Shield tunneling parameter self-adaptive regulation method and system based on intelligent control

Abstract

The embodiment of the application discloses a shield tunneling parameter self-adaptive regulation and control method and system based on intelligent control, which comprises the following steps: firstly, acquiring shield tunneling time series data and generating semantic labeled manipulation sentences through semantic labeling processing and dependency relationship pairing, obtaining semantic labeled manipulation sentence sequences according to the semantic labeled manipulation sentences, and performing context association learning processing based on an attention mechanism to generate context association representation vectors that fuse global causal correlation information; then, identifying and extracting causal trigger chains from the context association representation vector sequences and determining trigger delay time steps, and then determining a causal correlation mode; finally, determining the regulation and control action type and action execution time of each time step in turn by adopting a time step recursion strategy generation mode to obtain a tunneling parameter regulation and control instruction sequence, so that the causal transmission and time sequence dependence relationship between regulation and control actions is established at a global level, and the prediction ability for subsequent chain effects is realized.

Description

Technical Field

[0001] This application relates to the field of computer intelligent control technology, specifically to an adaptive control method and system for shield tunneling parameters based on intelligent control. Background Technology

[0002] Shield tunneling is the core construction method for large-scale tunnel projects. Its process involves the coordinated control of dozens of parameters, such as cutterhead rotation speed, propulsion speed, soil chamber pressure, and grouting volume.

[0003] Currently, the operation of tunnel boring machines (TBMs) mainly relies on manual experience or feedback control logic based on single parameter thresholds. Sensors installed at key parts of the TBM continuously generate massive amounts of time-series data at high frequency, recording snapshots of the equipment status at the moment each control action is triggered. The engineering data platform has accumulated a large number of historical control action event records, forming an operation log library covering various geological conditions. In existing technologies, the utilization of this time-series data typically remains at the level of statistical analysis or simple rule extraction, treating each control operation as an independent event. This fails to establish causal transmission and temporal dependencies between control actions at a global level, resulting in control strategies lacking the ability to predict subsequent chain effects, leading to frequent parameter fluctuations and insufficient adaptability to geological formations during the tunneling process. Summary of the Invention

[0004] This application provides a method and system for adaptive control of shield tunneling parameters based on intelligent control, which establishes causal transmission and temporal dependence between control actions at the global level, enabling the generated control strategy to predict subsequent chain effects and improve the problems of frequent parameter fluctuations and insufficient geological adaptability during the tunneling process.

[0005] This application provides an embodiment of a shield tunneling parameter adaptive control method based on intelligent control, including: Acquire shield tunneling time sequence data collected and cached by a sensor acquisition device in chronological order during the shield tunneling construction process; wherein, each control action event record line recorded sequentially in the shield tunneling time sequence data includes a control action event identifier field and a tunneling status monitoring parameter field, wherein the tunneling status monitoring parameter field stores a group of tunneling status monitoring parameters that are synchronously captured at the time of occurrence of the corresponding control action event; Based on the control action event identifier field and tunneling status monitoring parameter field corresponding to each control action event record line in the shield tunneling time series data, semantically tagged manipulation statements are generated through semantic tagging processing and dependency relationship pairing; wherein, all the semantically tagged manipulation statements are arranged in the original time order of the control action event record lines to form a semantically tagged manipulation statement sequence. The semantically labeled manipulation statement sequence is subjected to context association learning processing based on attention mechanism to generate attention weight coefficients between any two semantically labeled manipulation statements in the semantically labeled manipulation statement sequence. The attention weight coefficients are then used to perform weighted feature aggregation on the semantically labeled manipulation statement sequence to generate a context association representation vector that integrates global causal association information for each semantically labeled manipulation statement. The causal triggering chain between control action events is identified and extracted from the context association representation vector sequence, and the trigger delay time step between adjacent control action events on the causal triggering chain is determined. The causal association mode is determined based on the causal triggering chain and the trigger delay time step. Based on the aforementioned causal relationship pattern, a time-step recursive strategy is used to sequentially determine the control action type and execution time of each time step in the target tunneling period, thereby obtaining a sequence of tunneling parameter control instructions.

[0006] This application also provides an adaptive control system for shield tunneling parameters, including: processor; Storage device, on which computer programs are stored, When the computer program is executed by the processor, the processor implements any of the aforementioned intelligent control-based adaptive adjustment methods for shield tunneling parameters.

[0007] This application embodiment also provides a readable storage medium storing a program or instructions, which, when executed by a processor, implements the steps of the intelligent control-based shield tunneling parameter adaptive control method.

[0008] Therefore, the embodiments of this application have the following beneficial effects: This application reconstructs the discrete temporal sensing records in the shield tunneling process into a semantically labeled manipulation statement sequence with syntactic structure through semantic tagging and dependency relationship pairing, enabling the physical operation semantics to be uniformly formalized. Then, through context association learning processing based on attention mechanism, a context association representation vector integrating global causal association information is generated for each semantically labeled manipulation statement, completing the mapping from isolated event records to a causal semantic space. Furthermore, the causal triggering chain between control action events is identified and extracted from this semantic space, and a structured causal association pattern is formed by combining the trigger delay time step, realizing explicit modeling of the inherent causal temporal dynamics of tunneling control. Finally, based on the causal association pattern, a time step recursion strategy is adopted to generate a tunneling parameter control command sequence, enabling the parameter control process to leap from passive response to active adherence to causal laws, improving the foresight of the tunneling parameter control command sequence and the stability of the tunneling state. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an adaptive adjustment method for shield tunneling parameters based on intelligent control, provided in an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the basic structure of a shield tunneling parameter adaptive control system provided in an embodiment of this application.

[0011] Figure 3 This is a functional block diagram of an adaptive control device for tunnel boring machine parameters provided in an embodiment of this application. Detailed Implementation

[0012] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0013] See Figure 1 As shown, this figure is a flowchart of an adaptive control method for shield tunneling parameters based on intelligent control, provided in an embodiment of this application. This method is executed by an adaptive control system for shield tunneling parameters. Figure 1 As shown, the method includes steps 110-150.

[0014] This application embodiment analyzes massive amounts of time-series data collected in real time during the construction process to uncover causal relationships between control actions, and finally generates a sequence of tunneling parameter control commands adapted to the current geological conditions. This sequence of control commands can proactively adjust key tunneling parameters such as cutterhead rotation speed, propulsion speed, soil chamber pressure, and grouting volume, thereby achieving stable control of the tunneling process.

[0015] Step 110: Obtain the shield tunneling time sequence data collected and cached by the sensor acquisition device in chronological order during the shield tunneling construction process.

[0016] The control action event record line recorded sequentially in the shield tunneling time sequence data includes a control action event identifier field and a tunneling status monitoring parameter field. The tunneling status monitoring parameter field stores the tunneling status monitoring parameter group that is synchronously captured at the time of the corresponding control action event.

[0017] In this embodiment, the sensing and acquisition device consists of a programmable logic controller and an industrial sensor cluster, distributed in key components such as the cutterhead drive, propulsion cylinders, screw conveyor, and grouting pipelines. During the tunnel boring machine (TBM) excavation process, whenever an operator or automatic control system triggers a control action, such as increasing the propulsion cylinder pressure or increasing the cutterhead speed, the sensing and acquisition device immediately captures the action information at that moment and a snapshot of the current overall machine status. The generated data is written into a circular buffer as the basic unit of recording control action events.

[0018] The control action event identifier field is a structured text code that embeds the control action trigger timestamp and the control action event index number. The trigger timestamp is accurate to the millisecond level, and the event index number is a globally incrementing integer sequence.

[0019] The tunneling status monitoring parameter field stores a set of tunneling status monitoring parameters. This set of parameters is a high-dimensional floating-point vector that records the original values ​​of physical quantities such as cutterhead torque, propulsion speed, soil chamber pressure, screw speed, grouting pressure, cutterhead speed, penetration depth, and total thrust after analog-to-digital conversion by the sensors. The values ​​of each dimension are retained to four decimal places.

[0020] For example, in an exemplary but not limited control action event record line, the control action event identifier field is “T202X06XX-143022-010-0045”, meaning that the event index 0045 occurred at 14:30:22.010 milliseconds on June XX, 202X; the tunneling status monitoring parameter field is “Tq=1280.5632, AdV=35.2410, ChP=2.1568, ScR=8.9032, GtP=1.2510, CtR=2.6732, PtR=12.5643, ThF=15820.4511”, representing the values ​​of the above eight types of tunneling status monitoring parameters at the moment the control action is triggered. All control action event record lines are stored sequentially according to the chronological order of the trigger time, thus forming a complete shield tunneling time sequence data.

[0021] Step 120: Based on the control action event identifier field and tunneling status monitoring parameter field corresponding to each control action event record line in the shield tunneling time sequence data, semantically tagged manipulation statements are generated through semantic tagging processing and dependency relationship pairing.

[0022] All of the semantically labeled manipulation statements are arranged in the original time order of the control action event record line to form a semantically labeled manipulation statement sequence.

[0023] During tunnel boring machine (TBM) construction, the original time-series data is heterogeneous and highly noisy, and the operational intentions and equipment response patterns it contains cannot be directly calculated. This step transforms each record line into a standardized statement with a clear syntactic structure and semantic labels, thereby unifying the heterogeneous physical signal space into a symbolic space suitable for deep semantic computation.

[0024] Step 121: Read the control action event record line by line from the shield tunneling time sequence data. Perform field segmentation processing on the read control action event record line to separate the control action event identifier field and the tunneling status monitoring parameter field.

[0025] This application embodiment constructs a data reader based on a finite state machine. The reader sequentially addresses the record rows in the circular buffer according to time order. For each record row being read, the field segmentation module identifies the predefined top-level delimiter "|" and decomposes the record row into two independent parts: a control action event identifier field and a tunneling status monitoring parameter field. If a record row is missing a top-level delimiter due to signal interference, a data cleaning flag is triggered, and that record row is skipped directly. After the field segmentation process is completed, the two fields enter different semantic processing pipelines.

[0026] Step 122: Perform lexical analysis on the control action event identifier field, scan the predefined delimiters in the field, and extract the control action trigger timestamp and control action event index number.

[0027] For the event identifier field of the control action, the lexical parser performs a unidirectional scan from left to right. The field is separated by a hyphen "-", and the parser divides the field into five lexical units based on this separator. The first lexical unit is a pre-labeled operation category, fixed here as "T" to represent a control action. The second to fifth lexical units are parsed into strings combining year, month, day, hour, minute, second, and millisecond, respectively, based on data type determination. A timestamp reconstruction algorithm reassembles the scattered date and time lexical units into a standard control action trigger timestamp variable and a millisecond-level offset, ultimately forming a precise control action trigger timestamp. For the event index number, the parser reads the four-digit integer string after the last separator, converts it to an integer value through character code subtraction, and thus obtains the control action event index number.

[0028] For example, after performing lexical analysis on “T202X06XX-143022-010-0045”, the trigger time stamp of the control action is 202X-06-XX14:30:22.010, and the event index number is 45.

[0029] Step 123: Use the index number of the control action event to perform a search and match in the control action ontology knowledge base to obtain the concept identifier of the operation subject and the concept identifier of the operation type involved in the control action event. The concept identifier of the operation subject is determined as the operation subject code, and the concept identifier of the operation type is determined as the operation type code.

[0030] The control action ontology knowledge base is a domain knowledge graph stored in a graph database. Its internal nodes are conceptual identifiers of tunnel boring machine components and operational actions, while edges represent relational attributes such as "acts on" or "subclass". During retrieval, the control action event index number is used as the key. First, the index mapping table is queried to obtain the globally unique identifier within the knowledge base corresponding to that number. Then, based on this globally unique identifier, the node is located in the knowledge graph, and a one-hop neighbor query is performed starting from that node.

[0031] On the query path, if the edge type is "Action Type," then the conceptual identifier corresponding to the neighboring node is the operation type conceptual identifier, such as "TunnelFacePressureIncrease" or "ScrewConveyorSpeedDecrease." If the edge type is "Operation Target," then the conceptual identifier corresponding to the neighboring node is the operation subject conceptual identifier, such as "PressureControlCylinderGroup" or "ScrewConveyorMotor." The operation subject conceptual identifier is directly assigned the operation subject code, and the operation type conceptual identifier is directly assigned the operation type code. Both the operation subject code and the operation type code are atomic semantic symbols without carrying specific values.

[0032] Step 124: Perform numerical extraction and dimensional normalization on the tunneling status monitoring parameter fields to generate a normalized tunneling status monitoring parameter group. Each physical quantity value in the normalized tunneling status monitoring parameter group is mapped to a preset numerical range.

[0033] The tunneling status monitoring parameter field contains pairs of physical quantity abbreviations and numerical values, connected by an equal sign and separated by commas. The numerical extractor captures all physical quantity numerical strings sequentially based on the equal sign and comma separators, converts them into double-precision floating-point numbers through a floating-point parser, and stores them in an original parameter vector. The dimensions of this vector correspond to the categories of physical quantities such as torque, speed, and pressure.

[0034] The physical dimensions of different dimensions in the original parameter vector differ significantly; for example, the magnitude of torque is on the order of 10. 3 Newtons of rice, while the magnitude of earthen pressure is on the order of 10. 0Bar (a unit of pressure, 1 bar = 100,000 Pascals / Pa = 0.1 megapascals / MPa). During dimensional normalization, the normalization module pre-calculates the mean and standard deviation of each physical quantity dimension from a historical operational dataset, and constructs a vectorized mean offset vector and variance scaling vector accordingly. The normalization process employs the Z-score standardization method, which involves subtracting the mean offset of the corresponding dimension from each component of the original parameter vector, and then dividing the difference by the variance scaling of the corresponding dimension. After standardization, the physical quantity values ​​are projected onto a standard normal distribution interval with a mean of zero and a standard deviation close to one, thus forming a normalized set of tunneling status monitoring parameters.

[0035] Step 125: Input the operation type code into the value dimension selector to obtain a value dimension configuration list that matches the operation type code. The value dimension configuration list includes the value dimension name and dimension order that the operation type needs to describe.

[0036] The ThrustSpeed ​​dimension selector maintains a configuration dictionary indexed by operation type codes. For each operation type code already entered into the knowledge base, the configuration dictionary pre-stores a list of semantically strongly related ThrustSpeed ​​dimensions. When a thrust pressure-related operation type code is input to the ThrustSpeed ​​dimension selector, the selector performs an exact match search in the configuration dictionary, and the returned list of ThrustSpeed ​​dimension configurations may sequentially contain the three dimension names: "ThrustSpeed," "ChamberPressure," and "TotalThrustForce," along with their corresponding order indices. If an operation type code does not exist in the configuration dictionary, a preset default dimension order list is returned, which covers all available tunneling status monitoring parameter dimensions.

[0037] Step 126: Extract the parameter values ​​of the corresponding dimension from the normalized tunneling status monitoring parameter group according to the value dimension configuration list, and arrange them in dimension order to generate an operation value description column, which is an ordered value description tuple.

[0038] Value extraction follows the dimension order specified in the configuration manifest. In the normalized tunneling status monitoring parameter group, physical quantity abbreviations are used as keys, and normalized double-precision floating-point numbers are used as values, constructing a key-value pair array. The extractor traverses this array according to the dimension names in the configuration manifest. If a dimension name matches a key, the corresponding normalized parameter value is extracted. If the configuration manifest specifies "ThrustSpeed" and "ChamberPressure" in that order, the extractor captures the corresponding normalized values ​​sequentially and pushes them into an ordered list, ultimately forming an operational quantity description column containing multiple floating-point elements. The fixed order of the dimension values ​​in the operational quantity description column assigns a syntactic role to the numerical values.

[0039] Step 127: Combine the standardized dependency relationship template corresponding to the operation type code to perform dependency relationship analysis between the operation type code and the operation value description column, generate semantically labeled manipulation statements, and combine the semantically labeled manipulation statements into a semantically labeled manipulation statement sequence.

[0040] The core of this step is to endow unstructured numerical values ​​of each dimension with syntactic dependency structures, so that they form a tree-structured semantic expression with operation actions as the core and state parameters as modifiers.

[0041] Step 1271: Determine the standardized dependency relation template corresponding to the operation type code. The standardized dependency relation template includes the syntactic position of the operation type code as the dominator and the syntactic position of each value dimension in the operation value description column as the subject, as well as the dependency relation type marker between the dominator and the subject.

[0042] The standardized dependency relation template is a set of predefined dependency syntax trees. The root node of each syntax tree is an operation type code, and several dependency edges extend from the root node. Each dependency edge marks a specific dependency relation type. For example, "direct_object" indicates that the quantity dimension is the direct recipient of the action, "manner" indicates that the quantity dimension is modified as the manner of the action, and "result" indicates that the quantity dimension is the state result caused by the action. The template explicitly specifies the order of the quantity dimensions and their corresponding dependency roles. For example, in the template corresponding to the operation type code "propulsion pressure adjustment", the root node is "propulsion pressure adjustment", and it has three dependency edges pointing to the "propulsion speed" node (relation: manner), the "earthwork pressure" node (relation: result), and the "total thrust" node (relation: direct_object).

[0043] Step 1272: According to the dependency relationship type markers specified in the standardized dependency relationship template, establish directed dependency edges with dependency relationship type markers for the operation type code and the corresponding value dimensions in the operation value description column one by one, and attach the operation subject code as the topic label of the whole statement to the operation type code to obtain a semantically labeled manipulation statement with a complete semantic dependency structure.

[0044] The elements in the operation value description column output by the value dimension selector are mapped one by one, in order, to the dominated nodes pointed to by the respective dependency edges specified in the dependency relation template. Specifically, the first dominated node in the template corresponds to the first element in the operation value description column, the second dominated node corresponds to the second element, and so on, until all the edges specified in the template are filled with specific normalized parameter values.

[0045] This establishes a dependency syntax tree with the operation type code as the root and multiple parameter values ​​carrying actual values ​​as leaves. Then, the operation subject code is appended as a prefix attribute label to the attribute set of the root node, forming a topic label. At this point, a complete event semantic structure is encapsulated into a semantically labeled manipulation statement.

[0046] Step 1273: Temporarily store the semantically tagged manipulation statements generated by the current control action event record line, and determine whether there is a next control action event record line. If there is, return to continue reading; otherwise, execute the next sub-step.

[0047] The judgment logic is implemented by comparing the index number of the currently read control action event with the maximum index number of the circular buffer. If the current index number is less than the maximum index number, there is an unprocessed record line, the program pointer is incremented, and the process re-enters step 121 to read and process a new row of data. If the index number has reached the maximum, the loop ends.

[0048] Step 1274: Arrange all the temporarily stored semantically tagged manipulation statements in chronological order according to the time sequence of their corresponding control action trigger timestamps to generate the semantically tagged manipulation statement sequence.

[0049] After the loop completes, all generated semantically tagged manipulation statements are temporarily stored in a dynamic array structure. The sorting module iterates through this array, using the trigger timestamp of the control action inherited during the generation of each statement as the sort key, and calls a stable sorting algorithm based on merge sort to sort them in ascending order. After sorting, a sequence of semantically tagged manipulation statements is obtained, strictly arranged according to the event occurrence time.

[0050] Step 130: Perform attention-based context association learning processing on the semantically labeled manipulation statement sequence to generate attention weight coefficients between any two semantically labeled manipulation statements in the semantically labeled manipulation statement sequence, and use the attention weight coefficients to perform weighted feature aggregation on the semantically labeled manipulation statement sequence to generate a context association representation vector that integrates global causal association information for each semantically labeled manipulation statement.

[0051] Through the above operations, the originally discrete and isolated labeled statements are transformed into a dense vector sequence rich in global causal interaction information. Each vector no longer simply represents the semantics of an event at a certain moment, but integrates the antecedent and consequence information of that event in the entire operation sequence, providing a mathematical basis for accurately extracting causal chains.

[0052] Step 131: Input each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence into the statement embedding mapper, and convert the semantically labeled manipulation statement into a semantic initial vector of a set dimension through word embedding lookup table and multi-layer convolutional coding operation, and assemble all semantic initial vectors into a semantic initial vector sequence in sequence order.

[0053] The statement embedding mapper performs structured encoding on the semantically tagged manipulation statements input. First, the mapper expands the nodes on the statement's dependency syntax tree into a sequence of tokens using a depth-first traversal. Each token corresponds to an operation type code, an operation body code, or a specific normalized parameter value.

[0054] Then, each discrete token is mapped to a dense embedding vector using a word embedding lookup table, where the token's index in the lookup table is determined by its hash value. These embedding vectors are then passed through a convolutional coding module consisting of a one-dimensional convolutional layer, a batch normalization layer, and a rectified linear unit activation layer. The convolutional kernel slides along the token sequence direction to extract local syntactic features.

[0055] After several layers of convolutional encoding operations, the variable-length token features are finally aggregated into a fixed-length feature vector through global average pooling, which is the semantic initial vector of the set dimension. All semantic initial vectors are stacked along a new dimension according to their original order in the sequence, forming a semantic initial vector sequence of shape (sequence length, set dimension).

[0056] Step 132: Apply position encoding to the semantic initial vector sequence, add a position encoding vector generated by sine and cosine functions to each semantic initial vector to obtain a perceptual position semantic vector sequence carrying time position information. Each component value of the position encoding vector is related to the index of the vector dimension and the parity of the element position.

[0057] To explicitly inject order information into the model, this step generates a unique position encoding vector for each position in the sequence. For a position pos in the sequence, the component value of the i-th dimension of the position encoding vector depends on the functional relationship between pos and i.

[0058] Specifically, when i is an even index, this component is calculated using a sine function, the parameters of which are determined by the ratio of the position pos and the dimension index i to the total dimension of the vector; when i is an odd index, this component is calculated using a cosine function, with parameters in the same manner as described above. This generates a position encoding vector with the same dimension as the semantic initial vector. This position encoding vector is then element-wise added to the semantic initial vector at the corresponding position pos, yielding the perceptual position semantic vector carrying temporal position information. The perceptual position semantic vectors for all positions are then stacked along the sequence dimension to form a perceptual position semantic vector sequence.

[0059] Step 133: The perceived location semantic vector sequence is linearly transformed by the first linear projection array, the second linear projection array, and the third linear projection array respectively to generate the query vector sequence, the key vector sequence, and the value vector sequence. The internal parameters of the first linear projection array, the second linear projection array, and the third linear projection array are dynamically adjusted during the context association learning process.

[0060] The first, second, and third linear projection arrays are essentially three independent but structurally identical fully connected linear layers. Each linear projection array consists of a parameter matrix and a bias vector, with the parameter matrix having a size equal to the original vector dimension multiplied by the projection dimension. The perceptual location semantic vector sequence is input to each of these three linear projection arrays. Within each array, the perceptual location semantic vector sequence is multiplied by the parameter matrix, and the result is added to the bias vector. The output of the first linear projection array is the query vector sequence, the output of the second linear projection array is the key vector sequence, and the output of the third linear projection array is the value vector sequence. In each subsequent learning round, the parameter matrix and bias vector within these three vector sequences are dynamically adjusted using the gradient backpropagation algorithm to optimize the accuracy of subsequent attention calculations.

[0061] Step 134: For each query vector in the query vector sequence, perform a dot product operation with all key vectors in the key vector sequence to generate an attention score row. Combine the attention score rows corresponding to all query vectors into an attention score distribution table. The number of rows in the attention score distribution table is equal to the length of the semantic tagging manipulation statement sequence.

[0062] Taking the p-th query vector in the query vector sequence as an example, perform a dot product operation between it and the q-th key vector in the key vector sequence. That is, multiply the components of the two vectors along their corresponding dimensions, and then sum all the products term by term. The resulting scalar value is the original attention score of query vector p and key vector q. This process iterates through all values ​​of q from 1 to the sequence length L, resulting in an attention score row containing L scalars. This iteration is performed for each query vector, ultimately stacking the L attention score rows generated by all query vectors to form an L-row, L-column attention score distribution table. The horizontal axis index of this distribution table represents the source position of the key, and the vertical axis index represents the source position of the query.

[0063] Step 135: Perform a cyclic processing based on weighted feature aggregation according to the attention score distribution table, and use the final layer context understanding output vector sequence that has reached the preset number of cycles as the context association representation vector of each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence.

[0064] The purpose of the above iterative processing is to gradually extract and abstract deep feature vectors that incorporate global causal semantics through multi-layer attention networks and nonlinear transformations.

[0065] Step 1351: Process each element value in the attention score distribution table based on the scaling factor, apply normalized exponential transformation to each row of the scaled attention score distribution table to obtain the attention weight coefficient distribution table. Each row of the attention weight coefficient distribution table represents the attention weight of the semantic tagging manipulation statement at the corresponding query position to all key positions in the sequence. The scaling factor is the reciprocal of the square root of the key vector dimension.

[0066] To prevent the dot product result from being too large and causing the subsequent gradient calculation to be too small, a scaling factor is used to process the attention score distribution table. The scaling factor is calculated as follows: first, obtain the value of the key vector dimension, calculate the square root of the value, and then take the reciprocal of the square root as the scaling factor. Multiply each element value in the attention score distribution table by this scaling factor to reduce the score.

[0067] Subsequently, a normalized exponential transformation is performed on each row of the reduced attention score distribution table. Taking the p-th row as an example, firstly, the L scores in that row are each exponentially calculated to the base e of the natural constant, resulting in L exponential values. Then, the sum of these L exponential values ​​is calculated. Finally, each exponential value in that row is divided by this sum. After transforming all rows, the original distribution table is converted into an attention weight coefficient distribution table. At this point, the sum of the L values ​​in each row of this table is 1, and each value represents the weight contribution of the key position in the corresponding column to the construction of the context vector for the p-th query position.

[0068] Step 1352: Use the attention weight coefficient distribution table to perform weighted summation on the value vector sequence, take the attention weight coefficient of each row as the weight coefficient of the value vector at the corresponding query position, calculate the weighted sum of all value vectors at the query position, and obtain the attention convergence vector at that position.

[0069] For the p-th query position, obtain the p-th row of the attention weight coefficient distribution table, which contains L weight coefficients. Multiply the first weight coefficient element-wise with the value vector of the first position in the value vector sequence to obtain the first weighted vector; multiply the second weight coefficient element-wise with the second value vector to obtain the second weighted vector; and so on, until the L-th position. Summate these L weighted vectors element-wise along their corresponding dimensions; the sum is the attention convergence vector for the p-th query position. This process essentially extracts and aggregates information selectively from the value vectors of the entire sequence based on attention weights.

[0070] Step 1353: Perform an additive residual connection between the attention convergence vector at each location and the corresponding perceptual location semantic vector, and perform layer normalization on the additive result to obtain the intermediate representation vector sequence after residual adjustment.

[0071] To preserve low-level feature information before the original linear projection, the attention convergence vector at position p is element-wise added to its corresponding perceptual location semantic vector at position p. This residual connection facilitates the direct propagation of gradients to shallower layers. Subsequently, layer normalization is performed on all dimensional components of the vector obtained from the element-wise addition. The layer normalization method is as follows: first, the mean and variance of the vector are calculated; then, the mean is subtracted from each component of the vector, divided by the sum of the square root of the variance and a small constant, and finally multiplied by a learnable scaling parameter and added to a learnable translation parameter. This process is performed on vectors at all positions in the sequence to obtain a sequence of intermediate representation vectors after residual adjustment.

[0072] Step 1354: Input the intermediate representation vector sequence into the feedforward mapping structure to transform the intermediate representation vector sequence into a feedforward output vector sequence after nonlinear transformation; the feedforward mapping structure includes two linear mappings and one gated activation operation.

[0073] The feedforward mapping structure acts independently on the vector at each position in the intermediate representation vector sequence. This structure comprises an increasing-dimensional linear layer, a decreasing-dimensional linear layer, and a gated activation unit in between. The increasing-dimensional linear layer maps the intermediate representation vectors to a higher-dimensional space; the decreasing-dimensional linear layer maps the vectors from the higher-dimensional space back to the original dimension.

[0074] The gated activation unit consists of a linear gated path and an activation path, with Gaussian error linear units used as nonlinear activation functions on the activation path. The two output vectors are multiplied element-wise along their corresponding dimensions to achieve a gating effect before being fed into a dimensionality reduction linear layer. After passing through this module, the vector at each position undergoes an independent positional feedforward transformation, outputting a feedforward output vector sequence.

[0075] Step 1355: Perform residual connection and layer normalization processing on the feedforward output vector sequence again to obtain the context understanding output vector sequence of this layer. Then, use the context understanding output vector sequence of this layer as the input of the next isomorphic processing layer to perform the aforementioned attention processing process repeatedly until the preset number of iterations is reached.

[0076] The feedforward output vector sequence is element-wise added to the intermediate representation vector sequence before entering the feedforward structure to form a residual connection. The added result is then subjected to layer normalization again, calculated using the same method as step 1353. The processed vector sequence is the complete context understanding output vector sequence for this layer.

[0077] This application sets a fixed preset number of iterations, for example, 6. It checks if the counter for the currently completed processing levels is less than 6. If so, it uses the generated context understanding output vector sequence as the new input sequence, replacing the original perceptual location semantic vector sequence, and re-enters the linear projection process in step 133 to begin a new round of attention and feature transformation processing. If the count has reached 6, the loop stops.

[0078] Step 1356: Output the final layer context understanding output vector sequence that has reached the preset number of iterations, as the context association representation vector for each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence.

[0079] When the loop count counter reaches a preset total of 6, the final round generates a sequence of contextual understanding output vectors. Each vector within this sequence has undergone multiple rounds of self-attention and non-linear transformations, containing rich long-distance causal semantic interaction information. This entire vector sequence is output, with each vector corresponding to the statement in the same position in the original semantically labeled manipulation statement sequence, and named as the context-related representation vector of that statement.

[0080] Step 140: Identify and extract the causal triggering chain between control action events from the context association representation vector sequence, determine the trigger delay time step between adjacent control action events on the causal triggering chain, and determine the causal association mode based on the causal triggering chain and the trigger delay time step.

[0081] In order to reconstruct concrete event-level causal relationships from the representation space, this step uses discriminative learning and graph search algorithms to extract and structure causal chains.

[0082] Step 141: Extract the context association representation vector sequence, and sequentially use each context association representation vector as a causal source candidate vector according to the time step order, and use the context association representation vectors of all time steps after the causal source candidate vector as causal target candidate vectors.

[0083] This step establishes a candidate pair pool for subsequent causal determination. First, the sequence of context association vectors is loaded into an array structure with indices i from 0 to L-1. A nested loop is constructed to generate candidate pairs. The outer loop iterates through the entire sequence at index i, and the context association vector corresponding to the current index i is defined as the causal source candidate vector. The inner loop iterates from index i+1 to the end of the sequence, and the context association vector corresponding to the current index j is defined as the causal target candidate vector. Each pair (i, j) constitutes a candidate causal relationship pair with temporal constraints, awaiting causal determination.

[0084] Step 142: For each pair of causal source candidate vectors and causal target candidate vectors, concatenate them to generate the concatenated feature vector of the candidate pair. Input the concatenated feature vector into the causal binary classifier, and the causal binary classifier calculates the confidence score that the candidate pair has a causal relationship.

[0085] A causal binary classifier is a pre-trained neural network classifier. Its network structure consists of an input layer, two hidden fully connected layers, and an output layer, with the output layer employing an sigmoid growth curve activation function. When a source vector and a target vector arrive, they are concatenated along the feature dimension to form a concatenated feature vector twice the length of the original vectors. This concatenated feature vector serves as the input to the causal binary classifier, sequentially passing through matrix multiplication, bias addition, and non-linear activation of neurons in each layer. Finally, the output layer outputs a single scalar between 0 and 1, representing the confidence score of the candidate pair indicating a causal relationship. The closer the confidence score is to 1, the higher the probability of a causal relationship.

[0086] Step 143: Determine the causal association confidence score set based on the causal association confidence scores of all candidate pairs, and retain candidate pairs whose confidence scores exceed the preset causal discrimination interface to obtain a preliminary causal association edge set. Each edge in the preliminary causal association edge set connects a source event index and a target event index.

[0087] The confidence scores calculated for all candidate pairs in step 142 are collected into a set to form a causal association confidence score set. Then, a preset causal discrimination interface is applied as a threshold, which can be set to 0.75. All candidate pairs are iterated over; if their confidence score is greater than 0.75, the index tuple (i, j) of that candidate pair is retained as a directed causal edge, pointing from source event i to target event j. All retained edges constitute a preliminary causal association edge set, which can be viewed as a list storage structure of a directed graph.

[0088] Step 144: In the preliminary causal association edge set, take the earliest control action event of the time step as the starting event, and search and connect subsequent events along the directed edges in sequence to extract the complete causal trigger chain from the starting event to the event of the most recent time step. The event indexes on the complete causal trigger chain are arranged in a monotonically increasing order.

[0089] In a sparse directed graph consisting of a preliminary set of causal edges, a specific path search strategy is employed to find complete chains. First, the node with the smallest index *i* among all source events is identified as the starting event. Then, a depth-first search algorithm is used to explore all reachable adjacent nodes starting from this starting node, strictly following the direction of the directed edges. Among all paths found originating from the starting event, the path with the most nodes and the largest index of the endpoint node is selected as the complete causal triggering chain. The event indices corresponding to the nodes in this chain form a strictly monotonically increasing integer sequence.

[0090] Step 145: For each pair of adjacent events in the complete causal triggering chain, obtain the corresponding control action event index, calculate the index difference between the target event index and the source event index, and use the index difference as the trigger delay time step of the adjacent event pair.

[0091] After the complete causal triggering chain is represented as a sequence of node indices derived from the initial event based on causal relationships, a first-order difference operation is performed on the sequence.

[0092] Specifically, along the entire chain, the integer difference obtained by subtracting the index value of the preceding event from the index value of the subsequent event in the sequence represents the relative temporal distance between the two events. This difference is recorded as the trigger delay time step of the adjacent event pair. This time step is not a specific number of physical seconds, but an abstract time share in units of event recording intervals, reflecting the propagation delay from the issuance of an instruction to the generation of a causal effect.

[0093] Step 146: Pair the set of trigger delay time steps of all adjacent event pairs on the complete causal triggering chain with the event index sequence of the complete causal triggering chain to generate a causal triggering chain description structure containing path node information and path edge weight information.

[0094] Construct a causal triggering chain description structure, which is a data container with two main attributes. The first attribute is a list of node paths, whose value is the event index sequence of the complete causal triggering chain; the second attribute is a list of edge weights, whose value is the set of trigger delay time steps of all adjacent event pairs stored in sequence. The topology and strength of this causal relationship are encapsulated through key-value pairing.

[0095] Step 147: Determine whether there are multiple parallel causal triggering chains in the context association representation vector sequence. If so, perform the calculation operation of the trigger delay time step on each causal triggering chain to obtain multiple sets of causal triggering chain description structures.

[0096] In a complete causal discovery process, there may be multiple independent causal transmission paths. This step covers all paths by executing the search algorithm multiple times. In the preliminary causal association edge set generated in step 143, all nodes and edges contained in a complete chain are first marked, and these nodes and edges are temporarily hidden.

[0097] Then, among the remaining unmasked nodes, the node with the smallest number is repositioned as the new starting event, and the chain extraction and delay calculation operations of steps 144 to 146 are repeated to generate a new causal triggering chain description structure. This process is iterated until no new complete chain can be extracted. Finally, all the causal triggering chain description structures generated in the iterative process are merged into a set.

[0098] Step 148: Integrate all the obtained causal trigger chain description structures into the causal association pattern container, and output the causal association pattern container as the organized causal association pattern. Each causal trigger chain description structure in the causal association pattern container has a unique chain identifier.

[0099] The multiple sets of causal trigger chain description structures obtained in step 147 are sequentially imported into a causal association pattern container. This container contains an associative array, where the keys are automatically generated unique chain identifier strings, such as "Chain01" and "Chain02," and the values ​​are the corresponding causal trigger chain description structures. At this point, a causal association pattern containing multiple causal paths, each with refined nodes and propagation delays, is fully determined. This causal association pattern is used as the core logical basis and passed to the subsequent policy generation stage.

[0100] Step 150: Based on the causal relationship pattern, the control action type and action execution time of each time step in the target tunneling period are determined sequentially using a time step recursion strategy to obtain the tunneling parameter control instruction sequence.

[0101] Based on the discovered causal patterns of tunnel boring machine (TBM) operation, this step executes forward-looking control command generation, and iteratively outputs parameters to regulate actions at each time step through a sequence generation model.

[0102] Step 151: parse the causal association pattern, extract the common effect event shared by all causal triggering chains in the causal association pattern, and use the context association representation vector of the common effect event as the guiding target vector for strategy generation.

[0103] A topological overlay analysis is performed on all chains in the causal association pattern container. For each causal triggering chain in the container, the event index at the end of its path and its corresponding event node are extracted, and the intersection of the events at the ends of each chain is calculated. Typically, multiple chains converge on the same end event, such as ultimately leading to the result event "earth pressure instability alarm lifted". This shared event is the common effect event. Using the index of this common effect event, its corresponding vector is located and extracted from the context association representation vector sequence output in step 140. This vector is set as the guiding target vector in the policy generation process to guide the entire generation process toward the final stable goal.

[0104] Step 152: Extract the causal subnet from the earliest triggering event to the common effect event in the causal association pattern. For each event node in the causal subnet, extract the context association representation vector corresponding to the event node and assign a position label vector based on the topological sorting position of the event node in the causal subnet to obtain a set of node representation vectors carrying causal position identifiers.

[0105] Based on the common effect events and the earliest triggering event, forward and backward boundaries are defined in the directed edge set of the causal association pattern to segment causal subnets. The event nodes contained in the subnets constitute an ordered set. For each event node in this set, firstly, its corresponding context association representation vector output in step 140 is extracted; then, based on the node's order number in the subnet topology sorting, a sinusoidal position marker vector is generated, in the same way as in step 132; finally, the context association representation vector and the position marker vector are added element-wise to obtain a node representation vector carrying causal position identifiers. All nodes are processed in this way to form a set of node representation vectors carrying causal position identifiers.

[0106] Step 153: Input the set of node representation vectors carrying causal location identifiers into the message passing layer of the graph neural network to perform several rounds of synchronous message passing, update the node representation vector of each event node, so that each node representation vector absorbs the information of its causal neighboring nodes, and obtain the set of causal reinforced node representation vectors.

[0107] A graph attention network is used as the backbone computational structure of the message passing layer. In each round of message passing in the graph neural network, for any event node u in the subnet, all its neighbor nodes v are collected. For each pair (u, v), an attention coefficient is calculated, which is obtained by passing the current representation vectors of node u and node v through a single-layer feedforward network and an activation function. Subsequently, the representation vectors of all neighbor nodes v are weighted and summed using this coefficient to form an aggregated message from the neighbors. The representation vector of node u itself is concatenated with the aggregated message, and then processed through a linear transformation layer to obtain the updated representation vector of node u in this round. After a set number of rounds, such as 3 rounds of synchronous message passing, the representation vectors of all nodes have deeply absorbed the causal upstream and downstream information, and this output set is the causal reinforcement node representation vector set.

[0108] Step 154: Extract the vectors corresponding to the common effect events from the set of causal reinforcement node representation vectors and perform vector gating fusion with the guiding target vector to generate the policy initiation vector.

[0109] In the set of causal reinforcement node representation vectors, the reinforced vector corresponding to the common effect event node is located. The guiding target vector obtained in step 151 and the above-mentioned reinforced vector are input into a gating fusion module. The gating fusion module first calculates a gating coefficient, which is generated by concatenating the guiding target vector and the reinforced vector and passing it through a fully connected layer with an S-shaped growth curve activation function. Its value range is between 0 and 1. This gating coefficient is used to dynamically balance the proportion of the two information sources. The fusion process is as follows: the guiding target vector is multiplied by the gating coefficient, the reinforced vector is multiplied by (1 - the gating coefficient), and then the two results are added element by element. The sum is the policy initiation vector, which carries both the structural information of the historical causal subnet and reinforces the directional guidance of the final control target.

[0110] Step 155: Perform loop decoding processing based on the policy initiation vector and the fully connected network to obtain multiple instruction tuples, and perform serialization processing based on the instruction tuples to obtain a parameter control instruction sequence. The fully connected network includes a fully connected classifier, a fully connected regressor, and a fully connected interval predictor.

[0111] The loop decoding process is an autoregressive time series generation process. The policy initiation vector serves as the initial state of the entire generator, controlling the operation instructions to be executed at each subsequent time step.

[0112] Step 1551: After setting the policy initiation vector as the initial hidden state of the autoregressive policy decoding process, the decoding operation is performed step by step. At each decoding time step, the current hidden state is used to perform consistent attention matching with the set of causal reinforcement node representation vectors to generate a context instruction summary vector.

[0113] At the start of time step k, the hidden state vector from the previous step is held. The similarity score between this hidden state vector and each vector in the causal reinforcement node representation vector set output from step 153 is calculated; this score is the dot product of the two. These scores are converted into weights using a normalized exponential transform, and these weights are used to perform a weighted summation of all vectors in the causal reinforcement node representation vector set. The resulting weighted sum vector is the context instruction summary vector. This process allows each decoding step to dynamically focus on historical event information in the causal subnet most relevant to the current decision.

[0114] Step 1552: Based on the context instruction summary vector and the current hidden state, the control action type of the current time step is output by the fully connected classifier, the execution value of the control action of the current time step is output by the fully connected regressor, and the time span from the execution of the current time step action to the start of the next action is output by the fully connected interval predictor.

[0115] The context instruction summary vector and the current hidden state vector are concatenated along the dimensional direction, and the concatenation result is simultaneously fed into three parallel fully connected network branches.

[0116] The first branch is a fully connected classifier, whose output layer has the same number of neurons as the total number of all possible control action types. It is activated by a normalized exponential transformation and outputs a probability distribution. The index corresponding to the maximum probability is the control action type at the current time step k.

[0117] The second branch is a fully connected regressor, whose output layer is a single linear neuron that directly outputs a floating-point number representing the amount of the modulated action to be performed. The normalized range of this amount corresponds one-to-one with the normalization operation in step 124.

[0118] The third branch is a fully connected interval predictor, whose output layer is also a single neuron. It uses exponential linear unit activation to ensure that the output is positive, which represents the time span from the completion of the current action to the start of the next action.

[0119] Step 1553: Determine whether the type of control action generated in the current time step is a predefined termination action type. If yes, terminate the decoding process. If no, combine the type of control action generated in the current time step, the value of the control action execution, and the time span into an instruction tuple and store it in the instruction result list. Then, use the action generated in the current time step to embed and update the hidden state, and continue to the next decoding time step.

[0120] The globally unique index corresponding to the control action type output in step 1552 is compared with the predefined termination action index. If they are equal, a termination signal is sent to the decoding loop controller, and the decoding loop is exited. If they are not equal, the control action type index, the floating-point value of the control action execution amount, and the floating-point value of the time span generated in this time step k are packaged into a triplet to form an instruction tuple, which is then appended to the end of the instruction result list. Subsequently, to update the hidden state, the control action type index is first mapped to a dense embedding vector through an action embedding lookup table. Then, the embedding vector is added element-wise to the context instruction summary vector, and the result is used as the updated hidden state and passed to time step k+1.

[0121] Step 1554: After the decoding loop terminates, extract all instruction tuples sequentially from the instruction result list, determine the absolute execution time of each instruction tuple based on the time span within each instruction tuple, and obtain a sequence of instruction tuples carrying absolute execution time markers. Reorder the sequence of instruction tuples carrying absolute execution time markers according to the order of absolute execution times so that the instruction timeline monotonically increases.

[0122] After the decoding loop completes, the instruction result list stores instruction tuples arranged in the order of generation. Setting the starting absolute execution time as base 0, the instruction tuples in the instruction result list are traversed sequentially. For the m-th tuple, the time spans of all the preceding m-1 tuples are summed; the sum is the absolute execution time of the m-th tuple. Each tuple is bound to its calculated absolute execution time, forming a sequence of instruction tuples labeled with their absolute execution times. This sequence is then sorted in ascending order according to the values ​​of the absolute execution times to ensure the monotonicity and executability of the instruction timeline.

[0123] Step 1555: The sorted instruction tuple sequence is formatted and encapsulated. The control action type field, control action execution value field, and execution time field are serialized into structured instruction data blocks and used as the tunneling parameter control instruction sequence.

[0124] For each sorted instruction tuple, its control action type index, control action execution value (floating-point), and absolute execution time are extracted and encoded into a structured instruction data block. This data block uses a compact key-value pair format and contains three fields: `act_type` stores the integer type code, `act_val` stores the denormalized physical quantity value, and `act_ts` stores the millisecond-level timestamp. All structured instruction data blocks are packaged into a list in execution time order. This list is the final output sequence of tunneling parameter control instructions, which can be directly parsed and executed by the lower-level machine.

[0125] After obtaining the tunneling parameter control command sequence, the method further includes the step of compiling the tunneling parameter control command sequence into mechanical control pulse codes and performing steady-state compensation. To further ensure the physical compatibility and timing immunity of the commands at the execution level, the following pulse compilation and compensation steps are included: Step 210: Input the tunneling parameter control instruction sequence into the instruction compiler mapper, and convert the control action type code and control action execution value corresponding to each control instruction into mechanical control pulse code according to the time step sequence, so as to generate a mechanical control pulse code sequence that is time-aligned with the tunneling parameter control instruction sequence.

[0126] The instruction compiler mapper internally maintains a pulse protocol mapping table provided by the electronic control unit manufacturer. The key of the mapping table is the control action type code, and the value is a string of control bytecode structures required for that action. For each structured instruction data block in the tunneling parameter control instruction sequence, the mapper first finds the corresponding control bytecode template in the mapping table based on the `act_type` field, then fills the parameter value field of the template with the value of the `act_val` field, and performs additional calculations for cyclic redundancy check (CRC) codes, ultimately generating a complete mechanical control pulse code. All pulse codes are arranged in time-aligned sequence to form the mechanical control pulse code sequence.

[0127] Step 220: Obtain the causal triggering chain description structure corresponding to the causal association pattern, extract the trigger delay time step of each adjacent event pair in the causal triggering chain description structure, establish a delay-sensitive memory array with the trigger delay time step as the index dimension, read the mechanical control pulse code sequence bit by bit into the delay-sensitive memory array, so that each mechanical control pulse code is assigned to the corresponding index unit of the delay-sensitive memory array according to its corresponding time step.

[0128] Select the core chain used for decision-making from the causal relationship pattern container, and read the trigger delay time steps from its edge weight list. Construct a delay-sensitive memory array with a total capacity of the maximum delay step plus one, where each index cell can store a pulse code. Read each code of the mechanical control pulse code sequence in ascending order of time steps and store it in the array cell indexed by the time step number corresponding to that code. This allocation process continues until all pulse codes are placed in their corresponding delay slots.

[0129] Step 230: In the delay-sensitive memory array, taking the index unit corresponding to the starting event on the causal triggering chain as the starting point for reading, mechanical control pulse codes are extracted sequentially from each index unit along the path direction of the causal triggering chain, and spliced ​​into a causal orchestration pulse chain according to the extraction order.

[0130] The list of node paths in the causal triggering chain is used as the read address sequence. First, the original time step index corresponding to the starting event node is located. Then, according to the chronological order of events on the chain path, the corresponding index cell is addressed one by one in the delay-sensitive memory array. The mechanical control pulse code within that cell is read and appended to the end of a new pulse chain. The reading process skips redundant pulses that are not causally related in the middle, directly forming a concise causal orchestration pulse chain.

[0131] Step 240: Input the causal orchestration pulse chain into a multidimensional state comparator, parse the difference in action value and action time interval between adjacent pulse codes from the causal orchestration pulse chain, generate a pulse execution deviation distribution, and perform a logical comparison between the pulse execution deviation distribution and a preset steady-state deviation threshold range. When the pulse execution deviation distribution exceeds the preset steady-state deviation threshold range, generate a compensation pulse amplitude adjustment amount, and superimpose the compensation pulse amplitude adjustment amount into the mechanical control pulse code of the corresponding index unit to obtain the steady-state compensated causal orchestration pulse chain.

[0132] The multidimensional state comparator is an analytical digital filter. For two adjacent pulse codes A and B in a causal pulse train, their respective `act_val` values ​​are decoded and their absolute differences are calculated; this represents the action magnitude difference. Their respective `act_ts` values ​​are also decoded and their time differences are calculated; this represents the action time interval difference. These two are combined into a two-dimensional deviation vector. The two-dimensional deviation vectors of all adjacent pairs constitute the pulse execution deviation distribution. A preset steady-state deviation threshold range is set as a two-dimensional rectangular restricted area.

[0133] If any component of the deviation vector of a pair of pulses falls outside the forbidden zone of the corresponding dimension, compensation calculation is initiated: the excess of the action magnitude difference is multiplied by a proportional gain factor to obtain the compensation pulse amplitude adjustment. Backtracking to the corresponding index cell of the delay-sensitive memory array, the value of the `act_val` field of the pulse code in that cell is added to the compensation pulse amplitude adjustment to achieve steady-state compensation of the pulse amplitude. Finally, the pulse chain is re-extracted from the adjusted array according to step 230, thus obtaining the steady-state compensated causal orchestrated pulse chain.

[0134] After obtaining the sequence of tunneling parameter control commands, the method further includes steps of causal matching measurement and pattern correction of the command sequence. To ensure the causal matching quality between the output control commands and the original semantic events, and to achieve dynamic correction of causal correlation patterns, the method further includes causal consistency assessment and pattern write-back operations. Step 310: Align the sequence of tunneling parameter control instructions with the sequence of semantically tagged manipulation statements on the time axis to establish a one-to-one mapping relationship between the control instructions and the semantically tagged manipulation statements, and generate a set of instruction semantic pairing records.

[0135] The `act_ts` field of each instruction in the tunneling parameter control instruction sequence is extracted as the alignment key. Simultaneously, the trigger timestamp of the control action attached to each statement in the semantically tagged manipulation statement sequence is extracted as the alignment key. The two sequences are merged and aligned based on the time key. Within the allowed time tolerance window, a one-to-one mapping relationship is established for matching instructions and statements, i.e., a semantic pairing record for each instruction. These records are collected to generate an instruction semantic pairing record set.

[0136] Step 320: Using the context association representation vector corresponding to the common effect event in the context association representation vector sequence as the anchor point, extract the target control instruction that has a causal dependency on the anchor point in the instruction semantic pairing record set. Combine the operation value description column in the semantically labeled manipulation statement corresponding to the target control instruction with the control action execution value in the target control instruction to form a causal matching evaluation sample.

[0137] Search the instruction semantic pairing record set to find those pairings whose event indexes for the corresponding semantically tagged manipulation statements belong to the causal subnet of the common-effect events identified in step 151. Identify these found pairing records as targets. For each identified target, extract the operand value description column from its statement and the act_val value from the instruction, and concatenate these two to form a causal matching evaluation sample.

[0138] Step 330: Input the causal matching evaluation sample into the causal matching measurer. The causal matching measurer projects the operation value description column and the control action execution value into the same latent space through vector space mapping and calculates the directional deviation angle between the two in the latent space to generate the causal matching deviation index.

[0139] The causal matching measurer consists of a two-channel projection network. The operand description columns and the control action execution columns are projected upwards through two different fully connected layers to a 64-dimensional latent space, resulting in projection vectors A and B. The dot product of A and B is calculated; simultaneously, the magnitudes of A and B are calculated, which are the square roots of the sum of the squares of their respective components. Dividing the dot product by the product of the magnitudes yields the cosine similarity. The cosine similarity result is converted to radians using an inverse cosine transform; this angle is the directional deviation angle. After inverse normalization, the deviation angle becomes the causal matching deviation index; the larger the index, the higher the degree of causal logical mismatch.

[0140] Step 340: Adjust the causal association pattern backtrackingly according to the causal matching deviation index, select the causal triggering chain description structure associated with the anchor point from the causal association pattern container, correct the trigger delay time step of adjacent event pairs in the causal triggering chain description structure according to the direction of the causal matching deviation index, obtain the causal pattern adjustment scheme, and write the causal pattern adjustment scheme back to the causal association pattern container to update the causal association pattern.

[0141] When the causal matching deviation index exceeds a preset tolerance limit for several consecutive periods, a backtracking adjustment is triggered. The causal triggering chain description structure containing the common effect events is retrieved from the causal association pattern container. Based on the positive or negative directionality of the deviation index, it is determined whether the trigger delay should be extended or shortened. The adjustment mechanism is as follows: the original trigger delay time step *r* is multiplied by an adjustment factor negatively correlated with the deviation index. The corrected *r'* replaces the original *r*, forming the causal pattern adjustment scheme. Finally, the new edge weights in this adjustment scheme are written back to the corresponding chain description structure in the container, completing the online update of the causal association pattern.

[0142] In addition, embodiments of this application also provide specific algorithm types, training processes, and parameter configurations for sentence embedding mappers, feedforward mapping structures, causal binary classifiers, fully connected classifiers, fully connected regressors, fully connected interval predictors, compiler mappers, multidimensional state comparators, and causal matching metrics.

[0143] The sentence embedding mapper employs a semantic encoding structure based on token embedding and a cascaded one-dimensional convolutional neural network. Its main body consists of an embedding lookup layer, three one-dimensional convolutional modules, and a global average pooling layer stacked sequentially. The embedding lookup layer is a trainable weight matrix with rows equal to the vocabulary size and 256 columns, responsible for mapping discrete token indices to 256-dimensional dense vectors. Each of the three one-dimensional convolutional modules contains a convolutional layer with a kernel width of 3 and a stride of 1, a batch normalization layer, and a modified linear unit activation layer. The number of output channels for the three convolutional layers increases sequentially to 128, 256, and 512, respectively, to extract syntactic patterns from local to global levels layer by layer. During training, a corpus containing 500,000 labeled tunnel boring machine operation texts and their corresponding dependency syntax trees is used. The training objective is to reconstruct the parent-child node relationships in the dependency syntax trees. A contrastive loss function is used for optimization, and an adaptive moment estimation optimizer is selected. The initial learning rate is set to 0.0005, the batch size is 64, and the training cycle is 30 rounds. After training, the embedding lookup table and convolutional layer parameters are frozen and fixed, and the 512-dimensional vector after global average pooling is used as the output to form the semantic initial vector.

[0144] The feedforward mapping structure employs a positional feedforward network with a gated linear unit mechanism. This network, located after the self-attention sub-layer, contains two parallel linear projection branches. The first branch is the activation branch, which first expands the input vector dimension from 256 to 1024 using an upscaling linear layer, activates it using a Gaussian error linear unit function, and then compresses the 1024-dimensional vector back to 256 dimensions using a downscaling linear layer. The second branch is the gated branch, consisting of another set of independently parameterized upscaling and downscaling linear layers, but without using an activation function; the output is only used to form the gate signal. The downscaling results of the two branches are multiplied element-wise along their corresponding dimensions to achieve gate control. The training of this structure is embedded in the global context association learning task, sharing the same loss function and optimizer with the self-attention layer. The loss function is a labeled smoothed cross-entropy loss with a smoothing factor of 0.1, and the optimizer is an adaptive moment estimator with a batch size of 128. After training, the weight matrices and bias vectors of the up-dimensional linear layer and the down-dimensional linear layer are retained, and the parameter values ​​are distributed between -0.1 and 0.1, exhibiting an overall zero-mean normal distribution.

[0145] The causal binary classification discriminator employs a multilayer perceptron-based binary classification structure. The network input layer receives a 512-dimensional vector composed of concatenated causal source candidate vectors and causal target candidate vectors, which is then sequentially connected to two hidden fully connected layers and an output layer. The first hidden layer contains 256 neurons activated by Gaussian error linear units; the second hidden layer contains 128 neurons, also activated by Gaussian error linear units; the output layer has only one neuron activated by an sigmoid growth curve function, outputting a causal association confidence score. During training, positive and negative causal pairs labeled by experts are extracted from the tunnel boring machine's historical operation logs, with a positive-to-negative sample ratio of 1:3, for a total of 120,000 pairs. Training uses binary cross-entropy as the loss function, employs an adaptive moment estimate optimizer, a learning rate of 0.0001, a batch size of 256, and 20 training epochs, with a weight decay coefficient of 0.0001 for regularization. After training, the model possesses the ability to discriminate causal associations, and all connection weights and biases are fixed as inference parameters.

[0146] The fully connected classifier is a classification network that outputs the probability distribution of actions. Its input is a 512-dimensional vector obtained by concatenating the context instruction summary vector and the current hidden state vector. The network structure consists of two fully connected layers. The first layer has 256 neurons activated using modified linear units (MLUs), and the second layer is the output layer with the number of neurons equal to the total number of controllable action types, preset to 48, and an applied normalized exponential transform function to output the probability of each category. The training process is integrated with the decoder training. The loss function is cross-entropy loss. The sample distribution of all controllable action types is processed by class weight balancing. The optimizer is an adaptive moment estimator (AME), with a learning rate of 0.0003, a batch size of 64, and a training cycle of 40 epochs. After training, the classifier's weight matrix and bias vector are stored as fixed parameter configurations.

[0147] The fully connected regressor is a single-output regression prediction network. Its network structure is the same as the front end of the fully connected classifier. The input is a 512-dimensional vector, which passes through a linear modified activation layer with 256 neurons and is then connected to a linear neuron without an activation function. The output is a continuous floating-point value representing the magnitude of the control action. During training, mean squared error is used as the loss function, and the target value is the parameter value after normalization in step 124. The optimizer uses an adaptive moment estimator with a learning rate of 0.0002 and a batch size of 64, and is trained synchronously with the fully connected classifier for 40 epochs. After training, the network weights are retained, and the output values ​​are concentrated within a preset numerical range, which can be used as direct predictions of the magnitude.

[0148] The fully connected interval predictor is a regression network responsible for predicting time intervals. Its network structure is similar to that of a fully connected regressor, but the output layer activation uses an exponential linear unit function to ensure that the predicted time span is positive. The training objective is the true time span of the instruction tuple, the loss function is the smoothed absolute error loss, the optimizer is an adaptive moment estimator, the learning rate is 0.0002, the batch size is 64, and the training epochs are 40. After training, the network outputs a positive floating-point number, deltak, representing the execution interval share between adjacent control actions.

[0149] The compiler mapper employs a rule-based compilation engine based on a hash mapping protocol table. Its core is a statically maintained key-value mapping directory. This directory is constructed by the tunnel boring machine's electronic control unit designers based on the pulse width modulation protocol and bus communication protocol, using the control action type code as the key and a predefined mechanical control pulse code template as the value. The template reserves placeholder fields for parameters and checksum calculation bits. The compiler mapper does not involve trainable parameters; its accuracy is guaranteed by protocol integration and real-machine verification during the design phase. In the configured state, there are a total of 48 control action type codes, each corresponding to a binary pulse template with a length between 4 and 20 bytes. The set of all templates is stored in the mapper's read-only memory.

[0150] The multidimensional state comparator employs a discrete signal processing structure combining digital filtering and threshold discrimination, and does not contain trainable model parameters. Its core is a multidimensional threshold comparison logic unit, internally pre-defined as a steady-state deviation threshold range consisting of the difference in action magnitude and the difference in action time interval. This threshold range is determined based on the 99% confidence interval of hydraulic system step response tests and field trial tunneling data. Internally, the comparator uses a decoder to extract the `act_val` and `act_ts` fields from the pulse-coded stream, calculates the difference vector between adjacent pulse pairs, and then compares each component of the difference vector with the threshold interval dimension by dimension. Based on the comparison results, it triggers the compensation calculation logic. The compensation calculation uses a fixed proportional gain factor, which is an engineering constant set after modeling and analyzing the transfer functions of the propulsion cylinder and cutterhead drive system.

[0151] The causal matching measurer employs a semantic alignment metric structure based on a Siamese projection network. It consists of two structurally identical but parameter-independent fully connected projection branches. Each branch contains a linear transformation layer and a linear unit activation layer with leakage correction, projecting the input operand value description vector and the control action execution vector into a 64-dimensional latent space. In the latent space, the direction deviation angle is obtained as the causal matching deviation index by calculating the cosine of the angle between the two projection vectors and converting it to an inverse cosine angle. During training, 10,000 correctly matched samples and 2,000 randomly mismatched samples extracted from the causal triggering chain are used for comparative learning training. The loss function consists of two parts: a cosine distance minimization term for correctly matched pairs and an interval ranking loss term for mismatched pairs. The optimizer uses an adaptive moment estimation optimizer with a learning rate of 0.0001 and a training period of 15 epochs. After training, the weights of the two projection branches are fixed, giving the measurer the ability to reliably evaluate semantic consistency.

[0152] This application reconstructs discrete time-series sensor records from the tunnel boring machine (TBM) excavation process into a sequence of semantically labeled manipulation statements with syntactic structure through semantic tagging and dependency pairing. This allows for a unified formal expression of the semantics of physical operations. Then, through attention-based contextual learning, a contextual representation vector incorporating global causal information is generated for each semantically labeled manipulation statement, completing the mapping from isolated event records to a causal semantic space. Furthermore, causal trigger chains between control action events are identified and extracted from this semantic space, and combined with trigger delay time steps to form a structured causal relationship pattern, achieving explicit modeling of the inherent causal temporal dynamics of tunneling control. Finally, based on the causal relationship pattern, a time-step recursive strategy is used to generate a sequence of tunneling parameter control commands, enabling the parameter control process to shift from a passive response to an active adherence to causal laws, improving the foresight of the tunneling parameter control command sequence and the stability of the tunneling state.

[0153] In practical implementation, known technologies such as real-time data acquisition and ring buffer management in industrial control systems can be combined to address signal interference issues such as missing top-level delimiters in record rows. This can be achieved by configuring hardware filtering parameters and buffer verification mechanisms on the programmable logic controller (PLC) side to ensure complete separation between the control action event identifier field and the tunneling status monitoring parameter field. For dimensional normalization, the standard ranges and typical working intervals of each sensor channel in the tunnel boring machine's factory parameter manual and engineering construction history database can be referenced to establish initial configurations for the mean offset vector and variance scaling vector, achieving a unified mapping of different physical dimensions to preset numerical intervals. Regarding the construction of the control action ontology knowledge base, based on the tunnel boring machine component assembly drawings and operating procedure documents, the node and edge relationships of the operation subject concept identifier and operation type concept identifier can be predefined. An incremental graph database update mechanism can then continuously write the mapping relationship between the newly added control action event index number and the globally unique identifier into the index mapping table, ensuring the completeness of retrieval and matching in the semantically tagged manipulation statement generation process.

[0154] Furthermore, those skilled in the art can engineer the linear projection array parameters and position encoding schemes involved in the context association learning process based on the mature training framework of the Transformer model. Specifically, a conventional configuration of a deep self-attention network can be adopted, with a preset number of loops of 6 to 8 layers. Continuous temporal annotations of operation sequences in the shield tunneling history records can be used to pre-train the attention weights and the layer normalization parameters after residual connections. The training of the causal binary classifier can be achieved by manually extracting positive and negative causal event pairs from typical shield tunneling accident analysis reports and normal tunneling logs, obtaining the confidence score discrimination interface through supervised learning. For the synchronous autoregressive training of the fully connected classifier, regressor, and interval predictor, a standard teacher-forced strategy for sequence-to-sequence models can be adopted. Using the instruction sequences of the actual tunneling cycle as labels, the joint loss function of classification cross-entropy, mean squared error, and smoothing absolute error can be optimized to ensure the quality of instruction tuple generation and the physical feasibility of time span prediction.

[0155] Furthermore, by combining the fieldbus communication protocol specification and pulse width modulation standard, a pulse protocol mapping table for the compiler mapper can be constructed, mapping the 48 control action type codes to control bytecode templates recognizable by the tunnel boring machine's electronic control unit. The steady-state deviation threshold range of the action value difference and action time interval difference in the multi-dimensional state comparator can be determined by measuring the step response time of each hydraulic circuit and drive motor. The projection network training of the causal matching metric can be optimized based on ideal causal instruction-semantic pairing and random disturbance mismatch data generated by the construction simulation system, using cosine distance minimization and interval sorting contrastive loss. For the trigger delay time step, the adjustment factor in the causal mode adjustment scheme can be calibrated by combining the measured hysteresis characteristic curves of each actuator of the tunnel boring machine from receiving the action to its completion, realizing dynamic correction of causal logic relationships and container write-back updates.

[0156] See Figure 2 As shown in the figure, this is a schematic diagram of the basic structure of a shield tunneling parameter adaptive control system 20 provided in an embodiment of this application. The shield tunneling parameter adaptive control system 20 includes: Processor 201; Storage device 202, on which computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the aforementioned intelligent control-based adaptive adjustment methods for shield tunneling parameters.

[0157] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0158] See Figure 3 As shown in the figure, this is a functional block diagram of a shield tunneling parameter adaptive control device provided in an embodiment of this application. The shield tunneling parameter adaptive control device includes: The tunneling data acquisition module is used to acquire and cache the tunneling time sequence data collected and cached by the sensing acquisition device in chronological order during the tunneling construction process; wherein, each control action event record line recorded in the tunneling time sequence data includes a control action event identifier field and a tunneling status monitoring parameter field, and the tunneling status monitoring parameter field stores the tunneling status monitoring parameter group synchronously captured at the time of occurrence of the corresponding control action event; The semantic tagging processing module is used to generate semantically tagged manipulation statements based on the control action event identifier field and the tunneling status monitoring parameter field corresponding to each control action event record line in the shield tunneling time series data, through semantic tagging processing and dependency relationship pairing; wherein, all the semantically tagged manipulation statements are arranged in the original time order of the control action event record lines to form a semantically tagged manipulation statement sequence. The context association learning processing module is used to perform attention-based context association learning processing on the semantically labeled manipulation statement sequence, generate attention weight coefficients between any two semantically labeled manipulation statements in the semantically labeled manipulation statement sequence, and use the attention weight coefficients to perform weighted feature aggregation on the semantically labeled manipulation statement sequence to generate a context association representation vector that integrates global causal association information for each semantically labeled manipulation statement. The association pattern determination module is used to identify and extract the causal triggering chain between control action events from the context association representation vector sequence, determine the trigger delay time step between adjacent control action events on the causal triggering chain, and determine the causal association pattern based on the causal triggering chain and the trigger delay time step. The control instruction generation module is used to determine the control action type and action execution time of each time step in the target tunneling period in sequence according to the causal relationship pattern and the time step recursion strategy, so as to obtain the tunneling parameter control instruction sequence.

[0159] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0160] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0161] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0162] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0163] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for adaptive adjustment of shield tunneling parameters based on intelligent control, characterized in that, include: Acquire shield tunneling time sequence data collected and cached by a sensor acquisition device in chronological order during the shield tunneling construction process; wherein, each control action event record line recorded sequentially in the shield tunneling time sequence data includes a control action event identifier field and a tunneling status monitoring parameter field, wherein the tunneling status monitoring parameter field stores a group of tunneling status monitoring parameters that are synchronously captured at the time of occurrence of the corresponding control action event; Based on the control action event identifier field and tunneling status monitoring parameter field corresponding to each control action event record line in the shield tunneling time series data, semantically tagged manipulation statements are generated through semantic tagging processing and dependency relationship pairing; wherein, all the semantically tagged manipulation statements are arranged in the original time order of the control action event record lines to form a semantically tagged manipulation statement sequence. The semantically labeled manipulation statement sequence is subjected to context association learning processing based on attention mechanism to generate attention weight coefficients between any two semantically labeled manipulation statements in the semantically labeled manipulation statement sequence. The attention weight coefficients are then used to perform weighted feature aggregation on the semantically labeled manipulation statement sequence to generate a context association representation vector that integrates global causal association information for each semantically labeled manipulation statement. The causal triggering chain between control action events is identified and extracted from the context association representation vector sequence, and the trigger delay time step between adjacent control action events on the causal triggering chain is determined. The causal association mode is determined based on the causal triggering chain and the trigger delay time step. Based on the aforementioned causal relationship pattern, a time-step recursive strategy is used to sequentially determine the control action type and execution time of each time step in the target tunneling period, thereby obtaining a sequence of tunneling parameter control instructions.

2. The method according to claim 1, characterized in that, The method, based on the control action event identifier field and tunneling status monitoring parameter field corresponding to each control action event record line in the shield tunneling time series data, generates semantically tagged manipulation statements through semantic tagging processing and dependency relationship pairing, including: Read the control action event record line by line from the shield tunneling time sequence data, and perform field segmentation processing on the read control action event record line to separate the control action event identifier field and the tunneling status monitoring parameter field; Lexical analysis is performed on the control action event identifier field, and predefined delimiters within the field are scanned to extract the control action trigger timestamp and control action event index number; Using the index number of the control action event, a search and matching is performed in the control action ontology knowledge base to obtain the concept identifier of the operation subject and the concept identifier of the operation type involved in the control action event. The concept identifier of the operation subject is determined as the operation subject code, and the concept identifier of the operation type is determined as the operation type code. The tunneling status monitoring parameter fields are subjected to numerical extraction and dimensional normalization to generate a normalized tunneling status monitoring parameter group. Each physical quantity value in the normalized tunneling status monitoring parameter group is mapped to a preset numerical range. Input the operation type code into the value dimension selector to obtain a value dimension configuration list that matches the operation type code. The value dimension configuration list includes the value dimension name and dimension order that the operation type code needs to describe. According to the quantity dimension configuration list, the parameter values ​​corresponding to the quantity dimension are extracted from the normalized tunneling status monitoring parameter group, and arranged in the dimension order to generate an operation quantity description column, which is an ordered quantity description tuple; Combine the standardized dependency relationship template corresponding to the operation type code to perform dependency relationship analysis between the operation type code and the operation value description column, generate semantically labeled manipulation statements, and combine the semantically labeled manipulation statements into a semantically labeled manipulation statement sequence.

3. The method according to claim 2, characterized in that, The process involves performing dependency analysis between the operation type code and the operation value description column using the standardized dependency template corresponding to the operation type code, generating semantically labeled manipulation statements, and combining these statements into a semantically labeled manipulation statement sequence, including: Determine the standardized dependency relation template corresponding to the operation type code. The standardized dependency relation template includes the syntactic position of the operation type code as the dominant and the syntactic position of each value dimension in the operation value description column as the dominated, as well as the dependency relation type marker between the dominant and the dominated. According to the dependency relationship type marker specified in the standardized dependency relationship template, the operation type code and the corresponding value dimension in the operation value description column are established one by one with the dependency relationship type marker, and the operation subject code is attached to the operation type code as the topic label of the whole statement to obtain the semantically labeled manipulation statement with a complete semantic dependency structure. The semantically tagged manipulation statement generated by the current control action event record line is temporarily stored, and it is determined whether there is a next control action event record line. If it exists, it returns to continue reading; otherwise, it executes the next sub-step. Arrange all the temporarily stored semantically tagged manipulation statements in chronological order according to the timestamps of their corresponding control actions to generate the semantically tagged manipulation statement sequence.

4. The method according to claim 1, characterized in that, The process involves performing attention-based context association learning on the semantically labeled manipulation statement sequence to generate attention weight coefficients between any two semantically labeled manipulation statements in the sequence. These attention weight coefficients are then used to perform weighted feature aggregation on the semantically labeled manipulation statement sequence, generating a context association representation vector for each semantically labeled manipulation statement that incorporates global causal association information. This includes: Each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence is input into the statement embedding mapper. The semantically labeled manipulation statement is converted into a semantic initial vector of a set dimension through word embedding lookup table and multi-layer convolutional coding operation. All semantic initial vectors are arranged in sequence to form a semantic initial vector sequence. Position encoding is applied to the semantic initial vector sequence, and a position encoding vector generated by sine and cosine functions is added to each semantic initial vector to obtain a perceptual position semantic vector sequence carrying time position information. Each component value of the position encoding vector is related to the index of the vector dimension and the parity of the element position. The perceived location semantic vector sequence is linearly transformed by the first linear projection array, the second linear projection array and the third linear projection array respectively, to generate a query vector sequence, a key vector sequence and a value vector sequence. The internal parameters of the first linear projection array, the second linear projection array and the third linear projection array are dynamically adjusted during the context association learning process. For each query vector in the query vector sequence, perform a dot product operation with all key vectors in the key vector sequence to generate an attention score row. Combine the attention score rows corresponding to all query vectors into an attention score distribution table. The number of rows in the attention score distribution table is equal to the length of the semantic tagging manipulation statement sequence. Based on the attention score distribution table, a weighted feature aggregation-based iterative process is performed, and the final layer context understanding output vector sequence that has reached the preset number of iterations is used as the context association representation vector of each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence.

5. The method according to claim 4, characterized in that, The step of performing weighted feature aggregation-based iterative processing according to the attention score distribution table, and using the final layer context understanding output vector sequence after reaching a preset number of iterations as the context association representation vector for each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence, includes: The attention score distribution table is processed based on the scaling factor. A normalized exponential transformation is applied to each row of the scaled attention score distribution table to obtain the attention weight coefficient distribution table. Each row of the attention weight coefficient distribution table represents the attention weight of the semantic tagging manipulation statement at the corresponding query position to all key positions in the sequence. The scaling factor is the reciprocal of the square root of the key vector dimension. The value vector sequence is weighted and accumulated using the attention weight coefficient distribution table. The attention weight coefficient of each row is used as the weight coefficient of the value vector at the corresponding query position. The weighted sum of all value vectors at the query position is calculated to obtain the attention convergence vector at that position. The attention convergence vector at each location is added to the corresponding perceptual location semantic vector, and the residual connection is performed. The addition result is then subjected to layer normalization to obtain the intermediate representation vector sequence after residual adjustment. The intermediate representation vector sequence is input into the feedforward mapping structure, which transforms the intermediate representation vector sequence into a nonlinearly transformed feedforward output vector sequence; the feedforward mapping structure includes two linear mappings and one gated activation operation; The feedforward output vector sequence is subjected to residual connection and layer normalization processing again to obtain the current layer context understanding output vector sequence. The current layer context understanding output vector sequence is used as the input of the next level isomorphic processing layer to perform the aforementioned attention processing process repeatedly until the preset number of iterations is reached. The final layer context understanding output vector sequence, which has reached the preset number of iterations, is output as the context association representation vector for each semantically labeled manipulation statement in the semantically labeled manipulation statement sequence.

6. The method according to claim 1, characterized in that, The step of identifying and extracting causal triggering chains between control action events from the context association representation vector sequence, determining the trigger delay time step between adjacent control action events on the causal triggering chain, and determining the causal association pattern based on the causal triggering chain and the trigger delay time step includes: Extract the context association representation vector sequence, and sequentially use each context association representation vector as a causal source candidate vector according to the time step order. Then, use the context association representation vectors of all time steps after the causal source candidate vector as causal target candidate vectors. For each candidate pair, the causal source candidate vector and the causal target candidate vector are concatenated to generate the concatenated feature vector of the candidate pair. The concatenated feature vector is then input into the causal binary classifier, which calculates the confidence score that the candidate pair has a causal relationship. The causal association confidence score set is determined based on the causal association confidence scores of all candidate pairs, and candidate pairs whose confidence scores exceed the preset causal discrimination interface are retained to obtain a preliminary causal association edge set. Each edge in the preliminary causal association edge set connects a source event index and a target event index. In the initial set of causal association edges, the earliest control action event at the time step is taken as the starting event. Subsequent events are searched and connected along the directed edges in sequence to extract the complete causal triggering chain from the starting event to the event at the nearest time step. The event indices on the complete causal triggering chain are arranged in a monotonically increasing order. For each pair of adjacent events in the complete causal triggering chain, obtain the corresponding control action event index, calculate the index difference between the target event index and the source event index, and use the index difference as the trigger delay time step for each pair of adjacent events. The set of trigger delay time steps for each pair of adjacent events in the complete causal triggering chain is paired with the event index sequence of the complete causal triggering chain to generate a causal triggering chain description structure containing path node information and path edge weight information. Determine whether there are multiple parallel causal triggering chains in the context association representation vector sequence. If so, perform the calculation operation of the trigger delay time step on each causal triggering chain to obtain multiple sets of causal triggering chain description structures. All the obtained causal trigger chain description structures are integrated into the causal association pattern container, and the causal association pattern container is output as the organized causal association pattern. Each causal trigger chain description structure in the causal association pattern container has a unique chain identifier.

7. The method according to claim 1, characterized in that, Based on the aforementioned causal relationship pattern, a time-step recursive strategy is used to sequentially determine the type and execution time of the control action at each time step within the target tunneling period, resulting in a sequence of tunneling parameter control instructions, including: The causal association pattern is analyzed, and the common effect event shared by all causal triggering chains in the causal association pattern is extracted. The context association representation vector of the common effect event is used as the guiding target vector for strategy generation. Extract the causal subnet from the earliest triggering event to the common effect event in the causal association pattern. For each event node in the causal subnet, extract the context association representation vector corresponding to the event node. Combine the topological sorting position of the event node in the causal subnet with the position label vector to obtain a set of node representation vectors carrying causal position identifiers. The set of node representation vectors carrying causal location identifiers is input into the message passing layer of the graph neural network to perform several rounds of synchronous message passing, update the node representation vector of each event node, so that each node representation vector absorbs the information of its causal neighboring nodes, and obtains a set of causal reinforced node representation vectors. Extract the vectors corresponding to the common effect events from the set of causal reinforcement node representation vectors and perform vector gating fusion with the guiding target vector to generate a policy initiation vector; Based on the policy initiation vector and the fully connected network, a cyclic decoding process is performed to obtain multiple instruction tuples. Based on the instruction tuples, a parameter regulation instruction sequence is obtained by serialization. The fully connected network includes a fully connected classifier, a fully connected regressor, and a fully connected interval predictor.

8. The method according to claim 7, characterized in that, The process of performing loop decoding based on the strategy initiation vector and the fully connected network to obtain multiple instruction tuples, and then performing serialization processing based on the instruction tuples to obtain a parameter control instruction sequence, includes: After setting the policy initiation vector as the initial hidden state of the autoregressive policy decoding process, the decoding operation is executed step by step. At each decoding time step, the current hidden state is used to perform consistency attention matching with the set of causal reinforcement node representation vectors to generate a context instruction summary vector. Based on the context instruction summary vector and the current hidden state, the type of control action at the current time step is output by the fully connected classifier, the execution value of the control action at the current time step is output by the fully connected regressor, and the time span from the execution of the action at the current time step to the start of the next action is output by the fully connected interval predictor. Determine whether the type of control action generated in the current time step is a predefined termination action type. If yes, terminate the decoding process. If no, combine the type of control action generated in the current time step, the value of the control action execution, and the time span into an instruction tuple and store it in the instruction result list. Use the action generated in the current time step to embed and update the hidden state, and continue to the next decoding time step. After the decoding loop terminates, all instruction tuples are sequentially extracted from the instruction result list. The absolute execution time of each instruction tuple is determined according to the time span within each instruction tuple, resulting in a sequence of instruction tuples with absolute execution time markers. The sequence of instruction tuples with absolute execution time markers is then reordered according to the chronological order of the absolute execution times to make the instruction timeline monotonically increasing. The sorted instruction tuple sequence is formatted and encapsulated, and the control action type field, control action execution value field, and execution time field are serialized into structured instruction data blocks, which are then used as the tunneling parameter control instruction sequence.

9. A shield tunneling parameter adaptive control system, characterized in that, include: processor; A storage device storing a computer program, which, when executed by the processor, causes the processor to implement the intelligent control-based adaptive control method for shield tunneling parameters as described in any one of claims 1-8.

Images