to the dimension of the feature representation. The fully connected layers use ReLU activation to enhance non-linear expressiveness. Euclidean distance calculation considers feature vector normalization. The sigmoid function output ranges from 0 to 1; a similarity score closer to 1 indicates greater consistency with the standard pattern. The weight adjustment mechanism is implemented through backpropagation, enabling the model to adaptively learn feature importance during training. In some embodiments, the BMU can also integrate an attention mechanism, assigning different weights to different parts of the feature representation to focus on key regions. The training data for the BMU includes normal and anomalous cases from historical construction records. The training process uses a cross-entropy loss function to optimize model parameters. The inference phase of the BMU uses batch processing to improve computational efficiency. The output quality difference metric of the BMU is stored in a log file for subsequent analysis and visualization.

[0039] The causal analysis module uses the trolley operation frequency as a moderating variable. It processes quality variation metrics through a causal analysis framework, estimating the impact weight of each sensor data stream on construction quality. Operation frequency is obtained by statistically analyzing the number of trolley operation events per unit time, and is divided into low, medium, and high intervals. The causal analysis framework establishes a structural equation model, using quality variation metrics as the dependent variable, construction status as the treatment variable, and operation frequency as the moderating variable. Least squares estimation is used to calculate the marginal effect of each sensor data stream on changes in the treatment variable, and the average marginal effect of all samples is aggregated to obtain the impact weight of each sensor data stream. The interval division of operation frequency is determined based on quantiles of historical data; for example, low frequency corresponds to below the lower quartile, medium frequency to between quartiles, and high frequency to above the upper quartile. The structural equation model includes latent and observed variables. Least squares estimation uses gradient descent to solve for model parameters. The calculation of marginal effects involves partial derivative analysis, and the aggregation of impact weights uses an arithmetic mean to maintain robustness. In specific implementation, the structural equation model of the causal analysis module is defined as follows: Let... Indicates a measure of quality variation. Represents the construction state vector. Representing the operating frequency, the model takes the following form:

[0040]

[0041] in: It is the coefficient vector of the construction state. It is a coefficient representing the operating frequency. It is the coefficient of the interaction term. It is the error term, and the marginal effect is... After estimating the parameters using least squares, the average marginal effect is calculated as the influence weight for each sensor data stream. The causal analysis module includes a data preprocessing step, which standardizes quality variation measures and operation frequencies to eliminate dimensional influences. The structural equation model is fitted using iterative reweighted least squares, which addresses heteroscedasticity. The influence weights are output as a vector, with each element corresponding to the weight value of each sensor data stream. It can be understood that the causal inference capability of the causal analysis module helps identify key drivers of construction quality.

[0042] The integration of the pattern matching module and the causal analysis module is achieved through a data flow pipeline. The quality difference metric output by the pattern matching module serves as the input to the causal analysis module. The operation frequency data of the causal analysis module is acquired in real time from the event listening module. Data transmission between the pattern matching module and the causal analysis module employs a message queue mechanism, which ensures data order and integrity. The model update cycle of the pattern matching module is synchronized with the estimation frequency of the causal analysis module to avoid timing deviations. In some embodiments, the pattern matching module can also support online learning functionality, which dynamically adjusts model parameters based on new data to adapt to changes in construction conditions. The causal analysis module may also include a sensitivity analysis unit, which evaluates the robustness of model parameters to input changes. The allocation of computational resources between the pattern matching module and the causal analysis module is based on priority scheduling, which ensures that real-time critical tasks are executed first. The dual-branch structure of the pattern matching module can be accelerated using a heterogeneous computing architecture, which utilizes GPUs to process convolution operations. The least squares estimation of the causal analysis module is implemented as a distributed algorithm, which processes large-scale datasets. Optionally, the similarity score calculation in the pattern matching module can be replaced with cosine similarity, which reduces the impact of vector length. The frequency interval division in the causal analysis module can employ dynamic clustering, which adaptively adjusts interval boundaries. Optionally, the debugging interfaces for both the pattern matching and causal analysis modules provide parameter adjustment functions, allowing engineers to optimize model performance. The quality difference metric from the pattern matching module can be fed back to the data acquisition and processing module for feature representation optimization.

[0043] The implementation details of the pattern matching module include the design of convolutional layers. The number of filters in the convolutional layers is determined based on the number of channels in the input feature map. The same padding method is chosen for the convolutional layers to maintain spatial dimensions. The number of neurons in the fully connected layers is set based on the feature dimension reduction ratio. The dropout rate of the fully connected layers is used to prevent overfitting. Vectorization is used to improve the speed of Euclidean distance calculation. The sigmoid function is implemented using a lookup table method to optimize computational efficiency. Gradient pruning is used to stabilize training during weight adjustment. The inference time of the pattern matching unit must meet the time limit of the real-time control loop. The implementation details of the causal analysis module include the statistical method for operation frequency, which uses sliding window counting. The sliding window size is matched to the construction cycle. Variable selection for the structural equation model is based on feature importance ranking, which uses a recursive feature elimination algorithm. The convergence criterion for least squares estimation sets a residual threshold, which is determined based on the historical error distribution. Marginal effects are calculated independently for each sensor data stream. The normalization of influence weights ensures that the sum of the weights is one. The output of the causal analysis module is stored in a database for trend analysis and report generation. In practical implementation, cross-validation is used to validate the pattern matching and causal analysis modules. Cross-validation divides the data into training and test sets to evaluate the model's generalization ability. Performance metrics for both modules include accuracy and recall, reflecting the effectiveness of pattern classification and causal identification. The software implementation of these modules is based on a machine learning framework, which provides high-level APIs to simplify development. Hardware requirements for both modules include sufficient memory and processor speed to handle high-dimensional data streams. It can be understood that the collaborative work of the pattern matching and causal analysis modules enhances the intelligence level of construction quality monitoring.

[0044] The training process of the pattern matching module involves hyperparameter tuning, which uses either grid search or random search methods. Grid search traverses preset parameter combinations, while random search samples the parameter space. The loss function for the pattern matching module is either mean squared error or log loss, depending on the problem type. The optimizer for the pattern matching module uses the Adam algorithm, which adaptively adjusts the learning rate. The batch size for the pattern matching module is determined based on memory capacity, and the batch size affects training stability and speed. The model validation for the causal analysis module includes residual analysis and goodness-of-fit testing. Residual analysis checks for the randomness of the error term, and the goodness-of-fit test uses the R-squared index, which measures the proportion of variance explained by the model. The causal identification hypothesis testing for the causal analysis module uses instrumental variable methods, which handle endogeneity issues. The interpretation of the results in the causal analysis module considers confidence intervals, which provide the range of uncertainty in the estimate. In some embodiments, the pattern matching module can be extended to a multi-task learning architecture, which simultaneously predicts multiple quality indicators and shares feature representations. The causal analysis module can incorporate time series analysis techniques to capture dynamic causal relationships. The user interfaces of the pattern matching module and the causal analysis module display real-time quality scores and influence weight charts for easy operator monitoring.

[0045] Example 3: In specific implementation, the dynamic index of construction resistance is obtained by extracting low-frequency components related to construction resistance from the feature representation, performing spectral analysis on the low-frequency components, calculating the distribution ratio of spectral energy within a predetermined frequency band, and combining the displacement change rate in the trolley motion trajectory index with a weighted summation of the distribution ratio to generate the dynamic index of construction resistance. The trolley motion trajectory index is obtained by separating the trolley displacement trajectory data from the quality difference measurement, calculating the first and second derivatives of the displacement trajectory as motion features, performing time warping on the motion features, aligning trajectory segments from different construction stages, and then calculating the eigenvalues ​​of the covariance matrix of the trajectory segments to derive the trolley motion trajectory index. The dynamic index of construction resistance and the trolley motion trajectory index are used for the influence weight selection condition optimization of the feature optimization module. In practice, the low-frequency components of the dynamic index of construction resistance are extracted using a digital low-pass filter. The cutoff frequency of the digital low-pass filter is set below 10 Hz to capture the slowly changing components of construction resistance. The spectrum analysis uses a fast Fourier transform algorithm to convert the low-frequency components from the time domain to the frequency domain. The predetermined frequency band is determined based on historical resistance data, for example, the frequency band range is 0.1 Hz to 5 Hz. The distribution ratio is calculated as the ratio of energy in each frequency band to the total energy. The displacement change rate is obtained from the trolley motion trajectory index and used as the input variable for weighted summation.

[0046] The specific calculation process of the dynamic index of construction resistance involves a weighted summation formula, which is expressed as follows:

[0047]

[0048] in: It is a dynamic indicator of construction resistance. It is the energy distribution ratio of the i-th predetermined frequency band. These are the weighting coefficients for the corresponding frequency band. It is the rate of change of displacement. It is the scaling factor and weighting factor of the rate of change of displacement. The scaling factor is learned from the training data through regression analysis. Adjustments are made based on the correlation between construction resistance and motion trajectory in the formula. This indicates the number of predetermined frequency bands, which is typically set to 3 to 5 to balance computational complexity and index sensitivity.

[0049] In practice, the low-frequency component extraction steps include applying a Butterworth low-pass filter to the feature representation. The Butterworth low-pass filter is set to order 4 to provide a steep roll-off characteristic. Spectral analysis uses a Hanning window to reduce spectral leakage. The energy distribution ratio is calculated through numerical integration, which uses the trapezoidal rule to approximate the energy within the frequency band, and the rate of change of displacement is also considered. The real-time calculation is the difference value of the displacement sensor readings, with the difference interval set according to the sampling rate, for example, an interval of 0.1 seconds. In some embodiments, the calculation of the dynamic index of construction resistance can incorporate adaptive filtering technology. This adaptive filtering technology dynamically adjusts filter parameters to cope with changes in the construction environment. The boundaries of the predetermined frequency band can be dynamically updated based on real-time spectral characteristics, for example, by using clustering algorithms to automatically identify important frequency bands, and the coefficients of the weighted summation formula... and Periodic recalibration is possible, using the latest construction data to maintain the reliability of the indicators. Optionally, the output of the dynamic indicators of construction resistance can be smoothed using a moving average filter to eliminate short-term fluctuations, with the moving average window size matched to the construction cycle.

[0050] The acquisition of trolley motion trajectory indicators involves separating trolley displacement trajectory data from the quality difference metric. This data is obtained by parsing displacement-related fields in the quality difference metric. The first derivative of the displacement trajectory is calculated as instantaneous velocity, and the second derivative as instantaneous acceleration. Motion features include velocity and acceleration sequences. Time warping is performed using a dynamic time warping algorithm, which aligns trajectory segments from different construction stages to a common time axis. The aligned trajectory segments are used to calculate the covariance matrix. The eigenvalues ​​of the covariance matrix are derived through eigenvalue decomposition, with the largest eigenvalue representing the main direction of trajectory variation, serving as the trolley motion trajectory indicator. In practice, displacement trajectory data separation is based on sensor identifiers, which are embedded in the metadata of the quality difference metric. The first and second derivatives are calculated using the central difference method, which reduces numerical errors. The dynamic time warping algorithm optimizes the path cost function, which uses Euclidean distance to measure trajectory similarity. The covariance matrix is ​​calculated based on the aligned trajectory point set, and eigenvalue decomposition is solved iteratively using the QR algorithm.

[0051] The calculation example of the trolley motion trajectory index includes extracting a displacement array from the mass difference metric, the displacement array containing timestamp and displacement value pairs, obtaining the first derivative velocity array through the differential displacement array, obtaining the second derivative acceleration array through the differential velocity array, normalizing the motion feature array to ensure consistent scale across different segments, finding the optimal curvature path using a dynamic time warping algorithm, minimizing the cumulative distance of the optimal curvature path, reassembling the aligned trajectory segments into a matrix form, calculating the covariance matrix as the average of the outer products of the trajectory points, extracting eigenvalues ​​and eigenvectors through eigenvalue decomposition, and selecting the largest eigenvalue as the output for the trolley motion trajectory index.

[0052] The integration of dynamic construction resistance indicators and trolley trajectory indicators is achieved through a data fusion module. This module combines the two indicators into a joint feature vector, which is then input into a feature optimization module. The update frequency of the low-frequency components of the dynamic construction resistance indicators is synchronized with the calculation cycle of the trolley trajectory indicators, for example, once per second. The data fusion module uses a weighted concatenation method, with the weights adjusted based on the signal-to-noise ratio of the indicators. In practical implementation, the real-time calculation of both dynamic construction resistance indicators and trolley trajectory indicators requires low-latency processing. This low-latency processing is achieved through a pipelined architecture, which decomposes the indicator calculation task into parallel stages. The spectral analysis stage of the dynamic construction resistance indicators overlaps with the time warping stage of the trolley trajectory indicators to reduce the overall response time. Specific implementation details include filter design; the digital low-pass filter uses an infinite impulse response (IRR) structure, which saves computational resources; the Fast Fourier Transform (FFT) points for spectral analysis are set to 1024 to balance resolution and speed; the calculated energy distribution ratio is cached in a circular buffer, the size of which accommodates multiple update cycles; the weighting coefficient β for the displacement change rate is initialized with an empirical value obtained from historical data statistics, and then adjusted online. Specific implementation details for the trolley motion trajectory index include derivative calculation optimization; the derivative calculation uses the five-point Stogen formula to improve accuracy and reduce boundary effects; the dynamic time warping algorithm is accelerated using a lower bound technique to prune invalid paths in advance; and the eigenvalue calculation of the covariance matrix uses an iterative method, which avoids full matrix storage.

[0053] Application examples of dynamic construction resistance indicators include tunnel lining construction. When the trolley encounters changes in rock strata, the dynamic construction resistance indicator reflects the increase in resistance through the increase in energy of low-frequency components. Combined with the decrease in the rate of displacement change, a weighted summation formula outputs a high resistance value, triggering the system to adjust the trolley thrust. Application examples of trolley motion trajectory indicators include when the trolley's movement is unstable, the eigenvalues ​​of the covariance matrix of the trajectory segment show high variability, indicating the need for trajectory correction. In specific implementations, the output format of both dynamic construction resistance indicators and trolley motion trajectory indicators is standardized to floating-point arrays. These arrays are transmitted to subsequent modules via a communication protocol, which uses TCP / IP to ensure reliability. It can be understood that the implementation of dynamic construction resistance indicators and trolley motion trajectory indicators enhances the system's ability to perceive construction dynamics. Improvements to the dynamic construction resistance indicators include introducing multi-resolution analysis, which uses wavelet transform instead of Fourier transform to capture transient resistance events. Improvements to the trolley motion trajectory indicators include using deep learning models, which automatically learn trajectory features, reducing manual feature engineering.

[0054] See Figure 4This diagram is the core output carrier of the trolley motion trajectory index. The trolley motion trajectory index separates displacement trajectory data from mass difference measurement, calculates the first and second derivatives as motion features, aligns trajectory segments from different construction stages using a time warping algorithm, and derives the index through covariance matrix eigenvalue decomposition. The eigenvalues ​​of each trajectory segment in the diagram correspond to the main direction of trajectory variation; the eigenvalues ​​of segments 1 to 6 decrease, reflecting the gradual convergence of the degree of trajectory variation. This index is one of the weighting factors of the dynamic construction resistance index, participates in the adjustment of the influence weight threshold in the feature optimization module, supports the screening of high-influence weight features, and is a key basis for dynamically sensing the stability of trolley motion. Its changing trend is directly related to trajectory stability judgment, providing trajectory feature support for the phase interval division of the instruction generation module, reflecting the connecting role between system state perception and feature optimization.

[0055] Example 4: In specific implementation, the feature optimization module inputs the dynamic indicators of construction resistance and the trolley trajectory indicators into the pre-trained support vector machine model. The decision function output of the pre-trained support vector machine model is used to adjust the threshold of the influence weights. Based on the adjusted threshold, the influence weights of each sensor data stream are compared, and the feature representations corresponding to the sensor data streams with influence weights higher than the threshold are retained to form an optimized feature set. The pre-trained support vector machine model is trained using historical construction data, which includes the dynamic indicators of construction resistance, the trolley trajectory indicators, and the corresponding construction quality labels. The decision function of the support vector machine model maps the input indicators to a high-dimensional space based on the kernel function to find the optimal separating hyperplane. The output value of the decision function is converted into a probability score through the sigmoid function. The probability score is used as the basis for threshold adjustment. The threshold adjustment mechanism is dynamically updated according to the real-time construction environment. For example, when the dynamic indicators of construction resistance or the trolley trajectory indicators show an abnormal mode, the threshold is increased accordingly to filter more critical features. The process of generating the optimized feature set involves iteratively comparing the influence weights of each sensor data stream with the threshold. The influence weights come from the output of the causal analysis module. The output of the feature optimization module optimizes the feature set for the instruction generation module to drive instruction generation. It is understandable that the feature optimization module uses machine learning methods to optimize feature selection and improve the accuracy of system decisions.

[0056] In practice, the decision function of the pre-trained support vector machine model takes the following form:

[0057]

[0058] in: It is an input vector containing dynamic indicators of construction resistance and indicators of the trolley's motion trajectory. It is a kernel function that maps the input to a high-dimensional feature space. It is a weight vector. It is a bias term.

[0059] The threshold adjustment process of the feature optimization module specifically includes comparing the influence weight of each sensor data stream with the dynamic threshold. The influence weight comes from the output vector of the causal analysis module, and the dynamic threshold is calculated based on the probability score output by the decision function of the support vector machine model. The calculation method is as follows: ,in: This is the adjusted threshold. It is the initial threshold. It is the scaling factor. The decision function outputs probability scores, and the scaling factor is... Configuration based on construction scenario, such as high-risk scenarios. Increasing the value enhances the screening stringency, and optimizing the formation of the feature set involves traversing all sensor data streams and retaining those with influence weights above the threshold. The feature representation corresponding to the stream is extracted from the output of the data acquisition and processing module. The optimized feature set is stored in vector format for use by subsequent modules. See Table 1 for the threshold adjustment data.

[0060] Table 1: Threshold Adjustment Parameter Table

[0061] ;

[0062] In practical implementation, the feature optimization module includes software and hardware components. The software components implement the inference logic of the support vector machine model, and the inference logic uses optimization libraries such as LIBSVM to accelerate computation. The hardware components include application-specific integrated circuits (ASICs) or FPGAs. The FPGAs process the decision function computation in parallel. The real-time requirement for threshold adjustment is that the inference latency of the support vector machine model is less than milliseconds. The comparison operation affecting the weights uses vectorized instructions to improve efficiency. The size of the optimized feature set is dynamically variable, and the dynamic variable size is adjusted according to the number of retained streams. The input and output interfaces of the feature optimization module adopt a standardized protocol, which ensures compatibility with the causal analysis module and the instruction generation module.

[0063] In some embodiments, the feature optimization module can be extended to a multi-model ensemble framework. This framework combines multiple support vector machine (SVM) models to vote on the threshold. Threshold adjustment can introduce a time decay factor to reduce the influence of older data. The formation of the optimized feature set can include feature transformation steps, such as principal component analysis, for further dimensionality reduction. Optionally, the feature optimization module's debug mode allows manual threshold setting for testing and calibration. SVM model updates can be based on incremental learning, which adapts to changes in construction conditions. Optionally, the feature optimization module's logging function records the threshold adjustment history for offline analysis and optimization. A specific example of the feature optimization module is in tunnel lining construction. When the dynamic indicators of construction resistance show increased resistance and the trolley trajectory indicators show instability, the SVM model's decision function outputs a high-probability score, leading to a higher threshold. Only features from sensor flows with high influence weights are retained, forming an optimized feature set. This optimized feature set is used to generate precise driving instructions.

[0064] Example 5: In specific implementation, the instruction generation module partitions and merges the optimized feature set according to the temporal alignment relationship of the construction waveform, and generates driving instructions using an attention allocation mechanism. The partitioning operation of the instruction generation module extracts zero-crossing points and extreme points from the construction waveform signal, divides the waveform into multiple phase intervals, and groups the optimized feature set according to the time boundaries of the phase intervals. The feature representations within each group undergo time synchronization correction. The merging operation of the instruction generation module uses an attention allocation mechanism to generate driving instructions. The attention allocation mechanism calculates a self-attention score for the feature representation of each phase interval, and performs a weighted average of the feature representations based on the self-attention scores to obtain interval representations. All interval representations are input into a long short-term memory network for sequence modeling. The final hidden state of the long short-term memory network is mapped to driving instructions through a fully connected layer. Zero-crossing detection of construction waveform signals is achieved by identifying points where the signal value sign changes. Extreme point detection is achieved by locating local maximum and minimum values. The division of phase intervals ensures that each interval corresponds to a complete construction sub-process. Time synchronization correction uses a linear interpolation algorithm to align feature representations at different sampling times. The self-attention score calculation of the attention allocation mechanism is based on the transformation of query, key, and value matrices. The hidden state of the long short-term memory network captures the temporal dependencies between phase intervals. The fully connected layer converts the final hidden state into a driving instruction vector.

[0065] The specific implementation details of the instruction generation module include a zero-crossing detection algorithm, which uses a difference method to approximate the signal derivative and combines threshold judgment to avoid false detections caused by noise; extreme point detection, which locates peaks and troughs by comparing the values ​​of adjacent sampling points; the time boundaries of the phase interval are stored as a timestamp array; the optimized feature set grouping operation is allocated according to the interval in which the feature timestamps fall; and the linear interpolation algorithm for time synchronization correction uniformly resamples the feature sequence within the interval boundaries to ensure that the number of feature points in each interval is consistent. It can be understood that the temporal alignment accuracy of the instruction generation module directly affects the quality of the interval representation. The long short-term memory network modeling process of the instruction generation module includes inputting the interval representation sequence into the network sequentially; the forget gate, input gate, and output gate of the long short-term memory network controlling the information flow; finally, the hidden state summarizes the context information of the entire sequence; the weight matrix of the fully connected layer maps the hidden state to the driving instruction space; and the numerical range of the driving instructions is limited by the activation function.

[0066] In some embodiments, the instruction generation module can integrate an exception handling mechanism that skips the current cycle when an invalid phase interval is detected. The query, key, and value matrices of the attention allocation mechanism can share some projection weights, reducing the number of model parameters. The training of the Long Short-Term Memory network uses a backpropagation algorithm, which optimizes network parameters through gradient descent. Optionally, the driving instructions of the instruction generation module can be post-processed and smoothed using a low-pass filter to eliminate instruction jumps. The partitioning strategy can be dynamically adjusted according to the construction type, and this dynamic adjustment is achieved through configuration parameters.

[0067] An application example of the instruction generation module is as follows: During tunnel lining pouring, the construction waveform signal comes from a vibration sensor. Zero-crossing points and extreme points divide the pouring cycle into phase intervals such as filling, compaction, and stagnation. The optimized feature set within each interval is time-synchronized and then used to calculate the interval representation through a self-attention mechanism. The interval representation sequence is input into a long short-term memory network to predict the next trolley advance speed and grouting pressure instruction. In specific implementation, the software implementation of the instruction generation module is based on a deep learning framework. The deep learning framework provides high-level interfaces for the attention mechanism and the long short-term memory network. The threshold for zero-crossing detection is adaptively set according to the signal noise level. The neighborhood window size for extreme point detection is proportional to the waveform period. The number of interpolation points for time synchronization correction is fixed at 32 points to standardize the input. The number of layers in the long short-term memory network is set to 2 to balance complexity and performance.

[0068] See Figure 5This diagram serves as the core output of the attention allocation mechanism. The instruction generation module extracts zero-crossing points and extreme points from the construction waveform signal to divide it into phase intervals. For each interval, a self-attention score is calculated based on the optimized feature set. In the diagram, phases 1 to 5 represent the divided construction sub-process intervals. The attention score represents the feature weight percentage of each phase interval, with phase 5 receiving the highest score, corresponding to the core attention interval for instruction generation. These scores are used to weighted average the feature representations to obtain interval representations, which are then input into a long short-term memory network for sequence modeling, ultimately mapping to driving instructions. This data supports the context-adaptive instruction generation design based on waveform temporal alignment throughout the construction process. It reflects the dynamic focus of the attention mechanism on different construction stages and is a key basis for shifting instruction generation from instantaneous response to process-segment-level contextual understanding, ensuring the alignment of trolley actions with process requirements.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control system for an intelligent trolley used for tunnel lining construction, characterized in that, The control system includes: The event monitoring module is used to monitor the occurrence signals of trolley operation events and mark the reference time points of construction status changes in the continuously acquired sensor data stream; The data acquisition and processing module is used to extract multi-source sensor data within a fixed time interval before and after a reference time point, and to transform and process the multi-source sensor data to derive a characteristic representation of the construction status. The pattern matching module is used to feed the feature representation into the pattern matching unit, evaluate the pattern consistency between standard construction patterns and abnormal construction patterns, and generate a quality difference metric. The causal analysis module is used to use the trolley operation frequency as a moderating variable, process quality difference measurement through the causal analysis framework, and estimate the impact weight of each sensor data stream on construction quality. The feature optimization module is used to integrate the dynamic indicators of construction resistance and the trolley movement trajectory indicators, optimize the selection conditions of influence weights, screen out the feature representations corresponding to high influence weights, and form an optimized feature set. The instruction generation module is used to partition and merge the optimized feature set according to the timing alignment relationship of the construction waveform, and generate driving instructions using an attention allocation mechanism.

2. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The signal for monitoring the occurrence of trolley operation events includes: continuously receiving control pulse signals from the trolley actuator, detecting edge changes of the control pulse signals as event trigger markers; recording the precise timestamp of the event trigger marker, and comparing the time difference between the timestamp and the displacement mutation point in the displacement sensor data sequence; when the time difference is less than a set tolerance, confirming the timestamp as a reference time point for changes in construction status.

3. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The process of extracting multi-source sensor data within a fixed time interval before and after a reference time point includes: taking the reference time point as the center, extending forward and backward by a preset time duration to obtain the raw readings of the accelerometer, pressure sensor, and displacement sensor within that time window; performing a moving average filter on the raw readings to suppress noise, and then calculating the data variance and mean of each sensor channel as preliminary statistics; standardizing the preliminary statistics to ensure uniform data scale across channels, and then using principal component analysis for dimensionality reduction to generate a feature representation of the construction status.

4. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The process of feeding the feature representations into the pattern matching unit to evaluate the pattern consistency between standard construction patterns and abnormal construction patterns, and generating a quality difference metric, includes: constructing a dual-branch structure for the pattern matching unit, with each branch containing a convolutional layer and a fully connected layer to process the feature representations of the standard construction patterns and abnormal construction patterns, respectively; calculating the Euclidean distance between the output vectors of the two branches, and converting the Euclidean distance into a similarity score using the sigmoid function; adjusting the weights of the feature representations based on the similarity scores to maximize the similarity score in the standard pattern and minimize the similarity score in the abnormal pattern, thereby outputting a quality difference metric.

5. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The method of using trolley operation frequency as a moderating variable and processing quality difference measurement through a causal analysis framework to estimate the impact weight of each sensor data stream on construction quality includes: counting the number of trolley operation events per unit time and dividing the operation frequency into three intervals: low, medium, and high; establishing a structural equation model of the causal analysis framework, with quality difference measurement as the dependent variable, construction status as the processing variable, and operation frequency as the moderating variable; calculating the marginal effect of each sensor data stream when the processing variable changes through least squares estimation, and aggregating the average marginal effect of all samples to obtain the impact weight of each sensor data stream.

6. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The dynamic index of construction resistance is obtained by extracting low-frequency components related to construction resistance from the feature representation, performing spectral analysis on the low-frequency components, and calculating the distribution ratio of spectral energy within a predetermined frequency band; combining the displacement change rate in the trolley motion trajectory index, the distribution ratio is weighted and summed to generate the dynamic index of construction resistance.

7. The control system for an intelligent trolley used for tunnel lining construction according to claim 6, characterized in that, The trolley motion trajectory index is obtained in the following way: the trolley displacement trajectory data is separated from the quality difference measurement, and the first and second derivatives of the displacement trajectory are calculated as motion features; the motion features are time-normalized, the trajectory segments of different construction stages are aligned, and the eigenvalues ​​of the covariance matrix of the trajectory segments are calculated to derive the trolley motion trajectory index.

8. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The selection criteria for the optimized influence weights include: inputting the dynamic index of construction resistance and the trolley trajectory index into a pre-trained support vector machine model, and using the decision function output of the support vector machine model to adjust the threshold of the influence weights; based on the adjusted threshold, comparing the influence weights of each sensor data stream, retaining the feature representations corresponding to the sensor data streams with influence weights higher than the threshold, and forming an optimized feature set.

9. The control system for an intelligent trolley used for tunnel lining construction according to claim 1, characterized in that, The step of partitioning the optimized feature set according to the time alignment relationship of the construction waveform includes: extracting zero-crossing points and extreme points from the construction waveform signal, dividing the waveform into multiple phase intervals; grouping the optimized feature set according to the time boundaries of the phase intervals, and performing time synchronization correction on the feature representations within each group; The method of generating driving instructions using an attention allocation mechanism includes: calculating a self-attention score for the feature representation of each phase interval; performing a weighted average of the feature representation based on the self-attention score to obtain the interval representation; inputting all interval representations into a long short-term memory network for sequence modeling; and mapping the final hidden state of the network to driving instructions through a fully connected layer.

10. A smart trolley for tunnel lining construction, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the function of the control system of the intelligent trolley for tunnel lining construction as described in any one of claims 1 to 9.

Images