Computer-aided high-precision pressure detection method for aerospace gas cylinders

Abstract

The present invention belongs to the technical field of gas cylinder pressure detection, and provides a computer-aided high-precision pressure detection method for aerospace gas cylinders. The method first collects comprehensive data through monitoring equipment, high-precision sensors and pressure detection equipment, and performs feature engineering processing on the comprehensive data to extract high-quality feature data. Then, an AE-MSGRU model is used for pressure prediction. The model captures the short-term, medium-term and long-term dependencies of the gas cylinder pressure data respectively through a multi-level GRU layer, and combines the attention mechanism to accurately identify the key points in the pressure data, and further optimizes the time series feature data. Through the above steps, the method effectively improves the accuracy and reliability of aerospace gas cylinder pressure detection, ensures the safety of gas cylinders under different use conditions, and provides strong technical support for aerospace gas cylinder pressure detection.

Description

Technical Field

[0001] The invention relates to the technical field of gas cylinder pressure detection, in particular to a computer-aided high-precision pressure detection method for aerospace gas cylinders. Background Art

[0002] With the rapid development of the aerospace industry, the safety and reliability requirements for gas cylinders are becoming increasingly higher; traditional gas cylinder pressure detection methods usually rely on manual or traditional pressure measuring instruments, which have problems such as low detection accuracy, cumbersome operation, and poor real-time performance; during the use of aircraft, the pressure state of gas cylinders is crucial to safe operation; existing technologies often find it difficult to meet the needs of real-time monitoring and high-precision evaluation of pressure data, especially in high-pressure environments and rapidly changing conditions, these traditional methods often cannot provide sufficiently accurate pressure predictions and risk assessments; therefore, there is an urgent need for a high-precision, real-time, computer-aided pressure detection method to improve the safety performance and detection reliability of gas cylinders. Summary of the invention

[0003] In view of the above situation, the present invention provides a computer-aided high-precision pressure detection method for aerospace gas cylinders, aiming to solve the problems of insufficient pressure detection accuracy, poor real-time performance and limited data processing capability existing in the prior art; the method comprehensively collects gas cylinder data through high-precision pressure sensors and detection equipment, and generates high-quality feature data through feature engineering technology; an AE-MSGRU model is designed, which combines a multi-layer GRU structure with an attention mechanism to effectively capture the complex time series features in the gas cylinder pressure data; the model is implemented to accurately predict the gas cylinder pressure, and evaluate the qualification, ultimate pressure bearing capacity and pressure resistance performance of the gas cylinder based on the prediction results; this method significantly improves the accuracy and real-time performance of gas cylinder pressure detection, enhances the reliability of the system, and provides advanced technical support for the safe management of aerospace gas cylinders.

[0004] The present invention provides a computer-aided high-precision pressure detection method for aerospace gas cylinders, the method comprising the following steps:

[0005] Step S1: Data collection: using monitoring equipment, high-precision sensors and pressure detection equipment to collect initial pressure data, ultimate pressure data, withstand pressure data, transformed pressure data and environmental parameter data to obtain comprehensive data;

[0006] Step S2: Feature engineering: Use feature engineering methods to process comprehensive data to obtain high-quality feature data;

[0007] Step S3: Model design: Establish an AE-MSGRU model. The AE-MSGRU model includes an input layer, a multi-level GRU layer, an attention mechanism layer, and an output layer. The multi-level GRU layer includes a short-term GRU layer, a medium-term GRU layer, and a long-term GRU layer. Each layer of the AE-MSGRU model is composed of neurons. High-quality feature data is input into the AE-MSGRU model, and the pressure estimation value is obtained through layer-by-layer calculation.

[0008] Step S4: Model training and optimization: Use high-quality feature data to train the AE-MSGRU model, use the cross entropy loss function to evaluate the performance of the AE-MSGRU model, and use the stochastic gradient descent algorithm to optimize the parameters of the AE-MSGRU model to obtain a trained AE-MSGRU model;

[0009] Step S5: Pressure detection and qualification assessment: The pressure prediction value is obtained through the trained AE-MSGRU model, and the gas cylinder qualification test, gas cylinder ultimate pressure bearing capacity test and gas cylinder pressure resistance performance test are performed according to the pressure prediction value.

[0010] Furthermore, step S2 specifically includes the following steps:

[0011] Step S21: Extracting features of initial pressure data: extracting time features, constructing interactive features, and expanding polynomial features on the initial pressure data to generate a feature set of initial pressure data;

[0012] Step S22: Extraction of ultimate pressure bearing data features: performing pressure-time feature extraction, material strain feature construction and dynamic feature extraction on the ultimate pressure bearing data to generate an ultimate pressure bearing data feature set;

[0013] Step S23: Extracting withstand voltage data features: extracting key features from the withstand voltage data through trend feature analysis, periodic feature extraction and statistical feature calculation to generate a withstand voltage data feature set;

[0014] Step S24: transform pressure data feature extraction: applying frequency domain feature extraction and autoregressive feature analysis to generate a transform pressure data feature set;

[0015] Step S25: Environmental parameter data feature extraction: Generate an environmental parameter data feature set by constructing an environmental index and generating environmental pressure interaction features;

[0016] Step S26: Feature selection and dimensionality reduction: Use the feature importance analysis method to screen the initial pressure data feature set, the ultimate pressure data feature set, the pressure resistance data feature set, the transformed pressure data feature set and the environmental parameter data feature set, and finally form high-quality feature data.

[0017] Furthermore, step S3 specifically includes the following steps:

[0018] Step S31: Use neurons in the input layer to integrate high-quality feature data into a time series data set and transmit it to the multi-level GRU layer;

[0019] Step S32: The short-term dependency, the medium-term dependency and the long-term dependency are captured respectively by the short-term GRU layer, the medium-term GRU layer and the long-term GRU layer, and the time series features are extracted from the time series data set of the input layer to obtain the comprehensive time series feature data. The formula of the multi-level GRU layer is as follows;

[0020] Multi-level GRU layer formula:

[0021] Update Gate :

[0022] ;

[0023] in, ReLU activation function. Represents the input data, i.e., the time series dataset; The update gate weight matrix representing the input data; represents the hidden state of the previous time step; The update gate weight matrix representing the hidden state of the previous time step; represents the bias term of the update gate;

[0024] Reset Gate :

[0025] ;

[0026] in, The reset gate weight matrix representing the input data; The reset gate weight matrix representing the hidden state of the previous time step; represents the bias term of the reset gate;

[0027] Candidate status :

[0028] ;

[0029] in, represents the hyperbolic tangent activation function, The candidate state weight matrix representing the input data; The candidate state weight matrix representing the hidden state of the previous time step; A bias term representing the candidate state; It means element-wise multiplication;

[0030] Final State :

[0031] ;

[0032] in, Indicates the part that retains the hidden state of the previous time step, and the weight is controlled by the update gate; The weight of the part representing the new candidate state is also controlled by the update gate;

[0033] Step S33: introducing an attention mechanism layer, identifying the key points of the pressure data by analyzing the comprehensive time series feature data, assigning weights to the key points of the pressure data, further optimizing the comprehensive time series feature data, and obtaining the time series feature data optimized by the attention mechanism;

[0034] Step S34: The output layer converts the time series feature data optimized by the attention mechanism into air pressure values ​​through the ReLU activation function to generate a pressure estimation value.

[0035] Further, step S32 specifically includes the following steps:

[0036] Step S321: Capturing short-term dependencies: The short-term GRU layer processes the time series data set through a gating mechanism, focusing on the pressure change characteristics of the short-term time span, and generating an initial feature hidden state;

[0037] Step S322: Apply Dropout layer 1: Add Dropout layer 1 after the short-term GRU layer to randomly mask the output of neurons to obtain the initial feature optimized hidden state;

[0038] Step S323: mid-term dependency capture: the mid-term GRU layer further extracts the mid-term time span features in the initial feature optimization hidden state, captures the mid-term change trend of the cylinder pressure over time, and generates a mid-term feature hidden state;

[0039] Step S324: Apply Dropout layer 2: Add Dropout layer 2 after the mid-term GRU layer to randomly mask the output of neurons and obtain the mid-term feature optimized hidden state;

[0040] Step S325: long-term dependency capture: the long-term GRU layer deeply extracts the long-term time span features in the mid-term feature optimization hidden state, captures the deep trend features of the cylinder pressure change, and generates the deep feature hidden state;

[0041] Step S326: Apply Dropout layer 3: Add Dropout layer 3 after the long-term GRU layer to randomly mask the output of neurons, obtain the optimized deep feature hidden state, and finally generate comprehensive time series feature data.

[0042] Further, step S33 specifically includes the following steps:

[0043] Step S331: Calculate the self-attention score: perform linear transformation on the comprehensive time series feature data to generate the query matrix, key matrix and value matrix; obtain the initial attention score by calculating the dot product of the query matrix and the key matrix, and scale the initial attention score to obtain the attention score matrix; use the softmax function to normalize the attention score matrix to generate the attention weight matrix. The attention score calculation formula is as follows:

[0044] .Linear transformation:

[0045] , , ;

[0046] in, Represents comprehensive time series feature data; represents the query matrix, The weight matrix representing the query matrix; represents the key matrix, represents the weight matrix of the bond matrix; represents the value matrix, The weight matrix representing the value matrix;

[0047] .Attention score matrix:

[0048] ;

[0049] in, Represents the key matrix Dimensions; represents the transpose of the key matrix; represents the scaling factor;

[0050] .Attention weight matrix:

[0051] ;

[0052] in, represents the weight parameter, represents a regular term;

[0053] Step S332: weighted sum aggregation: apply the attention weight matrix to the value matrix to obtain a weighted value matrix; perform weighted summation on the weighted value matrix to generate weighted optimized time series feature data, the formula is as follows:

[0054] ;

[0055] in, Represents the time series feature data after weighted optimization; represents the weighted value matrix; Indicates the number of elements in the comprehensive time series feature data;

[0056] Step S333: residual connection and layer normalization: add the weighted optimized temporal feature data to the comprehensive temporal feature data to obtain the data after residual connection; perform layer normalization on the data after residual connection to finally obtain the temporal feature data after attention mechanism optimization, the formula is as follows:

[0057] ;

[0058] in, Represents the data after residual connection, Represents comprehensive time series feature data;

[0059] ;

[0060] in, Represents the time series feature data after the attention mechanism is optimized, Layer normalization function;

[0061] When calculating the self-attention score, the system calculates the attention score based on the relevance of the features and normalizes it through the softmax function, giving higher weights to features with strong relevance; then, in the weighting and aggregation steps, these weights are applied to the value matrix, ultimately making the key features occupy a more important position in the final time series feature data, thereby achieving the amplification and optimization of important information.

[0062] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0063] The present invention provides a computer-aided high-precision pressure detection method for aerospace gas cylinders, which achieves high precision and efficiency in aerospace gas cylinder pressure detection; firstly, through precise feature extraction and optimization processing, the high quality of data input is ensured, laying a solid foundation for subsequent modeling; secondly, the method divides the GRU network into three layers of short-term dependency, medium-term dependency and long-term dependency, and each layer focuses on capturing the features of a specific time span; compared with the traditional single-layer GRU model, it can more comprehensively analyze the pressure changes of gas cylinders on different time scales, significantly improving the accuracy and precision of pressure detection; by capturing dependencies at different time levels, this technology effectively improves the processing capability of gas cylinder pressure data and ensures high-precision pressure prediction.

[0064] The introduced attention mechanism further optimizes the feature extraction process by automatically identifying key points in the cylinder pressure data and assigning higher weights; by applying residual connections and layer normalization to the optimized time series feature data, the gradient vanishing and overfitting problems are solved, and the stability and robustness of the model are enhanced. This mechanism not only improves the detection accuracy of the cylinder's ultimate pressure bearing capacity and pressure resistance performance, but also ensures that the model can accurately identify and evaluate pressure changes when the pressure data fluctuates greatly, thereby improving the reliability of the overall detection.

[0065] The Dropout mechanism is reasonably applied after each layer of GRU units to effectively prevent the overfitting of the model during the training process; by randomly shielding some neuron outputs, the model complexity is reduced, the risk of overfitting is reduced, and the model shows stronger robustness in different application scenarios; in summary, the present invention can complete the pressure detection of aerospace gas cylinders more stably and accurately, ensure the safety of gas cylinders under various conditions of use, and provide reliable technical support for gas cylinder pressure detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a module schematic diagram of the AE-MSGRU model proposed in the present invention. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0068] Embodiment 1, the present invention provides a computer-aided high-precision pressure detection method for aerospace gas cylinders, the method comprising the following steps:

[0069] Step S1: Data collection: using monitoring equipment, high-precision sensors and pressure detection equipment to collect initial pressure data, ultimate pressure data, withstand pressure data, transformed pressure data and environmental parameter data to obtain comprehensive data;

[0070] Step S2: Feature engineering: Use feature engineering methods to process comprehensive data to obtain high-quality feature data;

[0071] Step S3: Model design: Establish an AE-MSGRU model. The AE-MSGRU model includes an input layer, a multi-level GRU layer, an attention mechanism layer, and an output layer. The multi-level GRU layer includes a short-term GRU layer, a medium-term GRU layer, and a long-term GRU layer. Each layer of the AE-MSGRU model is composed of neurons. High-quality feature data is input into the AE-MSGRU model, and the pressure estimation value is obtained through layer-by-layer calculation.

[0072] Step S4: Model training and optimization: Use high-quality feature data to train the AE-MSGRU model, use the cross entropy loss function to evaluate the performance of the AE-MSGRU model, and use the stochastic gradient descent algorithm to optimize the parameters of the AE-MSGRU model to obtain a trained AE-MSGRU model;

[0073] Step S5: Pressure detection and qualification assessment: The pressure prediction value is obtained through the trained AE-MSGRU model, and the gas cylinder qualification test, gas cylinder ultimate pressure bearing capacity test and gas cylinder pressure resistance performance test are performed according to the pressure prediction value.

[0074] Embodiment 2: This embodiment is based on the above embodiment, and step S1 specifically includes the following steps:

[0075] Initial pressure data: including the pressure value, temperature and environment of the cylinder during initial inflation;

[0076] Ultimate pressure bearing data: including the pressure value of the gas cylinder under the ultimate pressure, the duration of pressure bearing and the strain data of the gas cylinder material;

[0077] Pressure resistance data: including long-term pressure resistance performance data of gas cylinders under constant pressure, pressure fluctuations and changes in gas cylinder status;

[0078] Transformation pressure data: including the pressure change data of the gas cylinder at different flight altitudes and ambient temperatures and the dynamic response data of the cylinder material;

[0079] Environmental parameter data: including temperature, humidity and air pressure related to the environment in which the gas cylinder is located.

[0080] Embodiment 3: This embodiment is based on the above embodiment, and step S2 specifically includes the following steps:

[0081] Step S21: Extracting features of initial pressure data: extracting time features, constructing interactive features, and expanding polynomial features on the initial pressure data to generate a feature set of initial pressure data;

[0082] Initial pressure data feature extraction:

[0083] Temporal feature extraction: Extracting temporal features related to initial stress, including daily, weekly time periods, and seasonal cycles;

[0084] Interactive feature construction: The initial pressure value is combined with the temperature and humidity parameters to generate interactive features, including: initial pressure × ambient temperature, to capture the relationship between pressure and environment;

[0085] Polynomial features: Perform polynomial expansion on the initial pressure and temperature, including square and cubic features to capture nonlinear relationships.

[0086] Step S22: Extraction of ultimate pressure bearing data features: performing pressure-time feature extraction, material strain feature construction and dynamic feature extraction on the ultimate pressure bearing data to generate an ultimate pressure bearing data feature set;

[0087] Ultimate pressure data feature extraction:

[0088] Pressure-time feature extraction: calculate the pressure value and pressure duration, including the average, maximum and minimum values ​​of the pressure duration;

[0089] Material strain feature construction: Combine material strain data to generate strain coefficient and strain velocity to quantify the performance of gas cylinder materials under extreme pressure;

[0090] Dynamic feature extraction: Extract the pressure change rate and accumulated strain from the data under extreme pressure to help the model identify extreme situations.

[0091] Step S23: Extracting withstand voltage data features: extracting key features from the withstand voltage data through trend feature analysis, periodic feature extraction and statistical feature calculation to generate a withstand voltage data feature set;

[0092] Withstand voltage data feature extraction:

[0093] Trend characteristics: Apply trend analysis technology to the withstand voltage data to extract the upward and downward trends of long-term withstand voltage performance;

[0094] Periodicity characteristics: The periodicity of the withstand voltage data is analyzed by Fourier transform, and the main frequency components are extracted as periodic characteristics;

[0095] Statistical characteristics: Calculate the mean, standard deviation, maximum value, and minimum value of the withstand voltage data to quantify the stability of the withstand voltage performance.

[0096] Step S24: transform pressure data feature extraction: applying frequency domain feature extraction and autoregressive feature analysis to generate a transform pressure data feature set;

[0097] Transform pressure data feature extraction:

[0098] Frequency domain feature extraction: Use short-time Fourier transform to extract the frequency domain features of the transformed pressure data, including energy distribution and main frequency components;

[0099] Autoregressive features: Autoregressive analysis is performed on the transformed pressure data to extract autoregressive coefficients and capture the autocorrelation of the time series.

[0100] Step S25: Environmental parameter data feature extraction: Generate an environmental parameter data feature set by constructing an environmental index and generating environmental pressure interaction features;

[0101] Environmental parameter data feature extraction:

[0102] Environmental index construction: Based on temperature, humidity and air pressure environmental data, a comprehensive environmental index is constructed to reflect the comprehensive impact of environmental conditions on cylinder pressure;

[0103] Environmental pressure interaction features: Environmental parameters are combined with cylinder pressure data to generate interaction features, including environmental temperature × pressure change rate, to quantify the regulatory effect of the environment on pressure changes.

[0104] Step S26: Feature selection and dimensionality reduction: Use the feature importance analysis method to screen the initial pressure data feature set, the ultimate pressure data feature set, the pressure resistance data feature set, the transformed pressure data feature set and the environmental parameter data feature set, and finally form high-quality feature data.

[0105] Embodiment 4, according to Figure 1 This embodiment is based on the above embodiment, and step S3 specifically includes the following steps:

[0106] Step S31: Use neurons in the input layer to integrate high-quality feature data into a time series data set and transmit it to the multi-level GRU layer;

[0107] Step S32: The short-term dependency, the medium-term dependency and the long-term dependency are captured respectively by the short-term GRU layer, the medium-term GRU layer and the long-term GRU layer, and the time series features are extracted from the time series data set of the input layer to obtain the comprehensive time series feature data. The formula of the multi-level GRU layer is as follows;

[0108] Multi-level GRU layer formula:

[0109] Update Gate :

[0110] ;

[0111] in, ReLU activation function. Represents the input data, i.e., the time series dataset; The update gate weight matrix representing the input data; represents the hidden state of the previous time step; The update gate weight matrix representing the hidden state of the previous time step; represents the bias term of the update gate;

[0112] Reset Gate :

[0113] ;

[0114] in, The reset gate weight matrix representing the input data; The reset gate weight matrix representing the hidden state of the previous time step; represents the bias term of the reset gate;

[0115] Candidate status :

[0116] ;

[0117] in, represents the hyperbolic tangent activation function, The candidate state weight matrix representing the input data; The candidate state weight matrix representing the hidden state of the previous time step; A bias term representing the candidate state; It means element-wise multiplication;

[0118] Final State :

[0119] ;

[0120] in, Indicates the part that retains the hidden state of the previous time step, and the weight is controlled by the update gate; The weight of the part representing the new candidate state is also controlled by the update gate;

[0121] Step S33: introducing an attention mechanism layer, identifying the key points of the pressure data by analyzing the comprehensive time series feature data, assigning weights to the key points of the pressure data, further optimizing the comprehensive time series feature data, and obtaining the time series feature data optimized by the attention mechanism;

[0122] Step S34: The output layer converts the time series feature data optimized by the attention mechanism into air pressure values ​​through the ReLU activation function to generate a pressure estimation value.

[0123] Embodiment 5: This embodiment is based on the above embodiment, and step S32 specifically includes the following steps:

[0124] Step S321: Capturing short-term dependencies: The short-term GRU layer processes the time series data set through a gating mechanism, focusing on the pressure change characteristics of the short-term time span, and generating an initial feature hidden state;

[0125] After the model training is completed, the weight and bias parameters of the first layer are:

[0126] Update Gate :

[0127] Weight Matrix : ;

[0128] Weight Matrix : ;

[0129] Bias : ;

[0130] Reset Gate :

[0131] Weight Matrix : ;

[0132] Weight Matrix : ;

[0133] Bias : ;

[0134] Candidate status :

[0135] Weight Matrix : ;

[0136] Weight Matrix : ;

[0137] Bias : .

[0138] Step S322: Apply Dropout layer 1: Add Dropout layer 1 after the short-term GRU layer to randomly mask the output of neurons to obtain the initial feature optimized hidden state;

[0139] Step S323: mid-term dependency capture: the mid-term GRU layer further extracts the mid-term time span features in the initial feature optimization hidden state, captures the mid-term change trend of the cylinder pressure over time, and generates a mid-term feature hidden state;

[0140] After the model training is completed, the weight and bias parameters of the second layer are:

[0141] Update Gate :

[0142] Weight Matrix : ;

[0143] Weight Matrix : ;

[0144] Bias : ;

[0145] Reset Gate :

[0146] Weight Matrix : ;

[0147] Weight Matrix : ;

[0148] Bias : ;

[0149] Candidate status :

[0150] Weight Matrix : ;

[0151] Weight Matrix : ;

[0152] Bias : .

[0153] Step S324: Apply Dropout layer 2: Add Dropout layer 2 after the mid-term GRU layer to randomly mask the output of neurons and obtain the mid-term feature optimized hidden state;

[0154] Step S325: long-term dependency capture: the long-term GRU layer deeply extracts the long-term time span features in the mid-term feature optimization hidden state, captures the deep trend features of the cylinder pressure change, and generates the deep feature hidden state;

[0155] After the model training is completed, the weight and bias parameters of the third layer are:

[0156] Update Gate :

[0157] Weight Matrix : ;

[0158] Weight Matrix : ;

[0159] Bias : ;

[0160] Reset Gate :

[0161] Weight Matrix : ;

[0162] Weight Matrix : ;

[0163] Bias : ;

[0164] Candidate status :

[0165] Weight Matrix : ;

[0166] Weight Matrix : ;

[0167] Bias : .

[0168] Step S326: Apply Dropout layer 3: Add Dropout layer 3 after the long-term GRU layer to randomly mask the output of neurons, obtain the optimized deep feature hidden state, and finally generate comprehensive time series feature data.

[0169] Embodiment 6: This embodiment is based on the above embodiment, and step S33 specifically includes the following steps:

[0170] Step S331: Calculate the self-attention score: perform linear transformation on the comprehensive time series feature data to generate the query matrix, key matrix and value matrix; obtain the initial attention score by calculating the dot product of the query matrix and the key matrix, and scale the initial attention score to obtain the attention score matrix; use the softmax function to normalize the attention score matrix to generate the attention weight matrix. The attention score calculation formula is as follows:

[0171] .Linear transformation:

[0172] , , ;

[0173] in, Represents comprehensive time series feature data; represents the query matrix, The weight matrix representing the query matrix; represents the key matrix, represents the weight matrix of the bond matrix; represents the value matrix, The weight matrix representing the value matrix;

[0174] .Attention score matrix:

[0175] ;

[0176] in, Represents the key matrix Dimensions; represents the transpose of the key matrix; represents the scaling factor;

[0177] .Attention weight matrix:

[0178] ;

[0179] in, represents the weight parameter, represents a regular term;

[0180] Step S332: weighted sum aggregation: apply the attention weight matrix to the value matrix to obtain a weighted value matrix; perform weighted summation on the weighted value matrix to generate weighted optimized time series feature data, the formula is as follows:

[0181] ;

[0182] in, Represents the time series feature data after weighted optimization; represents the weighted value matrix, Indicates the number of elements in the comprehensive time series feature data;

[0183] Step S333: residual connection and layer normalization: add the weighted optimized temporal feature data to the comprehensive temporal feature data to obtain the data after residual connection; perform layer normalization on the data after residual connection to finally obtain the temporal feature data after attention mechanism optimization, the formula is as follows:

[0184] ;

[0185] in, Represents the data after residual connection, Represents comprehensive time series feature data;

[0186] ;

[0187] in, Represents the time series feature data after the attention mechanism is optimized, Layer normalization function;

[0188] When calculating the self-attention score, the system calculates the attention score based on the relevance of the features and normalizes it through the softmax function, giving higher weights to features with strong relevance; then, in the weighting and aggregation steps, these weights are applied to the value matrix, ultimately making the key features occupy a more important position in the final time series feature data, thereby achieving the amplification and optimization of important information.

[0189] Embodiment 7, based on the above embodiment, step S4 specifically includes the following steps:

[0190] Step S41: construct a time series dataset from high-quality feature data, and divide the dataset into a training set (80%) and a validation set (20%) in proportion;

[0191] Step S42: Initialize the multi-level GRU layer of the AE-MSGRU model, including the short-term GRU layer, the medium-term GRU layer and the long-term GRU layer, and randomly initialize the weights and bias parameters of the multi-level GRU layer;

[0192] Step S43: Model training, inputting the training set into the AE-MSGRU model, using the cross entropy loss function to calculate the loss of the AE-MSGRU model, and using the stochastic gradient descent algorithm to optimize the model parameters;

[0193] Cross entropy loss function:

[0194] ;

[0195] in, represents the cross entropy loss function, represents the true label, represents the predicted probability of the model, represents the number of samples in the training set, Indicates The true labels of samples, Indicates The predicted probability of samples;

[0196] The parameter update formula of the Adam optimization algorithm is:

[0197] ;

[0198] Among them, represents the iteration round, Indicates The model parameters of the round iteration, represents the learning rate, Represents the loss function Parameters The gradient of

[0199] Step S44: Model testing: input the test set into the AE-MSGRU model to evaluate the pressure prediction performance of the AE-MSGRU model.

[0200] Embodiment 8, this embodiment is based on the above embodiment, step S5: pressure detection and qualification evaluation: the pressure prediction value is obtained by the trained AE-MSGRU model, and the gas cylinder qualification detection, gas cylinder limit pressure bearing capacity detection and gas cylinder pressure resistance performance detection are performed according to the pressure prediction value. The specific detection results in this embodiment are as follows:

[0201] .Predicted air pressure value:

[0202] .

[0203] . Gas cylinder qualification test:

[0204] Setting the Standard:

[0205] Minimum safety pressure value: 140bar;

[0206] Maximum safety pressure value: 190bar;

[0207] Test results:

[0208] .

[0209] . Gas cylinder ultimate pressure bearing capacity test:

[0210] Setting the Standard:

[0211] The ultimate pressure bearing capacity of the gas cylinder: 190bar;

[0212] Test results:

[0213] .

[0214] . Gas cylinder pressure resistance test:

[0215] Setting the Standard:

[0216] Pressure resistance performance standard: pressure fluctuation within ±10%;

[0217] Test results:

[0218] .

[0219] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A computer-aided high-precision pressure detection method for aerospace gas cylinders, characterized by: The method comprises the following steps: Step S1: Data collection: collect initial pressure data, ultimate pressure bearing data, pressure resistance data, transformed pressure data and environmental parameter data to obtain comprehensive data; Step S2: Feature engineering: Use feature engineering methods to process comprehensive data to obtain high-quality feature data; Step S3: Model design: Establish an AE-MSGRU model. The AE-MSGRU model includes an input layer, a multi-level GRU layer, an attention mechanism layer, and an output layer. The multi-level GRU layer includes a short-term GRU layer, a medium-term GRU layer, and a long-term GRU layer. Each layer of the AE-MSGRU model is composed of neurons. High-quality feature data is input into the AE-MSGRU model, and the pressure estimation value is obtained through layer-by-layer calculation. Step S4: Model training and optimization: Use high-quality feature data to train the AE-MSGRU model, use the cross entropy loss function to evaluate the performance of the AE-MSGRU model, and use the stochastic gradient descent algorithm to optimize the parameters of the AE-MSGRU model to obtain a trained AE-MSGRU model; Step S5: Pressure detection and qualification assessment: The pressure prediction value is obtained through the trained AE-MSGRU model, and the gas cylinder qualification test, gas cylinder ultimate pressure bearing capacity test and gas cylinder pressure resistance performance test are performed according to the pressure prediction value; Step S2 specifically includes the following steps: Step S21: Extracting initial pressure data features: extracting time features, constructing interactive features, and expanding polynomial features on the initial pressure data to generate an initial pressure data feature set; Step S22: Extraction of ultimate pressure bearing data features: performing pressure-time feature extraction, material strain feature construction and dynamic feature extraction on the ultimate pressure bearing data to generate an ultimate pressure bearing data feature set; Step S23: Extracting withstand voltage data features: extracting key features from the withstand voltage data through trend feature analysis, periodic feature extraction and statistical feature calculation to generate a withstand voltage data feature set; Step S24: transform pressure data feature extraction: applying frequency domain feature extraction and autoregressive feature analysis to generate a transform pressure data feature set; Step S25: Environmental parameter data feature extraction: Generate an environmental parameter data feature set by constructing an environmental index and generating environmental pressure interaction features; Step S26: Feature selection and dimensionality reduction: Use the feature importance analysis method to screen the initial pressure data feature set, the ultimate pressure data feature set, the pressure resistance data feature set, the transformed pressure data feature set and the environmental parameter data feature set, and finally form high-quality feature data.

2. The computer-aided high-precision pressure detection method for aerospace gas cylinders according to claim 1, characterized in that: Step S3 specifically includes the following steps: Step S31: Use neurons in the input layer to integrate high-quality feature data into a time series data set and transmit it to the multi-level GRU layer; Step S32: capturing short-term dependencies, medium-term dependencies, and long-term dependencies respectively through the short-term GRU layer, the medium-term GRU layer, and the long-term GRU layer, extracting time series features from the time series data set of the input layer to obtain comprehensive time series feature data; Step S33: introducing an attention mechanism layer, identifying key points of pressure data by analyzing the comprehensive time series feature data, assigning weights to the key points of pressure data, further optimizing the comprehensive time series feature data, and obtaining time series feature data optimized by the attention mechanism; Step S34: The output layer converts the time series feature data optimized by the attention mechanism into air pressure values ​​through the ReLU activation function to generate a pressure estimation value.

3. The computer-aided high-precision pressure detection method for aerospace gas cylinders according to claim 2, characterized in that: Step S32 specifically includes the following steps: Step S321: Capturing short-term dependencies: The short-term GRU layer processes the time series data set through a gating mechanism, focusing on the pressure change characteristics of the short-term time span, and generating an initial feature hidden state; Step S322: Apply Dropout layer 1: Add Dropout layer 1 after the short-term GRU layer to randomly mask the output of neurons to obtain the initial feature optimized hidden state; Step S323: mid-term dependency capture: the mid-term GRU layer further extracts the mid-term time span features in the initial feature optimization hidden state, captures the mid-term change trend of the cylinder pressure over time, and generates a mid-term feature hidden state; Step S324: Apply Dropout layer 2: Add Dropout layer 2 after the mid-term GRU layer to randomly mask the output of neurons and obtain the mid-term feature optimized hidden state; Step S325: long-term dependency capture: the long-term GRU layer deeply extracts the long-term time span features in the mid-term feature optimization hidden state, captures the deep trend features of the cylinder pressure change, and generates the deep feature hidden state; Step S326: Apply Dropout layer 3: Add Dropout layer 3 after the long-term GRU layer to randomly mask the output of neurons, obtain the optimized deep feature hidden state, and finally generate comprehensive time series feature data.

4. The computer-aided high-precision pressure detection method for aerospace gas cylinders according to claim 2, characterized in that: Step S33 specifically includes the following steps: Step S331: Self-attention score calculation: perform linear transformation on the comprehensive time series feature data to generate a query matrix, a key matrix and a value matrix; obtain the initial attention score by calculating the dot product of the query matrix and the key matrix, and scale the initial attention score to obtain the attention score matrix; use the softmax function to normalize the attention score matrix to generate an attention weight matrix; Step S332: weighted sum aggregation: applying the attention weight matrix to the value matrix to obtain a weighted value matrix; performing weighted summation on the weighted value matrix to generate weighted optimized time series feature data; Step S333: residual connection and layer normalization: add the weighted optimized time series feature data to the comprehensive time series feature data to obtain the data after residual connection; perform layer normalization on the data after residual connection to finally obtain the time series feature data after attention mechanism optimization.

Images