BP neural network-based process abnormality diagnosis expert knowledge acquisition method

Abstract

This invention discloses a method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network, comprising: classifying a non-conforming product database, matching production information obtained from an MES system, and constructing a production information dataset; expanding the samples using Monte Carlo simulation, and training the BP neural network to obtain a control chart pattern recognition model M1; using M1 to identify control chart patterns in actual quality characteristic data; merging the identification results with fault causes and corrective measures to construct a basic database for process quality diagnosis; grouping by process, constructing a modified BP neural network, and training it with control chart anomaly patterns and fault causes as input and output, respectively, learning the nonlinear mapping relationship between the two, and storing the trained weights and thresholds as expert knowledge. This invention achieves self-learning and dynamic acquisition of process anomaly diagnosis knowledge, effectively solving the bottleneck problem of knowledge acquisition in expert systems, and improving the accuracy and intelligence level of diagnosis.

Description

Technical Field

[0001] This invention relates to the field of quality diagnosis technology for aerospace product manufacturing processes, specifically to a method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network. Background Technology

[0002] In recent years, scholars both domestically and internationally have frequently introduced technologies such as neural networks and expert systems into process quality diagnosis to analyze the sources of process quality fluctuations, thereby improving the capability and accuracy of process quality diagnosis. Expert systems, as one of the core technologies for process anomaly diagnosis, possess unique advantages in fault diagnosis, including interactivity, transparency, and flexibility. However, they still face a bottleneck in knowledge acquisition. The amount of knowledge is a key factor affecting the diagnosis of process anomalies in aerospace product assembly. Traditionally, expert knowledge can be summarized by experienced engineers within the field, but due to the limitations of individual knowledge, there is a lack of feasible means or tools to aggregate collective experience, resulting in a significant reduction in the effectiveness of process anomaly diagnosis. Summary of the Invention

[0003] The purpose of this invention is to provide a method for acquiring expert knowledge for process anomaly diagnosis based on BP neural network. The method aims to dynamically acquire expert knowledge for process anomaly diagnosis from the non-conforming product database of FRACAS (Failure Report Analysis and Corrective Action System) system, so as to realize machine self-learning and memory, as well as nonlinear knowledge solving.

[0004] To achieve the above objectives, the present invention employs the following technical solution: A method for acquiring expert knowledge for process anomaly diagnosis based on a backpropagation neural network includes: All non-conforming product records are retrieved from the aviation product non-conforming product database in the IQS system, classified according to different aviation product processing specialties and processes, and uniquely identified by the non-conforming product drawing number. The product manufacturing record barcode in the MES system is matched to obtain the production information of the processing / manufacturing batch in which each non-conforming quality characteristic belongs. The classification information is linked and integrated with the obtained production information to form a production information dataset. Based on the constructed production information dataset, the process quality characteristics corresponding to the sample measured values ​​of each processing / manufacturing batch are taken as the simulation object; the statistical characteristics of the sample measured values ​​are used as the benchmark, and the Monte Carlo data simulation method is used to expand the data to generate simulation sample data; after preprocessing the simulation sample data, it is used as the input of the first BP neural network for training to obtain model M1 for recognizing control chart pattern categories; For each production information record in the production information dataset, the set of sample measured values ​​corresponding to each record is preprocessed by standardization and encoding. Then, the processed data is input into the trained model M1 for control chart pattern recognition, and the actual control chart pattern recognition result corresponding to each non-conforming product record is obtained, forming the actual control chart pattern recognition result set. The actual control chart pattern recognition result set is combined and stored with the corresponding non-conforming quality information retrieved from the aviation product non-conforming product database of the IQS system to form a basic database for process quality diagnosis. For the basic database of process quality diagnosis, it is grouped according to different manufacturing processes of different processing specialties of aerospace products. In each group, a second BP neural network is constructed. The control chart anomaly pattern and the cause of failure are used as the input and output samples of the neural network for training. The nonlinear mapping relationship from the control chart anomaly pattern to the cause of failure is learned. After training, the weights and thresholds of each neural network are written into the expert database as expert knowledge for process anomaly diagnosis, thereby realizing the intelligent acquisition of expert knowledge for process anomaly cause diagnosis.

[0005] Furthermore, the process of constructing the production information dataset includes: Establish a data interface between the IQS system and the MES system; for each non-conforming product record after classification, use the "part drawing number" as a unique identifier in the non-conforming product database of the IQS system and match it with the product manufacturing record barcode in the MES system; each product manufacturing record barcode uniquely corresponds to the same part drawing number information; through matching, obtain the production information of the processing / manufacturing batch in which the non-conforming quality characteristics are located, including the total sample size and the set of sample measured values; Each processing specialty category and process information obtained from the classification is associated and integrated with the acquired manufacturing record barcodes, total sample size, and sample measured value set to form a production information record, thus constituting a production information dataset.

[0006] Furthermore, the Monte Carlo method is used to perform random simulation sampling to generate simulation sample data. : ; in, Sampling time; for Simulated measurement values ​​of process quality characteristics at any given time; The mean value of the quality characteristic measurements is taken here. ; for The deviation of the normal distribution at time points is taken here. ; Represents a normal distribution; This is an interference value; , The mean and standard deviation of the sample measured values ​​for each processing / manufacturing batch; For simulation sample data Standardization and encoding preprocessing are performed; the preprocessed simulation sample data is used as input, and the corresponding known control chart pattern category is used as output, which is then input into the first BP neural network for model training; The control chart modes include: normal mode, uptrend mode, downtrend mode, upward step mode, downward step mode, and cycle mode.

[0007] Furthermore, the encoding formula for the encoding process is as follows: ; in, It is about data The data obtained after encoding; This is the coding rate parameter; The values ​​are obtained after standardization of the sample measured values. Preprocessed data As input, the trained model M1 is used for forward computation to obtain the actual control chart pattern recognition result corresponding to each nonconforming product record. This constitutes the actual control chart pattern recognition result set. .

[0008] Furthermore, based on the product information in the MES system, a matching search is performed in the non-conforming product database of the IQS system to obtain the quality information of the non-conforming quality characteristics of aerospace products / parts; the actual control chart pattern recognition result set and the quality information of non-conforming quality characteristics are merged and stored to obtain the basic database for process quality diagnosis.

[0009] Furthermore, the basic database for process quality diagnosis is grouped according to different manufacturing processes of different processing specialties of aerospace products; in each group, a second BP neural network is constructed; the network adopts a supervised learning method and includes a three-layer structure of input layer, hidden layer and output layer; the transfer function between the input layer and the hidden layer adopts the smooth differentiable function Sigmoid; the transfer function between the hidden layer and the output layer adopts the hard limit function. During the forward propagation of the second BP neural network, the output value is calculated layer by layer from the input layer to the next: Let the input signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a control chart pattern recognition result. The number of input layer nodes is equal to the number of abnormal pattern categories in the control chart; the input signal... The control chart pattern category is derived from the control chart pattern category field of each record in the process quality diagnosis basic database; the control chart pattern category is encoded as follows: A binary vector of bits is used as input; Let the desired output signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a fault cause; the desired output signal The fault cause field is derived from the records in the process quality diagnosis basic database and uses one-hot encoding. This represents the number of output layer nodes, and its value is equal to the number of fault cause categories in this group. Let the actual output signal be ,in ; The error between the actual output and the expected output is calculated to obtain the total output error of a batch of samples. The error signal is propagated layer by layer from the output layer forward, and the weights are adjusted until the error is less than the set value. During the iteration process, the differential chain rule and the negative gradient descent method are used. The step size of the gradient descent method is dynamically controlled to control the learning convergence speed.

[0010] Furthermore, let , Let be the step size of the gradient descent method; then the dynamic control of the gradient descent step size is as follows: like ,but Remain unchanged; like Then adjust ; like Then make the following adjustments: If ,but Keep it unchanged, otherwise ; in, For the first The total output error of the next iteration; For the first The total output error of the next iteration; This is the step size of the gradient descent method used in the current iteration; This is the step size of the gradient descent method used in the previous iteration.

[0011] Furthermore, when it is necessary to diagnose anomalies in a specific process, the following operations are performed based on the real-time quality characteristic measurements of that process: Obtain real-time quality characteristic measurements for this process to form a set of measured values. ;right After standardization and encoding preprocessing, the data is input into the trained model M1 to identify the current control chart anomaly pattern for that process, thus obtaining the control chart anomaly pattern category. ; Based on the processing specialty and process to which the process belongs, the corresponding group of the pre-trained second BP neural network is retrieved from the constructed expert knowledge base; Identified control chart anomaly pattern categories As input to the second BP neural network, forward propagation calculation is performed, and the output of the second BP network is the diagnosed fault cause, its corresponding fault cause code, and corrective measures.

[0012] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network.

[0013] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network.

[0014] Compared with the prior art, the present invention has the following technical features: This invention automatically acquires process quality characteristic detection data and performs data standardization and encoding to reduce fluctuation noise caused by common factors in the process. It employs Monte Carlo simulation to obtain a large amount of sample data with similar quality characteristics to actual production, solving the problem of insufficient data for neural network model training. It uses a BP neural network for control chart pattern recognition to automatically classify process anomalies. By using control chart anomaly patterns and fault causes as input and output samples for neural network training, a nonlinear mapping relationship from control chart anomaly patterns to fault causes is derived. This solves the bottleneck problem of "knowledge acquisition" in expert systems, enabling self-learning and memorization of diagnostic knowledge, which is beneficial for improving the diagnostic capability and accuracy of process diagnostic expert systems. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a control chart pattern recognition result - upward trend pattern in one embodiment of the present invention. Detailed Implementation

[0016] Currently, aerospace product manufacturers have successively established FRACAS systems. FRACAS systems have gathered a large number of excellent fault analysis and solution cases. If these data are effectively mined, analyzed and utilized, it can make up for the "bottleneck" in acquiring diagnostic knowledge and improve the diagnostic accuracy of process anomaly early warning or expert systems.

[0017] This invention provides a method for acquiring expert knowledge for process anomaly diagnosis based on a backpropagation neural network. This method can dynamically acquire expert knowledge for process anomaly diagnosis from the non-conforming product database of the FRACAS system, and features machine self-learning and memory, as well as nonlinear knowledge solving. The specific steps of this invention are as follows: Step S1: Classify the non-conforming products database of aviation products and match it with the MES system to construct a production information dataset.

[0018] For all non-conforming product records in the aviation product non-conforming product database of the IQS (Integrated Quality System) system, they are classified according to different aviation product processing specialties and processes (parts), and the non-conforming product drawing number is used as a unique identifier. The product manufacturing record barcode of the MES (Manufacturing Execution System) system is matched to obtain the production information of the processing / manufacturing batch in which each non-conforming quality characteristic belongs. The classification information and the obtained production information are linked and integrated to form a production information dataset.

[0019] The aforementioned non-conforming quality characteristics refer to the collective term for a specific quality indicator (such as dimensions, geometric tolerances, surface quality, etc.) that exceeds the tolerance range specified in the design or process during the manufacturing process of aerospace products, resulting in the product or part being judged as a non-conforming product.

[0020] S101, classify the data according to processing specialty and process (part).

[0021] All non-conforming product records in the aviation product non-conforming product database are classified according to their respective processing specialty and process (part); processing specialty category ZY includes: sheet metal ( ),welding( ), heat meter ( ), additive manufacturing ), composite materials ), machining ( ),assembly( ), test flight ( ) etc., represented as: ; The process categories for each machining specialty are distinguished by the different processes (parts).

[0022] S102, Establish a data interface to obtain production information.

[0023] Establish a data interface between the IQS system and the MES system; for each non-conforming product record after classification, use the "part number" as a unique identifier in the non-conforming product database of the IQS system and match it with the product manufacturing record barcode in the MES system; each product manufacturing record barcode uniquely corresponds to the same part number information; through matching, obtain the production information of the processing / manufacturing batch where the non-conforming quality characteristics are located, including the total sample size. and the set of measured sample values .

[0024] ; in, For the total sample size, For the first Measured values ​​of a sample and All are positive integers.

[0025] The information on each processing specialty category and process (part) obtained from S101 is compared with the manufacturing record barcode, total sample size, and sample measured value set obtained from S102. The data is then integrated and linked together as a single production information record, thus forming a production information dataset.

[0026] In one embodiment of the present invention, in the non-conforming product database of the integrated quality system (IQS), the non-conforming part name is wall panel, the part drawing number is Jxx.a.100, the non-conforming quality characteristic number is TX001, and the non-conformity description is that the wall panel thickness is out of tolerance. Based on part number Jxx.a.100, the product manufacturing records in the MES system are matched through the data interface. The total number of production samples of the same batch of parts containing the non-conforming part is 25. The set of wall panel thickness measurement values ​​is test, in mm. Then test = {18.3, 18.4, 18.2, 18.4, 18.3, 18.6, 18.3, 18.5, 18.6, 18.4, 18.5, 18.5, 18.6, 18.5, 18.5, 18.6, 18.5, 18.6, 18.5, 18.6, 18.4, 18.5, 18.5, 18.6}.

[0027] Step S2: For the production information dataset, the Monte Carlo method is used for simulation expansion, and the first BP neural network is trained to obtain the control chart pattern recognition model M1.

[0028] For the constructed production information dataset, the set of sample measured values ​​for each processing / manufacturing batch is used. The corresponding process quality characteristics are used as simulation objects; the statistical characteristics of the sample measured values ​​are used as the benchmark, and the Monte Carlo data simulation method is used to expand the data to generate a large number of simulation sample data with similar quality characteristics; after preprocessing the simulation sample data, it is used as the input of the first BP neural network for training to obtain model M1 for identifying control chart pattern categories; this model establishes a mapping relationship from "process quality characteristic measurement value" to "control chart pattern category".

[0029] Process quality characteristics refer to the quantitatively measurable technical indicators (such as dimensions, geometric tolerances, surface quality, etc.) used to characterize the quality status of products or parts in the manufacturing process of aerospace products, as well as the measured values ​​generated during the processing / manufacturing process.

[0030] S201 uses the Monte Carlo method to generate simulation sample data.

[0031] For the aforementioned production information dataset, the set of sample measured values ​​for each processing / manufacturing batch is used. mean and standard deviation Based on this, the Monte Carlo method was used to perform random simulation sampling, generating a large amount of simulation sample data that closely matches actual production conditions. : ; in, Sampling time, ; for Simulated measurement values ​​of process quality characteristics at any given time; The mean value of quality characteristic measurements under controlled process conditions is taken here. ; For the process under controlled conditions The deviation of the normal distribution at time t, and , Let be the standard deviation under controlled conditions, which is taken here. ; Represents a normal distribution; For quality characteristic measurement interference values ​​caused by abnormal factors, when the simulation process is under control, .

[0032] S202, train the first BP neural network to obtain model M1.

[0033] The simulation sample data generated in step S201 Standardization and encoding preprocessing are performed (see S301); the preprocessed simulation sample data is used as input, and the corresponding known control chart pattern category is used as output, which is then input into the first BP neural network for model training; after training, model M1 for recognizing control chart pattern categories is obtained; there are 6 control chart pattern categories: normal mode, upward trend mode, downward trend mode, upward step mode, downward step mode, and periodic mode; the mathematical expressions for each mode are: a. Normal mode (NOR): The points on the control chart are randomly distributed; b. Periodic Pattern (CYC): ,in For amplitude, For periodicity; c. Upward / downward trend pattern (IT): ,in The slope represents the trend; "+" indicates an upward trend, and "-" indicates a downward trend. d. Upward (downward) step mode (US): ,in The moment when the step jump occurs. For step amplitude, For a unit step function, when hour, ,otherwise .

[0034] Step S3: For the set of measured values ​​of samples in the production information dataset, preprocess and input them into model M1 to obtain the actual control chart pattern recognition result set.

[0035] For each production information record in the production information dataset, the corresponding set of sample measured values Standardization and coding preprocessing are performed, and then the processed data is input into the trained model M1 for control chart pattern recognition to obtain the actual control chart pattern recognition result corresponding to each nonconforming product record, forming an actual control chart pattern recognition result set.

[0036] S301, for the set of measured sample values Then, standardization and encoding processes are performed sequentially.

[0037] a. Standardization process: ; in, These are the actual measured values ​​of the sample. for Standardized values; This is an estimate of the measured mean of the sample under controlled process conditions; here, the mean of the measured values ​​of this batch of samples is used. ; This is an estimate of the measured standard deviation of the sample under controlled process conditions; here, the standard deviation of the measured values ​​of this batch of samples is used. .

[0038] b. Encoding processing: The data encoding formula is as follows: Encoding reduces fluctuation noise caused by common factors in the process environment. ; in, It is about data The data obtained after encoding; The encoding rate parameter, in this example, has a value range of [value range missing]. ; The value is obtained after S301 standardization; when Falling into When there are segmented intervals, the encoded value is .

[0039] S302 uses model M1 for pattern recognition.

[0040] Preprocessed data As input, the trained model M1 is used for forward computation to obtain the actual control chart pattern recognition result corresponding to each non-conforming product record; the number of output layer nodes of model M1. The number of predefined control chart pattern categories; Model M1 is for The output result (corresponding to each production information record) is denoted as Its specific content is as follows: ; in, A manufacturing record barcode is used to uniquely identify a manufacturing batch; Assign a work number to identify the production task; Product model; For product batch shelf; This is the process number; Name of quality characteristic; Number the quality characteristics; For part or assembly drawing numbers; The control chart pattern category identified for model M1.

[0041] The output results of all production information records, denoted as W, together constitute the actual control chart pattern recognition result set. .

[0042] In one embodiment of the present invention, 25 quality characteristic measurement data parameters of the panel part are used as 25 input layer nodes of the M1 model, and 5 control chart anomaly pattern parameters are used as 5 output layer nodes of the M1 model. After identification by the machine learning algorithm, the quality characteristic control chart pattern recognition result of the non-conforming part is determined to be an "upward trend pattern". Figure 2 The output result W is denoted as: W={ Manufacturing record barcode-ZZJL001, dispatch number-N001, model-Jxx, batch / shelf number-0101, operation number-50, quality characteristic name-panel thickness, quality characteristic number-TX001, part / assembly drawing number-Jxx.a.100, control chart mode category-upward trend mode}.

[0043] Step S4: Merge the actual control chart pattern recognition results with the non-conforming product quality information to obtain the basic database for process quality diagnosis.

[0044] For the actual control chart pattern recognition result set This information is then merged and stored with the corresponding non-conforming quality information retrieved from the aviation product non-conforming product database of the IQS system to form a basic database for process quality diagnosis.

[0045] S401, retrieve non-conforming quality information.

[0046] Based on information such as product manufacturing record barcodes, work order numbers, and part drawing numbers from the MES system, a search and matching process is performed in the non-conforming product database of the IQS system to obtain quality information on the non-conforming quality characteristics of aerospace products / parts. Its specific content is as follows: ; in, Description of the cause of the fault; The code indicating the cause of the fault; Corrective measures were taken to address this fault.

[0047] S402, merge data to form a basic database.

[0048] The actual control chart pattern recognition result set Quality information related to non-conforming quality characteristics The data is merged and stored to obtain a basic database for process quality diagnosis, denoted as... The merging method is as follows: .

[0049] Each record in this database represents the "control chart pattern category" for a specific process (since this record originates from the non-conforming products database). Typically, the abnormal control chart mode (i.e., the five modes other than the normal mode in the control chart) is associated with its "fault cause" and "corrective action".

[0050] Step S5: Group the basic database for process quality diagnosis, construct the first BP neural network for training, obtain the nonlinear mapping relationship from abnormal control chart patterns to fault causes, and store it in the expert knowledge base.

[0051] Basic database for process quality diagnosis The components are grouped according to different manufacturing processes (parts) of different processing specialties of aviation products. In each group, a second BP neural network is constructed. The network is trained by using control chart anomaly patterns and fault causes as input and output samples, respectively, to learn the nonlinear mapping relationship from control chart anomaly patterns to fault causes. After training, the weights and thresholds of each neural network are written into an expert database as expert knowledge for process anomaly diagnosis, thereby realizing the intelligent acquisition of expert knowledge for process anomaly cause diagnosis.

[0052] S501, Data Packet and Network Structure Definition.

[0053] The basic database for process quality diagnosis is organized by different manufacturing processes (parts) in different processing specialties of aerospace products. Group the network into groups; construct a second BP neural network in each group; the network adopts a supervised learning approach and has a three-layer structure consisting of an input layer, a hidden layer, and an output layer.

[0054] Transfer function between input layer and hidden layer Using the smooth, differentiable function Sigmoid: ; in, This is the weighted input sum of the input layer nodes.

[0055] Transfer function between hidden layer and output layer Using hard limit function : ; in, This is the weighted sum of the inputs to the hidden layer nodes. The weights between neurons are obtained through training.

[0056] S502, forward propagation computation.

[0057] During the forward propagation of the second BP neural network, the output value is calculated layer by layer from the input layer to the back layer.

[0058] Let the input signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a control chart pattern recognition result. This represents the number of nodes in the input layer, and its value is equal to the number of abnormal pattern categories in the control chart; here, the control chart pattern recognition result corresponding to the input signal does not include normal patterns (NOR). Input signal Sourced from the basic database of process quality diagnosis Each record Field (Control Chart Pattern Category); will Control chart mode category coding is as follows The input is a binary vector of bits, and the encoding rule is as follows: The value is equal to the number of abnormal mode categories in the control chart (i.e., 5 abnormal modes). One-hot encoding is used, and each abnormal mode corresponds to a unique one. A binary vector with only one bit set to 1 and the rest set to 0.

[0059] Let the desired output signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a fault cause; the desired output signal From the same database records The field (fault cause) is encoded in the same way as the input signal, using one-hot encoding. The number of output layer nodes is equal to the number of fault cause categories in the group; the deduplication set of each fault cause is read from the non-conforming product database of the group.

[0060] Let the actual output signal be ,in .

[0061] Let the hidden layer be the first The output of each node is ,in ; The number of hidden layer nodes is determined through experiments or empirical formulas.

[0062] The weight matrix between the input layer and the hidden layer is The elements in column 0 Corresponding hidden layer Threshold for each node.

[0063] The weight matrix between the hidden layer and the output layer is The elements in column 0 Corresponding output layer number Threshold for each node.

[0064] In one embodiment of the present invention, the reasons for non-compliance of the wall panel thickness quality characteristics obtained after this step include:

[0065] The expected output signals of the sub-neural network are 8, D={d l}, l=1,2,…,8,d l The value can be 0 or 1, and the relationship between the reason code and the neural network output parameters is as follows:

[0066] S503, Error Calculation.

[0067] Calculate the error between the actual output and the expected output; the total output error for a single sample. for: ; in, For the first The actual output value of each output node; For the first The expected output value of each output node; This represents the number of nodes in the output layer.

[0068] Total output error for a batch of samples for: ; S504, Backpropagation and Weight Adjustment.

[0069] The error signal is propagated layer by layer forward from the output layer, and the weights are adjusted until the error is less than a set value; this process continues for several iterations. During the iterative process, the differential chain rule and the negative gradient descent method are used to calculate the th... Weights in the next iteration Changes: ; in, , This represents the number of nodes in the input layer. , This represents the number of hidden layer nodes. This is the step size for gradient descent. For the hidden layer The error term for each node is calculated using the following formula: ; in, For the hidden layer The output of each node; For the hidden layer Local error corresponding to each node; For the hidden layer The expected output corresponding to each node (here, the error signal propagated to the hidden layer).

[0070] S505, Dynamic Step Size Control.

[0071] During backpropagation, the gradient descent step size defined in step S504 is adjusted. Dynamic control is implemented to regulate the learning convergence speed. Let... The algorithm is as follows: like This indicates that the learning algorithm converges quickly. Remain unchanged; like This indicates that the learning algorithm is diverging and needs adjustment. ; like This indicates that the learning algorithm is converging slowly and requires further adjustments as follows: If ,but Keep it unchanged, otherwise .

[0072] in, For the first The total output error of the next iteration; For the first The total output error of the next iteration; This is the step size of the gradient descent method used in the current iteration; This is the step size of the gradient descent method used in the previous iteration, used to determine the historical state of the step size.

[0073] S506, acquires and stores expert knowledge.

[0074] Using the abnormal control chart pattern and the cause of failure as the input and output samples of the second BP neural network of each group, respectively, the network is trained according to steps S502 to S505 to obtain the nonlinear mapping relationship from the abnormal control chart pattern to the cause of failure. After training, the weights and thresholds of each group neural network are written into the database as an expert knowledge base for process abnormality diagnosis.

[0075] The result set JG must contain at least the following data: JG = {Description of nonconforming quality characteristic jg1, process number jg2, quality characteristic number jg3, control chart mode category jg4, cause of failure jg5, cause of failure code jg6, corrective action jg7} For example, according to the embodiment, the sub-neural network has 12 input layer nodes i, 20 hidden layer nodes s (using the simple formula: number of hidden layer nodes = number of input layer nodes + number of output layer nodes; the specific number needs to be adjusted and optimized according to the actual situation), 8 output layer nodes n, a maximum number of iterations of 12000, and a learning step size of... With a target error E of 0.01 and a value of 0.1, the fault cause output is "mold deformation" obtained through sample training. The non-conforming product database is retrieved; the cause code is G6, and the corrective action is "replace the mold". The machine learning result of the process anomaly cause diagnosis expert knowledge is recorded and denoted as JG. JG={Panel thickness out of tolerance, step 50, feature number TX001, upward trend mode, fault cause - mold deformation, cause code G6, corrective action - replace mold} This invention discloses a method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network. By leveraging the nonlinear pattern recognition capability and parallel processing capability of the neural network, it achieves rapid and accurate online fault diagnosis. It can effectively handle complex scenarios such as "the same fault may correspond to multiple abnormal symptoms" and "the same symptom may correspond to multiple fault causes," and can be widely applied to the process of diagnosing anomalies in production processes.

[0076] Step S6: When it is necessary to diagnose an anomaly in a specific process, perform the following operations based on the real-time quality characteristic measurement values ​​of that process: Obtain real-time quality characteristic measurements for this process to form a set of measured values. ;right After standardization and encoding preprocessing, the data is input into the trained model M1 to identify the current control chart anomaly pattern for that process, thus obtaining the control chart anomaly pattern category. ; Based on the processing specialty and process (part) to which the process belongs, the corresponding group of the trained second BP neural network is retrieved from the constructed expert knowledge base; Identified control chart anomaly pattern categories As input to the second BP neural network, forward propagation calculation is performed, and the output of the second BP network is the diagnosed fault cause, its corresponding fault cause code, and corrective measures.

[0077] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network, characterized in that, include: All non-conforming product records are retrieved from the aviation product non-conforming product database in the IQS system, classified according to different aviation product processing specialties and processes, and uniquely identified by the non-conforming product drawing number. The product manufacturing record barcode in the MES system is matched to obtain the production information of the processing / manufacturing batch in which each non-conforming quality characteristic belongs. The classification information is linked and integrated with the obtained production information to form a production information dataset. Based on the constructed production information dataset, the process quality characteristics corresponding to the sample measured values ​​of each processing / manufacturing batch are taken as the simulation object; the statistical characteristics of the sample measured values ​​are used as the benchmark, and the Monte Carlo data simulation method is used to expand the data to generate simulation sample data; after preprocessing the simulation sample data, it is used as the input of the first BP neural network for training to obtain model M1 for recognizing control chart pattern categories; For each production information record in the production information dataset, the set of sample measured values ​​corresponding to each record is preprocessed by standardization and encoding. Then, the processed data is input into the trained model M1 for control chart pattern recognition, and the actual control chart pattern recognition result corresponding to each non-conforming product record is obtained, forming the actual control chart pattern recognition result set. The actual control chart pattern recognition result set is combined and stored with the corresponding non-conforming quality information retrieved from the aviation product non-conforming product database of the IQS system to form a basic database for process quality diagnosis. For the basic database of process quality diagnosis, it is grouped according to different manufacturing processes of different processing specialties of aerospace products; in each group, a second BP neural network is constructed, and the abnormal control chart patterns and failure causes are used as input and output samples of the neural network for training, respectively, to learn the nonlinear mapping relationship from abnormal control chart patterns to failure causes. After training, the weights and thresholds of each neural network are written into the expert database as expert knowledge for process anomaly diagnosis, thereby realizing the intelligent acquisition of expert knowledge for process anomaly cause diagnosis.

2. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, The process of constructing the production information dataset includes: Establish a data interface between the IQS system and the MES system; for each non-conforming product record after classification, use the "part drawing number" as a unique identifier in the non-conforming product database of the IQS system and match it with the product manufacturing record barcode in the MES system; each product manufacturing record barcode uniquely corresponds to the same part drawing number information; through matching, obtain the production information of the processing / manufacturing batch in which the non-conforming quality characteristics are located, including the total sample size and the set of sample measured values; Each processing specialty category and process information obtained from the classification is associated and integrated with the acquired manufacturing record barcodes, total sample size, and sample measured value set to form a production information record, thus constituting a production information dataset.

3. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, The Monte Carlo method is used to simulate random sampling and generate simulation sample data. : ； in, Sampling time; for Simulated measurement values ​​of process quality characteristics at any given time; The mean value of the quality characteristic measurements is taken here. ; for The deviation of the normal distribution at time points is taken here. ; Represents a normal distribution; This is an interference value; , The mean and standard deviation of the sample measured values ​​for each processing / manufacturing batch; For simulation sample data Standardization and encoding preprocessing are performed; the preprocessed simulation sample data is used as input, and the corresponding known control chart pattern category is used as output, which is then input into the first BP neural network for model training; The control chart modes include: normal mode, uptrend mode, downtrend mode, upward step mode, downward step mode, and cycle mode.

4. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, The encoding formula for the encoding process is: ； in, It is about data The data obtained after encoding; This is the coding rate parameter; The values ​​are obtained after standardization of the sample measured values. Preprocessed data As input, the trained model M1 is used for forward computation to obtain the actual control chart pattern recognition result corresponding to each nonconforming product record. This constitutes the actual control chart pattern recognition result set. .

5. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, Based on the product information in the MES system, a matching search is performed in the non-conforming product database of the IQS system to obtain the quality information of non-conforming quality characteristics of aerospace products / parts; the actual control chart pattern recognition result set and the quality information of non-conforming quality characteristics are merged and stored to obtain the basic database for process quality diagnosis.

6. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, The basic database for process quality diagnosis is grouped according to different manufacturing processes of different processing specialties of aerospace products. In each group, a second BP neural network is constructed. The network adopts a supervised learning method and includes a three-layer structure of input layer, hidden layer and output layer. The transfer function between the input layer and the hidden layer adopts the smooth differentiable function Sigmoid. The transfer function between the hidden layer and the output layer uses a hard-limit function; During the forward propagation of the second BP neural network, the output value is calculated layer by layer from the input layer to the next: Let the input signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a control chart pattern recognition result. The number of input layer nodes is equal to the number of abnormal pattern categories in the control chart; the input signal... The control chart pattern category is derived from the control chart pattern category field of each record in the process quality diagnosis basic database; the control chart pattern category is encoded as follows: A binary vector of bits is used as input; Let the desired output signal be ,in ; The value can be 0 or 1, and the combination of these values ​​represents a fault cause; the desired output signal The fault cause field is derived from the records in the process quality diagnosis basic database and uses one-hot encoding. This represents the number of output layer nodes, and its value is equal to the number of fault cause categories in this group. Let the actual output signal be ,in ; Calculate the error between the actual output and the expected output, and obtain the total output error of a batch of samples; propagate the error signal layer by layer from the output layer forward, and adjust the weights until the error is less than the set value; During the iteration process, the differential chain rule and the negative gradient descent method are used; and the step size of the gradient descent method is dynamically controlled to control the learning convergence speed.

7. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, make , Let be the step size of the gradient descent method; then the dynamic control of the gradient descent step size is as follows: like ,but Remain unchanged; like Then adjust ; like Then make the following adjustments: If ,but Keep it unchanged, otherwise ; in, For the first The total output error of the next iteration; For the first The total output error of the next iteration; This is the step size of the gradient descent method used in the current iteration; This is the step size of the gradient descent method used in the previous iteration.

8. The method for acquiring expert knowledge for process anomaly diagnosis based on a BP neural network according to claim 1, characterized in that, When it is necessary to diagnose anomalies in a specific process, perform the following operations based on the real-time quality characteristic measurements of that process: Obtain real-time quality characteristic measurements for this process to form a set of measured values. ;right After standardization and encoding preprocessing, the data is input into the trained model M1 to identify the current control chart anomaly pattern for that process, thus obtaining the control chart anomaly pattern category. ; Based on the processing specialty and process to which the process belongs, the corresponding group of the pre-trained second BP neural network is retrieved from the constructed expert knowledge base; Identified control chart anomaly pattern categories As input to the second BP neural network, forward propagation calculation is performed, and the output of the second BP network is the diagnosed fault cause, its corresponding fault cause code, and corrective measures.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes a computer program, it implements the process anomaly diagnosis expert knowledge acquisition method based on BP neural network as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the process anomaly diagnosis expert knowledge acquisition method based on BP neural network as described in any one of claims 1-8.

Images