Evaluation Method for Shield Tunnel Cake Filtration Effect Based on Multimodal Sensing

Abstract

This invention belongs to the field of shield tunneling cake filtration technology, specifically a method for evaluating the filtration effect of shield tunneling cake based on multimodal sensing. The method includes step one: integrating a multimodal sensor cluster on a belt filter press to synchronously collect the following data every 0.5 seconds: cake moisture content output by a hyperspectral sensor, filtrate moisture content output by a capacitive moisture sensor, filter roller pressure output by a pressure sensor, initial slurry mass and remaining slurry mass output by a slurry weighing sensor in the weighing component, filtrate mass output by a filtrate weighing sensor, number and average thickness of cake cracks output by a vision sensor, slurry viscosity output by a viscosity sensor, and cake mass output by the weighing conveyor belt. The collected data is then transmitted in real-time to the central control room. This multimodal sensing-based method for evaluating the filtration effect of shield tunneling cake achieves comprehensive and complete perception of the filtration status by setting up a shield tunneling cake filtration effect evaluation system.

Description

Technical Field

[0001] This invention relates to the field of shield tunneling sludge cake filtration technology, and in particular to a method for evaluating the filtration effect of shield tunneling sludge cake based on multimodal sensing. Background Technology

[0002] Shield tunneling mud cake filtration is the core process of shield tunneling mud treatment. It achieves solid-liquid separation of mud through mechanical filtration, forming mud cake with low water content and clean water. It is an important technical means for green construction in modern tunnel engineering.

[0003] The existing method of filter cake moisture content control relies on manual testing, which cannot be monitored in real time. Manual sampling and testing are time-consuming and cannot guide current production. Because there is no real-time data feedback, operators can only set a fixed filter pressing time based on experience, which can easily lead to two outcomes:

[0004] Insufficient filter press: If the set time is too short, the moisture content of the sludge cake will exceed the standard, failing to meet the requirements for off-site transportation or resource utilization, requiring secondary rework, resulting in a double waste of energy and time;

[0005] Over-pressure filtration: If the set time is too long, the moisture content of the filter cake will be far below the standard value. Although the quality is qualified, the high-pressure pump, hydraulic station and other equipment will continue to run for an ineffective time, resulting in a large amount of energy waste. At the same time, it will aggravate the wear of the filter cloth and the fatigue damage of the filter plate, and shorten the service life of the equipment.

[0006] The inability to monitor changes in cake moisture content and filtration efficiency in real time, and the inability to reverse-engineer the filtration process based on monitoring data, and thus to optimize and control the process based on data, makes it impossible to achieve efficient filtration. Summary of the Invention

[0007] To address the technical problem that existing methods for evaluating the moisture content of shield tunneling cake filtration rely on manual detection, which cannot be monitored in real time and thus cannot achieve efficient filtration, this invention proposes a method for evaluating the filtration effect of shield tunneling cake based on multimodal sensing.

[0008] This invention proposes a method for evaluating the filter cake pressing effect of tunnel boring machines (TBMs) based on multimodal sensing, comprising the following steps: Step 1: Slurry injection and synchronous data acquisition throughout the entire cycle: The slurry feed pump is started to inject TBM slurry into the belt filter press. The data acquisition command in the central control room is initiated, and a multimodal sensor cluster integrated on the belt filter press synchronously acquires the following data at a frequency of 0.5 seconds / time: Moisture content of the filter cake output by the hyperspectral sensor. The filtrate moisture content output by the capacitive moisture sensor The pressure of the filter roller output by the pressure sensor The initial mud mass output by the mud weighing sensor in the weighing component. Residual mud quality The filtrate weighing sensor outputs the filtrate mass. The number of cracks in the mud cake output by the visual sensor With average thickness Viscosity sensor outputs mud viscosity And the output weighing conveyor belt output mud cake quality The collected data will be transmitted to the central control room in real time.

[0009] Step Two: Multi-Source Data Preprocessing: The central control room processes the acquired raw data through the signal preprocessing module, including pressure... ,quality , , and The data is denoised using discrete wavelet transform, and all the denoised data is normalized to eliminate dimensional differences.

[0010] Step 3: Dynamic Evaluation and Decision Fusion of Filtration Effect: The central control room calls a multi-parameter fusion algorithm model to achieve two-level fusion evaluation and decision-making based on the preprocessed data. The feature-level fusion model predicts the filtration endpoint and the final cake moisture content; the decision-level fusion model calculates the attention weight of each parameter and outputs the optimal pressing pressure for the next stage; the filtration efficiency score is calculated, with a score range of 0-100 points, and parameter maintenance, early warning, or closed-loop optimization is performed based on the score.

[0011] Step 4: Closed-loop optimization control: Based on the decision results and abnormal parameter location, perform underpressure optimization, overpressure early warning, moisture content / crack optimization, and chemical synergistic optimization operations;

[0012] Step 5: Filtration termination and model iteration: When the moisture content of the sludge cake stabilizes within the preset range for a preset duration and the remaining sludge mass approaches 0, a filtration termination command is sent and filter belt cleaning is initiated. All cycle data, optimization commands, and scores are recorded and stored in the database. The model parameters are periodically adjusted based on historical data to complete iterative optimization.

[0013] Preferably, the denoising process in step two employs Discrete Wavelet Transform (DWT), selecting a Sym4 wavelet basis and a decomposition level of 3 layers for pressure... ,quality , , and The high-frequency detail coefficients of the data are inversely reconstructed after soft or hard thresholding to filter out electromagnetic and mechanical vibration noise in the 10-100Hz range. Normalization is performed using Z-score normalization, calculated as follows:

[0014] ,in, For the normalized first Data points, The raw data collected by the sensor Data points, for pressure ,quality , , and Moisture content , viscosity Any type of raw monitoring data is collected in real time by the corresponding sensor. The pressure sensor collects the pressure of the filter roller, and the hyperspectral sensor collects the moisture content of the filter cake. The collection frequency is every 0.5 seconds, and the data is directly transmitted to the central control room for use as raw data. This is the average value of the sensor data over a preset 5-minute time window. The standard deviation of the sensor data within the preset time window is given.

[0015] Preferably, the feature-level fusion model in step three is an LSTM-autoencoder model, whose input features include the rate of change of water content. Pressure curve of filter roller Filtrate mass loss rate Output mud cake quality change rate The model training uses a loss function, calculated as follows:

[0016] ,in, The total loss of the model, , The weighting coefficients and > ;

[0017] The data reconstruction loss is calculated using the following formula: .in, The time step of the input feature sequence. This corresponds to 60 seconds of data collection. For a moment The original input feature vector contains the rate of change of water content at that moment. Filter roller pressure curve value Filtrate mass loss rate Output mud cake quality change rate Four core features The moment for model reconstruction The feature vectors of the model are encoded by the encoder. After dimensionality reduction and feature extraction, the vector is reconstructed by the decoder and used to compare it with the original vector. Compare and calculate the reconstruction loss;

[0018] The formula for calculating the predicted loss based on the final moisture content is as follows: ,in, The final moisture content of the filter cake, measured at the end of the filtration cycle, is determined by a hyperspectral sensor at the point of filtration termination when the moisture content meets the specified conditions. Stabilize at 18%-22% and maintain for 30 seconds. When ≈0, it can be directly measured. The final moisture content of the mud cake predicted by the model is obtained by the output layer after the model extracts deep features through the encoder based on the input temporal feature vector.

[0019] The decision-level fusion model in step three employs the Transformer multi-head attention mechanism, and the weight calculation formula is as follows: ,in, To optimize the target vector, This is a real-time state parameter vector, containing the current water content. Current pressure and pressing time , This is the basic adjustment vector corresponding to each state parameter. for The dimension of a vector , For query vector With key vector The product of the transposes of matrices is obtained through matrix multiplication.

[0020] The optimal pressing pressure in step three The calculation formula is:

[0021] ,in, The number of state parameters. The first one calculated through the attention mechanism Attention weights for each state parameter. In order to be with the first The basic adjustment amount of the pressing pressure corresponding to each state parameter This is the actual pressure value of the filter roller at the current moment, reflecting the current pressing intensity. It is collected in real time by the pressure sensor and obtained after noise reduction processing.

[0022] The formula for calculating the filter press efficiency score in step three is as follows:

[0023] ,in, The filter press efficiency is scored, with a value ranging from 0 to 100. , , , The weighting coefficients and , maximum, , , , , The preset target moisture content of the mud cake, This is the total time consumed in this filter press operation. The preset maximum allowable pressure filtration time, The total energy consumption for this filter press operation is calculated based on the filter press roller pressure and operating time. ,in, The average filter pressure, The power factor of the equipment;

[0024] The preset maximum allowable pressure filtration energy consumption, This refers to the total mass of the filter cake output from this filter press.

[0025] ≥90 points: Excellent pressure filtration effect, maintained execution parameters, keep current. The feed flow rate and reagent addition rate remain constant.

[0026] 70 points ≤ <90 points: The pressure filtration effect is good. Fine-tuning and optimization were performed, adjusting only the feed flow rate ±0.5m³ / h without changing the pressure;

[0027] 50 points ≤ <70 points: The pressure filtration effect is average. Perform moderate optimization, adjust the pressure by ±0.05MPa, and adjust in conjunction with the feed flow rate;

[0028] <50 points: Poor pressure filtration effect. Perform in-depth optimization, pause feeding, clean the filter belt, reset the pressure and reagent addition rate, and restart the pressure filtration.

[0029] Preferably, in step four, the calculation formula for adjusting the feed pump flow rate using a PID controller in the moisture content / crack optimization step is as follows:

[0030] ,in, For a moment The feed pump flow rate setpoint, For proportional gain, , For integral gain, , For differential gain, , From the start-up time of the filter press (time 0) to the current time The error integral is used to accumulate the error over a period of time and eliminate static deviation;

[0031] The error signal is calculated using the following formula:

[0032] ,in, The preset target rate of moisture content decrease, For based on Real-time calculation of the actual rate of moisture content decrease;

[0033] In step four, during the synergistic optimization of the pharmaceutical agents, the formula for calculating the dispersant addition rate is as follows:

[0034] ,in, For a moment The rate of addition of anionic dispersants. This is the proportional control coefficient. For a moment The measured mud viscosity, The preset safety threshold for mud viscosity, Ensure the addition rate is 0 when the viscosity is below the threshold.

[0035] In step four, the undervoltage optimization specifically involves... If the pressure remains below 0.8 MPa, extend the high-pressure pressing stage by 20% and increase the pressure gradient from 0.1 MPa / min to 0.15 MPa / min until... Stable at 1.0-1.2 MPa and ≤25%, ≤1 item;

[0036] The overvoltage warning is specifically as follows: If If the pressure is greater than 1.5MPa, reduce the pressure of the filter roller to 1.2MPa and reduce the filter belt tension by 10%. If the camera detects filter cloth damage, immediately stop the equipment and trigger a filter cloth replacement reminder.

[0037] In the aforementioned moisture content / crack optimization, if ≥3 cracks, reduce the pressure of the filter roller in the crack concentration area by 0.1MPa and increase the spray water volume of the filter belt by 10%;

[0038] Drug synergistic optimization: If When the pressure reaches >500 mPa·s, start the anionic dispersant addition device.

[0039] Preferably, the preset range in step five is 18%-22%, and the preset duration is 30 seconds; the periodic adjustment of model parameters specifically refers to monthly adjustments, and the adjustment targets include the LSTM-autoencoder model. , Coefficients and attention weights calculation parameters for the Transformer model;

[0040] The "recording of the entire filter press cycle data" in step five specifically includes: data within the entire cycle. , , , , , , , , , Time series data, and total output mass of mud cake Mud cake recovery rate ;

[0041] When storing in the database, and The associated storage is used to optimize the correlation prediction accuracy of cake quality-moisture content during subsequent model iterations.

[0042] Preferably, the weighing component in step one includes an input device and an output device. The input device includes a connecting frame, which is fixedly installed on the outer surface of the belt filter press frame. A sliding plate is fixedly installed on the outer surface of the connecting frame.

[0043] Preferably, a conveying assembly is rotatably connected to the outer surface of the connecting frame, a movable frame is slidably inserted into the outer surface of the connecting frame, the outer surface of the movable frame is fixedly installed with the outer surface of the synchronous belt of the conveying assembly, a lifting screw with gears is rotatably connected to the outer surface of the movable frame, a lifting motor with gears is fixedly installed on the outer surface of the movable frame, the gear of the lifting motor meshes with the gear of the lifting screw, a lifting sleeve is threadedly connected to the outer surface of the lifting screw, and the lifting sleeve is slidably inserted into the outer surface of the movable frame.

[0044] Preferably, a scraper with gears is rotatably connected to the outer surface of the lifting sleeve, the outer surface of the scraper is in contact with the outer surface of the slide plate, an electromagnet is fixedly installed on the outer surface of the lifting sleeve, the outer surface of the electromagnet is magnetically connected to the outer surface of the gears of the scraper, and a roller is rotatably connected to the outer surface of the scraper.

[0045] Preferably, the inner wall of the connecting frame is provided with a partition, the outer surface of the partition and the inner wall of the connecting frame form a U-shaped groove, a fixed cylinder assembly is fixedly installed on the inner wall of the partition, one end of the piston rod of the fixed cylinder assembly is slidably inserted into the inner wall of the connecting frame, a deflection rack is fixedly installed on the outer surface of both the connecting frame and the partition, the deflection rack meshes with the gear of the scraper, and the roller is slidably connected to the inner wall of the U-shaped groove.

[0046] Preferably, the output device further includes a hyperspectral sensor, which is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt via a bracket. The output weighing conveyor belt is located on one side of the belt filter press. The vision sensor is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt via a bracket. A support frame is fixedly mounted on the outer surface of the connecting frame. The outer surface of the support frame is fixedly mounted to the outer surface of the mud weighing sensor. A hopper with a mud feed pipe is fixedly mounted on the outer surface of the mud weighing sensor. The mud weighing sensor detects the initial mud mass and the remaining mud mass. The viscosity sensor is fixedly mounted on the wall of the feed pipe of the hopper. The capacitive moisture sensor is fixedly mounted on the inner wall of the collection tank of the belt filter press. The outer surface of the frame of the belt filter press is fixedly mounted to the outer surface of the filtrate weighing sensor. The outer surface of the filtrate weighing sensor is fixedly mounted to the lower surface of the collection tank of the belt filter press. The pressure sensor is fixedly mounted on the pressure detection interface of the filter roller bearing seat of the belt filter press.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. By setting up a shield tunneling cake pressing and filtration effect evaluation system, multi-modal high-frequency synchronous acquisition is achieved to realize full-dimensional and complete perception of the pressing and filtration status. Based on time-series characteristics, the pressing endpoint and final moisture content are predicted in advance. Combined with dual loss function optimization, the final moisture content is stabilized within the preset range of 18%-22%, avoiding the over-pressing and rupture of filter cake or under-pressing and excessive moisture content caused by the traditional experience-based judgment of the endpoint, thus improving the stability of product quality. Through the attention mechanism, the weight of each parameter is dynamically allocated to output the optimal pressing pressure, rather than a fixed pressure parameter. This can address local anomalies in a targeted closed-loop optimization, realize dynamic error correction in the pressing and filtration process, and reduce the defect rate. By storing full-cycle data and associating it with cake quality and moisture content, the model parameters are optimized monthly based on historical data, enabling the system to automatically adapt to the differentiated characteristics of mud in different shield tunneling projects (such as sandy mud and viscous mud) without the need for frequent manual calibration. Compared to the limitations of traditional fixed parameters adapted to a single type of slurry, the system's evaluation accuracy and optimization efficiency will continue to improve after long-term use, reducing reliance on operator experience, expanding applicable scenarios, and solving the technical problem that existing slurry cake filter presses rely on manual detection of moisture content, cannot monitor in real time, and cannot achieve efficient filter presses.

[0049] 2. By setting up weighing components, the initial mud quality, the remaining mud quality, and the quality of the mud cake can be detected. By alternating back and forth movement of the scraper, the scraper can scrape the mud remaining on the slide plate, preventing mud residue from affecting the mud quality data. Furthermore, the slide plate can transport the mud evenly to the belt filter press, improving the uniformity and effect of the filter press. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention;

[0051] Figure 2 This is a three-dimensional view of the hopper structure of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0052] Figure 3 This is a three-dimensional view of the slide plate structure of the shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0053] Figure 4 This is a three-dimensional view of the lifting screw structure of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0054] Figure 5 This is a three-dimensional view of the lifting sleeve structure of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0055] Figure 6 This is a three-dimensional view of the output weighing conveyor belt structure of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0056] Figure 7 This is a three-dimensional view of the filtrate weighing sensor structure of a shield tunneling sludge cake filtration effect evaluation method based on multimodal sensing proposed in this invention.

[0057] In the diagram: 1. Belt filter press; 2. Connecting frame; 21. Slide plate; 22. Conveying assembly; 23. Moving frame; 3. Lifting screw; 31. Lifting motor; 32. Lifting sleeve; 33. Scraper; 34. Adsorption electromagnet; 35. Roller; 4. Baffle; 41. Fixed cylinder assembly; 42. Deflecting rack; 5. Output weighing conveyor belt; 51. Hyperspectral sensor; 52. Vision sensor; 6. Support frame; 61. Slurry weighing sensor; 62. Hopper; 63. Viscosity sensor; 64. Filtrate weighing sensor; 65. Capacitive moisture sensor; 66. Pressure sensor. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0059] Example 1

[0060] Reference Figure 1 A method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing, involving simultaneous data acquisition during slurry injection and throughout the entire cycle: The slurry feed pump is started to inject shield tunneling slurry into the belt filter press 1. A data acquisition command is initiated from the central control room, and a cluster of multimodal sensors integrated on the belt filter press 1 synchronously acquires the following data at a frequency of 0.5 seconds / time: Moisture content of the filter cake output by the hyperspectral sensor 51. The filtrate moisture content output by the capacitive moisture sensor 65 Pressure sensor 66 outputs the pressure of the filter roller. The initial mud mass output by the mud weighing sensor 61 in the weighing component. Residual mud quality Filtrate weighing sensor 64 outputs filtrate mass The number of cracks in the mud cake output by the vision sensor 52 With average thickness Viscosity sensor 63 outputs mud viscosity. And the output weighing conveyor belt 5 output mud cake quality The collected data will be transmitted to the central control room in real time.

[0061] The belt filter press model 1 is a 3m wide double-roller driven type, suitable for shield tunneling projects with a single mud processing capacity of 5-10m³ / h. The filter roller diameter is 500mm, the roller surface material is wear-resistant rubber, and the filter belt is a polyester fiber composite filter cloth (pore size 5-10μm, air permeability 200L / (m²・min)), ensuring compatibility with multimodal sensors.

[0062] Central control room hardware configuration: It adopts an industrial control computer (IPC) and data acquisition card (DAQ) architecture. The IPC model is Advantech IPC-610L, and the DAQ card model is NicDAQ-9178. Data transmission adopts industrial Ethernet with a transmission delay of ≤100ms, ensuring real-time reception and storage of high-frequency data at 0.5 seconds / time.

[0063] Data storage module: Configured with a time-series database (InfluxDB) and a backup hard disk array. The time-series database is used to store real-time data throughout the entire lifecycle, sensor IDs, parameter values, and data quality indicators (0=normal, 1=abnormal). The hard disk array adopts a RAID5 redundancy architecture to ensure data storage security and traceability.

[0064] The central control room achieves multi-sensor synchronization via hardware trigger signals (the DAQ card sends a synchronization pulse every 0.5 seconds, and all sensors simultaneously collect data after receiving the pulse), avoiding time deviations caused by independent sensor timing (synchronization error ≤ ±10ms). After the collected data is transmitted to the IPC in real time, it must first undergo data validity verification (to determine whether the parameters are within a reasonable range: such as...). 0-100% 0-25MPa Invalid data (100-5000 mPa·s) is marked as abnormal and triggers a local audible and visual alarm (the alarm light is red and the buzzer frequency is 1kHz). At the same time, the backup sensor is automatically called (if the main pressure sensor is abnormal, the backup pressure sensor is switched to) to ensure continuous data acquisition.

[0065] Data preprocessing preparations: Before data is transmitted to the preprocessing module, quality parameters ( , , , Dynamic compensation is performed (because the inertial force generated by the conveyor belt movement will cause weighing errors, the compensation formula is...). ,in, The speed is 0.8 m / s to ensure the accuracy of the quality data.

[0066] Step Two: Multi-Source Data Preprocessing: The central control room processes the acquired raw data through the signal preprocessing module, including pressure... ,quality , , and The data is denoised using discrete wavelet transform, and all denoised data is normalized to eliminate dimensional differences.

[0067] In step two, the denoising process employs Discrete Wavelet Transform (DWT), selecting the Sym4 wavelet basis and a decomposition level of 3 layers for pressure. ,quality , , and The high-frequency detail coefficients of the data are inversely reconstructed after soft or hard thresholding to filter out electromagnetic and mechanical vibration noise in the 10-100Hz range. Normalization is performed using Z-score normalization, calculated as follows:

[0068] ,in, For the normalized first Data points, The raw data collected by the sensor Data points, for pressure ,quality , , and Moisture content , viscosity Any type of raw monitoring data is collected in real time by the corresponding sensor. Pressure sensor 66 collects the pressure of the filter roller, and hyperspectral sensor 51 collects the moisture content of the filter cake. The collection frequency is every 0.5 seconds, and the data is directly transmitted to the central control room for use as raw data. This is the average value of the sensor data over a preset 5-minute time window. This represents the standard deviation of the sensor data within a preset time window.

[0069] Step 3: Dynamic evaluation and decision fusion of filter press effect: The central control room calls the multi-parameter fusion algorithm model to realize two-level fusion evaluation and decision based on the preprocessed data. The feature-level fusion model predicts the filter press endpoint and the final cake moisture content; the decision-level fusion model calculates the attention weight of each parameter and outputs the optimal pressing pressure for the next stage; the filter press efficiency score is calculated, with a score range of 0-100 points, and parameter maintenance, early warning or closed-loop optimization is performed according to the score.

[0070] In step three, the feature-level fusion model is an LSTM-autoencoder model, whose input features include the rate of change of water content. Pressure curve of filter roller Filtrate mass loss rate Output mud cake quality change rate The model training uses a loss function, calculated as follows:

[0071] ,in, The total loss of the model, , The weighting coefficients and > , , The setting prioritizes the final moisture content prediction over data reconstruction, because the core indicator of filter press performance is... Does it meet the standard?

[0072] The data reconstruction loss is calculated using the following formula: .in, The time step of the input feature sequence. This corresponds to 60 seconds of data collection. For a moment The original input feature vector contains the rate of change of water content at that moment. Filter roller pressure curve value Filtrate mass loss rate Output mud cake quality change rate Four core features The moment for model reconstruction The feature vectors of the model are encoded by the encoder. After dimensionality reduction and feature extraction, the vector is reconstructed by the decoder and used to compare it with the original vector. Compare the calculation of reconstruction loss, It needs to be controlled at ≤0.1;

[0073] The formula for calculating the predicted loss based on the final moisture content is as follows: ,in, The final moisture content of the sludge cake measured at the end of the filtration cycle is determined by the hyperspectral sensor 51 at the time of filtration termination, provided that the moisture content meets the specified conditions. The value was directly measured when it stabilized between 18% and 22% for 30 seconds and M2≈0. The final moisture content of the mud cake predicted by the model is derived from the output layer after the model extracts deep features through the encoder based on the input temporal feature vector. It needs to be controlled at ≤1%;

[0074] Model training optimization: The Adam optimizer (learning rate 0.001, decay rate 0.0001) is used, with 50 training epochs and a batch size of 32. When the validation set loss does not decrease for 5 consecutive epochs, an early stopping mechanism is triggered to avoid overfitting.

[0075] For the refinement of the LSTM-autoencoder feature-level fusion model:

[0076] Model structure parameters: Encoder: Contains 3 layers of LSTM units, with 64 neurons in the first layer, 32 neurons in the second layer, and 16 neurons in the third layer, reducing the dimensionality to 16-dimensional feature vectors. The activation function is ReLU to avoid gradient vanishing. A Dropout layer is added to each layer with a dropout rate of 0.2 to prevent overfitting.

[0077] Decoder: Symmetrical to encoder, containing 3 layers of LSTM units (16→32→64), the output layer uses a linear activation function (reconstructing the input feature vector).

[0078] Input feature extraction methods:

[0079] Moisture content change rate The sliding window difference method is used for calculation, i.e. ,in, For a moment The measured moisture content of the mud cake was obtained by the hyperspectral sensor 51. The data was collected at all times and pre-processed for noise reduction. The hyperspectral sensor 51 analyzed the reflectance spectrum of the mud cake in the 400-1000nm band and calculated it using a partial least squares regression (PLSR) model. The measurement error was ≤±1%.

[0080] For a moment Before The measured moisture content of the mud cake over time, and They are from the same source (the same hyperspectral sensor 51) and have undergone the same noise reduction preprocessing.

[0081] To calculate the time interval for the rate of change, a sliding window difference method was used with a window length of 10 seconds (a fixed value), corresponding to 20 data points.

[0082] The value range is as follows: 0.3-0.6% / min in the initial stage of filter pressing (feeding stage), 0.8-1.2% / min in the middle stage (pressing stage), and 0.1-0.3% / min in the later stage (drying stage). If a negative value appears, it indicates that the moisture content has increased abnormally (such as the filter belt being damaged, causing the filtrate to flow back).

[0083] Filter roller pressure curve Take pressure data (600 points) within a 5-minute window, and obtain a smooth curve through polynomial fitting (3rd degree polynomial). Extract the slope and peak pressure of the fitted curve as... eigenvalues.

[0084] Fitting criteria: Based on 600 raw pressure data points within a 5-minute window, a third-order polynomial fitting was performed using the least squares method. The fitting formula is as follows:

[0085] ,in, , , , The fitting coefficients are used to ensure that the mean square error (MSE) between the fitted curve and the original data is ≤0.01MPa².

[0086] Slope of the fitted curve: pressure curve At the present moment The first derivative, i.e. It reflects the rate of change of pressure (rising, falling, or remaining constant).

[0087] If the slope is greater than 0, it indicates that the pressure is increasing (such as during the pressing stage); if the slope is approximately 0, it indicates that the pressure is stable (such as during the pressure holding stage); if the slope is less than 0, it indicates that the pressure is decreasing (such as during the pressure release stage).

[0088] Peak pressure: Pressure curve within a 5-minute window The maximum value, i.e. This reflects the maximum pressing intensity within the window. If the peak pressure continues to exceed 1.5 MPa, an overpressure warning needs to be triggered.

[0089] Filtrate mass loss rate Calculation method ,in, For a moment The cumulative mass of the filtrate is obtained by the electromagnetic flowmeter at the bottom of the filtrate collection tank. For a moment Before The cumulative mass of filtrate over time, and Same source (same flow meter).

[0090] With the rate of change of moisture content Consistency is maintained, with values ​​taken over 10 seconds (20 data points) to ensure a unified calculation benchmark for different rates of change, facilitating subsequent multi-parameter fusion analysis;

[0091] Value range: 5-10 kg / min during the middle stage of pressure filtration (pressing stage), and gradually decrease to 0-2 kg / min in the later stage (drying stage). If it is close to 0, it means that dehydration is basically completed. Unit: kg / min.

[0092] Output mud cake mass change rate :because For intermittent measurements (batch weighing after mud cake forming), linear interpolation was used to supplement the intermittent data before proceeding with the next step. The method of calculation, namely ,in, The cumulative mass of the output cake after linear interpolation. For a moment Before The cumulative mass of the mud cake is output after time interpolation. Consistent with the aforementioned rate of change, a value of 10 seconds (20 data points) is used to ensure a consistent calculation benchmark.

[0093] The decision-level fusion model in step three employs the Transformer multi-head attention mechanism, and the weight calculation formula is as follows: ,in, To optimize the target vector, This is a real-time state parameter vector, containing the current water content. The denoised and normalized value, the current pressure The denoised value and the pressing time The time elapsed since the filter press started up. This is the basic adjustment vector corresponding to each state parameter. for The dimension of a vector , For query vector With key vector The product of the transposes of the matrices is obtained through matrix multiplication.

[0094] Optimal pressing pressure in step three The calculation formula is:

[0095] ,in, The number of state parameters. The first one calculated through the attention mechanism Attention weights for each state parameter. In order to be with the first The basic adjustment amount of the pressing pressure corresponding to each state parameter This is the actual pressure value of the filter roller at the current moment, reflecting the current pressing intensity. It is collected in real time by the pressure sensor and obtained after noise reduction processing.

[0096] The requirement is 0.8MPa≤ ≤1.5MPa;

[0097] Lower limit 0.8MPa: Below this value, insufficient filter pressure leads to excessive moisture content in the filter cake (triggering underpressure optimization).

[0098] Upper limit 1.5MPa: If the pressure exceeds this value, the filter belt may be damaged due to excessive compression (triggering an overpressure warning).

[0099] To avoid Frequent fluctuations, for 5 consecutive outputs Smoothing formula:

[0100] ,in, for The smoothed pressure value at any given time. This is the smoothing coefficient, with a value of 0.8. time The calculated original optimal pressure, for The smoothed pressure value at any given moment.

[0101] The formula for calculating the filter press efficiency score in step three is:

[0102] ,in, The filter press efficiency is scored, with a value ranging from 0 to 100. , , , The weighting coefficients and , maximum, , , , , The preset target moisture content of the mud cake, This is the total time consumed in this filter press operation. The preset maximum allowable pressure filtration time, The total energy consumption for this filter press operation is calculated based on the filter press roller pressure and operating time. ,in, The average filter pressure, This is the power factor of the equipment.

[0103] The preset maximum allowable pressure filtration energy consumption, This represents the total mass of the filter cake output from this filter press.

[0104] A score of ≥90 indicates excellent pressure filtration performance and consistent execution parameters (maintaining current P, feed flow rate, and reagent addition rate).

[0105] 70 points ≤ <90 points: The pressure filtration effect is good, and fine-tuning optimization is performed (only the feed flow rate is adjusted by ±0.5m³ / h, without changing the pressure).

[0106] 50 points ≤ <70 points: The pressure filtration effect is average, and moderate optimization should be performed (adjust the pressure by ±0.05MPa, and adjust the feed flow rate accordingly).

[0107] <50 points: Poor pressure filtration effect. Perform in-depth optimization (pause feeding, clean the filter belt, reset the pressure and reagent addition rate, and restart the pressure filtration).

[0108] Step 4: Closed-loop optimization control: Based on the decision results and abnormal parameter location, perform underpressure optimization, overpressure early warning, moisture content / crack optimization, and reagent synergistic optimization operations.

[0109] In step four, during the moisture content / crack optimization, the calculation formula for adjusting the feed pump flow rate using a PID controller is as follows:

[0110] ,in, For a moment The feed pump flow rate setpoint, For proportional gain, , For integral gain, , For differential gain, It is obtained through the Ziegler-Nichols tuning method, that is, first... , Set to 0, increase Until the system exhibits constant amplitude oscillations (critical) =2.0), then press , ,in, The oscillation period is set to 10 seconds. Finally, through on-site debugging and fine-tuning, it was adjusted to the current value.

[0111] The sampling period of the PID controller is synchronized with the data acquisition frequency (0.5 seconds / time), and the output is the feed pump flow rate setpoint. , The adjustment range is 5-15 m³ / h (an over-flow alarm will be triggered if the flow exceeds the range).

[0112] From the start-up time of the filter press (time 0) to the current time The error integral is used to accumulate the error over a period of time and eliminate static deviation.

[0113] The error signal is calculated using the following formula:

[0114] ,in, The target moisture content reduction rate is set in stages: 0.5% / min for the feeding stage (0-5 minutes) (slow reduction to avoid the filter cake being too thin), 1.0% / min for the pressing stage (5-20 minutes) (rapid dehydration), and 0.2% / min for the drying stage (20-30 minutes) (slow drying to prevent cracking).

[0115] For based on The actual rate of moisture content decrease, calculated in real time, is used to calculate the rate within a 5-minute window using linear regression. The trend of change (the slope is) This avoids fluctuations in single-point differences.

[0116] when When the PID controller reaches >0.2% / min, it initiates adjustment; when When the output is ≤0.2% / min, the PID controller maintains a constant output.

[0117] In step four, during the synergistic optimization of the drug-pharmaceutical mixture, the formula for calculating the dispersant addition rate is:

[0118] ,in, For a moment The rate of addition of anionic dispersants. This is the proportional control coefficient. For a moment The measured mud viscosity, The preset safety threshold for mud viscosity, Ensure the addition rate is 0 when the viscosity is below the threshold.

[0119] Step four, undervoltage optimization, specifically involves... If the pressure remains below 0.8 MPa, extend the high-pressure pressing stage by 20% and increase the pressure gradient from 0.1 MPa / min to 0.15 MPa / min until... Stable at 1.0-1.2 MPa and ≤25%, ≤1 item;

[0120] The overvoltage warning is specifically as follows: If If the pressure is greater than 1.5MPa, reduce the pressure of the filter roller to 1.2MPa and reduce the filter belt tension by 10%. If the camera detects filter cloth damage, immediately stop the equipment and trigger a filter cloth replacement reminder.

[0121] In the optimization of moisture content / cracks, if ≥3 cracks, reduce the pressure of the filter roller in the crack concentration area by 0.1MPa and increase the spray water volume of the filter belt by 10%;

[0122] Drug synergistic optimization: If When the pressure reaches >500 mPa·s, start the anionic dispersant addition device.

[0123] Step 5: Filtration termination and model iteration: When the moisture content of the sludge cake stabilizes within the preset range for a preset duration and the remaining sludge mass approaches 0, a filtration termination command is sent and filter belt cleaning is initiated. All cycle data, optimization commands, and scores are recorded and stored in the database. The model parameters are periodically adjusted based on historical data to complete iterative optimization.

[0124] Quantitative definition of termination condition:

[0125] Moisture content of mud cake The value remained stable at 18%-22% for 30 consecutive seconds (60 data points, with each data point fluctuating within ±0.5%).

[0126] Remaining mud quality ≤5kg, and maintain ≤5kg for 10 consecutive seconds;

[0127] Filtrate flow rate: The electromagnetic flow meter at the outlet of the filtrate collection tank displays a flow rate ≤0.5L / min for 10 seconds;

[0128] All three conditions must be met simultaneously to trigger the filter termination command (to avoid misjudgment based on a single condition).

[0129] Post-termination operation procedure:

[0130] Turn off the mud feed pump and stop mud injection;

[0131] The filter press rollers continue to run for 5 minutes (air pressure stage) to ensure that the residual slurry on the filter belt is fully pressed.

[0132] Start the filter belt cleaning program: high-pressure water spray - hot air drying - tension calibration;

[0133] The output weighing conveyor belt 5 transports the mud cake to the designated storage area and records the total mass of the mud cake at the same time. (Weigh three times consecutively and take the average value);

[0134] Empty the filtrate collection tank and record the total mass of the filtrate (confirmed by the cumulative value of the flow meter).

[0135] Step five has a preset range of 18%-22% and a preset duration of 30 seconds; the model parameters are adjusted regularly, specifically monthly, and the adjustments include those for the LSTM-autoencoder model. , The coefficients and attention weights of the Transformer model are calculated.

[0136] When adjusting model parameters periodically (once a month), it is necessary to first select qualified filter press cycle data from the database to meet the requirements. For samples with a filtration efficiency score of ≥80 points and no equipment malfunction records, the number of selected data sets must be ≥50 (to ensure sample representativeness); adjustments were made to include:

[0137] LSTM-autoencoder model: During retraining, the newly selected data is divided into training and test sets in a 7:3 ratio, and adjustments are made. (Original 0.7) and The value of (original 0.3) (if the prediction error of the moisture content of the test set is >3%, then) Increase by 0.05 Reduce by 0.05 until the error is ≤3%).

[0138] Transformer model: Adjusting attention weight calculation parameters (such as query vector) (normalization coefficients) to ensure optimal pressing pressure The calculation error is ≤ ±0.05MPa;

[0139] After the model is updated, it needs to be validated offline (using nearly 10 sets of qualified data that were not used in training to test the model output). After the validation is successful, it is written to the algorithm module in the central control room through incremental update (to avoid system downtime caused by full update). During the update process, the original model backup needs to be retained (if the new model is abnormal, it will be automatically rolled back to the original model within 10 seconds).

[0140] Step five, recording the data for the entire filter press cycle, specifically includes: data within the entire cycle. , , , , , , , , , Time series data, and total output mass of mud cake Mud cake recovery rate ;

[0141] When storing in the database, and The associated storage is used to optimize the correlation prediction accuracy of cake quality-moisture content during subsequent model iterations.

[0142] Example 2

[0143] Reference Figures 2-7 As shown, the weighing components in step one include an input device and an output device. The input device includes a connecting frame 2, which is fixedly installed on the outer surface of the frame of the belt filter press 1. A sliding plate 21 is fixedly installed on the outer surface of the connecting frame 2.

[0144] A transmission assembly 22 is rotatably connected to the outer surface of the connecting frame 2. The transmission assembly 22 consists of a synchronous belt, a synchronous pulley, and a motor that drives the synchronous pulley to rotate. A movable frame 23 is slidably inserted into the outer surface of the connecting frame 2. The outer surface of the movable frame 23 is fixedly installed with the outer surface of the synchronous belt of the transmission assembly 22. The rotation of the synchronous belt drives the movable frame 23 to move on the connecting frame 2. A lifting screw 3 with gears is rotatably connected to the outer surface of the movable frame 23. A lifting motor 31 with gears is fixedly installed on the outer surface of the movable frame 23. The gear of the lifting motor 31 meshes with the gear of the lifting screw 3. A lifting sleeve 32 is threadedly connected to the outer surface of the lifting screw 3. The lifting sleeve 32 is slidably inserted into the outer surface of the movable frame 23.

[0145] Specifically, a scraper 33 with gears is rotatably connected to the outer surface of the lifting sleeve 32. The outer surface of the scraper 33 contacts the outer surface of the slide plate 21. The scraper 33 removes the residual mud on the slide plate 21 by moving. An electromagnet 34 is fixedly installed on the outer surface of the lifting sleeve 32. The outer surface of the electromagnet 34 is magnetically connected to the outer surface of the gear of the scraper 33. The electromagnet 34 fixes the gear on the scraper 33 to prevent the scraper 33 from deflecting. A roller 35 is rotatably connected to the outer surface of the scraper 33.

[0146] Specifically, the inner wall of the connecting frame 2 is provided with a partition 4. The outer surface of the partition 4 and the inner wall of the connecting frame 2 form a U-shaped groove. A fixed cylinder assembly 41 is fixedly installed on the inner wall of the partition 4. One fixed cylinder assembly 41 consists of two opposing cylinders. The partition 4 is fixed by the insertion between the cylinder and the connecting frame 2, without affecting the movement of the scraper 33 in the U-shaped groove. The piston rod of the cylinder rises and falls to facilitate the passage of the scraper 33. One end of the piston rod of the fixed cylinder assembly 41 is slidably inserted into the inner wall of the connecting frame 2. A deflection rack 42 is fixedly installed on the outer surface of both the connecting frame 2 and the partition 4. The deflection rack 42 meshes with the gear of the scraper 33. The roller 35 is slidably connected to the inner wall of the U-shaped groove.

[0147] Specifically, the output device also includes a hyperspectral sensor 51, which is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt 5 via a bracket. The output weighing conveyor belt 5 is located on one side of the belt filter press 1. A vision sensor 52 is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt 5 via a bracket. A support frame 6 is fixedly mounted on the outer surface of the connecting frame 2. The outer surface of the support frame 6 is fixedly mounted to the outer surface of the mud weighing sensor 61. A hopper 6 with a mud feed pipe is fixedly mounted on the outer surface of the mud weighing sensor 61. 2. The mud weighing sensor 61 detects the initial mud mass and the remaining mud mass. The viscosity sensor 63 is fixedly installed on the wall of the feed pipe of the feed hopper 62. The capacitive moisture sensor 65 is fixedly installed on the inner wall of the collection tank of the belt filter press 1. The outer surface of the frame of the belt filter press 1 is fixedly installed with the outer surface of the filtrate weighing sensor 64. The outer surface of the filtrate weighing sensor 64 is fixedly installed with the lower surface of the collection tank of the belt filter press 1. The pressure sensor 66 is fixedly installed on the pressure detection interface of the filter roller bearing seat of the belt filter press 1.

[0148] The hyperspectral sensor 51, model HyperspecVNIR400-1000nm, is fixed 1.5m directly above the output weighing conveyor belt 5 using an aluminum bracket. The lens is vertically downward, and the field of view covers the full width of the filter belt (3m). During data acquisition, the built-in diffuse reflection calibration plate (99% reflectivity) must be activated simultaneously to ensure the moisture content of the mud cake is within acceptable limits. Measurement error ≤ ±1%;

[0149] The CS220 capacitive moisture sensor (model 65) is embedded in the filtrate collection tank of a belt filter press, 300 mm from the tank wall. The sensor probe is in direct contact with the filtrate, and the measurement range is 0-100% (filtrate moisture content). With a resolution of 0.1%, measurement offset caused by tank wall vibration is avoided;

[0150] The pressure sensor 66, model PT124B-25MPa, uses a flange-mounted pressure detection interface on the filter roller shaft bearing seat (two sensors are configured for each filter roller, symmetrically distributed at both ends of the roller shaft), and has a measurement range of 0-25MPa (filter roller pressure). (Accuracy 0.2%FS), and the radial force interference generated by the rotation of the roller shaft must be shielded during data acquisition;

[0151] The output weighing conveyor belt 5 is a belt-type dynamic weighing module (measuring range 0-500kg, accuracy ±0.5kg). The output weighing conveyor belt 5 is installed 0.5m behind the discharge port of the belt filter press 1 (to measure the mass of the output sludge cake). );

[0152] The vision sensor 52, model Baslerac A2500-14gm, is fixed 1m directly above the output weighing conveyor belt 5 via an aluminum bracket. The lens is tilted at 45° (towards the direction of mud cake movement), with a resolution of 2592×1944 pixels. Combined with an LED strip light source (wavelength 650nm), it ensures that the minimum size of mud cake crack recognition is ≤0.5mm and the average thickness (H) measurement error is ≤±0.2mm.

[0153] The viscosity sensor 63 is a Brookfield DV2T, which is fixed to the wall of the feed pipe of the hopper 62 (500mm from the feed pipe inlet) by clamps. The probe is inserted into the pipe to a depth of 1 / 3 of its diameter, and the measurement range is 1-10000 mPa·s (mud viscosity). The measurement frequency is synchronized with the data acquisition frequency (0.5 seconds / time) to avoid probe clogging caused by slurry deposition;

[0154] The mud weighing sensor 61 is equipped with an OMEGALC305-10K compression weighing sensor. Two sets are evenly distributed along the bottom sides of the feed hopper 62 and are fixed between the feed hopper 62 and the support frame 6 by bolts. A buffer rubber pad is installed on the top of the sensor to ensure that the weight of the mud is vertically loaded onto the force-bearing surface of the sensor.

[0155] Working principle: The slurry in the slurry feed pipe is conveyed to the slide plate 21 through the hopper 62. The slurry weighing sensor 61 detects the weight of the slurry in the hopper 62. The conveying assembly 22 is activated, driving the moving frame 23 to move on the connecting frame 2. The moving frame 23 drives the scraper 33 to scrape the slurry on the slide plate 21. The scraper 33 moves within the U-shaped groove formed between the roller 35 on the scraper 21, the partition plate 4, and the connecting frame 2. When it encounters the piston rod of the fixed cylinder assembly 41, it scrapes the slurry on the slide plate 21 onto the belt filter press 1 below. When the scraper 33 on one side moves into position, the gear on the scraper 33 can mesh with the deflection rack 42, which can drive the scraper 33 to deflect. When the electromagnet 34 is energized, it magnetically connects to the gear on the scraper 33. The lifting motor 31 starts and drives the lifting screw 3 to rotate. The lifting screw 3 drives the lifting sleeve 32 to rise. The lifting sleeve 32 drives the scraper 33 to rise in the U-shaped groove. At the same time, the conveying component 22 drives the rising scraper 33 back. Another scraper 33 scrapes the mud on the slide plate 21. After crossing with the rising and returning scraper 33, it continues to move. When the returning scraper 33 encounters another deflecting rack 42, it drives the deflected scraper 33 to reset, so that the scraper 33 is perpendicular to the slide plate 21. The lifting screw 3 rotates and drives the scraper 33 to descend and contact the outer surface of the slide plate 21 for the next scraping action.

[0156] The viscosity sensor 63 on the slurry feed pipe detects the viscosity of the slurry, the capacitive moisture sensor 65 detects the moisture content in the collection tank, the pressure sensor 66 detects the pressure of the filter press roller, and the slurry cake pressed by the belt filter press 1 is weighed after entering the output weighing conveyor belt 5. The moisture content is detected by the hyperspectral sensor 51, and the visual sensor 52 detects the cracks and thickness of the slurry cake.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing, characterized in that: Step 1: Mud injection and full-cycle data synchronous acquisition: Start the mud feed pump to inject shield mud into the belt filter press, start the data acquisition command in the central control room, and integrate a multi-modal sensor cluster on the belt filter press (1) to synchronously acquire the following data at a frequency of 0.5 seconds / time: Moisture content of mud cake output by hyperspectral sensor (51) The filtrate moisture content output by the capacitive moisture sensor (65) The pressure of the filter roller output by the pressure sensor (66) The initial mud mass output by the mud weighing sensor (61) in the weighing component. Residual mud quality The filtrate weighing sensor (64) outputs the filtrate mass. The number of mud cake cracks output by the vision sensor (52) With average thickness Viscosity sensor (63) outputs mud viscosity And the output weighing conveyor belt (5) outputs the mud cake mass. The collected data will be transmitted to the central control room in real time. Step Two: Multi-Source Data Preprocessing: The central control room processes the acquired raw data through the signal preprocessing module, including pressure... ,quality , , and The data is denoised using discrete wavelet transform, and all the denoised data is normalized to eliminate dimensional differences. Step 3: Dynamic Evaluation and Decision Fusion of Filtration Effect: The central control room calls a multi-parameter fusion algorithm model to achieve two-level fusion evaluation and decision-making based on the preprocessed data. The feature-level fusion model predicts the filtration endpoint and the final cake moisture content; the decision-level fusion model calculates the attention weight of each parameter and outputs the optimal pressing pressure for the next stage; the filtration efficiency score is calculated, with a score range of 0-100 points, and parameter maintenance, early warning, or closed-loop optimization is performed based on the score. Step 4: Closed-loop optimization control: Based on the decision results and abnormal parameter location, perform underpressure optimization, overpressure early warning, moisture content / crack optimization, and chemical synergistic optimization operations; Step 5: Filtration termination and model iteration: When the moisture content of the sludge cake stabilizes within the preset range for a preset duration and the remaining sludge mass approaches 0, a filtration termination command is sent and filter belt cleaning is initiated. All cycle data, optimization commands, and scores are recorded and stored in the database. The model parameters are periodically adjusted based on historical data to complete iterative optimization.

2. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 1, characterized in that: The denoising process in step two employs discrete wavelet transform, selecting a Sym4 wavelet basis and a decomposition level of 3 layers, for pressure... ,quality , , and The high-frequency detail coefficients of the data are inversely reconstructed after soft or hard thresholding to filter out electromagnetic and mechanical vibration noise in the 10-100Hz range. Normalization is performed using Z-score normalization, calculated as follows: ,in, For the normalized first Data points, The raw data collected by the sensor Data points, for pressure ,quality , , and Moisture content , viscosity Any type of raw monitoring data is collected in real time by the corresponding sensor. The pressure sensor (66) collects the pressure of the filter roller, and the hyperspectral sensor (51) collects the moisture content of the filter cake. The collection frequency is every 0.5 seconds. The data is then directly transmitted to the central control room and used as raw data. This is the average value of the sensor data over a preset 5-minute time window. The standard deviation of the sensor data within the preset time window is given.

3. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 2, characterized in that: The feature-level fusion model in step three is an LSTM-autoencoder model, whose input features include the rate of change of water content. Pressure curve of filter roller Filtrate mass loss rate Output mud cake quality change rate The model training uses a loss function, calculated as follows: ,in, The total loss of the model, , The weighting coefficients and > ; The data reconstruction loss is calculated using the following formula: ,in, The time step of the input feature sequence. This corresponds to 60 seconds of data collection. For a moment The original input feature vector contains the rate of change of water content at that moment. Filter roller pressure curve value Filtrate mass loss rate Output mud cake quality change rate Four core features The moment for model reconstruction The feature vectors of the model are encoded by the encoder. After dimensionality reduction and feature extraction, the vector is reconstructed by the decoder and used to compare it with the original vector. Compare and calculate the reconstruction loss; The formula for calculating the predicted loss based on the final moisture content is as follows: ,in, The final moisture content of the filter cake measured at the end of the filter press cycle is determined by a hyperspectral sensor (51) at the time of filter press termination, provided that the moisture content meets the specified conditions. Stabilize at 18%-22% and maintain for 30 seconds. When ≈0, it can be directly measured. The final moisture content of the mud cake predicted by the model is obtained by the output layer after the model extracts deep features through the encoder based on the input temporal feature vector. The decision-level fusion model in step three employs the Transformer multi-head attention mechanism, and the weight calculation formula is as follows: ,in, To optimize the target vector, This is a real-time state parameter vector, containing the current water content. Current pressure and pressing time , This is the basic adjustment vector corresponding to each state parameter. for The dimension of a vector , query vector With key vector The product of the transposes of matrices is obtained through matrix multiplication. The optimal pressing pressure in step three The calculation formula is: ,in, The number of state parameters. The first one calculated through the attention mechanism Attention weights for each state parameter. In order to be with the first The basic adjustment amount of the pressing pressure corresponding to each state parameter The actual pressure value of the filter roller at the current moment reflects the current pressing intensity. It is collected in real time by the pressure sensor (66) and obtained after noise reduction processing. The formula for calculating the filter press efficiency score in step three is as follows: ,in, The filter press efficiency is scored, with a value ranging from 0 to 100. , , , The weighting coefficients and , maximum, , , , , The preset target moisture content of the mud cake, This is the total time consumed in this filter press operation. The preset maximum allowable pressure filtration time, The total energy consumption for this filter press operation is calculated based on the filter press roller pressure and operating time. ,in, The average filter pressure, The power factor of the equipment; The preset maximum allowable pressure filtration energy consumption, This refers to the total mass of the filter cake output from this filter press. ≥90 points: Excellent pressure filtration effect, maintained execution parameters, keep current. The feed flow rate and reagent addition rate remain constant. 70 points ≤ <90 points: The pressure filtration effect is good. Fine-tuning and optimization were performed, adjusting only the feed flow rate ±0.5m³ / h without changing the pressure; 50 points ≤ <70 points: The pressure filtration effect is average. Perform moderate optimization, adjust the pressure by ±0.05MPa, and adjust in conjunction with the feed flow rate; <50 points: Poor pressure filtration effect. Perform in-depth optimization, pause feeding, clean the filter belt, reset the pressure and reagent addition rate, and restart the pressure filtration.

4. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 3, characterized in that: In step four, during the moisture content / crack optimization, the calculation formula for adjusting the feed pump flow rate using a PID controller is as follows: ,in, For a moment The feed pump flow rate setpoint, For proportional gain, , For integral gain, , For differential gain, , From the start-up time of the filter press (time 0) to the current time The error integral is used to accumulate the error over a period of time and eliminate static deviation; The error signal is calculated using the following formula: ,in, The preset target rate of moisture content decrease, For based on Real-time calculation of the actual rate of moisture content decrease; In step four, during the synergistic optimization of the pharmaceutical agents, the formula for calculating the dispersant addition rate is as follows: ,in, For a moment The rate of addition of anionic dispersants. This is the proportional control coefficient. For a moment The measured mud viscosity, The preset safety threshold for mud viscosity, Ensure the addition rate is 0 when the viscosity is below the threshold. In step four, the undervoltage optimization specifically involves... If the pressure remains below 0.8 MPa, extend the high-pressure pressing stage by 20% and increase the pressure gradient from 0.1 MPa / min to 0.15 MPa / min until... Stable at 1.0-1.2 MPa and ≤25%, ≤1 item; The overvoltage warning is specifically as follows: If If the pressure is greater than 1.5MPa, reduce the pressure of the filter roller to 1.2MPa and reduce the filter belt tension by 10%. If the camera detects filter cloth damage, immediately stop the equipment and trigger a filter cloth replacement reminder. In the aforementioned moisture content / crack optimization, if ≥3 cracks, reduce the pressure of the filter roller in the crack concentration area by 0.1MPa and increase the spray water volume of the filter belt by 10%; Drug synergistic optimization: If When the pressure reaches >500 mPa·s, start the anionic dispersant addition device.

5. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 4, characterized in that: In step five, the preset range is 18%-22%, and the preset duration is 30 seconds; the periodic adjustment of model parameters specifically refers to monthly adjustments, and the adjustment targets include the LSTM-autoencoder model. , Coefficients and attention weights calculation parameters for the Transformer model; The "recording of the entire filter press cycle data" in step five specifically includes: data within the entire cycle. , , , , , , , , , Time series data, and total output mass of mud cake Mud cake recovery rate ; When storing in the database, and The associated storage is used to optimize the correlation prediction accuracy of cake quality-moisture content during subsequent model iterations.

6. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 1, characterized in that: The weighing component in step one includes an input device and an output device. The input device includes a connecting frame (2). The connecting frame (2) is fixedly installed on the outer surface of the frame of the belt filter press (1). A sliding plate (21) is fixedly installed on the outer surface of the connecting frame (2).

7. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 6, characterized in that: The outer surface of the connecting frame (2) is rotatably connected to the conveying component (22), and the outer surface of the connecting frame (2) is slidably inserted into the moving frame (23). The outer surface of the moving frame (23) is fixedly installed with the outer surface of the synchronous belt of the conveying component (22). The outer surface of the moving frame (23) is rotatably connected to the lifting screw (3) with gears. The outer surface of the moving frame (23) is fixedly installed with the lifting motor (31) with gears. The gear of the lifting motor (31) meshes with the gear of the lifting screw (3). The outer surface of the lifting screw (3) is threadedly connected to the lifting sleeve (32), and the lifting sleeve (32) is slidably inserted into the outer surface of the moving frame (23).

8. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 7, characterized in that: The outer surface of the lifting sleeve (32) is rotatably connected to a scraper (33) with gears. The outer surface of the scraper (33) is in contact with the outer surface of the slide plate (21). An electromagnet (34) is fixedly installed on the outer surface of the lifting sleeve (32). The outer surface of the electromagnet (34) is magnetically connected to the outer surface of the gear of the scraper (33). A roller (35) is rotatably connected to the outer surface of the scraper (33).

9. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 8, characterized in that: The inner wall of the connecting frame (2) is provided with a partition (4). The outer surface of the partition (4) forms a U-shaped groove with the inner wall of the connecting frame (2). A fixed cylinder assembly (41) is fixedly installed on the inner wall of the partition (4). One end of the piston rod of the fixed cylinder assembly (41) is slidably inserted into the inner wall of the connecting frame (2). A deflection rack (42) is fixedly installed on the outer surfaces of both the connecting frame (2) and the partition (4). The deflection rack (42) meshes with the gear of the scraper (33). The roller (35) is slidably connected to the inner wall of the U-shaped groove.

10. The method for evaluating the filter cake pressing effect of shield tunneling machines based on multimodal sensing according to claim 9, characterized in that: The output device also includes a hyperspectral sensor (51), which is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt (5) by a bracket. The output weighing conveyor belt (5) is located on one side of the belt filter press (1). The vision sensor (52) is fixedly mounted on the outer surface of the frame of the output weighing conveyor belt (5) by a bracket. A support frame (6) is fixedly mounted on the outer surface of the connecting frame (2). The outer surface of the support frame (6) is fixedly mounted to the outer surface of the mud weighing sensor (61). A hopper (6) with a mud feed pipe is fixedly mounted on the outer surface of the mud weighing sensor (61). 2) The mud weighing sensor (61) detects the initial mud mass and the remaining mud mass. The viscosity sensor (63) is fixedly installed on the wall of the feed pipe of the hopper (62). The capacitive moisture sensor (65) is fixedly installed on the inner wall of the collection tank of the belt filter press (1). The outer surface of the frame of the belt filter press (1) is fixedly installed on the outer surface of the filtrate weighing sensor (64). The outer surface of the filtrate weighing sensor (64) is fixedly installed on the lower surface of the collection tank of the belt filter press (1). The pressure sensor (66) is fixedly installed on the pressure detection interface of the filter roller bearing seat of the belt filter press (1).

Images