CN-121701219-B - Multi-step prediction and correction control method for shield tunneling machine posture based on latent variable enhanced Free transducer

CN121701219BCN 121701219 BCN121701219 BCN 121701219BCN-121701219-B

Abstract

The application belongs to the technical field of intelligent shield construction, and particularly discloses a shield machine posture multi-step prediction and correction control method based on a latent variable enhanced Free Transformer. The method comprises the steps of firstly obtaining shield construction process data, establishing a control quantity set, preprocessing, constructing error characteristics, change rate characteristics and unit mileage characteristics based on the preprocessed data, normalizing to form characteristic sequences, organizing the sequences into a token sequence with single time step, inputting a pre-trained Free transform model after position and mileage coding, aggregating full sequence information through a special block of a non-causal coder, calculating latent variable distribution parameters based on a hidden state sequence by the model, sampling to obtain the latent variables, mapping the latent variables into conditional vectors, broadcasting the conditional vectors to a decoding path, and finally parallelly outputting future shield attitude vectors, super-threshold probability and recommended control quantity by a multi-task output head. The application realizes the multi-step prediction and correction of working condition self-adaption, and can improve the safety and the intelligent level of shield construction.

Inventors

LI ZHONGWEI
LI JILONG
WANG ZHIBAO
CAI MINGCONG
WANG SISHUN
YU JUNHAN
ZHOU SHENGLI
ZHAO LIMIN
XIE XU
DING YANQI
FANG LEI
LIU TENG
YANG DONGHENG
WANG ZHIJIAN

Assignees

中铁十四局集团大盾构工程有限公司
中铁十四局集团有限公司
深圳市龙岗区轨道交通规划建设协调事务中心

Dates

Publication Date: 20260508
Application Date: 20260210

Claims (9)

1. The shield machine posture multi-step prediction and correction control method based on the latent variable enhanced Free transducer is characterized by comprising the following steps of: S1, acquiring shield construction process data, wherein the shield construction process takes time or mileage as an index and comprises a soil pressure field, a propulsion field, a residue soil field, a grouting field and an oil field; Wherein, the The soil pressure field comprises a soil pressure set value and a soil pressure actual value; The propulsion field comprises the total thrust and the propulsion speed of the jack; The residue soil field comprises the soil output; the grouting field comprises synchronous grouting pressure, synchronous grouting amount, secondary grouting pressure, secondary grouting amount and grouting date; The grease field comprises the dosage of shield tail grease shield tail grease pressure; s2, normalizing the preprocessed construction process data and the construction features based on the preprocessed construction process data construction error features, the change rate features and the unit mileage features to obtain a normalized feature sequence; S3, organizing the normalized feature sequence into token sequences, encoding each token, and inputting the encoded token sequences into a pre-trained Free transform model; the Free transform model is configured to generate latent variables based on the token sequence and inject the latent variables into a decoder; The specific steps of the latent variable injection decoder are as follows: s31, constructing at least one non-causal Free transform block as an encoder special block, wherein the encoder special block aggregates the output of a sharing module shared by an encoder and a decoder into full sequence information; S32, calculating latent variable distribution parameters based on a latent state sequence output by the encoder, and sampling according to the latent variable distribution parameters to obtain latent variables; S33, obtaining a condition vector with the same dimension as the Free transducer model through linear mapping, and broadcasting the condition vector to a decoding path; S4, the multi-task output heads of the Free transducer model based on the injection latent variables respectively output predicted shield attitude vectors, super-threshold probability and recommended control quantity.
2. The shield tunneling machine posture multi-step prediction and correction control method based on the latent variable enhanced Free transducer according to claim 1, characterized in that, In step S2, the error features include a soil pressure error including an observation error and a relative error: ; ; In the formula, Indicating soil pressure errors in time steps Is used for the observation of errors in the (c), Representing time steps Is used for controlling the actual value of the soil pressure, Representing time steps Is set with the set value of the soil pressure, Indicating soil pressure errors in time steps Is a relative error of (2); the change rate characteristics comprise shield attitude deviation change rate, soil pressure change rate, jack thrust change rate, propulsion speed change rate, soil output change rate and grouting amount change rate, and each change rate is represented by a first-order difference: ; In the formula, Expressed in time steps Is a scalar feature or a vector feature of (c), Representing the value of the same feature at the last time step, Representing the amount of variation of the feature between adjacent steps; the unit mileage features comprise grouting amount mileage features, soil discharge amount mileage features and propulsion amount mileage features, and the unit mileage features are calculated by dividing the change amount by the mileage change value.
3. The shield tunneling machine posture multi-step prediction and correction control method based on the latent variable enhanced Free transducer according to claim 2 is characterized in that, In step S3, the code of the token comprises a position code and a mileage code; sinusoidal position coding based on time step number when position coding is performed; When the mileage coding is performed, the cut mileage is mapped into continuous position embedding and spliced with the position coding.
4. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 3, wherein, In step S31, the shared module shared by the encoder and the decoder is composed of a plurality of Free transform blocks; The multivariable observation vector of each time step is converted into a representation vector consistent with the hiding dimension of the Free transducer model through linear mapping, and the representation vector of each time step is fused with the position code and the mileage code and then is input to the sharing module.
5. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 4, wherein the method comprises the steps of, In step S31, the non-causal Free transform block employs a non-causal self-attention mechanism, and does not use causal masking in calculating the attention weights so that each time step can access information for all positions in the token sequence.
6. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 5, wherein the method comprises the steps of, In step S32, pooling operation is performed along a time dimension based on the hidden state sequence, so as to obtain a global vector representing the current construction working condition; Based on the global vector, mean parameters and variance parameters of the latent variable Gaussian distribution are respectively generated through a learnable linear mapping to serve as distribution parameters.
7. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 6, wherein the method comprises the steps of, In step S32, the pooling operation is average pooling or attention pooling.
8. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 7, wherein, In step S4, a decoding module of the decoder consists of a plurality of causal Free transform blocks, and the decoder outputs future representation sequences according to preset quantity of time steps and synchronously generates prediction targets of the time steps; Returning the shield gesture of the future time step to the future representation sequence through a multi-layer perceptron or a linear layer; Defining the engineering threshold as a plane deviation limit or an elevation deviation limit, and outputting the super-threshold probability of the future time step according to the following formula: ; in the formula, Indicating future th The super-threshold risk probability of a step, The probability operator is represented by a representation of the probability operator, Indicating future th Step monitoring the value corresponding to the predicted random variable/predicted distribution sample of the target, Representing engineering thresholds; The recommended control quantity comprises a soil pressure set value, a propelling speed, a synchronous grouting pressure and a synchronous grouting quantity.
9. The method for multi-step prediction and correction control of the shield tunneling machine attitude based on the latent variable enhanced Free transducer according to claim 8, wherein the method comprises the steps of, The recommended control quantity output adopts the following supervision type output: ; in the formula, Representing future The recommended control quantity sequence of steps, A set of control amounts is represented and, Which is indicative of the output of the decoder, And The weights and offsets of the control recommendation output heads in the multi-task output heads are represented respectively.

Description

Multi-step prediction and correction control method for shield tunneling machine posture based on latent variable enhanced Free transducer Technical Field The application relates to the technical field of intelligent shield construction, in particular to a shield machine posture multi-step prediction and correction control method based on a latent variable enhanced Free transducer. Background In the tunnel construction process, the attitude (including horizontal deviation and elevation deviation) of the shield machine directly influences the segment assembly quality, the ground subsidence control and the construction safety, and the segment assembly quality, the ground subsidence control and the construction safety must be monitored and actively regulated and controlled in real time. For the shield attitude, the conventional engineering practice mainly depends on the experience of operators and combines sensor data to perform manual correction, so that the efficiency is low and hysteresis decision is easy to generate. In recent years, the field of intelligent shield construction attempts to introduce a data driving model to predict the shield posture and assist in decision-making. The improvement achieves a certain effect, but due to the complex shield construction, the prior art still has the following defects in the aspects of prediction and control: Firstly, the shield gesture is influenced by multiple factor coupling, wherein operation parameters such as grouting pressure, propulsion speed and the like have obvious time lag and accumulation effects on gesture adjustment, and a traditional time sequence model (such as LSTM and GRU) is difficult to accurately describe the long-range dynamic dependence, so that multi-step prediction errors are accumulated rapidly. Secondly, in actual construction, key environmental factors (such as stratum lithology change, underground water distribution and soil disturbance state in front of a cutter head) cannot be directly observed by a sensor, but have decisive influence on a shield-stratum interaction mechanism. Existing models are typically modeled based only on observable variables, such implicit conditions are not explicitly considered, and the generalization ability of the model is drastically reduced when geological conditions are abrupt or inter-regional construction occurs. Thirdly, the main flow method outputs a deterministic point predicted value in a multi-output mode, the uncertainty of the predicted result is not quantified, the deviation overrun risk cannot be estimated, and the risk sensitive control decision is difficult to support. Finally, the existing system generally predicts and controls the fracturing treatment, wherein a prediction module only provides future deviation estimation, a deviation correction strategy still depends on a rule base or manual intervention, and a plurality of integrated control proposal methods output operation parameters often neglect equipment physical constraint, so that recommended instructions lack constraint feasibility guarantee. Disclosure of Invention The application provides a shield machine posture multi-step prediction and correction control method based on a latent variable enhanced Free transducer, which solves the problems of error accumulation, weak generalization capability, prediction result quantification missing and instruction lack constraint in the process of shield machine posture prediction and correction control in the prior art. The technical scheme of the application is as follows: A shield machine posture multi-step prediction and correction control method based on a latent variable enhanced Free transducer comprises the following steps: s1, acquiring shield construction process data; s2, normalizing the preprocessed construction process data and the construction features based on the preprocessed construction process data construction error features, the change rate features and the unit mileage features to obtain a normalized feature sequence; S3, organizing the normalized feature sequence into token sequences, encoding each token, and inputting the encoded token sequences into a pre-trained Free transform model; the Free transform model is configured to generate latent variables based on the token sequence and inject the latent variables into a decoder; The specific steps of the latent variable injection decoder are as follows: s31, constructing at least one non-causal Free transform block as an encoder special block, wherein the encoder special block aggregates the output of a sharing module shared by an encoder and a decoder into full sequence information; S32, calculating latent variable distribution parameters based on a latent state sequence output by the encoder, and sampling according to the latent variable distribution parameters to obtain latent variables; S33, obtaining a condition vector with the same dimension as the Free transducer model through linear mapping, and broadcasting t