An aviation product processing quality prediction method considering multi-dimensional influencing factors
By constructing a multi-input quality prediction model, integrating horizontal and vertical factors, and using data standardization and LSTM networks, the problems of low accuracy and data silos in the prediction of aerospace product processing quality were solved, and high-precision prediction of multi-variety, small-batch products was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2023-05-11
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for predicting the processing quality of aerospace products fail to effectively consider multi-dimensional factors, resulting in low prediction accuracy, serious data silos, and an inability to uniformly predict product sizes of different orders of magnitude, as well as a lack of unified quality evaluation indicators.
By constructing a multi-input quality prediction model, acquiring and preprocessing aviation product processing-related data, integrating horizontal and vertical influencing factors, using LSTM layers and feature fusion layers, employing data standardization methods to unify quality evaluation indicators, and constructing a long short-term memory neural network for prediction.
It improved the accuracy of processing quality prediction, reduced production costs, enabled unified quality prediction for multiple varieties of small-batch products, and solved the problem of data silos.
Smart Images
Figure CN116485032B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of processing quality prediction, and in particular to a method for predicting the processing quality of aerospace products that considers multiple influencing factors. Background Technology
[0002] As a key development area of the manufacturing industry, the aviation industry is a concrete manifestation of a country's comprehensive national strength and core competitiveness. At the same time, the aviation manufacturing industry is a knowledge-intensive, technology-intensive, and multidisciplinary integrated industry. It is a typical representative of multi-variety, small-batch, and discrete manufacturing. Due to the extremely high complexity of aviation products, the large number of parts, and the high requirements for safety and reliability, it is very important to carry out quality management for complex aviation products.
[0003] Traditional quality management relies solely on post-production quality inspection, which cannot provide timely and accurate proactive quality control. In recent years, with the rapid development and application of technologies such as big data and artificial intelligence, using intelligent algorithms to deeply mine processing data to predict processing quality and achieve preventative product quality control has gradually become a research hotspot. However, current research on predicting the processing quality of complex aerospace products still faces the following challenges:
[0004] 1. The processing quality of aerospace products is affected by factors such as processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods. At the same time, the quality inspection results show a time-series change trend. Both the influence of various horizontal factors and the time-series change trend will affect the processing quality. However, the existing complex aerospace product processing quality prediction model does not consider multi-dimensional factors, resulting in low prediction accuracy.
[0005] 2. The processing quality of complex aerospace products is affected by multiple factors, and the relevant data is distributed across multiple systems. The data volume is huge and complex. Due to the different data collection, uploading and management methods of each system, they cannot be directly interconnected, forming data silos. As a result, existing methods for predicting the processing quality of complex aerospace products lack complete data that can be effectively correlated and integrated.
[0006] 3. Complex aerospace products are produced in a variety of small batches, with numerous product types and each product requiring complex dimensions but small batch sizes. Existing methods for predicting the processing quality of complex aerospace products cannot unify the product dimensions of different orders of magnitude, resulting in an inability to predict the processing quality of aerospace products in a unified manner. Summary of the Invention
[0007] The purpose of this invention is to address the problems of low prediction accuracy, lack of complete data for effective correlation and integration, and inability to uniformly predict the processing quality of aviation products in existing methods for predicting the processing quality of aviation products. Therefore, this invention proposes a method for predicting the processing quality of aviation products that considers multiple influencing factors.
[0008] A method for predicting the processing quality of aviation products that considers multiple influencing factors is as follows: obtain the correlation data of the processing quality of the aviation products to be predicted, input the correlation data of the processing quality of the aviation products to be predicted into the multi-input quality prediction model, and obtain the processing quality of the aviation products.
[0009] The data related to the processing quality of aerospace products includes: horizontal influencing factors and vertical influencing factors;
[0010] The horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods;
[0011] The longitudinal influencing factors are historical quality inspection results arranged by processing time;
[0012] The multi-input quality prediction model is obtained through the following method:
[0013] Step 1: Obtain relevant data on the processing quality of aviation products;
[0014] Step 2: Preprocess the data related to the processing quality of aviation products;
[0015] The preprocessing includes: feature encoding of the aviation product processing quality correlation data; and standardization of the label columns of the aviation product processing quality correlation data to obtain quality evaluation indicators.
[0016] Step 3: Perform sliding grouping on the preprocessed complex aerospace product processing quality correlation data, and divide the preprocessed aerospace product processing quality correlation data after sliding grouping into training set and test set;
[0017] Step 4: Build a quality prediction model. Use the training set to train the quality prediction model to obtain a well-trained quality prediction model. Then evaluate the well-trained quality prediction model and use the best-performing well-trained quality prediction model as the multi-input quality prediction model.
[0018] In the process of training the quality prediction model using the training set, the quality prediction model learns features from horizontal and vertical influencing factors and fuses the two types of features.
[0019] Furthermore, the aforementioned horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods, as detailed below:
[0020] The processing personnel include: age, work group, continuous working hours, length of service, and technical level.
[0021] The processing equipment includes: the manufacturer and model of the machine tool, cutting tool, and fixture, as well as the machine tool's usage time, the degree of cutting tool wear, the degree of cutting tool lubrication, the cutting parameters of the cutting tool, the clamping method of the fixture, and the clamping force.
[0022] The processing method includes: processing technology, processing pressure, and processing temperature;
[0023] The raw materials include: raw material batch, supplier, and quality grade;
[0024] The processing environment includes: ambient temperature, humidity, noise, and lighting;
[0025] The measurement method includes: measuring tools, measurement methods, and measurement accuracy.
[0026] Furthermore, the quality evaluation index is as follows:
[0027]
[0028] Furthermore, step three involves performing sliding grouping on the preprocessed complex aerospace product processing quality correlation data, and then dividing the preprocessed aerospace product processing quality correlation data after sliding grouping into a training set and a test set. This includes the following steps:
[0029] First, the data processed in step two is grouped using a sliding window with a size of timestep. The first timestep of label column data is used as the vertical influencing factor, and the (timestep+1)th feature column data is used as the horizontal influencing factor. The horizontal and vertical influencing factors are then used as input to predict the (timestep+1)th label column data, as shown in the following formula:
[0030] x1 t =[F t (1),F t (2)…F t (m)…F t (M)]
[0031] x2 t =[Y t-1 ,Y t-2 …Y t-timestep ]
[0032] x t =[x1 t x2 t ]
[0033] Where x1 t F represents the lateral feature input of sample t. t (m) represents the m-th horizontal influencing factor in sample t, where M is the total number of horizontal influencing factors, x2t Yt represents the longitudinal feature input of sample t, Yt-timestep represents the quality evaluation index of the time step before the current time of sample t, and x t It is the input of the quality prediction model; then, the preprocessed aviation product processing quality correlation data after sliding grouping is randomly shuffled, and the training set and test set are divided by stratified sampling within the range of quality evaluation index values.
[0034] In the preprocessed aerospace product processing quality correlation data after sliding grouping, each group of data is a sample.
[0035] Furthermore, the quality prediction model includes: a feature learning layer, a feature fusion layer, a fully connected layer, a Dropout layer, and an output layer;
[0036] The feature learning layer includes: a horizontal influencing factor feature acquisition unit and a vertical influencing factor feature acquisition unit;
[0037] The horizontal influencing factor feature acquisition unit learns quality data features from horizontal influencing factors in the training and test sets; the horizontal influencing factor feature acquisition unit includes: LSTM layer, fully connected layer, and Dropout layer;
[0038] The LSTM layer includes multiple LSTM networks;
[0039] The vertical influencing factor feature acquisition unit learns quality data features from the horizontal influencing factors in the training and test sets; the vertical influencing factor feature acquisition unit includes: attention mechanism, LSTM layer, fully connected layer, and Dropout layer;
[0040] The feature fusion layer: fuses features learned from the training set and fuses features learned from the test set;
[0041] The fully connected layer: connects the features output by the feature fusion layer into global features;
[0042] The Dropout layer: prevents the quality prediction model from overfitting;
[0043] The output layer outputs the predicted processing quality of aerospace products.
[0044] Furthermore, the output of the LSTM layer is as follows:
[0045] h t =o t ·Activation(c t )
[0046] c t =f t·c t-1 +i t ·C t
[0047] C t =Activation(ω c ·(h t-1 ,x t )+b c )
[0048] f t =RecurrentActivation(ω f ·(h t-1 ,x t )+b f )
[0049] i t =RecurrentActivation(ω i ·(h t-1 ,x t )+b i )
[0050] o t =RecurrentActivation(ω o ·(h t-1 ,x t )+b o )
[0051] Among them, f t It is the Gate of Oblivion, i t It's an input gate, o t It's an output gate, C t It is the process value for updating the cell state, c t It is the state of sample t, h t This is the final output of the LSTM layer. Both RecurrentActivation and Activation are non-linear activation functions, ω f ω i ω o ω c b represents the weight. f b i b o b c Indicates deviation.
[0052] Furthermore, the training quality prediction model employs the following loss function:
[0053]
[0054] Where t is the sample label, n is the total number of samples, and y tRepresents the true value of the sample. This represents the predicted value of the sample.
[0055] Furthermore, in step four, the trained quality prediction model is evaluated using the Mean Absolute Percentage Error (MAPE), as shown in the following formula:
[0056]
[0057] in, The mean of the true values of the samples is represented by MAPE. The best-performing trained quality prediction model is the one that corresponds to the minimum value of MAPE.
[0058] The beneficial effects of this invention are as follows:
[0059] The data-driven method for predicting the processing quality of complex aerospace products, which considers multiple influencing factors, subdivides these factors into horizontal factors (processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods) and vertical factors (historical quality inspection results arranged by processing time). By comprehensively considering the impact of these multiple influencing factors on processing quality in different dimensions and integrating them into the prediction model, the method improves the prediction effect and accuracy of processing quality and reduces processing costs.
[0060] This invention takes into account that complex aviation products are affected by multiple factors and that relevant data are distributed across multiple systems. Therefore, after obtaining relevant data from different systems, the relevant data is correlated and integrated through the business logic between the systems to accurately analyze quality problems and obtain complete data that is effectively correlated and integrated.
[0061] This invention considers that the production of complex aerospace products involves multiple varieties and small batches, with numerous product types and each product requiring complex dimensions but small batch sizes. Therefore, a unified quality evaluation index is needed to expand the data volume. Since the baseline values differ significantly between different dimensions, the measured values are not on the same order of magnitude. This invention uses a data standardization method that considers dimensional characteristics to map the quality inspection results to the range [-1,1], thereby unifying the quality evaluation index. This preserves the physical characteristics of the processed dimensions and unifies multiple dimensional data under the same evaluation standard, enabling this invention to uniformly predict the processing quality of products of different sizes.
[0062] The data-driven, multi-dimensional influencing factor-considered complex aerospace product processing quality prediction model of this invention learns quality data features from horizontal and vertical influencing factors and fuses the horizontal and vertical features through a model fusion layer. It also considers processing quality influencing factors such as people, machines, materials, methods, environment, and measurement, as well as the temporal characteristics of quality data, which effectively improves prediction accuracy, can improve the processing quality of complex aerospace products and reduce production costs, and has high practical value. Attached Figure Description
[0063] Figure 1 This is a flowchart of the present invention;
[0064] Figure 2 This is a schematic diagram of the structure of a multi-input quality prediction model;
[0065] Figure 3 This is a comparison chart of the predicted values and the actual values in the training set of this invention.
[0066] Figure 4 This is a comparison chart of predicted and actual values for the test set in an embodiment of the present invention. Detailed Implementation
[0067] Specific implementation method one: as follows Figure 1 As shown, this embodiment of a method for predicting the processing quality of aviation products considering multiple influencing factors includes: acquiring correlation data of the processing quality of complex aviation products to be predicted, inputting the correlation data of the processing quality of complex aviation products to be predicted into a multi-input quality prediction model, and obtaining the processing quality of complex aviation products.
[0068] The data related to the processing quality of aerospace products includes: horizontal influencing factors and vertical influencing factors;
[0069] The horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods;
[0070] The longitudinal influencing factors refer to the historical quality inspection results arranged by processing time.
[0071] The multi-input quality prediction model is obtained through the following method:
[0072] Step 1: Obtain correlation data on the processing quality of complex aerospace products;
[0073] Data on the correlation between multidimensional factors affecting processing quality and the processing quality of complex aerospace products is obtained from various digital systems in the workshop, including:
[0074] Horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods;
[0075] The specific data on processing personnel includes their age, work group, continuous working hours, length of service, and technical level, which reflects the impact of the processing personnel's mental state and skill level.
[0076] The processing equipment includes the manufacturer and model of machine tools, cutting tools, and fixtures, as well as the machine tool usage time, cutting tool wear, cutting tool lubrication, cutting tool cutting parameters, fixture clamping method, and clamping force, etc., to reflect the impact of the equipment's working parameters and working status.
[0077] The processing methods include processing technology, processing pressure, and processing temperature, which reflect the influence of processing parameters;
[0078] Raw materials include their batch, supplier, and quality grade, which reflects the impact of the quality of the raw materials.
[0079] The processing environment includes ambient temperature, humidity, noise, and lighting, reflecting the impact of environmental factors.
[0080] Measurement methods include measuring tools, measurement techniques, and measurement accuracy, which reflect the impact of measurement errors.
[0081] Longitudinal influencing factors: Historical quality inspection results arranged by processing time;
[0082] The processing quality of complex aerospace products is affected by multiple factors, and the relevant data is distributed across multiple systems. Due to the different methods of data collection, uploading, and management in each system, direct communication is not possible, resulting in data silos. Therefore, after obtaining relevant data from different systems, the relevant data is correlated and integrated through the business logic between the systems to accurately analyze quality issues.
[0083] Step 2: Preprocess the correlation data of complex aerospace product processing quality:
[0084] Step 21: Perform feature encoding on the processing quality correlation data of complex aerospace products;
[0085] The raw data related to the processing quality of complex aerospace products have complex feature formats and many non-numerical discrete features. A label encoding method is used to process the non-numerical discrete features.
[0086] Step 22: Standardize the label column (i.e., the quality inspection result column) of the complex aerospace product processing quality correlation data to obtain quality evaluation indicators:
[0087] Complex aerospace product manufacturing involves multiple varieties and small batches. Each product has numerous types, and the dimensions required for processing are complex while the batch size is small. Therefore, it is necessary to use the same quality evaluation indicators to uniformly predict data for different products and dimensions to increase the data volume. Because the baseline values differ significantly between different dimensions, the quality inspection results are not on the same order of magnitude. A reasonable method is needed to map the quality inspection results to the same range to unify the quality evaluation indicators. This invention employs a data standardization method that considers dimensional characteristics to standardize the quality inspection results and uses the standardized results as the quality evaluation indicators.
[0088] The data standardization method considering dimensional characteristics ensures that the results are mapped to the range [-1, 1] and conform to the physical characteristics of the processing dimensions. The calculation formula is as follows:
[0089]
[0090] Step 3: Perform sliding grouping on the preprocessed complex aerospace product processing quality correlation data, and then divide the preprocessed complex aerospace product processing quality correlation data after sliding grouping into training set and test set:
[0091] First, the data processed in step two is grouped using a sliding window with a size of timestep. The first timestep of label column data is used as the vertical influencing factor, and the (timestep+1)th feature column data is used as the horizontal influencing factor. The horizontal and vertical influencing factors are then used as input to predict the (timestep+1)th label column data, as shown in the following formula:
[0092] x1 t =[F t (1),F t (2)…F t (m)…F t (M)]
[0093] x2 t =[Y t-1 ,Y t-2 …Y t-timestep ]
[0094] x t =[x1 t x2 t ]
[0095] Where x1 t F represents the lateral feature input of sample t. t (m) represents the m-th horizontal influencing factor in sample t, where M is the total number of horizontal influencing factors, x2 t Yt represents the longitudinal feature input of sample t, Yt-timestep represents the quality evaluation index of the time step before the current time of sample t, and x t It is the input to the quality prediction model;
[0096] Then, the grouped data is randomly shuffled, and stratified sampling is performed within the range of quality evaluation index values to divide the data into training and test sets, with a ratio of 7:3 between the training and test sets.
[0097] In this context, a set of data in the training set and the test set constitutes a single sample.
[0098] Step 4: Construct a quality prediction model. Train the quality prediction model using the training set to obtain a well-trained quality prediction model. Then, evaluate the trained quality prediction model and select the best-performing model as the multi-input quality prediction model. This includes the following steps:
[0099] Step 41: Construct a quality prediction model;
[0100] The quality prediction model includes: a feature learning layer, a feature fusion layer, a fully connected layer, a Dropout layer, and an output layer;
[0101] The feature learning layer includes: a horizontal influencing factor feature acquisition unit and a vertical influencing factor feature acquisition unit;
[0102] The horizontal influencing factor feature acquisition unit learns quality data features from horizontal influencing factors in the training and test sets; the horizontal influencing factor feature acquisition unit includes: LSTM layer, fully connected layer, and Dropout layer;
[0103] The LSTM layer includes multiple LSTM networks;
[0104] The vertical influencing factor feature acquisition unit learns quality data features from the horizontal influencing factors in the training and test sets; the vertical influencing factor feature acquisition unit includes: attention mechanism, LSTM layer, fully connected layer, and Dropout layer;
[0105] The feature fusion layer fuses features learned from the training set and features learned from the test set. The feature fusion layer is implemented using the Concatenate function, as shown below:
[0106] merged=Concatenate([output1,output2])
[0107] Where output1 represents the horizontal features of the learned quality data, output2 represents the vertical features of the learned quality data, and Concatenate is the fusion layer function;
[0108] The fully connected layer is a Dense layer: it connects all the extracted local features into global features;
[0109] Dropout layer: Prevents overfitting of the network by modifying the number of neurons in the hidden layer;
[0110] The output layer is a Dense layer: used to output the results.
[0111] Long Short-Term Memory (LSTM) neural networks consist of chained units, each including a forget gate, an input gate, and an output gate. The calculation formula is as follows:
[0112] f t =RecurrentActivation(ω f ·(h t-1 ,x t )+bf )
[0113] i t =RecurrentActivation(ω i ·(h t-1 ,x t )+b i )
[0114] o t =RecurrentActivation(ω o ·(h t-1 ,x t )+b o )
[0115] C t =Activation(ω c ·(h t-1 ,x t )+b c )
[0116] c t =f t ·c t-1 +i t ·C t
[0117] h t =o t ·Activation(c t )
[0118] Among them, f t Represents the forget gate, i t Indicates the input gate, o t Indicates the output gate, C t To update the process value of the cell state, c t h represents the state of sample t. t This represents the final output of the LSTM layer. Both RecurrentActivation and Activation are non-linear activation functions, ω... f ω i ω o ω c b represents the weight. f b i b o b c Indicates weights and biases;
[0119] Among them, h t The output sample prediction value is then passed through the fully connected layer.
[0120] Step 42: Train the quality prediction model using the training set to obtain a well-trained quality prediction model;
[0121] The training quality prediction model uses the following loss function:
[0122] With mean square error As the loss function of the model, t is the sample label and n is the total number of samples.
[0123] Step 43: Evaluate the performance of the trained quality prediction model and select the best-performing trained quality prediction model as the multi-input quality prediction model.
[0124] The mean absolute percentage error (MAPE) was used as the evaluation index to assess the model and verify the performance of the quality prediction model.
[0125] The trained quality prediction model corresponding to the minimum MAPE value is selected as the multi-input quality prediction model; the calculation formula is as follows:
[0126]
[0127] Among them, y t Represents the true value of the sample. Indicates the sample predicted value. This represents the mean of the true values.
[0128] Example:
[0129] This invention uses an aircraft casing product as an example for illustration. Figure 1 A flowchart for predicting the processing quality of complex aerospace products.
[0130] Step 1: Obtain big data related to the processing quality of complex aerospace products. Extract multi-dimensional factors influencing processing quality from various digital systems in the workshop. These include horizontal influencing factors such as processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods; and vertical influencing factors such as historical quality inspection results arranged by processing time. Specifically, processing personnel include age, work group, continuous working hours, seniority, and technical level, reflecting the impact of personnel's mental state and skill level. Processing equipment includes the manufacturer and model of machine tools, cutting tools, and fixtures, as well as machine tool usage time, cutting tool wear, cutting tool lubrication, cutting tool parameters, fixture clamping methods, and clamping forces, reflecting the impact of equipment operating parameters and working conditions. Raw materials include batch, supplier, and quality grade, reflecting the impact of raw material quality. Processing methods include processing technology, processing pressure, and processing temperature, reflecting the impact of processing parameters. The processing environment includes ambient temperature, humidity, noise, and lighting, reflecting the impact of environmental factors. Measurement methods include measuring tools, measurement methods, and measurement accuracy, reflecting the impact of measurement errors.
[0131] The company's workshop utilizes various digital systems, including Manufacturing Execution System (MES), Quality Management System (QMS), Resource Management System (SAP), and Detailed Manufacturing Data and Process System (MDC). Product processing-related data is extracted from these systems, including quality inspection sheets, shift production schedules, work reporting schedules, and order and delivery status reports. This process obtains quality inspection information, shift production information, work reporting information, and order information for processed products. Data from different systems is then linked and integrated using keywords such as part number, size number, and work reporting time to obtain big data related to product processing quality. The data is then sorted using work reporting time as a unique ID.
[0132] Step 2, data preprocessing.
[0133] Feature encoding uses a label encoding method to handle non-numerical discrete features, such as encoding different processing workers as W1, W2, etc.
[0134] The quality inspection results are standardized using a data standardization method that considers dimensional characteristics, and the standardized results are used as a quality evaluation index. This method ensures that the results are mapped between [-1, 1] and conform to the physical characteristics of the processing dimensions. The calculation formula is as follows:
[0135]
[0136] If, in the example, a certain data point has a quality inspection result of 9.12, the baseline value for this dimension is 9.00, the lower limit of the dimension is 8.80, and the upper limit of the dimension is 9.20, then the standardized result is obtained by performing the following calculations:
[0137] Dimensional tolerance = 9.20 - 8.80 = 0.40
[0138]
[0139] Step 3: Group the data processed in Step 2 using a sliding window. The sliding window size is timestep. The data from the first timestep label columns are used as vertical influencing factors, and the data from the (timestep+1)th feature column are used as horizontal influencing factors. The horizontal and vertical influencing factors are then used as input to predict the (timestep+1)th label column data, as shown in the following formula:
[0140] x1 t =[F t (1),F t (2)…F t (m)…F t (M)]
[0141] x2 t =[Y t-1 ,Yt-2 …Y t-timestep ]
[0142] x t =[x1 t x2 t ]
[0143] Where x1 t F represents the lateral feature input of sample t. t (m) represents the m-th horizontal influencing factor in sample t, where M is the total number of horizontal influencing factors, x2 t Yt represents the longitudinal feature input of sample t, Yt-timestep represents the quality evaluation index of the time step before the current time of sample t, and x t It is the input to the quality prediction model;
[0144] The data completed by grouping is randomly shuffled, and the training set and test set are divided by stratified sampling according to the range of the quality evaluation index values.
[0145] In this embodiment, the horizontal feature inputs include processing workers, processing machine tools, processing orders, and processing technology, while the vertical feature inputs are the quality evaluation indicators for the first 50 time steps, i.e., timestep=50.
[0146] Step 4: Design a neural network to build a multi-input quality prediction model. The model structure is as follows: Figure 2 It includes: feature learning layer, feature fusion layer, fully connected layer, Dropout layer, and output layer;
[0147] The feature learning layer includes: a horizontal influencing factor feature acquisition unit and a vertical influencing factor feature acquisition unit;
[0148] The horizontal influencing factor feature acquisition unit learns quality data features from horizontal influencing factors in the training and test sets; the horizontal influencing factor feature acquisition unit includes: LSTM layer, fully connected layer, and Dropout layer;
[0149] The LSTM layer includes multiple LSTM networks;
[0150] The vertical influencing factor feature acquisition unit learns quality data features from the horizontal influencing factors in the training and test sets; the vertical influencing factor feature acquisition unit includes: attention mechanism, LSTM layer, fully connected layer, and Dropout layer;
[0151] The feature fusion layer: fuses features learned from the training set and fuses features learned from the test set;
[0152] The Long Short-Term Memory (LSTM) layer in the quality prediction model consists of chained units, each including a forget gate, an input gate, and an output gate. The calculation formula is as follows:
[0153] f t =RecurrentActivation(ω f ·(h t-1 ,x t )+b f )
[0154] i t =RecurrentActivation(ω i ·(h t-1 ,x t )+b i )
[0155] o t =RecurrentActivation(ω o ·(h t-1 ,x t )+b o )
[0156] C t =Activation(ω c ·(h t-1 ,x t )+b c )
[0157] c t =f t ·c t-1 +i t ·C t
[0158] h t =o t ·Activation(c t )
[0159] Among them, f t Represents the forget gate, i t Indicates the input gate, o t Indicates the output gate, C t To update the process value of the cell state, c t h represents the state of sample t. t ω represents the final output of the LSTM layer. f ω i ω o ω c b represents the weight. f b i b o bc ω represents weights and biases. Both RecurrentActivation and Activation are non-linear activation functions. RecurrentActivation is the activation function for the forget gate, input gate, and output gate. In this embodiment, it is set to Hard_sigmoid. Activation is the activation function for updating the unit state and the final output. In this embodiment, it is set to ReLU. ω and b represent weights and biases.
[0160] Step 5: Train the quality prediction model. Design a loss function, using mean squared error. The loss function of the model is used. The parameter settings of the quality prediction model are shown in Table 1 below.
[0161] Table 1 Training parameters of the multi-input quality prediction model
[0162]
[0163] Step 6: Evaluate the performance of the quality prediction model to obtain a multi-input quality prediction model.
[0164] The model was trained 1000 times according to the parameters in Table 1. The mean absolute percentage error (MAPE) was used as the evaluation metric to assess the model and verify its performance. The calculation formula is as follows:
[0165]
[0166] Where n represents the total number of samples in the test set, y t Represents the true value of the sample. Indicates the sample predicted value. This represents the mean of the true values. The comparisons of the true and predicted values in the training and test sets are shown below. Figure 3 , Figure 4 As shown in Table 2, the performance index results show that the quality prediction model achieved a prediction accuracy of 90.69% in the test set.
[0167] Table 2 Evaluation Index Results
[0168]
[0169] The above examples of the present invention are merely for illustrating the calculation model and calculation process of the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.
Claims
1. A method for predicting the processing quality of aerospace products considering multidimensional influencing factors, characterized in that... The specific process of the method is as follows: obtain the correlation data of the processing quality of the aviation product to be predicted, input the correlation data of the processing quality of the aviation product to be predicted into the multi-input quality prediction model, and obtain the processing quality of the aviation product. The data related to the processing quality of aviation products includes: horizontal influencing factors and vertical influencing factors; The horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods; The longitudinal influencing factors are historical quality inspection results arranged by processing time; The multi-input quality prediction model is obtained through the following method: Step 1: Obtain relevant data on the processing quality of aviation products; Step 2: Preprocess the data related to the processing quality of aviation products; The preprocessing includes: feature encoding of the aviation product processing quality correlation data; and standardization of the label columns of the aviation product processing quality correlation data to obtain quality evaluation indicators. The quality evaluation index is as follows: Step 3: Perform sliding grouping on the preprocessed complex aerospace product processing quality correlation data, and divide the preprocessed aerospace product processing quality correlation data after sliding grouping into training set and test set; Step 4: Build a quality prediction model. Use the training set to train the quality prediction model to obtain a well-trained quality prediction model. Then evaluate the well-trained quality prediction model and use the best-performing well-trained quality prediction model as the multi-input quality prediction model. The quality prediction model includes: a feature learning layer, a feature fusion layer, a fully connected layer, a Dropout layer, and an output layer; The feature learning layer includes: a horizontal influencing factor feature acquisition unit and a vertical influencing factor feature acquisition unit; The horizontal influencing factor feature acquisition unit learns quality data features from horizontal influencing factors in the training and test sets; the horizontal influencing factor feature acquisition unit includes: LSTM layer, fully connected layer, and Dropout layer; The LSTM layer includes multiple LSTM networks; The vertical influencing factor feature acquisition unit learns quality data features from the horizontal influencing factors in the training and test sets; the vertical influencing factor feature acquisition unit includes: attention mechanism, LSTM layer, fully connected layer, and Dropout layer; The feature fusion layer: fuses features learned from the training set and fuses features learned from the test set; The feature fusion layer is implemented using the Concatenate function; The fully connected layer: connects the features output by the feature fusion layer into global features; The Dropout layer: prevents the quality prediction model from overfitting; The output layer outputs the predicted processing quality of aerospace products; In the process of training the quality prediction model using the training set, the quality prediction model learns features from horizontal and vertical influencing factors and fuses the two types of features.
2. The method for predicting the processing quality of aerospace products considering multidimensional influencing factors according to claim 1, characterized in that: The horizontal influencing factors include: processing personnel, processing equipment, raw materials, processing methods, processing environment, and measurement methods, as detailed below: The processing personnel include: age, work group, continuous working hours, length of service, and technical level. The processing equipment includes: the manufacturer and model of the machine tool, cutting tool, and fixture, as well as the machine tool's usage time, the degree of cutting tool wear, the degree of cutting tool lubrication, the cutting parameters of the cutting tool, the clamping method of the fixture, and the clamping force. The processing method includes: processing technology, processing pressure, and processing temperature; The raw materials include: raw material batch, supplier, and quality grade; The processing environment includes: ambient temperature, humidity, noise, and lighting; The measurement method includes: measuring tools, measurement methods, and measurement accuracy.
3. The method for predicting the processing quality of aerospace products considering multidimensional influencing factors according to claim 2, characterized in that: Step three involves performing sliding grouping on the preprocessed complex aerospace product processing quality correlation data, and then dividing the preprocessed aerospace product processing quality correlation data after sliding grouping into a training set and a test set. This includes the following steps: First, the data processed in step two is grouped using a sliding window with a size of timestep. The first timestep of label column data is used as the vertical influencing factor, and the (timestep+1)th feature column data is used as the horizontal influencing factor. The horizontal and vertical influencing factors are then used as input to predict the (timestep+1)th label column data, as shown in the following formula: in, express The lateral feature input, express The There are 1,000 horizontal influencing factors, where M is the total number of horizontal influencing factors. Indicates sample Vertical feature input, , It is the input to the quality prediction model; Then, the preprocessed aviation product processing quality correlation data after sliding grouping is randomly shuffled, and stratified sampling is performed within the range of quality evaluation index values to divide the training set and test set. In the preprocessed aerospace product processing quality correlation data after sliding grouping, each group of data is a sample.
4. The method for predicting the processing quality of aerospace products considering multidimensional influencing factors according to claim 3, characterized in that: The output of the LSTM layer is as follows: in, It is the Gate of Oblivion. It's an input gate. It's an output gate. It is the process value for updating the cell state. yes state, yes The final output of the layer, and All are non-linear activation functions. Indicates deviation.
5. The method for predicting the processing quality of aerospace products considering multidimensional influencing factors according to claim 4, characterized in that: The training quality prediction model uses the following loss function: Where t is the sample label and n is the total number of samples. Represents the true value of the sample. This represents the predicted value of the sample.
6. The method for predicting the processing quality of aerospace products considering multidimensional influencing factors according to claim 5, characterized in that: In step four, the trained quality prediction model is evaluated using the Mean Absolute Percentage Error (MAPE), as shown in the following formula: in, This represents the mean of the true values of the sample. The trained quality prediction model corresponding to the minimum value is the best-performing trained quality prediction model.