WO2021143067A1

WO2021143067A1 - Method and apparatus for predicting workpiece quality, and computer device

Info

Publication number: WO2021143067A1
Application number: PCT/CN2020/099475
Authority: WO
Inventors: 林宏达
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-05-28
Filing date: 2020-06-30
Publication date: 2021-07-22
Also published as: CN111814385B; CN111814385A

Abstract

A method for predicting workpiece quality, relating to the field of medicine, and comprising: determining whether current data to be analyzed contains an indicator feature corresponding to a quality parameter of a defective workpiece (S1); if such a feature is contained, then inputting the current data to be analyzed into a prediction model for predictive analysis, the prediction model at least comprising two from among an Xgboost model, a RandomForest model, or a deep learning neural network model, the Xgboost model performing correction adjustment by means of a linear regression model (S2); according to a preset method, performing aggregation processing on the analysis results of the at least two models from among the Xgboost model, the RandomForest model, or the deep learning neural network model to obtain an aggregated result (S3); and on the basis of the aggregated result, determining the probability that the workpiece quality corresponding to the current data to be analyzed is substandard (S4). A loss function for an Xgboost model is designed on the basis of features of the Xgboost model and sparse data features during quality prediction, causing the Xgboost model to be more suitable for feature analysis of sparse data, and causing the method for predicting workpiece quality in the field of medicine to be industrially applicable.

Description

Method, device and computer equipment for predicting the quality of processed parts

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on May 28, 2020, the application number is 202010469626.2, and the invention title is "Methods, Apparatus and Computer Equipment for Predicting the Quality of Workpieces", the entire contents of which are incorporated by reference In this application.

Technical field

This application relates to the medical field, in particular to methods, devices and computer equipment for predicting the quality of processed parts.

Background technique

There are many types of processed parts in the medical field, including processed parts on diagnostic medical devices and processed parts on therapeutic medical devices. The quality and accuracy requirements are relatively high. There are also many factors that need to be referred to in the quality estimation of the above-mentioned processed parts. The existing prediction of whether the quality of the processed parts meets the standard, or whether the processed parts has defects or defects is not accurate enough, needs to be reviewed by a professional quality reviewer. The inventor realizes that this is not only necessary The reviewers carried out various related inspections on the processed parts, which resulted in a long time-consuming process and was unable to carry out general industrialization promotion.

technical problem

The main purpose of this application is to provide a method for predicting the quality of a machined part, which aims to solve the technical problem that the existing method for predicting the quality of a machined part cannot be industrialized and widely promoted.

Technical solutions

This application proposes a method for predicting the quality of processed parts, including:

Determine whether the current data to be analyzed contains the mark characteristics corresponding to the quality parameters of the defective workpiece;

If yes, input the current data to be analyzed into a prediction model for prediction analysis, where the prediction model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model, and the Xgboost model adopts a linear regression model Made corrections and adjustments;

Summarizing the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method, to obtain a summary result;

Determine the probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard according to the summary result.

This application also provides a device for predicting the quality of a processed part, including:

The first judging module is used to judge whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective processed part;

The input module is used to input the current to-be-analyzed data into a predictive model for predictive analysis if the current to-be-analyzed data contains the mark characteristics corresponding to the quality parameters of the defective workpiece, wherein the predictive model includes at least the Xgboost model and RandomForest Two of the model and the deep learning neural network model, the Xgboost model has been modified and adjusted through a linear regression model;

The summary module is configured to summarize the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method to obtain a summary result;

The second judgment module is used for judging the probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard according to the summary result.

The present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and the method for realizing the quality of a workpiece when the processor executes the computer program includes:

The present application also provides a computer-readable storage medium on which a computer program is stored. The method for realizing the quality of a workpiece when the computer program is executed by a processor includes:

Beneficial effect

This application designs the loss function of the Xgboost model based on the characteristics of the Xgboost model and the sparse data characteristics when the Xgboost model is used to predict the quality of processed parts. The logarithmic maximum similarity is used as the loss function and corrected by linear regression, so that The Xgboost model is more suitable for feature analysis of sparse data, so that the method of predicting the quality of processed parts can be widely promoted in industrialization. After training the Random Forest model, the deep learning neural network model, and the modified Xgboost model through the preprocessed feature data, respectively, the data samples to be tested are input into the above three models for analysis, and three analysis results are obtained. Integrate the above three results through the stacking model to realize the quality prediction of the processed parts corresponding to the sample data, and change the negative impact of the sparse feature of the data on the model.

Description of the drawings

Fig. 1 is a schematic flow chart of a method for predicting the quality of a processed part according to an embodiment of the present application;

Fig. 2 is a schematic structural diagram of a device for predicting the quality of a processed part according to an embodiment of the present application;

Fig. 3 is a schematic diagram of the internal structure of a computer device according to an embodiment of the present application.

The best mode of the present invention

1, a method for predicting the quality of a processed part according to an embodiment of the present application includes:

S1: Determine whether the current data to be analyzed contains the mark features corresponding to the quality parameters of the defective processed parts;

S2: If yes, input the current to-be-analyzed data into a predictive model for predictive analysis, where the predictive model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model, and the Xgboost model passes linearity The regression model has been revised and adjusted;

S3: According to a preset method, the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model are summarized and processed to obtain a summary result;

S4: Determine the probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard according to the summary result.

The current data to be analyzed in this embodiment is the feature data of the processed part, including product feature data such as height, width, length, weight, density, color uniformity, surface flatness, hardness, etc., as well as production time, production environment parameters, and batches of raw materials. Feature data of secondary processing information. By judging whether the current to-be-analyzed data contains the mark features corresponding to the quality parameters of the defective workpiece, and when the current to-be-analyzed data contains the mark features corresponding to the quality parameters of the defective workpiece, the predictive model is triggered to perform predictive analysis. The quality parameters of the above-mentioned defective workpiece are included in the characteristic data of the workpiece. For example, the density of the processed part has the greatest impact on its quality compliance, so it is necessary to search whether the current data to be analyzed includes low-density signature features. If there is a low-density signature feature, it is considered that there is a risk that the quality of the processed part does not meet the standard, which will trigger the prediction. Suppose the model analyzes and predicts all characteristic data. In addition to low density, the above-mentioned flag characteristics associated with the quality parameters of defective processed parts also include uneven surface, uneven color, and substandard size, etc., which vary according to the quality requirements of the processed parts to be tested. For example, a confidential small gear processing workpiece has very high requirements for its strength, but the special strength testing equipment will scrap the gear after inspecting the gear, so it needs to pass the length, width, height, weight and thermodynamics of the small gear Use characteristic data such as imaging to determine whether the gear's strength is up to standard.

Since the sample data in the model training of this embodiment is thousands of processed parts data, defective processed parts account for a small number, and the data structure features are sparse features, that is, most of the data is assigned a value of zero, which results in the failure to reflect the data during model training. The degree of discrimination affects the discriminative analysis effect of model training. The Xgboost model of this embodiment undergoes a specific correction process to make it meet the differential analysis of sparse data, and the deep learning neural network model meets the differential analysis of sparse data by designing a specific construction structure. Then, according to the stacking model, the analysis results of at least two of the above-mentioned Xgboost model, RandomForest model and deep learning neural network model are fused to highlight the main results of the machine learning model, and the error analysis content of the machine learning model is carried out in-depth Learn to correct. In this embodiment, the above-mentioned three models simultaneously analyze the current data to be analyzed in parallel, and then merge the analysis. The above fusion result is an evaluation of the risk score of the current data to be analyzed. The higher the score, the higher the probability that the quality of the processed parts predicted by the current data to be analyzed will not meet the standard. According to the characteristics of the Xgboost model and the sparse data characteristics when the Xgboost model is used to predict the quality of processed parts, the loss function of the Xgboost model is designed to use the logarithmic maximum similarity as the loss function and correct it through linear regression to make the Xgboost model It is more suitable for feature analysis of sparse data, so that the method of predicting the quality of processed parts can be widely promoted in industrialization. After training the Random Forest model, the deep learning neural network model, and the modified Xgboost model through the preprocessed feature data, respectively, the data samples to be tested are input into the above three models for analysis, and three analysis results are obtained. Integrate the above three results through the stacking model to realize the quality prediction of the processed parts corresponding to the sample data, and change the negative impact of the sparse feature of the data on the model.

Further, the loss function of the Xgboost model is constructed based on the logarithm maximum similarity, and the step S1 of judging whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece includes:

S11: Take the two-dimensional norm of the gradient matrix of the loss function

As a benchmark, a linear regression model is used to form the objective function of the Xgboost model

Wherein, the loss function is

y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result and the actual result obtained by the Xgboost model by inputting sample data, and α is the weight vector configuration ratio;

S12: Perform gradient optimization on the gradient of the Xgboost model according to the objective function, where the optimization direction of the gradient optimization is

X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio;

S13: Input the preprocessed sample data into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.

In this embodiment, too large learning parameters will cause the problem of too fast stride in the Xgboost optimization approaching process. In this case, it is easy to make the optimization direction and optimization number become infinite after optimization to a certain degree, which is expressed as Gradient explosion; if the learning parameter is too small, Xgboost will slowly approach the optimal result, but too slow optimization approach may make it fall into the trap of local minimum. If it falls into the local minimum, it will be overfitting. Pass The above loss function uses a linear regression model to form the objective function of the Xgboost model to achieve adaptive adjustments to the learning parameters in the Xgboost model, so as to select the most suitable learning parameters for the above sparse data analysis results through the linear regression model to ensure that the Xgboost model is suitable Optimization speed. Since the optimization approach method used by the Xgboost model is the second-order Taylor expansion of the loss function and the node value as the objective function, the shape of the loss function will largely determine the limit of the performance of the Xgboost model. Since the node activation method of the Xgboost model is sigmoid, for the input high-dimensional sparse structure data, during the optimization process of the Xgboost model, each optimization direction and optimization amount are mainly determined by the first derivative and the second derivative of the loss function , Using the logarithmic maximum similarity as the loss function will prevent the approaching stride of each optimization from being too large, avoiding the situation of f'->infi, that is, overfitting, and f'->0 will not appear In the case of gradient explosion, to ensure the smooth progress of the optimization.

The above-mentioned pre-processed sample data refers to the data after the sample data has passed the specified pre-processing method. The above-mentioned pre-processing method includes classifying the characteristic data of the processed parts in the sample data, using Monte Carlo tree search and other statistical methods to filter The feature data in model training is used to improve the development trend path that determines the quality of processed parts through various features and improve the accuracy of prediction. At the same time, the Random Froes model is used to filter the factors related to the quality of the processed parts, and the two are searched for intersection, and the quality parameter groups of the processed parts related to the substandard quality are obtained, and each quality parameter group corresponding to the development of the substandard product category is formed. The trend path, such as the trend path of non-compliant workpieces developed through density feature data, and the trend path of non-compliant workpieces developed through color feature data, weight feature data, and so on. Through further digging into the feature data of the logo features related to the substandard product types or quality parameter groupings, through normalization, regularization and other operations, the purpose of dimensionality reduction can be achieved. The pre-processed sample data includes the above-mentioned characteristic data and various trend paths that develop into substandard product categories. For example, the sample data is the characteristic data of the turbine blades. The turbine blades are huge processed parts. Whether the center of gravity of the turbine blades meets the requirements of the standard is the required prediction standard. However, in actual production, it is impossible to detect the center of gravity compliance of each turbine blade. . However, if the characteristic data such as the length, width, height and weight of the turbine blade are obtained, input the above model for predictive analysis. For example, through characteristic data such as length, width, height, weight, etc., it is developed into a trend path that the center of gravity of the turbine blade is not at the required position. For example, if low density is selected through the above two, it is easy to cause problems in the center of gravity, and the characteristic data of density will be paid special attention, and the density characteristic data will be used to predict the probability of the machined part not meeting the standard. When the analysis has multiple characteristic data For obvious influence trends, input multiple feature data into the prediction model for prediction at the same time.

Further, the deep learning neural network model includes an amplification layer, a deconstruction layer, and a learning layer. Before step S1 of judging whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece, it includes:

S1a: Select construction elements to respectively construct the amplification layer, deconstruction layer, and learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers connected in sequence Multilayer, the learning layer includes multiple hidden layers accumulated in sequence;

S1b: Connect the amplification layer, the deconstruction layer, and the learning layer in sequence to form the deep learning neural network model;

S1c: Input the preprocessed sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model.

In this embodiment, since the sample data is sparse structure data, it is necessary to perform deep detail enlargement, rearrangement, and deconstruction of the local features assigned to 0, so that the sample data is effective for model training, that is, the trained model responds to different samples. Data realizes differentiated analysis. The above-mentioned deep learning neural network model uses the amplification layer as the starting structure, the deconstruction layer as the intermediate structure, and the learning layer as the end structure. The above-mentioned amplification layer is composed of 4 hidden layers, which are 2^10ReLu, 2^11ReLu, 2^12ReLu and 2^13ReLu which are connected in sequence, 2^10ReLu is connected to the sample data input terminal, and 2^13ReLu is connected to the deconstruction layer. Since the sample data entering each hidden layer of the amplification layer is not subjected to convolution processing, only the sample data is partially enlarged to determine the distinguishing characteristics between the sample data, and the sample data is originally assigned to 0 local features for refinement distinguish. The above-mentioned deconstruction layer is composed of two convolutional layers, and each convolutional layer includes Bach Normalization*2^10 and Average Pooling which are sequentially connected. The sample data is rearranged and deconstructed through the deconstruction layer to obtain the local features with the original value of zero, and the association relationship to the entire sample data is distinguished after the magnification and refinement, and the value of non-zero is obtained. Input the processed sample data to the learning layer for learning and training, and learn and memorize the features of the processed sample data. The above-mentioned learning layer is composed of three hidden layers, respectively 2^10ReLu, 2^5ReLu and 2^4ReLu connected in sequence, 2^10ReLu is connected to the deconstruction layer, and 2^4ReLu is connected to the Softmax classifier. The preprocessing process of the above sample data is the same as above, and will not be repeated. The number of the aforementioned convolutional layers and hidden layers is determined according to the degree of optimization and the amount of calculation in the specific training process.

Further, before step S13 or S1c, it includes:

S101: Determine whether a correction instruction for the preprocessed sample data is received;

S102: If yes, the preprocessed sample data is partially enlarged and displayed according to the area selection instruction, where the area selection instruction is issued according to the mapping area when the user clicks on the screen and displayed at the sample data of the mapping area , The area selection instruction includes at least add and delete;

S103: Correct the sample data corresponding to the mapping area according to the type of the received area selection instruction;

S104: Restore the display state of the corrected sample data to the state before the partial enlarged display.

The sample data preprocessed in this embodiment can be manually revised. The preprocessed sample data is sample data processed through statistical processing or model processing, and it is presumed to be data that is beneficial to improving the prediction accuracy of the prediction model. , Improve the prediction accuracy after training the prediction model with sample data. The above-mentioned partial enlargement means that the connection relationship of the feature in the trend path of the product category that is developing into the substandard product category is not changed, and only the partial enlargement of the feature is performed, so that the feature can be accurately corrected.

Further, the step S3 of collecting the analysis results of at least two of the Xgboost model, the RandomForest model and the deep learning neural network model according to a preset method to obtain the summary result includes:

S30: After inputting the current data to be analyzed into the Xgboost model, RandomForest model, and deep learning neural network model, respectively, analysis results obtained respectively;

S31: Perform a weighted average on the matrix data corresponding to each analysis result according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result.

In this embodiment, the analysis results of the same input sample data are analyzed by the Xgboost model, the RandomForest model, and the deep learning neural network model, and the weighted average is used to obtain the summary result, so that the summary result can avoid the influence of the defects of each model, and the result is merged It is more in line with objective reality and the forecast results are more accurate. For example, the weights of the Xgboost model, the RandomForest model, and the deep learning neural network model are W1, W2, and W3, respectively. The prediction probabilities of the Xgboost model, RandomForest model, and deep learning neural network model for the current data to be analyzed are n1, n2, and n3, respectively, then the summary result is M, and the summary result M=W1*n1+W2*n2+W3*n3, where , N1, n2, n3 and the summary result M are all decimals between 0 and 1.

Further, performing a weighted average of the matrix data corresponding to each of the analysis results according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result before step S31 includes:

S311: Input sample data carrying tags into the Xgboost model, RandomForest model, and deep learning neural network model for training;

S312: Obtain feedback results of the Xgboost model, the RandomForest model, and the deep learning neural network model on the labeled sample data;

S313: Calculate the respective weights corresponding to the Xgboost model, RandomForest model, and deep learning neural network model through a linear regression model according to each of the feedback results and the assignment of the carrying tags.

In this embodiment, the Xgboost model, the RandomForest model, and the deep learning neural network model respectively give the feedback results of the sample data with tags x, y, and z. If there are multiple sample data with tags, the corresponding multiple Group x, y, z and the assignment t corresponding to each label, the assignment t corresponding to the label is 0 or 1, when t is 0, it is a label that does not meet the standard, and when t is 1, it is a label that meets the standard. There are as many sample data as there are combinations of x, y, z, and t to form a combination of W1*x+W2*y+W3*z=t, and the weights W1, W2, and W3 are calculated by linear regression model. In other embodiments of the present application, two models are used for summary analysis, and the summary process and principle of the two models are similar to the summary process and principle of the above three models, and will not be repeated.

Referring to Fig. 2, a device for predicting the quality of a processed part according to an embodiment of the present application includes:

The first judgment module 1 is used for judging whether the current data to be analyzed contains the mark features corresponding to the quality parameters of the defective processed parts;

The input module 2 is used to input the current to-be-analyzed data into a predictive model for predictive analysis if the current to-be-analyzed data contains the mark features corresponding to the quality parameters of the defective workpiece, wherein the predictive model includes at least the Xgboost model, Two of the RandomForest model and the deep learning neural network model, the Xgboost model has been modified and adjusted through a linear regression model;

The summary module 3 is configured to summarize the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method to obtain a summary result;

The second judgment module 4 is configured to judge the probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard according to the summary result.

Further, the loss function of the Xgboost model is constructed according to the logarithmic maximum similarity, and the device for predicting the quality of the processed part includes:

Forming module, used to calculate the two-dimensional norm of the loss function gradient matrix

Wherein, the loss function is

The optimization module is used to perform gradient optimization on the gradient of the Xgboost model according to the objective function, wherein the optimization direction of the gradient optimization is

The training module is used to input the preprocessed sample data into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.

Further, the deep learning neural network model includes an amplification layer, a deconstruction layer, and a learning layer. The device for predicting the quality of the processed part includes:

The selection module is used to select construction elements to respectively construct the amplification layer, deconstruction layer, and learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers in sequence A connected convolutional layer, where the learning layer includes multiple hidden layers accumulated in sequence;

The connection module is configured to sequentially connect the amplification layer, the deconstruction layer, and the learning layer to form the deep learning neural network model;

The determining module is used to input the preprocessed sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model.

Further, the device for predicting the quality of processed parts includes:

The third judgment module is used to judge whether a correction instruction to the preprocessed sample data is received;

The enlargement module is configured to, if a correction instruction for the preprocessed sample data is received, the preprocessed sample data is partially enlarged and displayed according to the area selection instruction, wherein the area selection instruction is based on the user's click The mapping area on the screen is sent out and displayed at the sample data of the mapping area, and the area selection instruction includes at least adding and deleting;

The correction module is configured to correct the sample data corresponding to the mapping area according to the type of the received area selection instruction;

The restoration module is used to restore the display state of the modified sample data to the state before the partial magnification display.

Further, the summary module 3 includes:

The first input unit is configured to input the current data to be analyzed into the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain analysis results respectively;

The summary unit is configured to perform a weighted average of the matrix data corresponding to each analysis result according to the weights corresponding to the Xgboost model, the RandomForest model and the deep learning neural network model to obtain the summary result.

Further, the summary module 3 includes:

The second input unit is used to input the sample data carrying the label into the Xgboost model, the RandomForest model and the deep learning neural network model for training;

An obtaining unit, configured to obtain feedback results of the Xgboost model, the RandomForest model, and the deep learning neural network model on the labeled sample data;

The calculation unit is used to calculate the respective weights of the Xgboost model, RandomForest model and deep learning neural network model through a linear regression model according to each of the feedback results and the assignment of the carrying tags.

Referring to FIG. 3, an embodiment of the present application also provides a computer device. The computer device may be a server, and its internal structure may be as shown in FIG. 3. The computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor designed by the computer is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer equipment is used to store all the data needed in the process of predicting the quality of the workpiece. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to realize the method of predicting the quality of the workpiece.

The processor executes the method for predicting the quality of the processed part, including: judging whether the current data to be analyzed contains a mark feature corresponding to the quality parameter of the defective processed part; if so, inputting the current to-be-analyzed data into a predictive model for predictive analysis, Wherein, the prediction model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model. The Xgboost model is modified and adjusted by a linear regression model; the Xgboost model, the RandomForest model, and the The analysis results of at least two models in the deep learning neural network model are summarized and processed to obtain a summary result; according to the summary result, the probability that the quality of the processed part corresponding to the current data to be analyzed is not up to standard is determined.

The above-mentioned computer equipment, based on the characteristics of the Xgboost model and the sparse data characteristics of the Xgboost model when used for the quality prediction of the processed parts, designed the loss function of the Xgboost model to use the logarithmic maximum similarity as the loss function and correct it through linear regression , Which makes the Xgboost model more suitable for feature analysis of sparse data, and enables the method of predicting the quality of processed parts to be widely promoted in industrialization. After training the Random Forest model, the deep learning neural network model, and the modified Xgboost model through the preprocessed feature data, respectively, the data samples to be tested are input into the above three models for analysis, and three analysis results are obtained. Integrate the above three results through the stacking model to realize the quality prediction of the processed parts corresponding to the sample data, and change the negative impact of the sparse feature of the data on the model.

In one embodiment, the loss function of the Xgboost model is constructed according to the logarithm maximum similarity, and the above-mentioned processor determines whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece before the step of: The two-dimensional norm of the function gradient matrix

Wherein, the loss function is

y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the Xgboost model's predicted result and the actual result obtained by inputting sample data, and α is the weight vector configuration ratio; the gradient of the Xgboost model is optimized according to the objective function, where the gradient is optimized The optimization direction is

X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio; input pre-processed sample data into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.

In an embodiment, the deep learning neural network model includes an amplification layer, a deconstruction layer, and a learning layer. Before the step of determining whether the current data to be analyzed contains the flag characteristics corresponding to the quality parameters of the defective workpiece, the above-mentioned processor includes: selecting The construction elements respectively construct the amplification layer, deconstruction layer, and learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers of convolutional layers connected in sequence, The learning layer includes multiple layers of hidden layers accumulated in sequence; sequentially connecting the amplification layer, deconstruction layer, and learning layer to form the deep learning neural network model; and inputting preprocessed sample data to the deep learning neural network model , To determine the model parameters of the deep learning neural network model.

In one embodiment, the processor inputs preprocessed sample data into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model before the step, or preprocess Before the step of inputting the latter sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model, the method includes: determining whether a correction instruction for the preprocessed sample data is received; if so , The preprocessed sample data is partially enlarged and displayed according to the area selection instruction, where the area selection instruction is issued according to the mapping area when the user clicks on the screen and is displayed at the sample data of the mapping area. The area selection instruction includes at least add and delete; according to the type of the received area selection instruction, the sample data corresponding to the mapping area is corrected; the display state of the corrected sample data is restored to the state before the partial enlarged display .

In one embodiment, the above-mentioned processor summarizes the analysis results of at least two of the Xgboost model, the RandomForest model and the deep learning neural network model according to a preset method, and the step of obtaining the summary result includes: combining all the analysis results of the Xgboost model, the RandomForest model, and the deep learning neural network model. The analysis results obtained after the current data to be analyzed are respectively input to the Xgboost model, RandomForest model, and deep learning neural network model; the matrix data corresponding to each analysis result is calculated according to the Xgboost model, RandomForest model and depth The weights corresponding to the neural network models are learned, and the weighted average is performed to obtain the summary result.

In one embodiment, the above-mentioned processor performs a weighted average on the matrix data corresponding to each of the analysis results according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result. , Including: inputting sample data carrying tags into the Xgboost model, RandomForest model, and deep learning neural network model for training; acquiring the Xgboost model, RandomForest model, and deep learning neural network model to perform training on the carrying tag The feedback results of the sample data; according to each of the feedback results and the assignment of the tags, the weights corresponding to the Xgboost model, the RandomForest model and the deep learning neural network model are calculated through a linear regression model.

Those skilled in the art can understand that the structure shown in FIG. 3 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.

The present application also provides a computer-readable storage medium. The computer-readable storage medium may be non-volatile or volatile. A computer program is stored thereon, and the computer program is executed by the processor to realize the prediction of the workpiece The quality method includes: judging whether the current data to be analyzed contains the mark features corresponding to the quality parameters of the defective processed parts; if so, inputting the current data to be analyzed into a predictive model for predictive analysis, wherein the predictive model at least includes Two of the Xgboost model, the RandomForest model and the deep learning neural network model. The Xgboost model is modified and adjusted by a linear regression model; at least one of the Xgboost model, the RandomForest model, and the deep learning neural network model is adjusted according to a preset method The analysis results of the two models are summarized and processed to obtain a summary result; according to the summary result, the probability that the quality of the processed part corresponding to the current data to be analyzed is not up to standard is judged.

The above-mentioned computer-readable storage medium, based on the characteristics of the Xgboost model and the sparse data characteristics when the Xgboost model is used for the quality prediction of processed parts, the loss function of the Xgboost model is designed to pass the logarithmic maximum similarity as the loss function and pass the linear Regression is modified to make the Xgboost model more suitable for feature analysis of sparse data, so that the method of predicting the quality of processed parts can be widely promoted in industrialization. After training the Random Forest model, the deep learning neural network model, and the modified Xgboost model through the preprocessed feature data, respectively, the data samples to be tested are input into the above three models for analysis, and three analysis results are obtained. Integrate the above three results through the stacking model to realize the quality prediction of the processed parts corresponding to the sample data, and change the negative impact of the sparse feature of the data on the model.

In an embodiment, the loss function of the Xgboost model is constructed according to the logarithm maximum similarity, and the above-mentioned processor determines whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece before the step of: The two-dimensional norm of the function gradient matrix

Wherein, the loss function is

y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result of the Xgboost model obtained by inputting sample data and the actual result, α is the weight vector configuration ratio; the gradient of the Xgboost model is optimized according to the objective function, where the gradient is optimized The optimization direction is

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by computer programs instructing relevant hardware. The above-mentioned computer programs can be stored in a non-volatile computer readable storage medium. Here, when the computer program is executed, it may include the procedures of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and/or volatile memory. Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not a limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous Link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

It should be noted that in this article, the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, It also includes other elements not explicitly listed, or elements inherent to the process, device, article, or method. If there are no more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, device, article, or method that includes the element.

The above are only the preferred embodiments of this application, and do not limit the scope of this application. Any equivalent structure or equivalent process transformation made using the content of the specification and drawings of this application, or directly or indirectly applied to other related The technical field is equally included in the scope of patent protection of this application.

Claims

A method for predicting the quality of machined parts, which includes:

Determine whether the current data to be analyzed contains the mark characteristics corresponding to the quality parameters of the defective workpiece;

If yes, input the current data to be analyzed into a prediction model for prediction analysis, where the prediction model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model, and the Xgboost model adopts a linear regression model Made corrections and adjustments;

Summarizing the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method, to obtain a summary result;

The probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard is determined according to the summary result.
The method for predicting the quality of a machined part according to claim 1, wherein the loss function of the Xgboost model is constructed based on logarithmic maximum similarity, and the judgment whether the current data to be analyzed contains a mark corresponding to the quality parameter of the defective machined part Before the characteristic steps, include:

Take the two-dimensional norm of the gradient matrix of the loss function
As a benchmark, a linear regression model is used to form the objective function of the Xgboost model
Wherein, the loss function is
y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result and the actual result obtained by the Xgboost model by inputting sample data, and α is the weight vector configuration ratio;

Perform gradient optimization on the gradient of the Xgboost model according to the objective function, where the optimization direction of the gradient optimization is
X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio;

The preprocessed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.
The method for predicting the quality of a machined part according to claim 1, wherein the deep learning neural network model includes an amplification layer, a deconstruction layer and a learning layer, and the judging whether the current data to be analyzed contains the quality parameter corresponding to the defective machined part Before the steps to mark features, include:

The construction elements are selected to construct the amplification layer, the deconstruction layer, and the learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers of convolutional layers connected in sequence , The learning layer includes multiple hidden layers accumulated in sequence;

Sequentially connecting the amplification layer, the deconstruction layer, and the learning layer to form the deep learning neural network model;

The preprocessed sample data is input to the deep learning neural network model to determine the model parameters of the deep learning neural network model.
The method for predicting the quality of a workpiece according to claim 2 or 3, wherein the pre-processed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the optimized Xgboost model Before the step of learning parameters, or before the step of inputting pre-processed sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model, it includes:

Judging whether a correction instruction to the preprocessed sample data is received;

If yes, the preprocessed sample data is partially enlarged and displayed according to the area selection instruction, where the area selection instruction is issued according to the mapping area when the user clicks on the screen and displayed at the sample data of the mapping area, so The area selection instructions include at least add and delete;

Correcting the sample data corresponding to the mapping area according to the type of the received area selection instruction;

Restore the display state of the corrected sample data to the state before the partial enlarged display.
The method for predicting the quality of a machined part according to claim 1, wherein the analysis results of at least two of the Xgboost model, RandomForest model, and deep learning neural network model are summarized and processed according to a preset method to obtain The steps to summarize the results include:

The analysis results obtained after inputting the current data to be analyzed into the Xgboost model, RandomForest model and deep learning neural network model respectively;

The matrix data corresponding to each analysis result is weighted and averaged according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result.
The method for predicting the quality of a workpiece according to claim 5, wherein the matrix data corresponding to each of the analysis results is weighted and averaged according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model. Before the step of summarizing the results, it includes:

Input the sample data carrying labels into the Xgboost model, RandomForest model and deep learning neural network model for training;

Acquiring the feedback results of the Xgboost model, the RandomForest model, and the deep learning neural network model on the sample data with tags;

According to each of the feedback results and the assignment of the tags, the weights corresponding to the Xgboost model, the RandomForest model and the deep learning neural network model are calculated through a linear regression model.
A device for predicting the quality of processed parts, which includes:

The first judging module is used to judge whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective processed part;

The input module is used to input the current to-be-analyzed data into a predictive model for predictive analysis if the current to-be-analyzed data contains the mark characteristics corresponding to the quality parameters of the defective workpiece, wherein the predictive model includes at least the Xgboost model and RandomForest Two of the model and the deep learning neural network model, the Xgboost model has been modified and adjusted through a linear regression model;

The summary module is configured to summarize the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method to obtain a summary result;

The second judgment module is used for judging the probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard according to the summary result.
The device for predicting the quality of a processed part according to claim 7, wherein the loss function of the Xgboost model is constructed based on logarithmic maximum similarity, and the device comprises:

Forming module, used to calculate the two-dimensional norm of the gradient matrix of the loss function
As a benchmark, a linear regression model is used to form the objective function of the Xgboost model
Wherein, the loss function is
y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result and the actual result obtained by the Xgboost model by inputting sample data, and α is the weight vector configuration ratio;

The optimization module is used to perform gradient optimization on the gradient of the Xgboost model according to the objective function, wherein the optimization direction of the gradient optimization is
X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio;

The training module is used to input the preprocessed sample data into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.
A computer device includes a memory and a processor, the memory stores a computer program, wherein the method for realizing the quality of a workpiece when the processor executes the computer program includes:

Determine whether the current data to be analyzed contains the mark characteristics corresponding to the quality parameters of the defective workpiece;

If yes, input the current data to be analyzed into a prediction model for prediction analysis, where the prediction model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model, and the Xgboost model adopts a linear regression model Made corrections and adjustments;

Summarizing the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method, to obtain a summary result;

The probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard is determined according to the summary result.
The computer device according to claim 9, wherein the loss function of the Xgboost model is constructed according to the logarithmic maximum similarity, and the step of judging whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece ,include:

Take the two-dimensional norm of the gradient matrix of the loss function
As a benchmark, a linear regression model is used to form the objective function of the Xgboost model
Wherein, the loss function is
y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result and the actual result obtained by the Xgboost model by inputting sample data, and α is the weight vector configuration ratio;

Perform gradient optimization on the gradient of the Xgboost model according to the objective function, where the optimization direction of the gradient optimization is
X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio;

The preprocessed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.
The computer device according to claim 9, wherein the deep learning neural network model includes an amplification layer, a deconstruction layer, and a learning layer, and the step of judging whether the current data to be analyzed contains a mark feature corresponding to a quality parameter of a defective workpiece Before, including:

The construction elements are selected to construct the amplification layer, the deconstruction layer, and the learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers of convolutional layers connected in sequence , The learning layer includes multiple hidden layers accumulated in sequence;

Sequentially connecting the amplification layer, the deconstruction layer, and the learning layer to form the deep learning neural network model;

The preprocessed sample data is input to the deep learning neural network model to determine the model parameters of the deep learning neural network model.
The computer device according to claim 10 or 11, wherein the pre-processed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model Before the step, or before the step of inputting preprocessed sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model, the method includes:

Judging whether a correction instruction to the preprocessed sample data is received;

If yes, the preprocessed sample data is partially enlarged and displayed according to the area selection instruction, where the area selection instruction is issued according to the mapping area when the user clicks on the screen and displayed at the sample data of the mapping area, so The area selection instructions include at least add and delete;

Correcting the sample data corresponding to the mapping area according to the type of the received area selection instruction;

Restore the display state of the corrected sample data to the state before the partial enlarged display.
8. The computer device according to claim 9, wherein the step of performing summary processing on the analysis results of at least two of the Xgboost model, RandomForest model, and deep learning neural network model according to a preset method, to obtain the summary result ,include:

The analysis results obtained after inputting the current data to be analyzed into the Xgboost model, RandomForest model and deep learning neural network model respectively;

The matrix data corresponding to each analysis result is weighted and averaged according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result.
The computer device according to claim 13, wherein the matrix data corresponding to each of the analysis results are weighted and averaged to obtain the summary result according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model. Before the steps, include:

Input the sample data carrying labels into the Xgboost model, RandomForest model and deep learning neural network model for training;

Acquiring the feedback results of the Xgboost model, the RandomForest model, and the deep learning neural network model on the labeled sample data;

According to each of the feedback results and the assignment of the tags, the weights corresponding to the Xgboost model, the RandomForest model and the deep learning neural network model are calculated through a linear regression model.
A computer-readable storage medium having a computer program stored thereon, wherein the method for predicting the quality of a processed part when the computer program is executed by a processor includes:

Determine whether the current data to be analyzed contains the mark characteristics corresponding to the quality parameters of the defective workpiece;

If yes, input the current data to be analyzed into a prediction model for prediction analysis, where the prediction model includes at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model, and the Xgboost model adopts a linear regression model Made corrections and adjustments;

Summarizing the analysis results of at least two of the Xgboost model, the RandomForest model, and the deep learning neural network model according to a preset method, to obtain a summary result;

The probability that the quality of the processed part corresponding to the current data to be analyzed does not meet the standard is determined according to the summary result.
15. The computer-readable storage medium according to claim 15, wherein the loss function of the Xgboost model is constructed based on the logarithmic maximum similarity, and the judgment whether the current data to be analyzed contains the mark feature corresponding to the quality parameter of the defective workpiece Before the steps, include:

Take the two-dimensional norm of the gradient matrix of the loss function
As a benchmark, a linear regression model is used to form the objective function of the Xgboost model
Wherein, the loss function is
y refers to the real result, x refers to the input sample data, θ refers to the weight of each function in the Xgboost model, P(|) is the conditional probability, ω is the weight of each variable in the linear regression model, J(ω; X , Y) is the difference between the predicted result and the actual result obtained by the Xgboost model by inputting sample data, and α is the weight vector configuration ratio;

Perform gradient optimization on the gradient of the Xgboost model according to the objective function, where the optimization direction of the gradient optimization is
X refers to the input sample data, ω is the weight of each variable in the linear regression model, J(ω; X, y) is the difference between the predicted result and the actual result obtained by the Xgboost model through the input sample data, α is Weight vector configuration ratio;

The preprocessed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the learning parameters of the optimized Xgboost model.
The computer-readable storage medium according to claim 15, wherein the deep learning neural network model includes an amplification layer, a deconstruction layer, and a learning layer, and the judgment whether the current data to be analyzed contains a mark corresponding to the quality parameter of the defective processed part Before the characteristic steps, include:

The construction elements are selected to construct the amplification layer, the deconstruction layer, and the learning layer of the deep learning neural network model, wherein the amplification layer includes multiple layers of hidden layers accumulated in sequence, and the deconstruction layer includes multiple layers of convolutional layers connected in sequence , The learning layer includes multiple hidden layers accumulated in sequence;

Sequentially connecting the amplification layer, the deconstruction layer, and the learning layer to form the deep learning neural network model;

The preprocessed sample data is input to the deep learning neural network model to determine the model parameters of the deep learning neural network model.
The computer-readable storage medium according to claim 16 or 17, wherein the pre-processed sample data is input into the Xgboost model for gradient optimization for training, so as to determine the optimized Xgboost model Before the step of learning parameters, or before the step of inputting pre-processed sample data into the deep learning neural network model to determine the model parameters of the deep learning neural network model, it includes:

Judging whether a correction instruction to the preprocessed sample data is received;

If yes, the preprocessed sample data is partially enlarged and displayed according to the area selection instruction, where the area selection instruction is issued according to the mapping area when the user clicks on the screen and displayed at the sample data of the mapping area, so The area selection instructions include at least add and delete;

Correcting the sample data corresponding to the mapping area according to the type of the received area selection instruction;

Restore the display state of the corrected sample data to the state before the partial enlarged display.
The computer-readable storage medium according to claim 15, wherein the analysis results of at least two of the Xgboost model, RandomForest model, and deep learning neural network model are summarized according to a preset method to obtain a summary The results of the steps include:

The analysis results obtained after inputting the current data to be analyzed into the Xgboost model, RandomForest model and deep learning neural network model respectively;

The matrix data corresponding to each analysis result is weighted and averaged according to the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model to obtain the summary result.
The computer-readable storage medium according to claim 19, wherein the matrix data corresponding to each of the analysis results is weighted and averaged according to the weights corresponding to the Xgboost model, RandomForest model, and deep learning neural network model. Before describing the steps to summarize the results, include:

Input the sample data carrying labels into the Xgboost model, RandomForest model and deep learning neural network model for training;

Acquiring the feedback results of the Xgboost model, the RandomForest model, and the deep learning neural network model on the labeled sample data;

According to each of the feedback results and the assignment of the tags, the weights corresponding to the Xgboost model, the RandomForest model, and the deep learning neural network model are calculated through a linear regression model.