Drilling fluid performance evaluation method based on multi-source measurement data fusion

Through the fusion of multi-source measurement data, using technologies such as generative adversarial networks, time crystal optimization algorithms and autoencoders, the problems of insufficient data and model stability in traditional drilling fluid performance evaluation methods are solved, and more accurate and stable evaluation results are achieved.

CN118690155BActive Publication Date: 2024-11-22STAR MAP (TIANJIN) ENERGY TECHNOLOGY DEVELOPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411176400.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-22
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

Traditional drilling fluid performance evaluation methods have insufficient data samples, poor model generalization capabilities, gradient disappearance or explosion problems, and insufficient feature extraction, which affects the accuracy and stability of the evaluation.

Method used

Using a method based on multi-source measurement data fusion, the data is expanded by generating adversarial networks, and the time crystal optimization algorithm enhances model stability. The autoencoder combines the reverse reinforcement learning strategy to extract key features and use higher-order neural networks for classification.

Benefits of technology

It improves the generalization ability and accuracy of the model, enhances training stability, effectively captures and expresses data characteristics, and improves classification accuracy and robustness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure QLYQS_1
    Figure QLYQS_1
  • Figure QLYQS_2
    Figure QLYQS_2
  • Figure QLYQS_8
    Figure QLYQS_8
Patent Text Reader

Abstract

The drilling fluid performance evaluation method based on multi-source measurement data fusion belongs to a data processing method, and includes the following steps: S1, data collection and annotation; S2, data expansion using a generative adversarial network algorithm based on quantum state reconstruction; S3, optimizing neural network parameters using a time crystal optimization algorithm to achieve feature extraction; S4, using an autoencoder as a feature dimension reduction model for model training; S5, using a high-order neural network as a classifier model for model training; S6, using the trained model to process new samples to achieve drilling fluid performance evaluation. The present invention improves the generalization ability and accuracy of the model, reduces the risk of gradient disappearance or explosion, can more effectively learn and express key features of data, and is more robust when facing different data structures.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to a data processing method, and in particular relates to a drilling fluid performance evaluation method based on multi-source measurement data fusion. Background Art

[0002] Drilling fluid plays a vital role in oil drilling and production, and its performance directly affects drilling efficiency and safety. Traditional drilling fluid performance evaluation methods usually rely on limited data collected on the ground and face problems such as small data samples, low data utilization, and insufficient model generalization ability. In addition, existing evaluation technologies may have deficiencies in data processing and feature extraction, affecting the accuracy and reliability of the final evaluation.

[0003] In view of the multi-source measurement data fusion evaluation method, the Chinese invention patent with application number CN202310640997.6 proposes a carbon efficiency ratio measurement method and a low-carbon performance evaluation method for low-ambient temperature air source heat pump (chilled water) units. The carbon efficiency ratio is the carbon dioxide equivalent emitted by the low-temperature air source heat pump (chilled water) unit to provide unit heat. The carbon efficiency ratio is obtained by calculating the ratio of the carbon footprint of the low-temperature air source heat pump (chilled water) unit to the sum of the functional units of the low-temperature air source heat pump (chilled water) unit. Among them, the carbon footprint is the total amount of greenhouse gas carbon emissions in the low-temperature air source heat pump (chilled water) unit system expressed in carbon dioxide equivalent, and the sum of functional units is the sum of heat provided by the low-temperature air source heat pump (chilled water) unit during its service life. This measurement method takes into account both the total carbon emissions of the product and the functional contribution of the product, and distributes the carbon emissions to each functional unit through the ratio, which is scientific and reasonable. The present invention also discloses a method for evaluating the low-carbon performance of a low-ambient temperature air source heat pump (chilled water) unit using the carbon efficiency ratio. The Chinese invention patent with application number CN202211450578.8 proposes a method for evaluating the insulation withstand voltage performance of XLPE cables based on the polarization response conversion factor. The method first uses a non-destructive polarization / depolarization current method (PDC) to measure the polarization / depolarization current of the cable under test, and then measures the low-frequency complex dielectric constant of the cable. The polarization response conversion factor is calculated from the measured data, and finally the withstand voltage level evaluation coefficient of the XLPE cable is calculated, thereby realizing a non-destructive evaluation of the withstand voltage level of the XLPE cable. The method of the present invention can provide a reference for the selection of withstand voltage tests and insulation status evaluation of XLPE cables before commissioning. The Chinese invention patent with application number CN202311774843.2 proposes a performance evaluation model and a performance evaluation method for an energy storage system. The energy storage system performance evaluation model of the present application includes a battery module, a signal monitoring and processing module, a control module and a thermal management module; the signal monitoring and processing module is used to collect the status information of the battery cell, and output the first system parameter and the battery cell temperature; the control module controls the working mode of the battery module and outputs the liquid cooling state parameter used to control the cooler in the thermal management module; the thermal management module is used to output the second system parameter and adjust the temperature of the battery module. The present application can characterize the performance of the energy storage system under different strategies, output the first system parameter, the second system parameter and the battery cell temperature, and the technical indicators for evaluation are relatively rich, thereby effectively reducing the testing cost and labor cost, improving the efficiency of product development, effectively improving the risk identification ability, and reducing the risk of failure.

[0004] The above technical solutions are innovative, but the following problems still need to be further solved: 1. Traditional methods may not fully utilize the available multi-source data, resulting in insufficient accuracy and comprehensiveness of the evaluation results. 2. In the case of limited data samples, traditional machine learning models often have poor generalization capabilities due to insufficient training samples. 3. Some neural network structures may encounter gradient vanishing or exploding problems, affecting the stability of training and model performance. 4. Traditional feature extraction methods may not be able to fully capture all key data features, limiting classification performance. Summary of the invention

[0005] The purpose of the present invention is to propose a drilling fluid performance evaluation method based on multi-source measurement data fusion, which increases the diversity and quantity of training samples by generating data expanded by adversarial networks, helping to improve the generalization ability and accuracy of the model; the stability of the model in the training process is enhanced by the time crystal optimization algorithm, reducing the risk of gradient disappearance or explosion; the autoencoder combined with the inverse reinforcement learning strategy can more effectively learn and express the key features of the data, which is beneficial to the subsequent classification accuracy; the high-order neural network makes the model more robust when facing different data structures.

[0006] The technical solution adopted by the present invention is:

[0007] The drilling fluid performance evaluation method based on multi-source measurement data fusion includes the following steps:

[0008] S1. Data collection and annotation:

[0009] The data comes from data collected on the ground, and the collected data is stored in a standardized data format, specifically in JSON format; the collected data is manually annotated;

[0010] S2, data expansion:

[0011] A generative adversarial network algorithm based on quantum state reconstruction is used for data expansion. The training process of the algorithm is as follows: S21, initialize the parameters of the generator and the discriminator; S22, input drilling fluid performance data: convert classical data into quantum data through quantum data encoding; S23, the generator tries to generate new drilling fluid data: use random noise as the input of the generator, and then output the generated data; S24, the discriminator receives data from the generator and real drilling fluid data, and then correctly classifies these data sources into generated data and real data; S25, in adversarial training, the parameters of the generator and the discriminator are updated alternately. The goal of the generator is to deceive the discriminator so that the discriminator misjudges the generated data as real data; the discriminator strives to improve its own judgment ability, and the update process is constrained by the loss function L; S26, after multiple rounds of adversarial training, the quantum layer in the generator is optimized. By adjusting the parameters of the quantum layer, the representation ability of quantum data is optimized and the characteristic simulation of the data is enhanced; S27, iterative steps S22-S26 are repeated until the preset stop iteration condition is met, indicating that the model training is completed;

[0012] S3. Feature extraction model training:

[0013] The expanded data is input into the feature extraction model for training the feature extraction model, and a time crystal optimization algorithm is used to optimize the neural network parameters. The training process of the algorithm is as follows: S31, initialize the parameters of the neural network model, and define the initial state parameters of the neural network, including weights and biases, and determine the cycle of the algorithm. The cycle of the algorithm determines the time interval for each parameter update, wherein the parameter initialization is set by random initialization; S32, in each cycle, according to the current network state and historical performance, a time crystal state is mapped, and this state represents a periodic change pattern of the network parameters; S33, based on time The parameter adjustment rules of the time crystal state are used, and the parameter update of each cycle is based on the current time crystal state; S34, external perturbations are periodically used to simulate the dynamic response of the physical time crystal when it is affected by external factors, so as to adjust the parameter update strategy of the next cycle to enhance the adaptability and robustness of the network; S35, the feedback system is used to monitor the network performance after each cycle, and the cycle is adjusted according to the performance feedback to optimize the overall training process; S36, after multiple cycles, the long-term performance of the network is evaluated and the loss adjustment function is updated; S37, iterative steps S32-S36 are repeated until the preset stop iteration condition is met, indicating that the model training is completed;

[0014] S4. Feature dimensionality reduction model training:

[0015] The feature-extracted data is input into the feature dimensionality reduction model for training. The autoencoder is used as the feature dimensionality reduction model. The training process of the autoencoder algorithm is as follows: S41. The weights and biases of the autoencoder are initialized by small random numbers to destroy the symmetry so that the training process can proceed smoothly; S42. In the forward propagation process, the feature-extracted data is input into the encoder and converted into a low-dimensional hidden representation. The decoder attempts to reconstruct the original input and outputs the reconstructed data; S43. The reconstruction loss and L2 regularization loss are used to form a total loss function to evaluate the performance of the model; S44. Repeat the iterative steps S42-S43 until the preset stop iteration condition is met, indicating that the model training is completed;

[0016] S5. Classifier model training:

[0017] The data after feature dimension reduction is input into the classifier model to train the classifier model. A high-order neural network is used as the classifier model. The training process of the high-order neural network classification algorithm is as follows: S51. The weights and biases of the high-order neural network are initialized using the Glorot method; S52. The data after feature dimension reduction is loaded into the high-order neural network model. The data is forward propagated in the network through multiple hidden layers. Each hidden layer uses the ReLU activation function to increase the nonlinear processing capability; S53. The loss function is used to evaluate the difference between the model output and the actual label, and the model parameters are updated through back propagation; S54. The adaptive sub-gradient method is applied to update the weights, and the learning rate of each parameter is adjusted to cope with the sparsity of the data; S55. Repeat the iterative steps S51-S54 until the preset stop iteration condition is met, indicating that the model training is completed;

[0018] S6. Drilling fluid performance evaluation:

[0019] The trained model is used to process new samples to achieve drilling fluid performance evaluation. Suppose a new drilling fluid sample data is collected, which is first processed by the feature extraction model, and then further compressed by the data dimension reduction model to improve processing efficiency, and finally the classification result is obtained through the classifier model. The evaluation classification categories include: level I performance, level II performance, level III performance, level IV performance, and level V performance.

[0020] Furthermore, the data expansion in step S2 specifically adopts the following training process:

[0021] In the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a generative adversarial network algorithm based on quantum state reconstruction for data expansion, and the generative adversarial network algorithm based on quantum state reconstruction includes two components: a generator (G) and a discriminator (D). The generator is responsible for generating new data that is as close to the actual drilling fluid performance as possible, and the task of the discriminator is to distinguish whether the input data comes from a real data set or data generated by the generator. In view of the problem that the traditional generative adversarial network may have low generation quality on specific types of data, the present invention adopts a topologically encoded loss function improvement, utilizes the characteristics of the quantum state, and reconstructs the topological structure of the data through a quantum computing model to ensure the diversity and accuracy of the generated data.

[0022] S201, initialize the parameters of the generator G and the discriminator D, where the weight initialization method of the generator G is expressed as:

[0023] ;

[0024] The weight initialization method of the discriminator D is expressed as:

[0025] ;

[0026] In the formula, represents a normal distribution with a mean of 0 and a standard deviation of 1; is the number of nodes in the generator input layer; is the weight of the generator G; is the weight of the discriminator D;

[0027] S202, the input drilling fluid performance data is firstly encoded by quantum data, and the classical data is converted into quantum data q, and the encoding function The implementation is expressed as:

[0028] ;

[0029] In the formula, represents quantum superposition; q is quantum data; xe is the original drilling fluid performance data; is the encoding function; The original drilling fluid performance data xe The quantum probability distribution of the elements converted into; It is quantum ground state;

[0030] The calculation method of the quantum probability distribution converted from the i-th element of the original drilling fluid performance data xe is expressed as:

[0031] ;

[0032] In the formula, is the temperature parameter; is the i-th element of the original drilling fluid performance data xe; is the jth element of the original drilling fluid performance data xe; e is the natural exponential function;

[0033] S203, the generator attempts to generate new drilling fluid data, specifically using random noise z as the input of the generator G, and outputting the generated data, which can be expressed as:

[0034] ;

[0035] In the formula, are the parameters of the generator; z is random noise; is a generator function; is the data generated by the generator;

[0036] S204, the discriminator receives the data from the generator and the real drilling fluid data. The task of the discriminator D is to correctly classify these data sources into generated data and real data, which can be expressed as:

[0037] ;

[0038] Where y is the output of the discriminator, are the parameters of the discriminator;

[0039] S205. In adversarial training, the parameters of the generator and the discriminator are updated alternately. The goal of the generator is to deceive the discriminator so that it misjudges the generated data as real data; the discriminator strives to improve its own judgment ability. The update process is constrained by the loss function L, which is expressed as:

[0040] ;

[0041] Where z is the input noise; Generator functions; is the discriminator function; is the real data distribution; is the noise distribution; It is the expected operation; is a symbol that obeys a specific distribution; and It is the dynamically adjusted weight factor of the discriminator and the generator, which is set by dynamic update;

[0042] S206. After multiple rounds of adversarial training, the quantum layer in the generator is optimized by adjusting the parameters of the quantum layer. , optimize the representation ability of quantum data, enhance the characteristic simulation of data, and the parameter adjustment method is expressed as:

[0043] ;

[0044] In the formula, is the loss function, is the parameter of the adjusted quantum layer, is the quantum layer parameter learning rate, Express The gradient of

[0045] S207, repeating iterative steps S202-S206 until the preset stop iteration condition is met, indicating that the model training is completed.

[0046] Furthermore, the feature extraction model training in step S3 has the following specific process:

[0047] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention uses a 4-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a time crystal optimization algorithm to optimize the parameters of the neural network to achieve the training of the neural network feature bar model. Inspired by the theoretical physical structure, the time crystal shows a periodic structure in the time dimension, and can maintain a stable period of internal dynamics even under external influences. The time crystal optimization algorithm simulates the periodic changes of the time crystal in the time dimension to adjust and optimize the parameters of the neural network.

[0048] S301, initializing the parameters of the neural network model, defining the initial parameters of the neural network, including weight w and bias b, and determining the cycle Ts of the algorithm, wherein the cycle of the algorithm determines the time interval for each parameter update, wherein the initialization of the parameters is set in a random initialization manner;

[0049] S302. In each period Ts, a time crystal state is mapped according to the current network state and historical performance. This state represents a periodic change pattern of network parameters, which is expressed as:

[0050] ;

[0051] In the formula, and are the weight and bias of the tth iteration respectively, is the loss adjustment function, is a periodic state adjusted according to the loss function; Indicates the current iteration number;

[0052] S303, parameter adjustment rules based on the time crystal state, the parameter update of each cycle is performed according to the current time crystal state, and the parameter adjustment amount calculation method is expressed as:

[0053] ;

[0054] In the formula, and are the weight and bias of the tth iteration respectively, and are the parameter adjustments for weight and bias, respectively. is the learning rate, and is the gradient of the loss function with respect to weights and biases;

[0055] The gradient calculation method of the loss function with respect to the weight is expressed as:

[0056] ;

[0057] In the formula, is the ith input of the neural network, is the true value of the i-th sample; is the network's response to the i-th input The predicted value is obtained by extracting features through the preset Softmax function and classifying the feature vector;

[0058] The gradient calculation method of the loss function with respect to the bias is expressed as:

[0059] ;

[0060] In the formula, ndc is the number of samples input in the current batch;

[0061] S304, periodically adopt external disturbances to simulate the dynamic response of the physical time crystal when it is affected by external factors, so as to adjust the parameter update strategy of the next period to enhance the adaptability and robustness of the network, which is expressed as:

[0062] ;

[0063] In the formula, is the amplitude adjustment function, r is the loss change rate, is the external disturbance, is the disturbance intensity;

[0064] Amplitude adjustment function The calculation of dynamically adjusts the amplitude based on the loss change rate, and the calculation method is expressed as:

[0065] ;

[0066] In the formula, is the basic amplitude, and is the adjustment coefficient, which is used to control the sensitivity and amplitude of the amplitude adjustment;

[0067] The calculation method of loss change rate r can be expressed as:

[0068] ;

[0069] In the formula, and are the losses of the previous cycle and the current cycle respectively;

[0070] S305, using the feedback system to monitor the network performance after each cycle, and adjusting the cycle T according to the performance feedback to optimize the overall training process, which is expressed as:

[0071] ;

[0072] In the formula, is the learning rate parameter that adjusts the cycle length, is the adjusted period value, is the loss of the current cycle, is the target loss, used to calculate the benchmark for periodic updates;

[0073] S306. After multiple cycles, evaluate the long-term performance of the network and update the loss adjustment function, expressed as:

[0074] ;

[0075] In the formula, is the adaptive rate parameter, is the loss adjustment function before updating, is the updated loss adjustment function, is the loss of the previous cycle, is the loss of the current cycle;

[0076] S307, repeating iterative steps S302-S306 until the preset stop iteration condition is met, indicating that the model training is completed.

[0077] Furthermore, the feature dimension reduction model training in step S4 has the following specific process:

[0078] The feature-extracted data is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention uses an autoencoder as the feature dimensionality reduction model, and the autoencoder consists of two parts: an encoder and a decoder. The encoder is responsible for converting high-dimensional input data into low-dimensional codes, and the decoder attempts to reconstruct the original input data from this low-dimensional code. In order to improve the performance of the autoencoder, the present invention utilizes an inverse reinforcement learning strategy. The encoding process not only learns the compressed representation of the data, but also learns how to optimize the encoding process by maximizing the expected reward, thereby better capturing the key information in the data.

[0079] S401, the weights and biases of the autoencoder are initialized by small random numbers to destroy the symmetry so that the training process can proceed smoothly, wherein the initialization method of the weights and biases of the autoencoder is expressed as:

[0080] ;

[0081] In the formula, and are the weights and biases of the autoencoder of the pth layer respectively; is the element in the i-th row and j-th column of the weight matrix of the p-th layer of the autoencoder; It is the autoencoder The number of neurons in the layer; are random values ​​generated from a standard normal distribution;

[0082] S402, in the forward propagation process, the data after feature extraction Input to the encoder and converted into a low-dimensional hidden representation , the decoder tries to reconstruct the original input and outputs the reconstructed data, expressed as:

[0083] ;

[0084] In the formula, is the linear transformation output of the decoder layer p; is the activation output of the decoder layer p; For decoder The activation output of the layer; is the ReLU activation function;

[0085] S403, use reconstruction loss and L2 regularization loss to form the total loss function , to evaluate the performance of the model, the calculation method is expressed as:

[0086] ;

[0087] In the formula, is the total loss; represents the square of the L2 norm; xd is the input data of the autoencoder, that is, the feature vector obtained after feature extraction; is the activation output of the i-th neuron in the p-th layer of the decoder; is the output reconstructed by the autoencoder; is the weight of sparse regularization;

[0088] S404, repeating iterative steps S402-S403 until a preset stop iteration condition is met, indicating that the model training is completed.

[0089] Furthermore, the classifier model training in step S5 has the following specific process:

[0090] The data after feature dimension reduction is input into the classifier model for training the classifier model. The present invention adopts a high-order neural network as the classifier model to effectively capture and learn complex nonlinear relationships. The present invention adopts an adaptive sub-gradient optimization algorithm to improve the traditional gradient descent method, and can more effectively process sparse data by adjusting the learning rate of each parameter.

[0091] S501. Initialize the weights and biases of the high-order neural network using the Glorot method, expressed as:

[0092] ;

[0093] In the formula, is the element in the i-th row and j-th column of the q-th layer weight matrix of the high-order neural network; and are the weights and biases of the qth layer of the high-order neural network respectively; and is the number of neurons in the previous layer and the current layer; re is uniformly distributed A random number drawn from

[0094] S502: The data after feature dimension reduction Loaded into the high-order neural network model, the data is propagated forward in the network, passing through multiple hidden layers, each of which uses the ReLU activation function to increase nonlinear processing capabilities, expressed as:

[0095] ;

[0096] In the formula, is the linear transformation output of the qth layer; is the output of the qth layer after being processed by the ReLU activation function; is the first Layer output;

[0097] S503, using the loss function to evaluate the difference between the model output and the actual label, and updating the model parameters through back propagation, expressed as:

[0098] ;

[0099] In the formula, is the probability distribution predicted by the model; C is the total number of sample categories; is the true label distribution; is the cross entropy loss function, evaluating the predicted probability The actual probability The difference between is the gradient of the loss function with respect to the weights; is the gradient of the loss function with respect to the network output;

[0100] The calculation method is expressed as:

[0101] ;

[0102] In the formula, is the probability predicted by the model, is the true label;

[0103] is the derivative of the ReLU activation function, indicating that ReLU The derivative is 1 when it is greater than 0, otherwise it is 0, expressed as:

[0104] ;

[0105] It is a linear transformation Relative to weight The gradient of is approximately equal to the output of the previous layer , expressed as:

[0106] ;

[0107] S504, applying the adaptive sub-gradient method to update the weights, adjusting the learning rate of each parameter to cope with the sparsity of the data, expressed as:

[0108] ;

[0109] In the formula, is the weight parameter of the high-order neural network before updating, is the weight parameter of the updated high-order neural network, is the global learning rate; It is the accumulated square of gradient before update, which is used to adjust the learning rate; To avoid division by zero, a small constant;

[0110] The update is implemented according to the square of the gradient, and the calculation method is expressed as:

[0111] ;

[0112] In the formula, is the updated accumulated square gradient;

[0113] S505. Repeat iterative steps S501-S504 until a preset stop iteration condition is met, indicating that the model training is completed.

[0114] Further, in step S205, and The update strategy is expressed as:

[0115] ;

[0116] In the formula, is the classification accuracy of the discriminator on the most recent batch of data; is the learning rate factor; , are the weight factors of the discriminator and generator at the tth iteration respectively; In the The weight factors of the discriminator and generator for each iteration;

[0117] The expected operation is to batch the samples and The average calculation is obtained, which is expressed as:

[0118] ;

[0119] Where mcs is the batch size; From the real data distribution The i-th sample drawn from ; is the noise distribution The i-th noise vector extracted from ; is a batch of fake data produced by the generator.

[0120] Furthermore, in step S206, the gradient Calculated by the chain rule, it is expressed as:

[0121] ;

[0122] In the formula, is the partial derivative of the loss function L with respect to the quantum data q; is the quantum data to quantum layer parameters The partial derivative of .

[0123] Further, in step S302, the loss adjustment function Exponential decay is used to reflect the impact of loss on the state, so that high loss corresponds to lower state impact, thereby slowing down the fluctuation of parameters. The calculation method is expressed as:

[0124] ;

[0125] In the formula, is the loss function, is a hyperparameter that affects the impact of historical weights on the current update;

[0126] Loss Function The calculation method is expressed as:

[0127] ;

[0128] In the formula, is the true value of the i-th sample, is the network's response to the i-th input The predicted value is calculated by the preset Softmax function using the weight W and bias b of the current neural network.

[0129] Furthermore, in step S303, the updating method of weights and biases can be expressed as:

[0130] ;

[0131] In the formula, and are the weight and bias parameters for the t+1th time, is the resonance factor, which depends on the parameter change rate of the adjacent layer and the historical change of the parameters of this layer, expressed as:

[0132] ;

[0133] In the formula, and They are -1st floor and +1 layer weight update amount, It is the current weights of the layer, is a hyperparameter that adjusts the resonance response.

[0134] Further, in step S403, based on the total loss function, the gradient descent method is used to update the parameters; for the weight parameters of the autoencoder, the gradient of the loss function is The calculation method is expressed as:

[0135] ;

[0136] In the formula, is the sign function of the activation output, used for L1 regularization;

[0137] The gradient descent method is used to update weights and biases, and the update method is expressed as:

[0138] ;

[0139] In the formula, and are the weights and biases of the p-th layer of the updated autoencoder respectively; is the weight adjustment based on the importance of dynamic features; and are the weight and bias of the pth layer of the autoencoder before updating; is the learning rate at the tth iteration; and is the gradient of the loss function with respect to weights and biases;

[0140] The calculation method of the weight adjustment based on the importance of dynamic features is expressed as:

[0141] ;

[0142] In the formula, is the dynamic feature learning rate, is the Softmax function, represents the importance score of the p-th layer feature, calculated as the loss function The absolute value of the partial derivative with respect to this feature; is the weight adjustment based on the importance of dynamic features;

[0143] Feature Importance Scoring The calculation method is expressed as:

[0144] ;

[0145] In the formula, is a function of taking absolute value; is the symbol of partial derivative; It is the partial derivative of the loss function with respect to the input features of the p-th layer.

[0146] Beneficial effects of the present invention:

[0147] 1. The present invention uses quantum computing technology to improve the traditional generative adversarial network and generates high-quality data samples through quantum data encoding and the characteristics of quantum states.

[0148] 2. The present invention utilizes the concept of time crystals in physics and simulates the changes of time crystals in the time dimension through periodic parameter updates to enhance the stability and training effect of the neural network.

[0149] 3. In the process of feature dimensionality reduction, the present invention not only learns the compressed representation of the data, but also optimizes the encoding process to maximize the expected reward, which helps to capture the key information in the data.

[0150] 4. The present invention adopts an adaptive sub-gradient method for the training of high-order neural networks, and effectively handles the sparsity of data and improves the learning efficiency of the model by adjusting the learning rate of each parameter.

[0151] Based on the above innovations, the technical effects are as follows:

[0152] 1. The data expanded by generating adversarial networks increases the diversity and quantity of training samples, helping to improve the generalization ability and accuracy of the model.

[0153] 2. The time crystal optimization algorithm enhances the stability of the model during training and reduces the risk of gradient disappearance or explosion.

[0154] 3. The autoencoder combined with the inverse reinforcement learning strategy can more effectively learn and express the key features of the data, which is beneficial to the subsequent classification accuracy.

[0155] 4. The adaptive learning rate adjustment of high-order neural networks is optimized according to the characteristics of the data, making the model more robust when facing different data structures. DETAILED DESCRIPTION

[0156] The drilling fluid performance evaluation method based on multi-source measurement data fusion includes the following steps:

[0157] S1. Data collection and annotation

[0158] The data in the present invention is derived from data collected on the ground. The collected data is stored in a standardized data format, specifically in JSON format. In this embodiment, the attributes of the data include:

[0159] ax: timestamp, indicating the specific time of data collection;

[0160] ay: sensor ID, used to identify the sensor from which the data comes;

[0161] az: temperature in degrees Celsius;

[0162] aa: pressure, in Pascals;

[0163] ab: pH value, no unit;

[0164] ac: electrical conductivity, in Siemens per meter;

[0165] ad: viscosity, in poise;

[0166] ae: density in kilograms per cubic meter;

[0167] af: water content, expressed as a percentage;

[0168] ag: The concentration of chemical components, the specific chemicals depend on the formulation of the drilling fluid.

[0169] It should be emphasized that this embodiment is only used to illustrate a data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0170] Manually label the collected data, and the labeling categories include: Level I performance, Level II performance, Level III performance, Level IV performance, and Level V performance;

[0171] The Grade I performance refers to the highest performance level of drilling fluid, which can remain stable under extreme geological and temperature conditions, that is, it has excellent chemical stability, extremely low corrosion and superior lubricity under high temperature and high pressure environment; the Grade II performance is slightly lower than Grade I, but can still cope with most drilling environments, performs well under a certain range of temperature and pressure conditions, and is suitable for most conventional drilling operations; the Grade III performance is a medium level of drilling fluid performance, suitable for mild geological environments and operating conditions, and may degrade under extreme conditions. The Grade IV performance is limited to a certain extent, suitable for shallower drilling or less complex geological conditions, and may not work effectively under complex or harsh conditions; the Grade V performance is the lowest performance level, usually only suitable for very specific and non-harsh conditions, and may require frequent replacement and maintenance.

[0172] S2. Data expansion

[0173] In the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a generative adversarial network algorithm based on quantum state reconstruction for data expansion, and the generative adversarial network algorithm based on quantum state reconstruction includes two components: a generator (G) and a discriminator (D). The generator is responsible for generating new data that is as close to the actual drilling fluid performance as possible, and the task of the discriminator is to distinguish whether the input data comes from a real data set or data generated by the generator. In view of the problem that the traditional generative adversarial network may have low generation quality on specific types of data, the present invention adopts a topologically encoded loss function improvement, utilizes the characteristics of the quantum state, and reconstructs the topological structure of the data through a quantum computing model to ensure the diversity and accuracy of the generated data.

[0174] The training process of the algorithm is as follows:

[0175] S201, initialize the parameters of the generator G and the discriminator D, where the weight initialization method of the generator G is expressed as:

[0176] ;

[0177] The weight initialization method of the discriminator D is expressed as:

[0178] ;

[0179] In the formula, represents a normal distribution with a mean of 0 and a standard deviation of 1; nds is the number of nodes in the generator input layer; is the weight of the generator G; is the weight of the discriminator D;

[0180] S202, the input drilling fluid performance data is firstly encoded by quantum data, and the classical data is converted into quantum data q, and the encoding function The implementation is expressed as:

[0181] ;

[0182] In the formula, represents quantum superposition; q is quantum data; xe is the original drilling fluid performance data; is the encoding function; is the quantum probability distribution converted from the i-th element of the original drilling fluid performance data xe; is the i-th quantum ground state;

[0183] The calculation method of the quantum probability distribution converted from the i-th element of the original drilling fluid performance data xe is expressed as:

[0184] ;

[0185] In the formula, is the temperature parameter; is the i-th element of the original drilling fluid performance data xe; The jth element of the original drilling fluid performance data xe; e is a natural exponential function; preferably, Set to 3.

[0186] S203, the generator attempts to generate new drilling fluid data, specifically using random noise z as the input of the generator G and outputting the generated data , which can be expressed as:

[0187] ;

[0188] In the formula, are the parameters of the generator; z is random noise; is a generator function; is the data generated by the generator;

[0189] S204, the discriminator receives the data from the generator and the real drilling fluid data. The task of the discriminator D is to correctly classify these data sources into generated data and real data, which can be expressed as:

[0190] ;

[0191] Where y is the output of the discriminator, are the parameters of the discriminator;

[0192] S205. In adversarial training, the parameters of the generator and the discriminator are updated alternately. The goal of the generator is to deceive the discriminator so that it misjudges the generated data as real data; the discriminator strives to improve its own judgment ability. The update process is constrained by the loss function L, which is expressed as:

[0193] ;

[0194] Where z is the input noise; Generator functions; is the discriminator function; is the real data distribution; is the noise distribution; It is the expected operation; is a symbol that obeys a specific distribution; and It is the dynamically adjusted weight factor of the discriminator and the generator, which is set by dynamic update;

[0195] and The update strategy is expressed as:

[0196] ;

[0197] In the formula, is the classification accuracy of the discriminator on the most recent batch of data; is the learning rate factor; , are the weight factors of the discriminator and generator at the tth iteration respectively; are the weight factors of the discriminator and generator at the t+1th iteration respectively; preferably, Set to 0.005.

[0198] The expected operation is to batch the samples and The average calculation is obtained, which is expressed as:

[0199] ;

[0200] Where mcs is the batch size; From the real data distribution The i-th sample drawn from ; is the noise distribution The i-th noise vector extracted from ; is a batch of fake data produced by the generator;

[0201] S206. After multiple rounds of adversarial training, the quantum layer in the generator is optimized by adjusting the parameters of the quantum layer. , optimize the representation ability of quantum data, enhance the characteristic simulation of data, and the parameter adjustment method is expressed as:

[0202] ;

[0203] In the formula, is the loss function, is the parameter of the adjusted quantum layer, is the quantum layer parameter learning rate, Express Preferably, Set to 0.004.

[0204] gradient Calculated by the chain rule, it is expressed as:

[0205] ;

[0206] In the formula, is the partial derivative of the loss function L with respect to the quantum data q; is the quantum data to quantum layer parameters The partial derivative of .

[0207] S207, repeating iterative steps S202-S206 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0208] S3. Feature extraction model training

[0209] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention uses a 4-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a time crystal optimization algorithm to optimize the parameters of the neural network to achieve the training of the neural network feature bar model. Inspired by the theoretical physical structure, the time crystal shows a periodic structure in the time dimension, and can maintain a stable period of internal dynamics even under external influences. The time crystal optimization algorithm simulates the periodic changes of the time crystal in the time dimension to adjust and optimize the parameters of the neural network.

[0210] The training process of the neural network model based on time crystal optimization is as follows:

[0211] S301, initializing the parameters of the neural network model, defining the initial parameters of the neural network, including weight w and bias b, and determining the cycle T of the algorithm, which determines the time interval between each parameter update, wherein the initialization of the parameters is set in a random initialization manner;

[0212] S302. In each period Ts, a time crystal state is mapped according to the current network state and historical performance. This state represents a periodic change pattern of network parameters, which is expressed as:

[0213] ;

[0214] In the formula, and are the weight and bias of the tth iteration respectively, is the loss adjustment function, is a periodic state adjusted according to the loss function; Indicates the current iteration number; preferably, the parameter is set to 0.5.

[0215] Loss Adjustment Function Exponential decay is used to reflect the impact of loss on the state, so that high loss corresponds to lower state impact, thereby slowing down the fluctuation of parameters. The calculation method is expressed as:

[0216] ;

[0217] In the formula, is the loss function, is a hyperparameter that affects the impact of historical weights on the current update;

[0218] Loss Function The calculation method is expressed as:

[0219] ;

[0220] In the formula, is the true value of the i-th sample; is the network's response to the i-th input The predicted value is obtained by extracting the features through the preset Softmax function and classifying the feature vector.

[0221] S303, parameter adjustment rules based on the time crystal state, the parameter update of each cycle is performed according to the current time crystal state, and the parameter adjustment amount calculation method is expressed as:

[0222] ;

[0223] In the formula, and are the weight and bias of the tth iteration respectively, and are the parameter adjustments for weight and bias, respectively. is the learning rate, and is the gradient of the loss function with respect to the weights and biases; preferably, Set to 0.001.

[0224] The gradient calculation method of the loss function with respect to the weight is expressed as:

[0225] ;

[0226] In the formula, is the ith input of the neural network, is the true value of the i-th sample; is the network's response to the i-th input The predicted value is obtained by extracting features through the preset Softmax function and classifying the feature vector;

[0227] The gradient calculation method of the loss function with respect to the bias is expressed as:

[0228] ;

[0229] In the formula, ndc is the number of samples input in the current batch;

[0230] The update method of weights and biases can be expressed as:

[0231] ;

[0232] In the formula, and are the weight and bias parameters for the t+1th time, is the resonance factor, which depends on the parameter change rate of the adjacent layer and the historical change of the parameters of this layer, and can be expressed as:

[0233] ;

[0234] In the formula, and They are -1st floor and +1 layer weight update amount, It is the current weights of the layer, is a hyperparameter for adjusting the resonance response. Preferably, Set to 0.05.

[0235] S304, periodically adopt external disturbances to simulate the dynamic response of the physical time crystal when it is affected by external factors, and adjust the parameter update strategy of the next period to enhance the adaptability and robustness of the network, which is expressed as:

[0236] ;

[0237] In the formula, is the amplitude adjustment function, r is the loss change rate, is the external disturbance, is the disturbance intensity; preferably, Set to 0.3.

[0238] Amplitude adjustment function The calculation of dynamically adjusts the amplitude based on the loss change rate, and the calculation method is expressed as:

[0239] ;

[0240] In the formula, is the basic amplitude, and is an adjustment coefficient used to control the sensitivity and amplitude of the amplitude adjustment; preferably, Set to 0.5, and 0.4 and 2 respectively.

[0241] The calculation method of loss change rate r can be expressed as:

[0242] ;

[0243] In the formula, and are the losses of the previous cycle and the current cycle respectively;

[0244] S305, using the feedback system to monitor the network performance after each cycle, and adjusting the cycle T according to the performance feedback to optimize the overall training process, which is expressed as:

[0245] ;

[0246] In the formula, is the learning rate parameter that adjusts the cycle length, is the adjusted period value, is the loss of the current period, is the target loss, used to calculate the benchmark for periodic updates; preferably, Set to 0.0005.

[0247] S306. After multiple cycles, evaluate the long-term performance of the network and update the loss adjustment function, expressed as:

[0248] ;

[0249] In the formula, is the adaptive rate parameter, is the loss adjustment function before updating, is the updated loss adjustment function, is the loss of the previous cycle, is the loss of the current cycle;

[0250] S307, repeating iterative steps S302-S306 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0251] S4. Feature dimensionality reduction model training

[0252] The feature-extracted data is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention uses an autoencoder as the feature dimensionality reduction model, and the autoencoder consists of two parts: an encoder and a decoder. The encoder is responsible for converting high-dimensional input data into low-dimensional codes, and the decoder attempts to reconstruct the original input data from this low-dimensional code. In order to improve the performance of the autoencoder, the present invention utilizes an inverse reinforcement learning strategy. The encoding process not only learns the compressed representation of the data, but also learns how to optimize the encoding process by maximizing the expected reward, thereby better capturing the key information in the data.

[0253] The training process of the autoencoder algorithm is as follows:

[0254] S401, the weights and biases of the autoencoder are initialized by small random numbers to destroy the symmetry so that the training process can proceed smoothly, wherein the initialization method of the weights and biases of the autoencoder is expressed as:

[0255] ;

[0256] In the formula, and are the weights and biases of the autoencoder of the pth layer respectively; is the element in the i-th row and j-th column of the weight matrix of the p-th layer of the autoencoder; It is the autoencoder The number of neurons in the layer; are random values ​​generated from a standard normal distribution;

[0257] S402, in the forward propagation process, the data after feature extraction Input to the encoder and converted into a low-dimensional hidden representation , the decoder tries to reconstruct the original input and outputs the reconstructed data, expressed as:

[0258] ;

[0259] In the formula, is the linear transformation output of the decoder layer p; is the activation output of the decoder layer p; For decoder The activation output of the layer; is the ReLU activation function;

[0260] S403, use reconstruction loss and L2 regularization loss to form the total loss function , to evaluate the performance of the model, the calculation method is expressed as:

[0261] ;

[0262] In the formula, is the total loss; represents the square of the L2 norm; xd is the input data of the autoencoder, that is, the feature vector obtained after feature extraction; is the activation output of the i-th neuron in the p-th layer of the decoder; is the output reconstructed by the autoencoder; is the weight of sparse regularization; preferably, Set to 0.3.

[0263] Based on the total loss function, the gradient descent method is used to update the parameters; for the weight parameters of the autoencoder, its gradient with respect to the loss function The calculation method is expressed as:

[0264] ;

[0265] In the formula, is the sign function of the activation output, used for L1 regularization.

[0266] The gradient descent method is used to update weights and biases, and the update method is expressed as:

[0267] ;

[0268] In the formula, and are the weights and biases of the p-th layer of the updated autoencoder respectively; is the weight adjustment based on the importance of dynamic features; and are the weight and bias of the pth layer of the autoencoder before updating; is the learning rate at the tth iteration; and is the gradient of the loss function with respect to weights and biases;

[0269] The calculation method of the weight adjustment based on the importance of dynamic features is expressed as:

[0270] ;

[0271] In the formula, is the dynamic feature learning rate, is the Softmax function, represents the importance score of the p-th layer feature, calculated as the loss function The absolute value of the partial derivative with respect to this feature; is the weight adjustment based on the importance of dynamic features;

[0272] Feature Importance Scoring The calculation method is expressed as:

[0273] ;

[0274] In the formula, is a function of taking absolute value; is the symbol of partial derivative; It is the partial derivative of the loss function with respect to the input features of the p-th layer.

[0275] S404. Repeat iterative steps S402-S403 until the preset stop iteration condition is met, which means that the model training is completed; in this embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations, which is set to 1000 times.

[0276] S5. Classifier model training

[0277] The data after feature dimension reduction is input into the classifier model for training the classifier model. The present invention adopts a high-order neural network as the classifier model to effectively capture and learn complex nonlinear relationships. The present invention adopts an adaptive sub-gradient optimization algorithm to improve the traditional gradient descent method, and can more effectively process sparse data by adjusting the learning rate of each parameter.

[0278] The training process of the high-order neural network classification algorithm is as follows:

[0279] S501. Initialize the weights and biases of the high-order neural network using the Glorot method, expressed as:

[0280] ;

[0281] In the formula, is the element in the i-th row and j-th column of the q-th layer weight matrix of the high-order neural network; and are the weights and biases of the qth layer of the high-order neural network respectively; and is the number of neurons in the previous layer and the current layer; re is uniformly distributed A random number drawn from

[0282] S502: The data after feature dimension reduction Loaded into the high-order neural network model, the data is propagated forward in the network, passing through multiple hidden layers, each of which uses the ReLU activation function to increase nonlinear processing capabilities, expressed as:

[0283] ;

[0284] In the formula, is the linear transformation output of the qth layer; is the output of the qth layer after being processed by the ReLU activation function; is the first Layer output;

[0285] S503, using the loss function to evaluate the difference between the model output and the actual label, and updating the model parameters through back propagation, expressed as:

[0286] ;

[0287] In the formula, is the probability distribution predicted by the model; C is the total number of sample categories; is the true label distribution; is the cross entropy loss function, evaluating the predicted probability The actual probability The difference between is the gradient of the loss function with respect to the weights; is the gradient of the loss function with respect to the network output;

[0288] The calculation method is expressed as:

[0289] ;

[0290] In the formula, is the probability predicted by the model, is the true label;

[0291] is the derivative of the ReLU activation function, indicating that ReLU The derivative is 1 when it is greater than 0, otherwise it is 0, expressed as:

[0292] ;

[0293] It is a linear transformation Relative to weight The gradient of is approximately equal to the output of the previous layer , expressed as:

[0294] ;

[0295] S504, applying the adaptive sub-gradient method to update the weights, adjusting the learning rate of each parameter to cope with the sparsity of the data, expressed as:

[0296] ;

[0297] In the formula, is the weight parameter of the high-order neural network before updating, is the weight parameter of the updated high-order neural network, is the global learning rate; It is the accumulated square of gradient before update, which is used to adjust the learning rate; To avoid division by a small constant of zero; preferably, Set to 0.005, Set to .

[0298] The update is implemented according to the square of the gradient, and the calculation method is expressed as:

[0299] ;

[0300] In the formula, is the updated accumulated square gradient;

[0301] S505. Repeat iterative steps S501-S504 until a preset stop iteration condition is met, indicating that the model training is completed; in this embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0302] S6. Drilling fluid performance evaluation

[0303] The trained model is used to process new samples to achieve drilling fluid performance evaluation. In this embodiment, a new piece of drilling fluid sample data is collected, which is first processed by a feature extraction model, and then further compressed by a data dimension reduction model to improve processing efficiency, and finally a classification result is obtained by a classifier model. In this embodiment, the evaluation classification categories include: level I performance, level II performance, level III performance, level IV performance, and level V performance.

Claims

1. Drilling fluid performance evaluation method based on multi-source measurement data fusion, It is characterized in that The following steps are involved: S1. Data collection and annotation: The data comes from data collected on the ground, and the collected data is stored in a standardized data format, specifically in JSON format; the collected data is manually annotated; S2, data expansion: A generative adversarial network algorithm based on quantum state reconstruction is used for data augmentation. The training process of the algorithm is as follows: S21, initialize the parameters of the generator and the discriminator; S22, input drilling fluid performance data: convert classical data into quantum data through quantum data encoding; S23, the generator tries to generate new drilling fluid data: use random noise as the input of the generator, and then output the generated data; S24, the discriminator receives the data from the generator and the real drilling fluid data, and then correctly classifies these data sources into generated data and real data; S25, in adversarial training, the parameters of the generator and the discriminator are updated alternately. The goal of the generator is to deceive the discriminator so that the discriminator misjudges the generated data as real data; The discriminator strives to improve its own judgment ability, and the updating process is constrained by the loss function L; S26, after multiple rounds of adversarial training, the quantum layer in the generator is optimized, and the representation ability of quantum data is optimized by adjusting the parameters of the quantum layer, and the characteristic simulation of the data is enhanced; S27, iterative steps S22-S26 are repeated until the preset stop iteration condition is met, indicating that the model training is completed; S3. Feature extraction model training: The expanded data is input into the feature extraction model for training the feature extraction model, and a time crystal optimization algorithm is used to optimize the neural network parameters. The training process of the algorithm is as follows: S31, initialize the parameters of the neural network model, and define the initial state parameters of the neural network, including weights and biases, and determine the cycle of the algorithm. The cycle of the algorithm determines the time interval for each parameter update, wherein the parameter initialization is set by random initialization; S32, in each cycle, according to the current network state and historical performance, a time crystal state is mapped, and this state represents a periodic change pattern of the network parameters; S33, based on time The parameter adjustment rules of the time crystal state are used, and the parameter update of each cycle is based on the current time crystal state; S34, external perturbations are periodically used to simulate the dynamic response of the physical time crystal when it is affected by external factors, so as to adjust the parameter update strategy of the next cycle to enhance the adaptability and robustness of the network; S35, the feedback system is used to monitor the network performance after each cycle, and the cycle is adjusted according to the performance feedback to optimize the overall training process; S36, after multiple cycles, the long-term performance of the network is evaluated and the loss adjustment function is updated; S37, iterative steps S32-S36 are repeated until the preset stop iteration condition is met, indicating that the model training is completed; S4. Feature dimensionality reduction model training: The feature-extracted data is input into the feature dimensionality reduction model for training. The autoencoder is used as the feature dimensionality reduction model. The training process of the autoencoder algorithm is as follows: S41. The weights and biases of the autoencoder are initialized by small random numbers to destroy the symmetry so that the training process can proceed smoothly; S42. In the forward propagation process, the feature-extracted data is input into the encoder and converted into a low-dimensional hidden representation. The decoder attempts to reconstruct the original input and outputs the reconstructed data; S43. The reconstruction loss and L2 regularization loss are used to form a total loss function to evaluate the performance of the model; S44. Repeat the iterative steps S42-S43 until the preset stop iteration condition is met, indicating that the model training is completed; S5. Classifier model training: The data after feature dimension reduction is input into the classifier model to train the classifier model. A high-order neural network is used as the classifier model. The training process of the high-order neural network classification algorithm is as follows: S51. The weights and biases of the high-order neural network are initialized using the Glorot method; S52. The data after feature dimension reduction is loaded into the high-order neural network model. The data is forward propagated in the network through multiple hidden layers. Each hidden layer uses the ReLU activation function to increase the nonlinear processing capability; S53. The loss function is used to evaluate the difference between the model output and the actual label, and the model parameters are updated through back propagation; S54. The adaptive sub-gradient method is applied to update the weights, and the learning rate of each parameter is adjusted to cope with the sparsity of the data; S55. Repeat the iterative steps S51-S54 until the preset stop iteration condition is met, indicating that the model training is completed; S6. Drilling fluid performance evaluation: The trained model is used to process new samples to achieve drilling fluid performance evaluation. Suppose a new drilling fluid sample data is collected, which is first processed by the feature extraction model, and then further compressed by the data dimension reduction model to improve processing efficiency, and finally the classification result is obtained through the classifier model. The evaluation classification categories include: level I performance, level II performance, level III performance, level IV performance, and level V performance.

2. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 1, It is characterized in that The data expansion in step S2 specifically adopts the following training process: S201, initialize the parameters of the generator G and the discriminator D, where the weight initialization method of the generator G is expressed as: ; The weight initialization method of the discriminator D is expressed as: ; In the formula, represents a normal distribution with a mean of 0 and a standard deviation of 1; is the number of nodes in the generator input layer; is the weight of the generator G; is the weight of the discriminator D; S202, the input drilling fluid performance data is firstly encoded by quantum data, and the classical data is converted into quantum data q, and the encoding function The implementation is expressed as: ; In the formula, represents quantum superposition; q is quantum data; xe is the original drilling fluid performance data; is the encoding function; The original drilling fluid performance data xe The quantum probability distribution of the elements converted into; It is quantum ground state; Original drilling fluid performance data xe The calculation method of the quantum probability distribution converted into elements is expressed as: ; In the formula, is the temperature parameter; The original drilling fluid performance data xe elements; is the jth element of the original drilling fluid performance data xe; e is the natural exponential function; S203, the generator attempts to generate new drilling fluid data, specifically using random noise z as the input of the generator G, and outputting the generated data, which can be expressed as: ; In the formula, are the parameters of the generator; z is random noise; is a generator function; The data generated for the generator; S204, the discriminator receives the data from the generator and the real drilling fluid data. The task of the discriminator D is to correctly classify these data sources into generated data and real data, which can be expressed as: ; In the formula, is the output of the discriminator, are the parameters of the discriminator; S205. In adversarial training, the parameters of the generator and the discriminator are updated alternately. The goal of the generator is to deceive the discriminator so that it misjudges the generated data as real data; the discriminator strives to improve its own judgment ability. The update process is constrained by the loss function L, which is expressed as: ; Where z is the input noise; Generator functions; is the discriminator function; is the real data distribution; is the noise distribution; It is the expected operation; is a symbol that obeys a specific distribution; and It is the dynamically adjusted weight factor of the discriminator and the generator, which is set by dynamic update; S206. After multiple rounds of adversarial training, the quantum layer in the generator is optimized by adjusting the parameters of the quantum layer. , optimize the representation ability of quantum data, enhance the characteristic simulation of data, and the parameter adjustment method is expressed as: ; In the formula, is the loss function, is the parameter of the adjusted quantum layer, is the quantum layer parameter learning rate, Express The gradient of S207, repeating iterative steps S202-S206 until the preset stop iteration condition is met, indicating that the model training is completed.

3. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 1, It is characterized in that The feature extraction model training in step S3 has the following specific process: S301, initializing the parameters of the neural network model, defining the initial parameters of the neural network, including weight w and bias b, and determining the cycle Ts of the algorithm, wherein the cycle of the algorithm determines the time interval for each parameter update, wherein the initialization of the parameters is set in a random initialization manner; S302. In each period Ts, a time crystal state is mapped according to the current network state and historical performance. This state represents a periodic change pattern of network parameters, which is expressed as: ; In the formula, and Respectively The weights and biases for the iterations, is the loss adjustment function, is a periodic state adjusted according to the loss function; Indicates the current iteration number; S303, parameter adjustment rules based on the time crystal state, the parameter update of each cycle is performed according to the current time crystal state, and the parameter adjustment amount calculation method is expressed as: ; In the formula, and Respectively The weights and biases for the iteration, and are the parameter adjustments for weight and bias, respectively. is the learning rate, and is the gradient of the loss function with respect to weights and biases; The gradient calculation method of the loss function with respect to the weight is expressed as: ; In the formula, The neural network Inputs, It is The true value of samples; Is the network Input The predicted value is obtained by extracting features through the preset Softmax function and classifying the feature vector; The gradient calculation method of the loss function with respect to the bias is expressed as: ; In the formula, ndc is the number of samples input in the current batch; S304, periodically adopt external disturbances to simulate the dynamic response of the physical time crystal when it is affected by external factors, so as to adjust the parameter update strategy of the next period to enhance the adaptability and robustness of the network, which is expressed as: ; In the formula, is the amplitude adjustment function, r is the loss change rate, is the external disturbance, is the disturbance intensity; Amplitude adjustment function The calculation of dynamically adjusts the amplitude based on the loss change rate, and the calculation method is expressed as: ; In the formula, is the basic amplitude, and is the adjustment coefficient, which is used to control the sensitivity and amplitude of the amplitude adjustment; The calculation method of loss change rate r can be expressed as: ; In the formula, and are the losses of the previous cycle and the current cycle respectively; S305, using the feedback system to monitor the network performance after each cycle, and adjusting the cycle according to the performance feedback to optimize the overall training process, which is expressed as: ; In the formula, is the learning rate parameter that adjusts the cycle length, is the adjusted period value, is the loss of the current period, is the target loss, used to calculate the benchmark for periodic updates; S306. After multiple cycles, evaluate the long-term performance of the network and update the loss adjustment function, expressed as: ; In the formula, is the adaptive rate parameter, is the loss adjustment function before updating, is the updated loss adjustment function, is the loss of the previous cycle, is the loss of the current cycle; S307, repeating iterative steps S302-S306 until the preset stop iteration condition is met, indicating that the model training is completed.

4. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 1, It is characterized in that The specific process of feature dimension reduction model training in step S4 is as follows: S401, the weights and biases of the autoencoder are initialized by small random numbers to destroy the symmetry so that the training process can proceed smoothly, wherein the initialization method of the weights and biases of the autoencoder is expressed as: ; In the formula, and are the weights and biases of the autoencoder of the pth layer respectively; is the weight matrix of the pth layer of the autoencoder Line Elements of a column; It is the autoencoder The number of neurons in the layer; are random values ​​generated from a standard normal distribution; S402, in the forward propagation process, the data after feature extraction Input to the encoder and converted into a low-dimensional hidden representation , the decoder tries to reconstruct the original input and outputs the reconstructed data, expressed as: ; In the formula, is the linear transformation output of the decoder layer p; is the activation output of the decoder layer p; For decoder The activation output of the layer; is the ReLU activation function; S403, using reconstruction loss and Regularization loss together constitutes the total loss function , to evaluate the performance of the model, the calculation method is expressed as: ; In the formula, is the total loss; express The square of the norm; xd is the input data of the autoencoder, that is, the feature vector obtained after feature extraction; is the activation output of the i-th neuron in the p-th layer of the decoder; is the output reconstructed by the autoencoder; is the weight of sparse regularization; S404, repeating iterative steps S402-S403 until a preset stop iteration condition is met, indicating that the model training is completed.

5. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 1, It is characterized in that The specific process of classifier model training in step S5 is as follows: S501. Initialize the weights and biases of the high-order neural network using the Glorot method, expressed as: ; In the formula, is the element in the i-th row and j-th column of the q-th layer weight matrix of the high-order neural network; and are the weights and biases of the qth layer of the high-order neural network respectively; and is the number of neurons in the previous layer and the current layer; re is uniformly distributed A random number drawn from S502: The data after feature dimension reduction Loaded into the high-order neural network model, the data is propagated forward in the network, passing through multiple hidden layers, each of which uses the ReLU activation function to increase nonlinear processing capabilities, expressed as: ; In the formula, is the linear transformation output of the qth layer; is the output of the qth layer after being processed by the ReLU activation function; is the first Layer output; S503, using the loss function to evaluate the difference between the model output and the actual label, and updating the model parameters through back propagation, expressed as: ; In the formula, is the probability distribution predicted by the model; C is the total number of sample categories; is the true label distribution; is the cross entropy loss function, evaluating the predicted probability The actual probability The difference between is the gradient of the loss function with respect to the weights; is the gradient of the loss function with respect to the network output; The calculation method is expressed as: ; In the formula, is the probability predicted by the model, is the true label; is the derivative of the ReLU activation function, indicating that ReLU The derivative is 1 when it is greater than 0, otherwise it is 0, expressed as: ; It is a linear transformation Relative to weight The gradient of is approximately equal to the output of the previous layer , expressed as: ; S504, applying the adaptive sub-gradient method to update the weights, adjusting the learning rate of each parameter to cope with the sparsity of the data, expressed as: ; In the formula, is the weight parameter of the high-order neural network before updating, is the weight parameter of the updated high-order neural network, is the global learning rate; It is the accumulated square of gradient before update, which is used to adjust the learning rate; To avoid division by zero, a small constant; The update is implemented according to the square of the gradient, and the calculation method is expressed as: ; In the formula, is the updated accumulated square gradient; S505. Repeat iterative steps S501-S504 until a preset stop iteration condition is met, indicating that the model training is completed.

6. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 2, It is characterized in that In step S205, and The update strategy is expressed as: ; In the formula, is the classification accuracy of the discriminator on the most recent batch of data; is the learning rate factor; , are the weight factors of the discriminator and generator at the tth iteration respectively; In the The weight factors of the discriminator and generator for this iteration; The expected operation is to batch the samples and The average calculation is obtained, which is expressed as: ; Where mcs is the batch size; From the real data distribution The i-th sample drawn from ; From the noise distribution The i-th noise vector extracted from ; is a batch of fake data produced by the generator.

7. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 2, It is characterized in that In step S206, the gradient Calculated by the chain rule, it is expressed as: ; In the formula, is the partial derivative of the loss function L with respect to the quantum data q; is the quantum data to quantum layer parameters The partial derivative of .

8. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 3, It is characterized in that In step S302, the loss adjustment function Exponential decay is used to reflect the impact of loss on the state, so that high loss corresponds to lower state impact, thereby slowing down the fluctuation of parameters. The calculation method is expressed as: ; In the formula, is the loss function, is a hyperparameter that affects the impact of historical weights on the current update; Loss Function The calculation method is expressed as: ; In the formula, is the true value of the i-th sample, is the network's response to the i-th input The predicted value is calculated by the preset Softmax function using the weight W and bias b of the current neural network.

9. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 3, It is characterized in that In step S303, the updating method of weights and biases is expressed as: ; In the formula, and Respectively The weight and bias parameters of each time, is the resonance factor, which depends on the parameter change rate of the adjacent layer and the historical change of the parameters of this layer, expressed as: ; In the formula, and They are -1st floor and +1 layer weight update amount, It is the current weights of the layer, is a hyperparameter that adjusts the resonance response.

10. The drilling fluid performance evaluation method based on multi-source measurement data fusion according to claim 4, It is characterized in that In step S403, based on the total loss function, the gradient descent method is used to update the parameters; for the weight parameters of the autoencoder, the gradient of the loss function is The calculation method is expressed as: ; In the formula, is the sign function of the activation output, used for L1 regularization; The gradient descent method is used to update weights and biases, and the update method is expressed as: ; In the formula, and are the weights and biases of the p-th layer of the updated autoencoder respectively; is the weight adjustment based on the importance of dynamic features; and are the weight and bias of the pth layer of the autoencoder before updating; is the learning rate at the tth iteration; and is the gradient of the loss function with respect to weights and biases; The calculation method of the weight adjustment based on the importance of dynamic features is expressed as: ; In the formula, is the dynamic feature learning rate, is the Softmax function, represents the importance score of the p-th layer feature, calculated as the loss function The absolute value of the partial derivative with respect to this feature; is the weight adjustment based on the importance of dynamic features; Feature Importance Scoring The calculation method is expressed as: ; In the formula, is a function of taking absolute value; is the symbol of partial derivative; It is the partial derivative of the loss function with respect to the input features of the p-th layer.