Intelligent decision algorithm updating method applied to material proportioning regulation and server

Abstract

The embodiment of the application discloses an intelligent decision algorithm updating method and a server applied to material proportioning regulation, which first acquires real-time process state data, calls a current proportioning strategy generator to generate an initial material proportioning instruction sequence and extracts embedding layer weights and attention distribution; then determines input parameters of an adversarial disturbance synthesizer, generates process state confusion vectors and disturbance synthesizer state features, and element-by-element superimposes to generate adversarial enhanced process state data, and further inputs a formula instruction mapping model to obtain a disturbed material proportioning instruction and calculate a decision boundary offset measure; then, according to the decision boundary offset measure, a performance gain parameter is calculated, and the embedding layer weights, the attention distribution and the disturbance synthesizer weights are updated; finally, the current proportioning strategy generator is reconstructed to obtain an updated proportioning strategy generator, and real-time process state data is input to generate an optimized material proportioning instruction sequence, so as to realize adaptive evolution of the material proportioning strategy in adversarial training.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence algorithm technology, specifically to an intelligent decision-making algorithm update method and server applied to material proportioning control. Background Technology

[0002] In the slurry mixing ratio control scenario of tunnel shield excavation, the generation of material mixing ratio instructions usually relies on the mixing strategy model trained based on historical data. Such models collect real-time process status data such as slurry density, viscosity, sand content, slurry pressure, propulsion speed, cutterhead torque and formation permeability coefficient, and use a sequence-to-sequence neural network architecture to predict the material type and mixing ratio parameters at each sampling time, and output the material mixing ratio instruction sequence.

[0003] To improve the adaptability of material proportioning strategies to complex operating conditions, some technical solutions employ data augmentation to expand the training samples. However, the perturbations applied by data augmentation lack targeted perception of the decision boundary structure within the proportioning strategy model, making it difficult to generate effective adversarial samples. While adapting to new operating conditions, they are prone to losing their ability to proportion materials in historical scenarios. Therefore, how to achieve adaptive evolution of material proportioning strategies in adversarial training to ensure the accuracy and generalization ability of proportioning commands under complex process fluctuations is a technical problem that needs to be solved. Summary of the Invention

[0004] This application provides an intelligent decision-making algorithm update method and server for material proportioning control.

[0005] This application provides an intelligent decision-making algorithm update method for material proportioning control, including: Acquire real-time process status data for material ratio control scenarios, call the current ratio strategy generator, generate an initial material ratio instruction sequence based on the real-time process status data, and extract the embedding layer weights and attention distribution of the current ratio strategy generator; The input parameters of the adversarial perturbation synthesizer are determined based on the real-time process state data, the embedding layer weights, and the attention distribution. The process state confusion vector and perturbation synthesizer state features are then generated via the adversarial perturbation synthesizer. The process state confusion vector and the real-time process state data are superimposed element by element to generate adversarial enhanced process state data. The adversarial enhanced process state data and the perturbation synthesizer state features are input into the formula instruction mapping model to obtain the perturbation material ratio instruction. The deviation of the perturbation material ratio instruction from the initial material ratio instruction sequence is calculated to obtain the decision boundary offset metric. The performance gain parameter of the influence of the process state confusion vector on the decision boundary is calculated based on the decision boundary offset metric. The embedding layer weights and attention distribution of the current ratio strategy generator are updated based on the performance gain parameter. When updating the embedding layer weights, an elastic weight consolidation strategy is applied to constrain the weight update magnitude according to the difference between the current embedding layer weights and the historical embedding layer weights. The perturbation synthesizer weights in the adversarial perturbation synthesizer are updated based on the performance gain parameter. The current proportioning strategy generator is reconstructed using the updated embedding layer weights and the updated attention distribution to obtain the updated proportioning strategy generator. The real-time process status data is then input into the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

[0006] This application also provides an intelligent decision-making algorithm update server, 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 intelligent decision-making algorithm update methods applied to material proportioning control.

[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 decision-making algorithm update method applied to material proportioning control.

[0008] Therefore, the embodiments of this application have the following beneficial effects: By constructing a game-theoretic update closed loop between the current proportioning strategy generator and the adversarial perturbation synthesizer, the time-frequency domain representation of real-time process state data, the embedded layer weight response tensor, and the attention distribution structure pattern are uniformly transformed into input parameters of the adversarial perturbation synthesizer, enabling the process state confusion vector to accurately capture the weak areas of the decision boundary of the current proportioning strategy generator; the adversarial enhanced process state data is mapped to the material proportioning instruction space through the formula instruction mapping model, and the performance gain parameter of the influence of the process state confusion vector on the decision boundary is obtained by calculating the decision boundary offset metric between the perturbed material proportioning instruction and the initial material proportioning instruction sequence, and the embedded layer weights, attention distribution of the current proportioning strategy generator, and perturbation synthesizer weights of the adversarial perturbation synthesizer are updated according to the performance gain parameter. When updating the embedded layer weights, an elastic weight consolidation strategy is introduced, which constrains the weight update amplitude according to the difference between the current embedded layer weights and the historical embedded layer weights, thereby strengthening the robustness of the proportioning strategy while avoiding catastrophic forgetting.

[0009] As a result, the reconstructed updated proportioning strategy generator continuously outputs optimized material proportioning instruction sequences, realizing the adaptive evolution of material proportioning strategies in adversarial training, and significantly improving the accuracy and generalization ability of proportioning instructions under complex process state fluctuations. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an intelligent decision-making algorithm update method for material proportioning control, provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the basic structure of an intelligent decision-making algorithm update server provided in an embodiment of this application.

[0012] Figure 3 This is a functional block diagram of an intelligent decision-making algorithm update device provided in an embodiment of this application. Detailed Implementation

[0013] 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.

[0014] See Figure 1 As shown, this figure is a flowchart of an intelligent decision-making algorithm update method for material proportioning control provided in an embodiment of this application. This method is executed by an intelligent decision-making algorithm update server. Figure 1 As shown, the method includes steps 110-150.

[0015] The intelligent decision-making algorithm update method for material proportioning control provided in this application can be applied to the slurry proportioning control scenario in tunnel shield excavation engineering. During shield excavation, parameters such as slurry density, viscosity, sand content, and slurry pressure need to be dynamically adjusted according to changes in geological conditions. The accuracy of the proportioning instructions directly affects the stability of the excavation face and the slurry's muck-carrying efficiency. This method is executed on an intelligent decision-making algorithm update server. By establishing a game-theoretic update loop between the proportioning strategy generator and the adversarial perturbation synthesizer, the proportioning strategy generator continuously strengthens its adaptability to process state fluctuations during adversarial training, thereby outputting a more robust material proportioning instruction sequence.

[0016] Step 110: Obtain real-time process status data of the material ratio control scenario, call the current ratio strategy generator, generate an initial material ratio instruction sequence based on the real-time process status data, and extract the embedding layer weights and attention distribution of the current ratio strategy generator.

[0017] Real-time process status data is collected by sensor arrays deployed at various monitoring points of the tunnel boring machine. The sensor arrays include mud densitometers, viscometers, sand content detectors, slurry pressure transmitters, propulsion speed encoders, cutterhead torque sensors, and a formation permeability coefficient estimation module.

[0018] Each sensor sends its readings to the intelligent decision-making algorithm update server at a fixed sampling frequency. The server performs timestamp alignment on the received multi-source sensor data streams, merging the sensor readings within the same millisecond time window into a single process parameter snapshot vector. The dimension of this vector equals the total number of sensors, with each dimension corresponding to the instantaneous value of a process parameter. Following the chronological order of sampling, all process parameter snapshot vectors are arranged into a process status data sequence indexed by time nodes. This sequence serves as the input to the current proportioning strategy generator.

[0019] Step 111: The real-time process status data is sorted sequentially according to the acquisition time to construct a process status data sequence indexed by time nodes.

[0020] After receiving the raw data streams from various sensors, the server first performs timestamp parsing on each data stream, extracting the acquisition time information carried by each sensor record. Using a globally unified clock source as a reference, the millisecond-precision acquisition times are converted into monotonically increasing time node identifiers, each uniquely corresponding to one sampling time. For the same time node identifier, the server aggregates all process parameter values ​​belonging to that time into a multi-dimensional vector. The components of this vector correspond to mud density, viscosity, sand content, slurry pressure, propulsion speed, cutterhead torque, and formation permeability coefficient, respectively. If a sensor does not generate a valid reading at a given time node, the missing components are filled in using linear interpolation from adjacent times. All multi-dimensional vectors corresponding to all time nodes are arranged in ascending order of time node identifiers, forming a process status data sequence indexed by time nodes. The data structure of this sequence is a 2D matrix consisting of the number of time steps multiplied by the process parameter dimension.

[0021] Step 112: Using a preset dimension expansion mapping operator, perform feature space transformation on the process parameter snapshots at each time node in the process state data sequence to generate a process state feature vector sequence with a unified embedding dimension.

[0022] The preset dimension expansion mapping operator consists of a set of learnable projection parameters, which are stored as a matrix of input dimensions multiplied by a unified embedding dimension. The dimension expansion mapping operator is loaded and applied to the process state data sequence during the preprocessing stage of the current proportioning strategy generator. For the process parameter snapshot vector at each time point in the process state data sequence, the server performs matrix multiplication with the projection parameter matrix, linearly mapping the original process parameter dimension space to the unified embedding dimension space.

[0023] After mapping, each time node corresponds to a process state feature vector with a dimension equal to the unified embedding dimension. The process state feature vectors of all time nodes are arranged in chronological order to form a process state feature vector sequence. The value of the unified embedding dimension is determined in the current matching strategy generator design stage, for example, set to 256 or 512, so that subsequent embedding layers can directly receive dimension-matched input.

[0024] Step 113: Input the process state feature vector sequence into the embedding layer of the current ratio strategy generator, perform a linear projection operation on the process state feature vector sequence using the set of embedding layer weight parameters stored in the embedding layer to obtain the process state embedding vector sequence, and read out and cache the set of embedding layer weight parameters as embedding layer weights.

[0025] The current matching strategy generator's embedding layer is the first layer of the network, storing a set of embedding layer weight parameters. This set is in the form of a square matrix with one uniform embedding dimension multiplied by another uniform embedding dimension. The process state feature vector sequence is fed into the embedding layer in batches. The embedding layer performs a linear projection operation on each process state feature vector in the sequence, that is, it performs matrix multiplication between the process state feature vector and the embedding layer weight parameter set, and the result is the process state embedding vector.

[0026] The process state feature vectors at all time points are traversed to obtain a process state embedding vector sequence, where each vector maintains a uniform embedding dimension. While the server performs the linear projection operation, it reads all values ​​of the embedding layer weight parameter set from the embedding layer's parameter storage area via the parameter reading interface and copies them to a temporary parameter buffer. This buffer is specifically used to store the embedding layer weights extracted in the current round, for use in subsequent steps to determine the input parameters of the adversarial perturbation synthesizer and update the weight consolidation.

[0027] Step 114: The process state embedding vector sequence is processed by the multi-head attention calculation layer of the current ratio strategy generator to generate an attention weight distribution subgraph for each attention head, and all attention weight distribution subgraphs are combined into an attention distribution.

[0028] After further processing using a multi-head attention mechanism, the server obtains the complete attention distribution result. This step involves the following sub-steps: Step 1141: Input the process state embedding vector sequence into the multi-head attention computation layer of the current ratio strategy generator, and use the trained first query weight parameter group, first key weight parameter group and first value weight parameter group to perform a linear transformation on the process state embedding vector sequence to obtain the query vector sequence, key vector sequence and value vector sequence respectively.

[0029] The multi-head attention computation layer of the current matching strategy generator has three pre-trained sets of weight parameters: the first query weight parameter set, the first key weight parameter set, and the first value weight parameter set. Each set of weight parameters corresponds to one attention head. For each time position in the process state embedding vector sequence, the server performs matrix multiplication with the first query weight parameter set of each attention head on the corresponding process state embedding vector, generating the query vector of that attention head at that time position. After traversing all time positions, a query vector sequence for that attention head is formed.

[0030] Using the same processing method, the process state embedded vector sequence is multiplied by the first key weight parameter group and the first value weight parameter group respectively to generate the key vector sequence and value vector sequence for each attention head. Each attention head independently has a complete set of query vector sequence, key vector sequence, and value vector sequence, and the parameters are not shared between the sets of vector sequences.

[0031] Step 1142: For each attention head, calculate the association strength between the query vector at each time position in the query vector sequence and the key vectors at all time positions in the key vector sequence, and scale the association strength using a scaling factor to obtain the original attention score sequence.

[0032] For a single attention head, after obtaining the query vector sequence and key vector sequence for that attention head, the server calculates the association strength time-wise. Taking the query vector at the p-th time position in the query vector sequence as an example, the server calculates the dot product between this query vector and the key vector at the q-th time position in the key vector sequence. The dot product result is the initial association value between the query vector and the key vector at that time pair. After traversing all time positions in the key vector sequence, a set of initial association value sequences corresponding to the query vector is obtained. The length of this initial association value sequence is equal to the number of time steps in the key vector sequence.

[0033] The server multiplies each value in the initial sequence of associations by a scaling factor, which is the reciprocal of the square root of the quotient of the query vector dimension in the attention head divided by the number of attention heads. The scaled sequence of values ​​is the original attention score sequence corresponding to the query vector, where each value corresponds to the association strength score at one time position in the key vector sequence.

[0034] Step 1143: For each attention head, perform an exponential normalization transformation on the original attention score sequence to generate an attention weight distribution subgraph at each time position of the attention head, and use the attention weight distribution subgraph to perform weighted fusion on the value vector sequence to obtain a single-head attention output vector sequence.

[0035] The server performs an exponential normalization transformation on the original attention score sequence. First, each score value in the sequence is used as the exponent of the exponential function e to obtain the exponential value corresponding to each score value. Then, the sum of all exponential values ​​is calculated, and each exponential value is divided by the sum to obtain the normalized attention weight corresponding to each time position. The normalized attention weights at all time positions constitute a row vector of the attention weight distribution subgraph of the attention head at the p-th time position in the query vector sequence.

[0036] Each vector in the single-head attention output vector sequence is obtained by weighted fusion of the value vector sequence and attention weights. The weighted fusion calculation process is as follows: multiply the value vector at each time position in the value vector sequence by the corresponding attention weight and then sum them. The sum is the single-head attention output vector of that attention head at query position p. Traverse all time positions in the query vector sequence to generate the complete single-head attention output vector sequence.

[0037] Step 1144: Concatenate the single-head attention output vector sequences of all attention heads along the feature dimension to obtain a multi-attention head concatenated vector sequence, and perform parameterized linear fusion processing on the multi-attention head concatenated vector sequence to generate an attention fusion vector sequence.

[0038] The server concatenates the single-head attention output vectors generated by all attention heads at the same query time position along the feature dimension, forming a concatenated vector with a length equal to the dimension of the single-head attention output vector multiplied by the number of attention heads. After performing this concatenation operation on all query time positions, a sequence of multi-attention head concatenated vectors is obtained.

[0039] The server then uses a set of parameterized linear fusion weight matrices to perform linear fusion on each concatenated vector in the multi-attention head concatenation vector sequence, compressing the dimension of the concatenated vectors back to a unified embedding dimension. The parameterized linear fusion weight matrix is ​​in the form of the concatenated vector dimension multiplied by the unified embedding dimension, and the fused vector sequence is the attention fusion vector sequence.

[0040] Step 1145: Perform a residual connection between the attention fusion vector sequence and the process state embedding vector sequence, and perform layer normalization on the result after residual connection to obtain the context representation sequence.

[0041] The server performs an element-wise addition operation on the attention fusion vector sequence and the process state embedding vector sequence. During the addition, the values ​​of the attention fusion vector and the process state embedding vector in the corresponding dimensions are added pairwise according to the same time position.

[0042] The summed vector sequence enters the layer normalization operation. For each time point, the layer normalization operation performs the following steps: calculates the mean of each dimension's value, calculates the standard deviation of each dimension's value, subtracts the mean from each dimension's value, divides by the standard deviation, multiplies by a learnable scaling parameter, and adds a learnable translation parameter. The vector sequence output by the layer normalization operation is the context representation sequence.

[0043] Step 1146: Input the context representation sequence into the instruction decoding layer of the current proportioning strategy generator. Perform type tendency analysis on the context representation at each time position through the material type discrimination branch in the instruction decoding layer to obtain the material type identifier. Perform numerical prediction on the context representation at each time position through the proportioning parameter regression branch in the instruction decoding layer to obtain the proportioning ratio parameter.

[0044] The instruction decoding layer includes parallel material type discrimination branches and proportion parameter regression branches. The material type discrimination branch consists of a fully connected classification head, which takes a context representation vector at a given time location as input and outputs a material type preference vector. The dimension of this vector is equal to the total number of selectable material types, and each dimension stores the preference score for the corresponding material type. The server performs a soft maximization operation on this preference vector, converting the preference scores of each dimension into a probability distribution, and selects the material type with the highest probability as the material type identifier for that time location.

[0045] The proportioning parameter regression branch consists of a fully connected regression head. This head receives the context representation vector at the same time point as input and outputs a real scalar. This scalar is mapped to the interval between 0 and 1 by a sigmoid activation function, and then multiplied by a preset maximum proportioning value to obtain the proportioning parameter at that time point. The proportioning parameter indicates the percentage of the selected material type in the total material.

[0046] Step 1147: Combine the material type identifier and the proportion parameter at the same time position into a material proportion tuple, and concatenate all the material proportion tuples in time node order to generate the initial material proportion instruction sequence.

[0047] The server independently performs combined operations at each time point, encapsulating the material type identifier and proportion parameter into a binary structure, which is the material proportion tuple. One material proportion tuple corresponds to one time node, representing the specific material allocation plan to be executed at that time node. Following the ascending order of the time node identifiers, all material proportion tuples are concatenated into a sequence, which is the initial material proportion instruction sequence.

[0048] Step 1148: After generating attention weight distribution subgraphs for all attention heads in the multi-head attention calculation layer, combine all attention weight distribution subgraphs into an attention distribution.

[0049] During the processing in step 1143, the server has temporarily stored the row vectors of the attention weight distribution subgraphs generated by each attention head at each query position in the temporary attention cache area.

[0050] After the single-head attention output vector sequences of all attention heads have been generated, the server extracts all attention weight distribution subgraph data from the temporary attention buffer and organizes it into a 3D tensor according to the attention head number dimension, query position time dimension, and key position time dimension. This 3D tensor is the attention distribution. The first dimension of the attention distribution indexes the attention head, the second dimension indexes the query position, and the third dimension indexes the key position. Any element in the tensor represents the attention weight value of a specified attention head at a specified query position to a specified key position.

[0051] Step 120: Determine the input parameters of the adversarial perturbation synthesizer based on the real-time process state data, the embedding layer weights, and the attention distribution, and generate the process state confusion vector and perturbation synthesizer state features via the adversarial perturbation synthesizer.

[0052] After extracting the embedding layer weights and attention distribution, the server uses real-time process state data, embedding layer weights, and attention distribution as raw materials to construct the input parameters required for the adversarial perturbation synthesizer. This process involves time-frequency domain decomposition of the real-time process state data, construction of the response tensor of the embedding layer weights, and extraction of the structural patterns of the attention distribution.

[0053] Step 121: Perform time-frequency domain decomposition on the real-time process status data to extract the time-series fundamental component and high-frequency fluctuation component as a dual-channel representation of the process status.

[0054] The server performs time-frequency domain decomposition on the time series of each process parameter dimension in the real-time process status data. The decomposition uses a sliding window short-time Fourier transform, with a window length of 32 time steps and a sliding step size of 8 time steps. After weighting the process parameter value sequence within each window using a Hanning window function, a fast Fourier transform is performed to obtain the corresponding frequency domain coefficient sequence. Coefficients in the frequency domain coefficient sequence with frequencies near 0 Hz are classified as the time-series fundamental component, reflecting the trend of the process parameter within that window. Coefficients in the frequency domain coefficient sequence with frequencies higher than a preset cutoff frequency threshold are classified as high-frequency fluctuation components, reflecting the instantaneous oscillation characteristics of the process parameter.

[0055] After performing the above decomposition on all process parameter dimensions and all windows, the time-series fundamental components of each dimension are stacked along the channel dimension to form a time-series fundamental characteristic tensor, and the high-frequency fluctuation components of each dimension are stacked along the channel dimension to form a high-frequency fluctuation characteristic tensor. The time-series fundamental characteristic tensor and the high-frequency fluctuation characteristic tensor together constitute a dual-channel representation of the process state.

[0056] Step 122: Perform an outer product operation between the dual-channel representation of the process state and the embedding layer weights to generate an embedding space response tensor. The embedding space response tensor is used to characterize the degree of influence of process state changes on the embedding layer decision boundary of the current proportioning strategy generator.

[0057] The server processes the time-series fundamental wave feature tensor and the high-frequency fluctuation feature tensor in the dual-channel representation of the process state separately. Taking the time-series fundamental wave feature tensor as an example, the server flattens the time-series fundamental wave feature tensor along the time dimension, rearranging the feature values ​​of each time step and each dimension into a single fundamental wave flattened vector. The fundamental wave flattened vector is then subjected to an outer product operation with the embedding layer weight parameter set. In the matrix generated by the outer product operation, the element in the m-th row and n-th column is equal to the product of the m-th component of the fundamental wave flattened vector and the n-th component of the flattened embedding layer weight parameter set. Each row vector of this matrix is ​​then reshaped into a tensor according to the original shape of the embedding layer weights to obtain the embedding space response sub-tensor corresponding to the time-series fundamental wave.

[0058] The high-frequency fluctuation feature tensor is processed in the same way to obtain the embedding space response sub-tensor corresponding to the high-frequency fluctuation. The two embedding space response sub-tensors are stacked along the newly added modal dimension to generate the complete embedding space response tensor.

[0059] Step 123: Perform structural analysis on the attention distribution, extract the attention concentration pattern and attention dispersion pattern to generate an attention pattern descriptor, which includes the response peak position and response distribution width of each attention head.

[0060] The server performs structural analysis on each attention head in the attention distribution. For the attention weight distribution subgraph corresponding to the h-th attention head in the attention distribution, the server scans the attention weight vector along the key position dimension, locates the key position index with the largest attention weight value in the vector, and records this index as the peak response position of the attention head at the corresponding query position.

[0061] Simultaneously, the server calculates the dispersion of the attention weight vector by using the attention weights as a probability mass function, calculating the expected value of the key position index and the second-order central moment. The square root of the second-order central moment is the response distribution width of the attention head at that query position. After traversing all attention heads and all query positions, the server arranges the response peak positions of each attention head at each query position into a response peak position matrix by attention head number and query position number, and arranges the response distribution widths in the same way into a response distribution width matrix. The response peak position matrix and the response distribution width matrix are concatenated along the newly added channel dimension, and the resulting output is the attention pattern descriptor.

[0062] Step 124: Channel-align the embedded spatial response tensor and the attention pattern descriptor and concatenate them as input parameters for the adversarial perturbation synthesizer.

[0063] The server checks the feature dimension consistency between the embedding space response tensor and the attention pattern descriptor. When the two do not match in spatial dimensions other than the channel dimension, the server adjusts the spatial dimension of the attention pattern descriptor to match that of the embedding space response tensor through bilinear interpolation.

[0064] After adjustment, the server concatenates the embedded spatial response tensor and attention pattern descriptor along the channel dimension. The number of channels in the concatenated tensor is equal to the sum of the number of channels in the embedded spatial response tensor and the number of channels in the attention pattern descriptor. This concatenated tensor is the input parameter of the adversarial perturbation synthesizer.

[0065] Step 125: Input the input parameters of the adversarial perturbation synthesizer into the perturbation encoder of the adversarial perturbation synthesizer, perform hierarchical feature extraction through multiple stacked coding transformation layers, and generate a perturbation hidden state sequence. Each coding transformation layer includes a linear mapping unit and a gated activation unit.

[0066] The perturbation encoder of the adversarial perturbation synthesizer consists of several stacked coding transformation layers, each containing a linear mapping unit and a gated activation unit. The linear mapping unit performs a fully connected linear transformation on the input, and the transformation matrix is ​​in the form of the input feature dimension multiplied by the output feature dimension.

[0067] The gated activation unit adopts a gated linear unit structure, dividing the features output by the linear mapping unit into two parts along the channel dimension. A sigmoid activation function is applied to one part, while no activation is applied to the other part. The two parts are multiplied element-wise and used as the output of the gated activation unit. The input parameters are processed by the first coding transform layer to generate the first-level hidden state features. The first-level hidden state features are used as the input to the second coding transform layer, and so on. The features output after all coding transform layers are the perturbed hidden state sequence.

[0068] Step 126: The perturbation hidden state sequence is passed to the temporal state evolution module of the adversarial perturbation synthesizer to read the historical perturbation synthesizer state features generated in the previous iteration, and the perturbation synthesizer state features of the current iteration are generated based on the interaction between the perturbation hidden state sequence and the historical perturbation synthesizer state features.

[0069] The temporal state evolution module internally maintains a state buffer to store the perturbation synthesizer state features generated in the previous iteration. The server reads the historical perturbation synthesizer state features from the state buffer and sends the perturbation hidden state sequence and the historical perturbation synthesizer state features together to the gate control loop unit. The gate control loop unit calculates the reset gate vector and the update gate vector based on the current time step input of the perturbation hidden state sequence and the historical perturbation synthesizer state features.

[0070] The reset gate vector determines which dimensions of information from the historical perturbation synthesizer state features are retained, while the update gate vector determines the mixing ratio of the current time step candidate state with the historical perturbation synthesizer state features. After recursive calculations by the gated recurrent unit over all time steps, the hidden state output of the last time step is the perturbation synthesizer state feature for the current iteration. Simultaneously, the server writes the current iteration's perturbation synthesizer state feature back to the state buffer, overwriting the historical perturbation synthesizer state features.

[0071] Step 127: Decouple the state features of the disturbance synthesizer by dimension, separate the process condition modulation component and the ratio interference injection component, and locally adjust the fluctuation range of the real-time process state data through the process condition modulation component to generate a condition adjustment signal.

[0072] The state characteristics of the perturbation synthesizer are input into two parallel fully connected branch networks. The first fully connected branch network consists of one fully connected layer and outputs a process condition modulation component. The dimension of the process condition modulation component is the same as the dimension of the process parameters in the real-time process state data.

[0073] Then, the values ​​of each dimension of the process condition modulation component are multiplied element by element with the condition fluctuation range parameter of the corresponding process parameter at the corresponding time position in the real-time process status data. The condition fluctuation range parameter is taken as the difference between the maximum and minimum values ​​of the process parameter in the most recent 8 time steps. The result of the multiplication is the condition adjustment signal.

[0074] Step 128: Combine the ratio interference injection component with the operating condition adjustment signal point by point to obtain the initial process state confusion vector. Perform amplitude constraint processing on the initial process state confusion vector to keep the numerical change amplitude of the initial process state confusion vector within a preset disturbance intensity range, and generate the process state confusion vector.

[0075] The second fully connected branch network consists of one fully connected layer and outputs a ratio interference injection component. The dimension of the ratio interference injection component is the same as the dimension of the process parameters in the real-time process status data. The server adds the ratio interference injection component and the operating condition adjustment signal time-wise and dimension-wise to obtain the initial process status confusion vector.

[0076] Subsequently, a truncation function is applied to constrain the amplitude of the initial process state confusion vector. The truncation function's processing rules are as follows: for each value in the initial process state confusion vector, if the value is greater than the preset upper limit of disturbance intensity, the value is replaced with the upper limit value; if the value is less than the preset lower limit of disturbance intensity, the value is replaced with the lower limit value; if the value is between the upper and lower limits, the original value remains unchanged. The truncated vector is the generated process state confusion vector.

[0077] Step 130: Perform element-wise superposition processing on the process state confusion vector and the real-time process state data to generate adversarial enhanced process state data. Input the adversarial enhanced process state data and the perturbation synthesizer state features into the formula instruction mapping model to obtain the perturbed material proportioning instruction, and calculate the degree of deviation between the perturbed material proportioning instruction and the initial material proportioning instruction sequence to obtain the decision boundary offset metric.

[0078] After the server generates the process state confusion vector, it superimposes this vector as an adversarial perturbation term onto the original real-time process state data to construct adversarially enhanced process state data within the decision boundary neighborhood. This adversarially enhanced process state data, along with the perturbation synthesizer state features, is then input into the formulation instruction mapping model to generate perturbed material proportioning instructions. By quantifying the deviation of the perturbed material proportioning instructions from the initial material proportioning instruction sequence, a decision boundary offset metric is obtained to measure the sensitivity of the current proportioning strategy's decision boundary.

[0079] Step 131: Perform a term-by-term summation operation on the process state confusion vector and each process parameter snapshot contained in the real-time process state data at the same time position to obtain adversarial enhancement process state data.

[0080] The process status obfuscation vector and the real-time process status data have completely identical time steps and process parameter dimensions. The server aligns the two vectors one by one according to the time position index. For time position t, the process parameter snapshot vector at that moment is extracted from the real-time process status data, and the obfuscation vector fragment corresponding to that moment is extracted from the process status obfuscation vector. The corresponding dimension values ​​of the two vectors are then added pairwise.

[0081] The addition operation iterates through all process parameter dimensions, including mud density, viscosity, sand content, slurry pressure, propulsion speed, cutterhead torque, and formation permeability. The calculation result for each dimension is equal to the sum of the original process parameter value and the perturbation value. After the calculation is completed by overlaying all time positions, the generated new matrix is ​​the data for the enhanced process state.

[0082] Step 132: Input the adversarial enhancement process state data and the perturbation synthesizer state features into the conditional feature encoder of the recipe instruction mapping model. Extract the local receptive field of the adversarial enhancement process state data through the shared feature extraction layer in the conditional feature encoder to generate an enhanced state feature map. Then, transform the perturbation synthesizer state features into a conditional modulation vector through the conditional injection layer in the conditional feature encoder and superimpose it onto the enhanced state feature map to obtain conditional fusion features.

[0083] The conditional feature encoder of the recipe instruction mapping model receives adversarial enhancement process state data and perturbation synthesizer state features as dual inputs. The shared feature extraction layer consists of several two-dimensional convolutional kernel groups. Each convolutional kernel group contains multiple 3x3 convolutional kernels. The convolutional kernels slide along the time and feature dimensions on the adversarial enhancement process state data. At each sliding position, the weights of the convolutional kernels are calculated as a weighted sum of the values ​​within the corresponding receptive field. The output channels of all convolutional kernel groups are concatenated along the channel dimension to form the enhanced state feature map.

[0084] The conditional injection layer contains one fully connected transform unit that maps the perturbation synthesizer state features to a conditional modulation vector. The length of the conditional modulation vector is equal to the number of channels in the enhanced state feature map. For each channel of the enhanced state feature map, the server multiplies the values ​​of all spatiotemporal positions of that channel by the modulation coefficient of the corresponding channel in the conditional modulation vector. The modulated feature map is the conditional fusion feature.

[0085] Step 133: Input the conditional fusion feature into the manifold mapping module of the recipe instruction mapping model. The manifold mapping module contains a mapping network composed of multiple fully connected residual blocks.

[0086] After receiving the conditional fusion features, the manifold mapping module flattens them into a one-dimensional vector and feeds it into a mapping network consisting of multiple fully connected residual blocks. Each fully connected residual block contains two fully connected layers and one residual connection. The first fully connected layer maps the input vector to an intermediate dimension, applies a modified linear unit activation function, and then passes it to the second fully connected layer. The second fully connected layer maps the intermediate dimension back to the input vector dimension. The output vector is then element-wise added to the original input vector of the residual block, and the result is normalized by the layer and used as the output of the residual block. Multiple fully connected residual blocks are stacked in series, with the output of the previous residual block serving as the input of the next residual block.

[0087] Step 134: Using the mapping network, the conditional fusion features are transformed by coordinate transformation based on the set of manifold basis vectors stored in the manifold mapping module, and the conditional fusion features are mapped to the low-dimensional manifold subspace of the material proportioning instruction space to generate manifold latent variables.

[0088] The manifold mapping module pre-sets a set of learnable manifold basis vectors, which contains several basis vectors, the number of which equals the dimension of the low-dimensional manifold subspace. After being transformed layer by layer by a mapping network composed of fully connected residual blocks, the mapping network outputs an intermediate vector with the same dimension as each basis vector in the manifold basis vector set.

[0089] The server calculates the inner product of the intermediate vector and each basis vector in the manifold basis vector set, resulting in a set of manifold coordinate coefficients. Each manifold coordinate coefficient corresponds to the projection intensity of the intermediate vector onto the direction of the corresponding basis vector. All manifold coordinate coefficients are combined into manifold latent variables, the dimension of which is equal to the dimension of the low-dimensional manifold subspace. The manifold latent variables constitute the low-dimensional manifold representation of the conditional fusion features in the material proportioning instruction space.

[0090] Step 135: Input the manifold latent variable into the instruction reconstruction decoder of the recipe instruction mapping model, and perform spatial reconstruction of the manifold latent variable through a deconvolution upsampling layer and a channel attention adjustment layer to generate the disturbance material ratio instruction.

[0091] The instruction reconstruction decoder consists of alternating stacks of deconvolutional upsampling layers and channel attention adjustment layers. The deconvolutional upsampling layer receives manifold latent variables and gradually recovers the spatial dimension through transposed convolution operations. Each upsampling layer expands the temporal dimension of the features by a factor of 2, and batch normalization is applied after upsampling.

[0092] The channel attention adjustment layer applies a channel attention mechanism to the upsampled feature map. The calculation process is as follows: global average pooling is performed on the feature map along the spatial dimension to obtain the channel description vector; the channel description vector is then compressed to one-quarter of the original number of channels through a two-layer fully connected bottleneck structure, and then expanded back to the original number of channels. The output of the bottleneck structure is activated by a sigmoid function to generate the attention adjustment coefficients for each channel; each channel of the feature map is multiplied by its corresponding attention adjustment coefficient. After processing by all deconvolutional upsampling layers and channel attention adjustment layers, the final output reconstructed feature map is the disturbance material proportioning instruction.

[0093] Step 136: Create a decision boundary analysis unit, and input the disturbing material proportioning instruction and the initial material proportioning instruction sequence into the decision boundary analysis unit to extract the material type distribution vector and proportioning parameter distribution vector in the disturbing material proportioning instruction, and extract the material type benchmark distribution vector and proportioning parameter benchmark distribution vector in the initial material proportioning instruction sequence.

[0094] The server creates a decision boundary analysis unit, which is a data processing module dedicated to calculating the offset metric between two sets of proportioning instructions. After receiving the sequence of disturbing material proportioning instructions and the initial material proportioning instruction sequence, the decision boundary analysis unit parses the material type identifier and proportioning ratio parameter of the disturbing material proportioning instruction at each time position. It encodes the material type identifiers at all time positions into a sequence of uniquely encoded vectors. The average of this sequence along the time dimension yields a material type distribution vector, where each dimension represents the frequency of the corresponding material type at each time position of the disturbing material proportioning instruction. Simultaneously, the proportioning ratio parameters at all time positions are categorized by material type and averaged to obtain a proportioning parameter distribution vector. The initial material proportioning instruction sequence is processed in the same way to obtain the material type baseline distribution vector and the proportioning parameter baseline distribution vector.

[0095] Step 137: Calculate the type offset between the material type distribution vector and the material type reference distribution vector, and calculate the parameter offset between the proportioning parameter distribution vector and the proportioning parameter reference distribution vector.

[0096] When calculating the type offset, the server subtracts the material type distribution vector from the material type baseline distribution vector dimension by dimension to obtain the type difference vector. The squares of each dimension of the type difference vector are then summed, and the square root of the sum is taken. The resulting scalar is the type offset. Similarly, when calculating the parameter offset, the ratio parameter distribution vector is subtracted from the ratio parameter baseline distribution vector dimension by dimension to obtain the parameter difference vector. The squares of each dimension of the parameter difference vector are then summed, and the square root of the sum is taken. The resulting scalar is the parameter offset.

[0097] Step 138: Weighted fusion of the type offset and the parameter offset to obtain a comprehensive offset characterization; cumulative statistics of the comprehensive offset characterization over time; construct an offset trend evaluation curve based on the rate of change and acceleration of change of the comprehensive offset characterization between adjacent time nodes; and determine the decision boundary offset metric based on the offset trend evaluation curve.

[0098] The server multiplies the category offset and parameter offset by preset category offset weighting coefficients and parameter offset weighting coefficients respectively, and then sums them to obtain the comprehensive offset representation. The comprehensive offset representation is accumulated statistically along the time dimension, and the difference between the comprehensive offset representations of adjacent time nodes is calculated as the first-order change, and the difference between adjacent first-order changes is calculated as the second-order change. The comprehensive offset representation, first-order change, and second-order change at each time node are concatenated into a 3D vector sequence. After performing moving average smoothing on this vector sequence, the offset trend evaluation curve is obtained. The decision boundary offset metric is taken as the cumulative value of the comprehensive offset representation of the offset trend evaluation curve at the last time node.

[0099] Step 140: Calculate the performance gain parameter of the influence of the process state confusion vector on the decision boundary based on the decision boundary offset metric; update the embedding layer weights and attention distribution of the current ratio strategy generator based on the performance gain parameter; when updating the embedding layer weights, apply an elastic weight consolidation strategy to constrain the weight update magnitude according to the difference between the current embedding layer weights and the historical embedding layer weights; and update the perturbation synthesizer weights in the adversarial perturbation synthesizer based on the performance gain parameter.

[0100] After obtaining the decision boundary offset metric, the performance gain parameter of the process state confusion vector in this round of the game is calculated based on it. The performance gain parameter is then decomposed into a guiding component that guides the update of the matching strategy generator and a guiding component that guides the update of the adversarial perturbation synthesizer. When updating the embedding layer weights, the elastic weight consolidation strategy applies constraints by referring to historical embedding layer weights to prevent key decision parameters from being overwritten or forgotten under large perturbations.

[0101] Step 141: Input the decision boundary offset metric into the performance evaluation function. With the goal of maximizing the decision boundary offset metric, generate a performance gain parameter for the process state confusion vector through gradient backpropagation. The performance gain parameter indicates the effect of the process state confusion vector on pushing away from the decision boundary in the current game round.

[0102] The performance evaluation function is defined as a monotonically increasing transformation of the decision boundary offset metric, which is mapped using a logarithmic function with a temperature coefficient. After substituting the decision boundary offset metric into the performance evaluation function, the server calculates the gradient of the output value relative to the process state confusion vector. This gradient is backpropagated along the computational path in step 130 back to the process state confusion vector, yielding a set of partial derivatives for each dimension of the confusion vector. This set of partial derivatives is the performance gain parameter, and its magnitude reflects the contribution of perturbations in each dimension to the decision boundary offset metric.

[0103] Step 142: Decompose the performance gain parameter into a strategy generator-guided component and a perturbation synthesizer-guided component. The strategy generator-guided component is used to enhance the robustness of the current ratio strategy generator, and the perturbation synthesizer-guided component is used to guide the adversarial perturbation synthesizer to generate a confusion vector.

[0104] The server processes the performance gain parameter in two branches. The policy generator guiding component is the product of the performance gain parameter and a preset sign inversion coefficient of -1, ensuring that the direction of the policy generator guiding component is opposite to the gradient direction of the performance gain parameter, guiding the matching policy generator to update in the direction of canceling perturbations. The perturbation synthesizer guiding component is the performance gain parameter itself or its proportionally scaled result, maintaining the same gradient direction as the performance gain parameter, guiding the adversarial perturbation synthesizer to update in the direction of enhancing perturbations.

[0105] Step 143: Construct a policy update loss function using the policy generator-guided component, and calculate the first gradient of the policy update loss function with respect to the embedding layer weights, and the second gradient with respect to the attention weight parameters in the attention distribution.

[0106] The policy update loss function consists of the policy generator guiding component and the current-ratio policy generator parameters. The server flattens the policy generator guiding component into a vector, calculates the inner product of this vector with itself, and uses the result of the inner product as the loss value of the policy generator.

[0107] Subsequently, the partial derivative of the loss value with respect to each element in the set of embedding layer weight parameters is calculated, and all partial derivatives constitute the first gradient. Simultaneously, the partial derivative of the loss value with respect to each element in the attention distribution tensor is calculated, and all partial derivatives constitute the second gradient.

[0108] Step 144: Before updating the embedding layer weights by applying the first gradient, read the weight change constraint threshold corresponding to the embedding layer weights from the constraint parameter record of the elastic weight consolidation strategy.

[0109] The elastic weight consolidation strategy maintains a constraint parameter record table on the server. This table stores the weight change constraint thresholds corresponding to each parameter in the embedding layer weights. The initial value of the weight change constraint threshold is determined based on the statistical analysis of the update magnitude of the embedding layer weights in historical game rounds, and is taken as a preset multiple of the standard deviation of the historical update magnitude of that parameter. The server reads the weight change constraint thresholds corresponding to each dimension from the constraint parameter record table one by one, according to the dimension index of the embedding layer weight to be updated.

[0110] Step 145: Scale the first gradient dimension by dimension according to the weight change constraint threshold, and obtain the updated embedding layer weights by adding the scaled first gradient to the original embedding layer weights.

[0111] For each element in the first gradient, the server calculates the ratio of the absolute value of that element to the corresponding weight change constraint threshold. When the ratio is greater than 1, the element is multiplied by the quotient of the weight change constraint threshold divided by the absolute value of the element, scaling the element to within the constraint threshold range; when the ratio is less than or equal to 1, the element's original value remains unchanged. The scaled first gradient is then multiplied by the learning rate and added element-wise to the original embedding layer weights to obtain the updated embedding layer weights.

[0112] Step 146: Update the attention distribution based on the second gradient. Apply the second gradient to the query weight parameters, key weight parameters, and value weight parameters in the multi-head attention calculation layer through the attention weight redistribution process to obtain the updated attention distribution.

[0113] The server expands the second gradient along the three dimensions of the attention distribution tensor. Based on the functional relationship between the attention distribution tensor and the query weight parameters, key weight parameters, and value weight parameters in the multi-head attention computation layer, it applies the chain rule to pass the second gradient layer by layer to the query weight parameters, key weight parameters, and value weight parameters, respectively. This yields the query weight gradient tensor, key weight gradient tensor, and value weight gradient tensor. After multiplying these three gradient tensors by the learning rate, they are subtracted element-wise from the corresponding original weight parameters to obtain the updated query weight parameters, key weight parameters, and value weight parameters. Using the updated three sets of weight parameters, the forward computation of steps 1141 to 1144 is re-executed to generate the updated attention distribution.

[0114] Step 147: Construct a perturbation enhancement loss function using the perturbation synthesizer guiding component, and calculate the perturbation synthesizer gradient of the perturbation enhancement loss function relative to the coding transform layer parameters, the temporal fusion parameter set, and the manifold mapping module parameters in the adversarial perturbation synthesizer.

[0115] The perturbation enhancement loss function consists of the perturbation synthesizer guiding component and the output of the adversarial perturbation synthesizer. The server flattens the perturbation synthesizer guiding component into a vector, calculates the inner product of this vector with itself, and takes the negative value, which is used as the perturbation enhancement loss value. The partial derivatives of this loss value with respect to the linear mapping unit weights and gated activation unit parameters of each coding transform layer in the perturbation encoder, the partial derivatives with respect to the gated loop unit parameters in the temporal state evolution module, and the partial derivatives with respect to the manifold mapping module parameters in the recipe instruction mapping model are calculated. All partial derivatives are combined to form the perturbation synthesizer gradient.

[0116] Step 148: Update the perturbation synthesizer weights in the adversarial perturbation synthesizer according to the perturbation synthesizer gradient. The update operation includes gradient descent adjustment of the linear mapping unit weights, gated activation unit parameters, temporal fusion parameter set, and conditional injection layer parameters in the conditional feature encoder of the coding transform layer.

[0117] The server multiplies each part of the gradient in the perturbation synthesizer gradient by the learning rate and then subtracts it element by element from the corresponding perturbation synthesizer weight parameters.

[0118] The specific update scope includes: the weight matrix of the linear mapping unit in each coding transform layer of the perturbation encoder, and the branch transform parameters of the gated activation unit; the reset gate weight matrix, update gate weight matrix, candidate state weight matrix, and corresponding bias vector of the gated loop unit in the temporal state evolution module; and the fully connected transform weight matrix and bias vector of the conditional injection layer in the conditional feature encoder. The updated perturbation synthesizer weights cover the original parameters, completing the parameter update of the adversarial perturbation synthesizer in this round.

[0119] Step 150: Reconstruct the current proportioning strategy generator using the updated embedding layer weights and the updated attention distribution to obtain the updated proportioning strategy generator. Input the real-time process status data into the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

[0120] After updating the embedding layer weights and attention distribution, the updated parameters are reassembled into the network structure of the current proportioning strategy generator to form the updated proportioning strategy generator. Under the same real-time process state data input, the updated proportioning strategy generator can output a more robust optimized material proportioning instruction sequence.

[0121] Step 151: Read the updated embedding layer weights and the updated attention distribution, and cache the updated embedding layer weights and the updated attention distribution in the parameter assembly cache area.

[0122] The server reads the updated embedding layer weights and attention distribution from the memory space of the parameter update module, copies the embedding layer weights in matrix format to the embedding layer weight cache slots in the parameter assembly cache, and copies the query weight parameters, key weight parameters, and value weight parameters from the attention distribution to the attention weight cache slots in the parameter assembly cache. The parameter assembly cache is a contiguous memory area pre-allocated before the update operation.

[0123] Step 152: Obtain the internal structure definition information of the current proportioning strategy generator. The internal structure definition information is used to describe the dimension information and storage layout of the embedded layer parameter interface of the current proportioning strategy generator.

[0124] The internal structure definition information is stored in the model metadata of the current matching strategy generator in the form of a structured description document. The server parses the model metadata and extracts the dimensional information of the embedding layer parameter interface, including the input feature dimension, output embedding dimension, and the memory row-major or column-major storage layout identifier of the parameter matrix. At the same time, it extracts the number of attention heads in the multi-head attention module of the attention calculation layer, the query vector dimension, key vector dimension, and value vector dimension of each attention head, as well as the offset address of each weight parameter in memory.

[0125] Step 153: Based on the internal structure definition information, tensor reshaping is performed on the updated embedding layer weights in the parameter assembly cache to generate an embedding weight parameter package with a storage layout that conforms to the embedding layer parameter interface.

[0126] The server extracts the tensor shape and memory layout required by the embedding layer parameter interface from the internal structure definition information. When the shape or layout of the updated embedding layer weight matrix in the parameter assembly cache is inconsistent with the interface requirements, a tensor reshaping operation is performed: if the shape is inconsistent, the shape is transformed by adjusting the number of rows and columns of the matrix to ensure that the total number of elements remains unchanged; if the memory layout is inconsistent, the row-first layout and column-first layout are converted to each other by matrix transposition. After reshaping, the embedding layer weights are encapsulated into an embedding weight parameter package, with a dimension checksum attached to the header of the parameter package.

[0127] Step 154: parse the updated attention distribution in the parameter assembly cache, decompose the updated query weight tensor, the updated key weight tensor, and the updated value weight tensor, and align the formats of the updated query weight tensor, the updated key weight tensor, and the updated value weight tensor to generate an attention weight parameter package.

[0128] The updated attention distribution is stored in the parameter assembly cache as a pending assembly form. The server, based on the attention head numbering order and the dimension of each attention head's weight tensor, sequentially decomposes the attention distribution into the updated query weight tensor, updated key weight tensor, and updated value weight tensor corresponding to each attention head. Dimension validation and format alignment are performed on each of the three sets of weight tensors for each attention head. If the dimension of one attention head's weight tensor is inconsistent with the single-header dimension required by the interface, adjustment is made through dimension padding or truncation. The aligned weight tensors are then packaged into an attention weight parameter package according to the attention head numbering order.

[0129] Step 155: Initialize the parameter loading channel, write the embedded weight parameter package into the embedding layer parameter storage area of ​​the current ratio strategy generator through the parameter loading channel, and write the attention weight parameter package into the attention calculation layer parameter storage area of ​​the current ratio strategy generator.

[0130] The server initializes one parameter loading channel, which serves as a data path connecting the parameter assembly cache area and the internal parameter storage area of ​​the current ratio strategy generator. The parameter loading channel first writes the parameter values ​​from the embedded weight parameter package segment by segment according to the offset address of the embedded layer parameter storage area, following the memory arrangement specified by the storage layout identifier.

[0131] After the embedding layer parameters are written, the parameter loading channel groups the query weight tensor, key weight tensor, and value weight tensor in the attention weight parameter package according to the attention head, and writes them sequentially into the parameter slots corresponding to each attention head in the attention calculation layer parameter storage area.

[0132] Step 156: Perform dimension verification on the embedding layer output tensor and the attention calculation layer output tensor of the current matching strategy generator. When the dimensions of the embedding layer output tensor and the attention calculation layer output tensor are consistent with the input dimensions of the instruction decoding layer, convert the current matching strategy generator after writing the parameters into the updated matching strategy generator.

[0133] The server inputs a set of test process status data with dimensions conforming to the interface requirements into the current ratio strategy generator after parameter writing. It captures the embedding layer output tensor and the attention calculation layer output tensor. The server checks the last dimension length of both the embedding layer output tensor and the attention calculation layer output tensor. If both are equal to the input dimension of the instruction decoding layer, the parameter assembly is considered successful. The server then marks the current ratio strategy generator after parameter writing as the updated ratio strategy generator and releases the parameter assembly buffer.

[0134] Step 157: The real-time process status data is processed using the embedded layer, attention calculation layer, and instruction decoding layer of the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

[0135] The server uses the raw real-time process status data obtained in step 110 as input to the updated proportioning strategy generator. The real-time process status data flows sequentially through the embedding layer, attention calculation layer, and instruction decoding layer of the updated proportioning strategy generator. The calculation logic of each layer is the same as in steps 112 to 114, but the weight parameters used in each layer have been replaced with the updated parameters. The final material proportioning instruction sequence output by the instruction decoding layer is the optimized material proportioning instruction sequence.

[0136] Step 1571: Arrange the real-time process status data into process status data batches according to the acquisition time sequence and input them into the updated proportioning strategy generator. The process status data batches are then projected and transformed by the embedding weight parameter package through the embedding layer of the updated proportioning strategy generator, and an optimized embedding vector sequence is output.

[0137] The server arranges the real-time process status data according to the acquisition time sequence, organizing it into process status data batches, with the time step of each batch maintaining its original length. After the process status data batch enters the embedding layer of the updated proportioning strategy generator, the embedding layer calls the written embedding weight parameter package to perform projection transformation. The process parameter snapshot vector at each time node is multiplied by the embedding weight parameter package matrix, outputting an optimized embedding vector. The optimized embedding vectors of all time nodes constitute an optimized embedding vector sequence.

[0138] Step 1572: The attention calculation layer of the updated matching strategy generator performs multi-head attention fusion and residual normalization on the optimized embedding vector sequence based on the attention weight parameter package, and outputs the optimized context representation sequence.

[0139] After the optimized embedding vector sequence enters the attention calculation layer, the attention calculation layer calls the query weight tensor, key weight tensor, and value weight tensor in the attention weight parameter package. Following the same multi-head attention calculation process as steps 1141 to 1145, the optimized embedding vector sequence is subjected to linear transformation, association strength calculation, exponential normalization, multi-head concatenation, parameterized linear fusion, residual connection, and layer normalization processing, and the optimized context representation sequence is output.

[0140] Step 1573: Read the optimization context representation sequence through the instruction decoding layer of the updated proportioning strategy generator, output the optimized material type identifier sequence through the material category prediction branch, and output the optimized proportioning parameter sequence through the proportioning value regression branch.

[0141] The material category prediction branch of the instruction decoding layer receives the optimization context representation vector at each time position in the optimization context representation sequence, performs material type tendency analysis and soft maximization selection at each time position, and outputs an optimized material type identifier sequence. The proportion numerical regression branch synchronously receives the optimization context representation vector, performs proportion ratio parameter regression prediction at each time position, and outputs an optimized proportion ratio parameter sequence.

[0142] Step 1574: Combine the optimized material type identifier sequence and the optimized proportion parameter sequence according to their time positions to generate an optimized material proportion instruction sequence.

[0143] The server encapsulates the optimized material type identifier and optimized ratio parameter output at each time position into an optimized material ratio tuple, and concatenates all optimized material ratio tuples in ascending order of time position index. The resulting complete sequence is the optimized material ratio instruction sequence.

[0144] Based on steps 110-150, the method further includes: Step 210: Obtain the feedback process status data generated after the execution of the optimized material ratio instruction sequence, align the feedback process status data with the real-time process status data in time sequence, and extract the feedback embedding vector sequence and the real-time embedding vector sequence through the embedding layer of the updated ratio strategy generator.

[0145] After the optimized material proportioning instruction sequence is issued to the mud mixing execution unit of the tunnel boring machine, the mixing execution unit adjusts the mud proportioning according to the instruction. The sensor array continues to monitor the adjusted process status and collects feedback process status data. After the server obtains the feedback process status data, it aligns it with the real-time process status data used in step 110 according to the time node. When the time node offset between the two exceeds the preset window, a dynamic time warping algorithm is used to adjust the correspondence.

[0146] After alignment, the feedback process status data and the real-time process status data are respectively sent to the embedding layer of the updated proportioning strategy generator. The embedding layer uses the embedding weight parameter package to perform projection transformation on the two sets of data, and outputs the feedback embedding vector sequence and the real-time embedding vector sequence respectively.

[0147] Step 220: Use the multi-head attention calculation layer of the updated ratio strategy generator to perform joint attention processing on the feedback embedding vector sequence and the real-time embedding vector sequence to generate a cross-temporal interactive attention distribution, and parse the process state offset sensitivity pattern from the cross-temporal interactive attention distribution.

[0148] The server concatenates the feedback embedding vector sequence and the real-time embedding vector sequence along the time dimension to form a joint embedding vector sequence. The multi-head attention computation layer performs self-attention computation on the joint embedding vector sequence. During computation, the attention weights between the query vector from the feedback embedding vector part and the key vector from the real-time embedding vector part constitute a cross-temporal interactive attention distribution. This distribution reflects the temporal dependency between the feedback process state and the real-time process state. The server extracts a transition matrix of attention weights between the feedback time region and the real-time time region from the cross-temporal interactive attention distribution. The column normalization result of the transition matrix constitutes a process state offset sensitivity pattern. The values ​​of each column in this pattern indicate the offset response strength of each time position of the feedback process state to each time position of the real-time process state.

[0149] Step 230: Using the process state offset sensitivity mode as an adjustment factor, the decision boundary offset metric is directionally modulated to obtain an updated fitness index. Based on the updated fitness index, the weight change constraint threshold in the elastic weight consolidation strategy is dynamically reconfigured to adjust the constraint strength of the embedded layer weight update in subsequent iterations.

[0150] The server flattens the process state offset sensitivity pattern matrix into a single adjustment factor vector, calculates the average value of each dimension of the adjustment factor vector, and multiplies the decision boundary offset metric determined in step 138 by this average value to obtain the updated fitness index.

[0151] Subsequently, the server dynamically reconfigures the weight change constraint threshold in the elastic weight consolidation strategy based on the update fitness index: when the update fitness index is greater than the preset fitness upper limit, the weight change constraint threshold is reduced by a preset shrinking ratio to tighten the parameter update constraint; when the update fitness index is less than the preset fitness lower limit, the weight change constraint threshold is increased by a preset expanding ratio to relax the parameter update constraint.

[0152] Step 240: Input the feedback process status data as the new round of real-time process status data into the updated proportioning strategy generator, and adjust the perturbation intensity of the process status confusion vector generated by the adversarial perturbation synthesizer according to the reconfigured weight change constraint threshold.

[0153] The server assigns the feedback process status data to the real-time process status data variable for the new iteration, initiating the next round of game-theoretic update processes from steps 110 to 150. In step 128 of the next round, after the adversarial perturbation synthesizer generates the initial process status confusion vector, the upper and lower limits of the perturbation intensity used by the truncation function are adaptively adjusted based on the reconfigured weight change constraint threshold: when the weight change constraint threshold increases, the perturbation intensity range narrows synchronously; when the weight change constraint threshold decreases, the perturbation intensity range widens synchronously.

[0154] Optionally, the method further includes: Step 310: Collect the process state confusion vector, the performance gain parameter, and the decision boundary offset metric corresponding to the process state confusion vector generated in multiple consecutive game update cycles to form a game trajectory data stream.

[0155] After each round of steps 110 to 150, the server stores the generated process state confusion vector, performance gain parameter, and decision boundary offset metric as a game trajectory record in the game trajectory database. After accumulating a preset number of game update cycles, all records are concatenated in cycle index order to form a game trajectory data stream. Each record in the game trajectory data stream contains a timestamp, full data of the process state confusion vector, full data of the performance gain parameter, and a scalar of the decision boundary offset metric.

[0156] Step 320: Perform temporal causal convolution processing on the game trajectory data stream to extract the perturbation evolution feature sequence, and perform feature pointwise convolution processing to extract the gain response feature sequence.

[0157] The server applies two types of convolution processing to the game trajectory data stream. The temporal causal convolution processing uses a causal convolution kernel that slides unidirectionally along the time dimension of the game trajectory data stream. The output of the causal convolution kernel at each position depends only on the process state obfuscation vector data at the current and historical positions. The outputs at all sliding positions constitute a perturbation evolution feature sequence. The feature-pointwise convolution processing uses a pointwise convolution kernel that performs a channel-wise linear combination of the performance gain parameters along the feature dimension of the game trajectory data stream. The outputs at all time positions constitute a gain response feature sequence.

[0158] Step 330: Align the perturbation evolution feature sequence and the gain response feature sequence across sequences to generate a perturbation-gain coupling representation sequence, and generate a directed association topology based on the perturbation-gain coupling representation sequence. The directed association topology uses perturbation states as nodes and gain-driven transition relationships as edges, and configures the hidden state representations corresponding to each node as the node association vector of the directed association topology.

[0159] The server performs cross-sequence mutual information alignment on the perturbation evolution feature sequence and the gain response feature sequence. It calculates pairwise mutual information values ​​between the vectors of the perturbation evolution features and the vectors of the gain response features at each time position. Using the mutual information matrix as the alignment basis, the gain response feature sequence is reordered to match the temporal pattern of the perturbation evolution feature sequence. The reordered gain response feature sequence and the perturbation evolution feature sequence are then concatenated at their corresponding time positions to generate the perturbation-gain coupling representation sequence.

[0160] Furthermore, a directed relational topology is generated based on the perturbation gain coupling representation sequence: the coupling vector at each time position of the perturbation gain coupling representation sequence is used as a node, the direction of the directed edge between nodes is determined by the time sequence, the weight of the directed edge is taken as the cosine similarity of the relational vectors of adjacent time positions, and the relational vector of each node is initialized as the corresponding coupling vector.

[0161] Step 340: When the current ratio strategy generator is updated due to the acquisition of new real-time process status data, the matching perturbation node is determined by mapping the new real-time process status data to the directed association topology, the node association vector corresponding to the matching perturbation node is obtained, and the node association vector is trend-modulated by using the weight of the transition edge between nodes. The modulated node association vector is used as the initial historical perturbation synthesizer state feature of the time-series state evolution module in the adversarial perturbation synthesizer to accelerate the generation of process status confusion vector.

[0162] When new real-time process status data arrives and triggers a new round of updates, the server converts the new real-time process status data into an embedding vector through an embedding layer. It then calculates the Euclidean distance between the embedding vector and the node association vectors of each node in the directed associative topology, selecting the node with the smallest Euclidean distance as the matching perturbation node. The server reads the node association vector stored in the matching perturbation node and, based on the transition edge weights of the directed edges originating from the matching perturbation node, multiplies the node association vector by the direction vector corresponding to the largest transition edge weight to obtain the modulated node association vector. This modulated node association vector is then written into the state buffer of the temporal state evolution module in the adversarial perturbation synthesizer as the initial historical perturbation synthesizer state features for use in step 126.

[0163] In the above technical solution, the current proportioning strategy generator is a pre-trained sequence-to-sequence decision model, and its overall architecture adopts a Transformer encoder-decoder structure. The embedding layer is essentially a trainable linear projection layer, with parameters consisting of a weight matrix multiplied by the input dimension and a uniform embedding dimension. During the pre-training phase, end-to-end training is performed using the backpropagation algorithm. During training, historical process status data of tunnel boring machines and corresponding manually labeled material proportioning instructions are used as supervision signals. The optimizer is Adam, with an initial learning rate of 0.0001, a batch size of 32, and 100 training epochs. The early stopping condition is set to terminate training if the verification loss does not decrease after 10 consecutive epochs. After pre-training, the embedding layer weight matrix is ​​serialized and stored.

[0164] The multi-head attention computation layer contains 8 parallel attention heads. Each attention head is equipped with an independent set of first query weight parameters, first key weight parameters, and first value weight parameters. Each set of weight parameters is a trainable parameter matrix. The matrix form is the unified embedding dimension multiplied by the single-head attention dimension. The single-head attention dimension is the quotient of the unified embedding dimension divided by the number of attention heads.

[0165] During pre-training, the query weight parameter group, key weight parameter group, and value weight parameter group all participate in forward attention calculation and backward gradient update. The training hyperparameters are consistent with those of the embedding layer. After training, the weight parameters of each group are saved in the form of floating-point tensors. The instruction decoding layer includes a parallel material type discrimination branch and a ratio parameter regression branch. The material type discrimination branch is a fully connected classification head with a unified embedding dimension as the input dimension and the total number of selectable material types as the output dimension. The classification head weight matrix is ​​trained using the cross-entropy loss function. During training, the material type labels are encoded using one-hot encoding. After training convergence, the weight matrix and bias vector of the fully connected classification head are fixed together.

[0166] The matching parameter regression branch consists of a fully connected regression head with an input dimension equal to the unified embedding dimension and an output dimension of 1. A sigmoid activation function is applied after the regression head to map the predicted values ​​to the 0-1 range. Mean squared error loss is used during training, and the weight vector and bias scalar of the regression head are saved after training. The layer normalization operation synchronously updates its internal learnable scaling and translation parameters during training. The scaling and translation parameters are both equal to the unified embedding dimension. Initially, the scaling parameters are initialized with all 1s, and the initial translation parameters are initialized with all 0s. These two parameters are retained within the layer normalization module after training.

[0167] In this embodiment, the adversarial perturbation synthesizer is updated alternately with the matching strategy generator in the game update strategy. The perturbation encoder consists of six stacked encoding transformation layers, each containing a linear mapping unit and a gated activation unit. The linear mapping unit is essentially a fully connected layer, and its weight matrix is ​​the product of the current layer's input feature dimension and the current layer's output feature dimension. During the initialization phase of the adversarial perturbation synthesizer, the weight matrix of the linear mapping unit is randomly initialized using a Xavier uniform distribution, and the bias vector is initialized with all zeros. The gated activation unit adopts a gated linear unit structure, containing no additional trainable parameters. The gating operation is achieved by dividing the output of the linear mapping unit into two parts along the channel dimension, applying sigmoid activation to one part, and then multiplying it element-wise with the other part.

[0168] The temporal state evolution module is implemented using a gated recurrent unit (GRU). Each GRU contains a reset gate weight matrix, an update gate weight matrix, a candidate state weight matrix, and three corresponding bias vectors. Before the adversarial perturbation synthesizer is first invoked, the weight matrices of each GRU are randomly initialized using an orthogonal initialization method. The reset gate bias and update gate bias are initialized to -1 to retain historical state information during the early training phase, while the candidate state bias is initialized to 0. All trainable parameters of the adversarial perturbation synthesizer are updated online in the game-theoretic update strategy using gradient descent adjustments as described in steps 147 and 148, without using an offline pre-trained model.

[0169] It can be understood that the formula instruction mapping model is a third functional model independent of the proportioning strategy generator and the adversarial perturbation synthesizer, responsible for converting adversarial enhancement process state data into perturbation material proportioning instructions. The conditional feature encoder includes a shared feature extraction layer and a conditional injection layer. The shared feature extraction layer consists of three sets of convolutional kernels stacked sequentially. Each set of convolutional kernels contains 64 3x3 two-dimensional convolutional kernels. The convolution operation slides along the time dimension and the feature dimension. The convolutional kernel weights are assigned using the He initialization method. Each set of convolutional kernels is followed by a batch normalization layer and a modified linear unit activation layer. The batch normalization layer learns scaling and offset parameters during training, and these parameters are saved along with the convolutional kernel weights after training.

[0170] The conditional injection layer contains one fully connected transform unit, which maps the perturbation synthesizer state features to a conditional modulation vector with the same number of channels as the enhanced state feature map. The training of the fully connected weight matrix and the overall training of the recipe instruction mapping model are carried out simultaneously.

[0171] The manifold mapping module internally maintains a set of manifold basis vectors, with 64 vectors and a dimension of 256 for each vector. The basis vectors are randomly sampled from a standard normal distribution during module initialization. During training, the values ​​of each basis vector are continuously adjusted through gradient backpropagation. After training, the manifold basis vector set is frozen and saved. The mapping network consists of four cascaded fully connected residual blocks. Each fully connected residual block contains two fully connected layers. The first fully connected layer maps the input dimension to 512, and the second fully connected layer maps 512 back to the input dimension. The residual connections element-wise add the original input of the residual block to the output of the second fully connected layer. The weights of the fully connected layers are initialized using Xavier algorithms.

[0172] The instruction reconstruction decoder consists of three alternately stacked deconvolutional upsampling layers and three channel attention adjustment layers. The transposed convolutional kernels of the deconvolutional upsampling layers are 4x4 in size with a stride of 2. Each upsampling expands the time dimension by a factor of 2. The bottleneck structure of the channel attention adjustment layers first compresses the number of channels to one-quarter of the original number of channels and then expands them back to the original number of channels. The weights of the two fully connected layers in the bottleneck structure are learned during training.

[0173] The formula instruction mapping model is pre-trained offline. The training data uses adversarial enhanced process status data and manually reviewed and confirmed material proportioning instruction pairs. The loss function is a weighted sum of material type classification loss and proportioning parameter regression loss. The optimizer is Adam, with an initial learning rate of 0.0002, a batch size of 16, and 200 training epochs. After training, all convolutional kernel weights, fully connected layer weights, manifold basis vector sets, batch normalization layer parameters, and channel attention adjustment layer parameters are persistently stored.

[0174] Furthermore, the updated matching strategy generator is a new instance of the current matching strategy generator after the parameter update in step 140. Its network structure is completely identical to the current matching strategy generator, including an embedding layer, a multi-head attention computation layer, and an instruction decoding layer. The weight parameters of the embedding layer are filled with the embedding layer weights updated by the elastic weight fixing policy constraint in step 145. The embedding layer no longer retains the original weight values ​​from the pre-training stage. The query weight parameters, key weight parameters, and value weight parameters of the multi-head attention computation layer are filled with the parameters updated by the attention weight redistribution process in step 146. The instruction decoding layer includes a material category prediction branch and a ratio numerical regression branch. The material category prediction branch corresponds to the material type discrimination branch before the update, and the ratio numerical regression branch corresponds to the ratio parameter regression branch before the update. The fully connected weights of the two branches are used as part of the overall network parameters in the game update. They receive gradient updates synchronously with the embedding layer and the attention calculation layer. The update mechanism is the same as the training process of the corresponding branches, but offline pre-training data is no longer used. Instead, the loss function is updated and adjusted online according to the strategy in step 140. The learnable scaling parameters and translation parameters of the layer normalization are also fine-tuned synchronously with other network parameters during the game update process.

[0175] In steps 310-340, the temporal causal convolution processing is implemented using 1D causal convolution with a kernel length of 3 and a kernel count of 64. The kernel weights are initialized using a Glorot uniform distribution when the game trajectory data stream is first constructed. The convolution operation slides unidirectionally from left to right along the time dimension. The output at each time point depends only on the input at the current and historical time points. The kernel weights and biases are kept fixed during online model operation through basic parameter configurations that are independent of the processed data.

[0176] The feature pointwise convolution processing is implemented using pointwise convolution with a kernel size of 1. The number of kernels equals the channel dimension of the gain response feature sequence. The kernel weights are randomly initialized upon first use and periodically recalibrated as the game progresses. The recalibration method involves updating the kernel values ​​using the performance gain parameter sequence from the most recent 10 rounds of game trajectory data stream through least-squares fitting. The cross-temporal interactive attention processing reuses the computational logic of the multi-head attention calculation layer of the updated matching strategy generator. The difference lies in that the input key vector sequence and query vector sequence are derived from the feedback embedding vector sequence and the real-time embedding vector sequence, respectively. The generation process of the cross-temporal interactive attention distribution does not require additional training parameters; it uses the attention weight parameters already existing in the updated matching strategy generator to complete the calculation.

[0177] In addition, the decision boundary analysis unit is a data processing module without training parameters. Internally, it only contains predefined rules for calculating class offsets and parameter offsets. These rules are the square root of the sum of squared vector differences, respectively. The offset weight coefficients are manually set during method initialization and remain unchanged. The performance evaluation function is defined as a logarithmic function with a temperature coefficient. Its functional form is the natural logarithm of the decision boundary offset metric multiplied by the temperature coefficient. The temperature coefficient is set at method startup and remains constant throughout the game iterations. The performance evaluation function itself does not contain trainable parameters; it only serves as the starting point for the backpropagation of the gradient of the process state confusion vector.

[0178] This application embodiment constructs a game-theoretic update closed loop between the current proportioning strategy generator and the adversarial perturbation synthesizer. It unifies the time-frequency domain representation of real-time process state data, the embedding layer weight response tensor, and the attention distribution structure pattern into input parameters for the adversarial perturbation synthesizer, enabling the process state confusion vector to accurately capture weak areas of the decision boundary of the current proportioning strategy generator. The adversarial-enhanced process state data is mapped to the material proportioning instruction space via a formula instruction mapping model. By calculating the decision boundary offset metric between the perturbed material proportioning instruction and the initial material proportioning instruction sequence, the performance gain parameter of the process state confusion vector's impact on the decision boundary is obtained. Based on this performance gain parameter, the embedding layer weights, attention distribution of the current proportioning strategy generator, and the perturbation synthesizer weights of the adversarial perturbation synthesizer are updated respectively. An elastic weight consolidation strategy is introduced when updating the embedding layer weights, constraining the weight update amplitude based on the difference between the current and historical embedding layer weights, thus enhancing the robustness of the proportioning strategy while avoiding catastrophic forgetting.

[0179] As a result, the reconstructed updated proportioning strategy generator continuously outputs optimized material proportioning instruction sequences, realizing the adaptive evolution of material proportioning strategies in adversarial training, and significantly improving the accuracy and generalization ability of proportioning instructions under complex process state fluctuations.

[0180] See Figure 2 As shown in the figure, this is a schematic diagram of the basic structure of an intelligent decision-making algorithm update server 20 provided in an embodiment of this application. The intelligent decision-making algorithm update server 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 intelligent decision-making algorithm update methods applied to material proportioning control.

[0181] 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.

[0182] See Figure 3 As shown, this figure is a functional block diagram of an intelligent decision-making algorithm update device provided in an embodiment of this application. The intelligent decision-making algorithm update device includes: The status data acquisition module is used to acquire real-time process status data of the material ratio control scenario, call the current ratio strategy generator, generate an initial material ratio instruction sequence based on the real-time process status data, and extract the embedding layer weights and attention distribution of the current ratio strategy generator. An adversarial perturbation synthesis module is used to determine the input parameters of the adversarial perturbation synthesizer based on the real-time process state data, the embedding layer weights, and the attention distribution, and to generate a process state confusion vector and perturbation synthesizer state features via the adversarial perturbation synthesizer; The element overlay processing module is used to perform element-by-element overlay processing on the process state confusion vector and the real-time process state data to generate adversarial enhanced process state data. The adversarial enhanced process state data and the perturbation synthesizer state features are input into the formula instruction mapping model to obtain the perturbation material ratio instruction, and the deviation of the perturbation material ratio instruction from the initial material ratio instruction sequence is calculated to obtain the decision boundary offset metric. The proportioning strategy update module is used to calculate the performance gain parameter of the influence of the process state confusion vector on the decision boundary based on the decision boundary offset metric, update the embedding layer weights and attention distribution of the current proportioning strategy generator based on the performance gain parameter, apply an elastic weight consolidation strategy to constrain the weight update magnitude according to the difference between the current embedding layer weights and the historical embedding layer weights when updating the embedding layer weights, and update the perturbation synthesizer weights in the adversarial perturbation synthesizer according to the performance gain parameter. The proportioning instruction generation module is used to reconstruct the current proportioning strategy generator using the updated embedding layer weights and the updated attention distribution to obtain the updated proportioning strategy generator. The real-time process status data is input into the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

[0183] 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.

[0184] 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.

[0185] 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.

[0186] 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.

[0187] 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 updating an intelligent decision-making algorithm applied to material proportioning control, characterized in that, The method includes: Acquire real-time process status data for material ratio control scenarios, call the current ratio strategy generator, generate an initial material ratio instruction sequence based on the real-time process status data, and extract the embedding layer weights and attention distribution of the current ratio strategy generator; The input parameters of the adversarial perturbation synthesizer are determined based on the real-time process state data, the embedding layer weights, and the attention distribution. The process state confusion vector and perturbation synthesizer state features are then generated via the adversarial perturbation synthesizer. The process state confusion vector and the real-time process state data are superimposed element by element to generate adversarial enhanced process state data. The adversarial enhanced process state data and the perturbation synthesizer state features are input into the formula instruction mapping model to obtain the perturbation material ratio instruction. The deviation of the perturbation material ratio instruction from the initial material ratio instruction sequence is calculated to obtain the decision boundary offset metric. The performance gain parameter of the influence of the process state confusion vector on the decision boundary is calculated based on the decision boundary offset metric. The embedding layer weights and attention distribution of the current ratio strategy generator are updated based on the performance gain parameter. When updating the embedding layer weights, an elastic weight consolidation strategy is applied to constrain the weight update magnitude according to the difference between the current embedding layer weights and the historical embedding layer weights. The perturbation synthesizer weights in the adversarial perturbation synthesizer are updated based on the performance gain parameter. The current proportioning strategy generator is reconstructed using the updated embedding layer weights and the updated attention distribution to obtain the updated proportioning strategy generator. The real-time process status data is then input into the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

2. The method according to claim 1, characterized in that, The step of invoking the current proportioning strategy generator to generate an initial material proportioning instruction sequence based on the real-time process status data, and extracting the embedding layer weights and attention distribution of the current proportioning strategy generator, includes: The real-time process status data is sorted sequentially according to the acquisition time to construct a process status data sequence indexed by time nodes; By using a preset dimension expansion mapping operator, the process parameter snapshots at each time point in the process state data sequence are transformed in feature space to generate a process state feature vector sequence with a unified embedding dimension. The process state feature vector sequence is input into the embedding layer of the current ratio strategy generator. The process state feature vector sequence is linearly projected using the set of embedding layer weight parameters stored in the embedding layer to obtain the process state embedding vector sequence. The set of embedding layer weight parameters is read out as the embedding layer weights and cached. The process state embedding vector sequence is processed by the multi-head attention calculation layer of the current ratio strategy generator to generate an attention weight distribution subgraph for each attention head, and all attention weight distribution subgraphs are combined into an attention distribution.

3. The method according to claim 2, characterized in that, The process state embedding vector sequence is processed by the multi-head attention calculation layer of the current ratio strategy generator to generate an attention weight distribution subgraph for each attention head, and all attention weight distribution subgraphs are combined into an attention distribution, including: The process state embedding vector sequence is input into the multi-head attention computation layer of the current ratio strategy generator. The process state embedding vector sequence is linearly transformed using the trained first query weight parameter group, first key weight parameter group and first value weight parameter group to obtain the query vector sequence, key vector sequence and value vector sequence respectively. For each attention head, the association strength between the query vector at each time position in the query vector sequence and the key vector at all time positions in the key vector sequence is calculated, and the association strength is scaled by a scaling factor to obtain the original attention score sequence. For each attention head, an exponential normalization transformation is performed on the original attention score sequence to generate an attention weight distribution subgraph at each time position of the attention head, and the value vector sequence is weighted and fused using the attention weight distribution subgraph to obtain a single-head attention output vector sequence. The single-head attention output vector sequences of all attention heads are concatenated along the feature dimension to obtain a multi-attention head concatenated vector sequence, and parameterized linear fusion processing is performed on the multi-attention head concatenated vector sequence to generate an attention fusion vector sequence. The attention fusion vector sequence and the process state embedding vector sequence are residually concatenated, and the result after residual concatenation is subjected to layer normalization to obtain the context representation sequence. The context representation sequence is input into the instruction decoding layer of the current proportioning strategy generator. The material type discrimination branch in the instruction decoding layer performs type tendency analysis on the context representation at each time position to obtain the material type identifier. The proportioning parameter regression branch in the instruction decoding layer performs numerical prediction on the context representation at each time position to obtain the proportioning ratio parameter. The material type identifier and the proportion parameter at the same time position are combined into a material proportion tuple, and all the material proportion tuples are concatenated in time node order to generate the initial material proportion instruction sequence; After generating attention weight distribution subgraphs for all attention heads in the multi-head attention computation layer, all attention weight distribution subgraphs are combined into an attention distribution.

4. The method according to claim 1, characterized in that, The step of determining the input parameters of the adversarial perturbation synthesizer based on the real-time process state data, the embedding layer weights, and the attention distribution, and generating a process state confusion vector and perturbation synthesizer state features via the adversarial perturbation synthesizer, includes: The real-time process status data is decomposed in the time-frequency domain to extract the time-series fundamental component and high-frequency fluctuation component as a dual-channel representation of the process status. The dual-channel representation of the process state is multiplied by the weight of the embedding layer to generate an embedding space response tensor. The embedding space response tensor is used to characterize the degree of influence of the process state change on the embedding layer decision boundary of the current ratio strategy generator. Structural analysis is performed on the attention distribution to extract attention concentration patterns and attention dispersion patterns to generate attention pattern descriptors. The attention pattern descriptors include the response peak positions and response distribution widths of each attention head. The embedded spatial response tensor and the attention pattern descriptor are channel-aligned and concatenated as input parameters for the adversarial perturbation synthesizer; The input parameters of the adversarial perturbation synthesizer are input into the perturbation encoder of the adversarial perturbation synthesizer, and hierarchical feature extraction is performed through multiple stacked coding transformation layers to generate a perturbation hidden state sequence. Each coding transformation layer contains a linear mapping unit and a gated activation unit. The perturbation hidden state sequence is passed to the temporal state evolution module of the adversarial perturbation synthesizer to read the historical perturbation synthesizer state features generated in the previous iteration, and the perturbation synthesizer state features of the current iteration are generated based on the interaction between the perturbation hidden state sequence and the historical perturbation synthesizer state features. The state characteristics of the disturbance synthesizer are decoupled by dimension, and the process condition modulation component and the ratio interference injection component are separated. The process condition modulation component is used to locally adjust the fluctuation range of the real-time process state data to generate a condition adjustment signal. The initial process state confusion vector is obtained by combining the ratio interference injection component with the operating condition adjustment signal point by point. The amplitude of the initial process state confusion vector is constrained so that the numerical change amplitude of the initial process state confusion vector is kept within a preset disturbance intensity range, thereby generating the process state confusion vector.

5. The method according to claim 1, characterized in that, The process involves element-wise superposition of the process state confusion vector and the real-time process state data to generate adversarial enhanced process state data. This adversarial enhanced process state data and the perturbation synthesizer state features are then input into a formula instruction mapping model to obtain perturbed material proportioning instructions. Finally, the deviation between the perturbed material proportioning instructions and the initial material proportioning instruction sequence is calculated to obtain a decision boundary offset metric, including: The process state confusion vector and each process parameter snapshot contained in the real-time process state data are summed at the same time position to obtain the adversarial enhancement process state data. The adversarial enhancement process state data and the perturbation synthesizer state features are input into the conditional feature encoder of the recipe instruction mapping model. The shared feature extraction layer in the conditional feature encoder performs local receptive field extraction on the adversarial enhancement process state data to generate an enhanced state feature map. The perturbation synthesizer state features are transformed into conditional modulation vectors and superimposed on the enhanced state feature map through the conditional injection layer in the conditional feature encoder to obtain conditional fusion features. The conditional fusion features are input into the manifold mapping module of the recipe instruction mapping model, and the manifold mapping module contains a mapping network composed of multiple fully connected residual blocks; Using the mapping network, the conditional fusion features are transformed by coordinate transformation based on the set of manifold basis vectors stored in the manifold mapping module, and the conditional fusion features are mapped to a low-dimensional manifold subspace of the material proportioning instruction space to generate manifold latent variables; The manifold latent variables are input into the instruction reconstruction decoder of the formula instruction mapping model, and the manifold latent variables are spatially reconstructed through a deconvolution upsampling layer and a channel attention adjustment layer to generate the disturbance material ratio instruction; Create a decision boundary analysis unit, and input the disturbing material proportioning instruction and the initial material proportioning instruction sequence into the decision boundary analysis unit to extract the material type distribution vector and proportioning parameter distribution vector in the disturbing material proportioning instruction, and extract the material type benchmark distribution vector and proportioning parameter benchmark distribution vector in the initial material proportioning instruction sequence; Calculate the type offset between the material type distribution vector and the material type reference distribution vector, and calculate the parameter offset between the proportioning parameter distribution vector and the proportioning parameter reference distribution vector; The type offset and the parameter offset are weighted and fused to obtain a comprehensive offset characterization. The comprehensive offset characterization is accumulated and statistically analyzed over time. An offset trend evaluation curve is constructed based on the rate of change and acceleration of change of the comprehensive offset characterization between adjacent time nodes. The decision boundary offset metric is determined based on the offset trend evaluation curve.

6. The method according to claim 1, characterized in that, The process includes calculating the performance gain parameter of the impact of the process state confusion vector on the decision boundary based on the decision boundary offset metric, updating the embedding layer weights and attention distribution of the current proportioning strategy generator based on the performance gain parameter, applying an elastic weight consolidation strategy to constrain the weight update magnitude according to the difference between the current embedding layer weights and historical embedding layer weights when updating the embedding layer weights, and updating the perturbation synthesizer weights in the adversarial perturbation synthesizer according to the performance gain parameter, including: The decision boundary offset metric is input into the performance evaluation function. With the goal of maximizing the decision boundary offset metric, a performance gain parameter for the process state confusion vector is generated by gradient backpropagation. The performance gain parameter indicates the effect of the process state confusion vector on pushing away from the decision boundary in the current game round. The performance gain parameter is decomposed into a strategy generator-guided component and a perturbation synthesizer-guided component. The strategy generator-guided component is used to enhance the robustness of the current ratio strategy generator, and the perturbation synthesizer-guided component is used to guide the adversarial perturbation synthesizer to generate a confusion vector. The policy update loss function is constructed using the policy generator-guided component, and the first gradient of the policy update loss function with respect to the embedding layer weights and the second gradient with respect to the attention weight parameters in the attention distribution are calculated. Before updating the embedding layer weights using the first gradient, the weight change constraint threshold corresponding to the embedding layer weights is read from the constraint parameter record of the elastic weight consolidation strategy. The first gradient is scaled dimension by dimension according to the weight change constraint threshold. The updated embedding layer weights are obtained by adding the scaled first gradient to the original embedding layer weights. The attention distribution is updated based on the second gradient. The second gradient is applied to the query weight parameters, key weight parameters, and value weight parameters in the multi-head attention calculation layer through the attention weight redistribution process to obtain the updated attention distribution. The perturbation enhancement loss function is constructed using the perturbation synthesizer guiding component, and the perturbation enhancement loss function is calculated as the perturbation synthesizer gradient with respect to the coding transform layer parameters, the temporal fusion parameter set, and the manifold mapping module parameters in the adversarial perturbation synthesizer. The perturbation synthesizer weights in the adversarial perturbation synthesizer are updated based on the perturbation synthesizer gradient. The update operation includes gradient descent adjustment of the linear mapping unit weights, gated activation unit parameters, temporal fusion parameter set, and conditional injection layer parameters in the conditional feature encoder of the coding transform layer.

7. The method according to claim 1, characterized in that, The updated material proportioning strategy generator is reconstructed using the updated embedding layer weights and the updated attention distribution to obtain an updated material proportioning strategy generator. The real-time process status data is then input into the updated material proportioning strategy generator to generate an optimized material proportioning instruction sequence, including: Read the updated embedding layer weights and the updated attention distribution, and cache the updated embedding layer weights and the updated attention distribution in the parameter assembly cache area; Obtain the internal structure definition information of the current proportioning strategy generator, wherein the internal structure definition information is used to describe the dimension information and storage layout of the embedding layer parameter interface of the current proportioning strategy generator; Based on the internal structure definition information, the updated embedding layer weights in the parameter assembly cache are reshaped using tensors to generate an embedding weight parameter package with a storage layout that conforms to the embedding layer parameter interface. The updated attention distribution in the parameter assembly cache is parsed, and the updated query weight tensor, updated key weight tensor, and updated value weight tensor are decomposed. The updated query weight tensor, updated key weight tensor, and updated value weight tensor are then format-aligned to generate an attention weight parameter package. Initialize the parameter loading channel, and write the embedded weight parameter package into the embedding layer parameter storage area of ​​the current ratio strategy generator through the parameter loading channel, and write the attention weight parameter package into the attention calculation layer parameter storage area of ​​the current ratio strategy generator; The dimensions of the embedding layer output tensor and the attention calculation layer output tensor of the current matching strategy generator are verified. When the dimensions of the embedding layer output tensor and the attention calculation layer output tensor are consistent with the input dimensions of the instruction decoding layer, the current matching strategy generator after writing the parameters is converted into the updated matching strategy generator. The real-time process status data is processed using the embedded layer, attention calculation layer, and instruction decoding layer of the updated proportioning strategy generator to generate an optimized material proportioning instruction sequence.

8. The method according to claim 7, characterized in that, The process of using the embedded layer, attention calculation layer, and instruction decoding layer of the updated proportioning strategy generator to process the real-time process status data to generate an optimized material proportioning instruction sequence includes: The real-time process status data is arranged into process status data batches according to the acquisition time sequence and input into the updated ratio strategy generator. The process status data batches are then projected and transformed by the embedding layer of the updated ratio strategy generator using the embedding weight parameter package, and an optimized embedding vector sequence is output. The attention calculation layer of the updated matching strategy generator performs multi-head attention fusion and residual normalization on the optimized embedding vector sequence based on the attention weight parameter package, and outputs an optimized context representation sequence. The updated proportioning strategy generator reads the optimization context representation sequence through the instruction decoding layer, outputs the optimized material type identifier sequence through the material category prediction branch, and outputs the optimized proportioning ratio parameter sequence through the proportioning value regression branch. The optimized material type identifier sequence and the optimized proportion parameter sequence are combined according to their time positions to generate an optimized material proportion instruction sequence.

9. A smart decision-making algorithm update server, 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 decision-making algorithm update method for material proportioning control as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the intelligent decision-making algorithm update method for material proportioning control as described in any one of claims 1-8.

Images