Intelligent factory production quality detection method and system fused with artificial intelligence

Abstract

The invention discloses a smart factory production quality detection method and system fused with artificial intelligence, and relates to the technical field of smart factory production quality detection.The smart factory production quality detection method fused with artificial intelligence comprises the steps that a production quality detection request containing a client ID and a current product category of a smart factory is received, matching and determining configuration parameters of the acquisition equipment according to a preset category and parameter mapping table; receiving original detection data acquired by the smart factory according to the acquisition equipment configuration parameters, and performing data processing on the original detection data to obtain a feature vector set; performing weighted fusion on the feature vector set through the dynamic weight to form a fused feature vector; and inputting the fusion feature vector into a pre-constructed artificial intelligence-based defect identification model, and outputting a defect identification result by the artificial intelligence-based defect identification model. According to the invention, the accuracy of defect identification and the process intelligence level in production quality detection of the smart factory are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent factory production quality detection, and in particular to an intelligent factory production quality detection method and system fusing artificial intelligence. BACKGROUND

[0002] In the field of intelligent factory production quality detection, the existing technology usually carries out basic detection and preliminary defect identification for single type of data (such as only image data or only size data), but in the face of complex multi-type data detection scenarios of intelligent factory, there is a lack of systematic and intelligent processing scheme, and there are obvious shortcomings in data fusion, equipment adaptation and defect identification.

[0003] In addition, part of the existing technology has tried to carry out quality detection work around electrical signals, sizes, images and other multi-type data, but in actual application, there are still many deficiencies, which are difficult to meet the efficient and accurate production quality detection needs of intelligent factory, and the specific defects are as follows:

[0004] (1) Rough data fusion: when processing electrical signals, sizes, images and other multi-type raw detection data, simple splicing or fixed weight fusion is usually used, without considering the dynamic value difference of different data in reflecting product defects, which cannot fully tap the feature value of multi-type data, resulting in one-sided and insufficient precision of features used for defect identification.

[0005] (2) Poor equipment adaptability: in view of the product category difference corresponding to different client IDs, the existing technology generally adopts unified configuration, lacks flexible and accurate acquisition device configuration parameter adaptation mechanism, and is difficult to match the diversified product detection scenarios of intelligent factory, so that the quality of raw detection data acquisition is affected, and high-quality data support cannot be provided for subsequent defect identification.

[0006] (3) Inaccurate defect identification: due to unscientific data fusion, poor equipment parameter adaptability, and the fact that part of the scheme does not deeply apply the accurate defect identification model based on artificial intelligence, the product defect feature extraction is not comprehensive and accurate when identifying defects, which is prone to miss detection and false detection, and cannot meet the needs of intelligent factory for detection accuracy and intelligence. SUMMARY

[0007] The purpose of the present application is to provide an intelligent factory production quality detection method and system fusing artificial intelligence, which improves the accuracy and intelligent level of defect identification in intelligent factory production quality detection.

[0008] To achieve the above objectives, this application provides a smart factory production quality inspection method integrating artificial intelligence, comprising the following steps: S1: Receiving a production quality inspection request from the smart factory containing a client ID and the current product category, matching and determining the configuration parameters of the acquisition device according to a preset mapping table of categories and parameters, and sending the configuration parameters of the acquisition device to the smart factory according to the client ID; S2: Receiving raw inspection data collected by the smart factory according to the configuration parameters of the acquisition device, processing the raw inspection data to obtain a feature vector set; wherein, the raw inspection data includes at least: raw electrical signals, raw dimensional data, and raw image data; the feature vector set includes: instantaneous feature vectors, dimensional feature vectors, and image feature vectors; S3: Weighting and fusing the feature vector set through dynamic weights to form a fused feature vector; S4: Inputting the fused feature vector into a pre-constructed artificial intelligence-based defect recognition model, the artificial intelligence-based defect recognition model outputs the defect recognition result and sends it to the smart factory.

[0009] As described above, the sub-step of receiving a production quality inspection request from a smart factory containing a client ID and the current product category, and matching and determining the acquisition device configuration parameters according to a preset category and parameter mapping table is as follows: S11: Using the current product category in the production quality inspection request as the query keyword, the preset category and parameter mapping table is traversed; wherein, the preset category and parameter mapping table includes: multiple standard product categories, each standard product category corresponding to a set of standard acquisition configuration parameters, and the standard acquisition configuration parameters include at least: electrical acquisition configuration parameters, size acquisition configuration parameters, and image acquisition configuration parameters; S12: The standard acquisition configuration parameters corresponding to the standard product category that directly matches the query keyword are determined as the acquisition device configuration parameters.

[0010] As described above, the sub-steps for receiving raw detection data collected by the smart factory according to the configuration parameters of the acquisition equipment, processing the raw detection data, and obtaining a set of feature vectors are as follows: S21: Extract features from the raw electrical signals to obtain instantaneous feature vectors; wherein, the raw electrical signals include: time-domain current signals, time-domain voltage signals, and time-domain resistance signals; instantaneous feature vector = [transient fluctuation characteristics of current signals, transient drop / spiking characteristics of voltage signals, transient jump characteristics of resistance signals]; S22: Extract features from the raw dimensional data to obtain dimensional feature vectors; wherein, the raw dimensional data includes: raw length, raw spacing, and raw thickness; dimensional feature vector = [length deviation rate, spacing deviation rate, thickness deviation rate]; S23: Extract features from the raw image data to obtain image feature vectors; S24: Form a set of feature vectors based on the instantaneous feature vectors, dimensional feature vectors, and image feature vectors.

[0011] As described above, the sub-steps for extracting features from the original electrical signal to obtain the instantaneous feature vector are as follows: S211: Input the time-domain current signal into a pre-built current feature extraction model, and the current feature extraction model outputs the transient fluctuation features of the current signal; S212: Input the time-domain voltage signal into a pre-built voltage feature extraction model, and the voltage feature extraction model outputs the transient drop / spiking features of the voltage signal; S213: Input the time-domain resistance signal into a pre-built resistance feature extraction model, and the resistance feature extraction model outputs the transient jump features of the resistance signal; S214: Form the instantaneous feature vector based on the transient fluctuation features of the current signal, the transient drop / spiking features of the voltage signal, and the transient jump features of the resistance signal.

[0012] As shown above, based on machine learning or deep learning techniques, the transient fluctuation characteristics of the current signal, the transient drop / peak characteristics of the voltage signal, and the transient jump characteristics of the resistance signal can be arranged in a preset order to form an instantaneous feature vector.

[0013] As shown above, the sub-steps for extracting features from the original size data to obtain the size feature vector are as follows: S221: Input the original length into the pre-built length feature extraction model, and the length feature extraction model outputs the length deviation rate; S222: Input the original spacing into the pre-built spacing feature extraction model, and the spacing feature extraction model outputs the spacing deviation rate; S223: Input the original thickness into the pre-built thickness feature extraction model, and the thickness feature extraction model outputs the thickness deviation rate; S224: Form the size feature vector based on the length deviation rate, spacing deviation rate, and thickness deviation rate.

[0014] As shown above, feature vectors are obtained by extracting features from the original image data through a pre-built convolutional neural network based on an attention mechanism.

[0015] As described above, the sub-step of weighted fusion of the feature vector set using dynamic weights to form a fused feature vector is as follows: S31: Input the current product category corresponding to the feature vector set into a pre-built AI-based dynamic weight model, which outputs the corresponding dynamic weights. The dynamic weights include at least: instantaneous feature weights, size feature weights, and image feature weights; S32: Construct a fused feature vector based on the dynamic weights and the feature vector set; where the expression for the fused feature vector is: Frh = η ss ·F elec +η tx ·F img +η cc ·F size Where Frh is the fused feature vector; ηss For instantaneous feature weights, η cc For the size feature weights, η tx For image feature weights, η ss +η tx +η cc =1; F elec F is the instantaneous feature vector; size F is the size feature vector; img This is the image feature vector.

[0016] As described above, the sub-steps for inputting the fused feature vector into a pre-built AI-based defect identification model and outputting the defect identification result are as follows: S41: Input the fused feature vector into the pre-built AI-based defect identification model, and output the current defect probability of multiple defect types through the AI-based defect identification model; S42: Analyze the current defect probability of the corresponding defect type using the defect probability threshold for each defect type to generate sub-defect results; if the current defect probability is greater than or equal to the defect probability threshold, the generated sub-defect result is defective; if the current defect probability is less than the defect probability threshold, the generated sub-defect result is defect-free; S43: Summarize all sub-defect results to generate the defect identification result; if the sub-defect result for each defect type is defect-free, the generated defect identification result only includes: defect-free; if there are one or more defect categories whose sub-defect results are defective, the generated defect identification result includes not only defective but also the corresponding defect type.

[0017] This application also provides a smart factory production quality inspection system integrating artificial intelligence, comprising: at least one smart factory for producing electronic components and an artificial intelligence inspection center; wherein each smart factory is equipped with a smart factory client and a data acquisition subsystem; the smart factory client is used to send production quality inspection requests; receive data acquisition device configuration parameters; send raw inspection data; and receive defect identification results; the data acquisition subsystem is used to complete the configuration according to the configuration instructions sent by the smart factory client based on the data acquisition device configuration parameters, collect raw inspection data, and send the raw inspection data to the smart factory client; the artificial intelligence inspection center is used to execute the above-mentioned smart factory production quality inspection method integrating artificial intelligence.

[0018] The beneficial effects achieved by this application are as follows:

[0019] (1) The intelligent factory production quality inspection method and system integrating artificial intelligence of this application can effectively process and integrate multiple types of raw inspection data, construct a more comprehensive and accurate integrated feature vector, and improve the comprehensiveness and accuracy of defect identification.

[0020] (2) The intelligent factory production quality inspection method and system integrating artificial intelligence of this application adapts the configuration parameters of the acquisition equipment according to the client ID and the current product category, which enhances the adaptability of the inspection process to different scenarios and products, improves the intelligence and personalization level of intelligent factory production quality inspection, ensures the efficient transmission and application of defect identification results, and helps intelligent factories to control product quality in a timely manner. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 A schematic diagram of one embodiment of a smart factory production quality inspection system that integrates artificial intelligence;

[0023] Figure 2 A flowchart of one embodiment of a smart factory production quality inspection method that integrates artificial intelligence. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this application provides a smart factory production quality inspection system that integrates artificial intelligence, including: at least one smart factory 1 for producing electronic components and an artificial intelligence inspection center 2.

[0026] Each smart factory 1 is equipped with a smart factory client and a data acquisition subsystem.

[0027] Smart Factory Client: Used to send production quality inspection requests; receive configuration parameters of the data acquisition equipment; send raw inspection data; and receive defect identification results.

[0028] The data acquisition subsystem is used to complete the configuration according to the configuration instructions sent by the smart factory client based on the configuration parameters of the acquisition device, acquire the raw test data, and send the raw test data to the smart factory client.

[0029] Artificial Intelligence Testing Center 2: Used to perform the following smart factory production quality testing methods that integrate artificial intelligence.

[0030] Furthermore, the artificial intelligence detection center 2 includes at least: a transceiver unit, a configuration unit, a data processing unit, and a defect identification unit.

[0031] Transceiver Unit: Receives production quality inspection requests from the smart factory containing the client ID and the current product category, and sends the production quality inspection requests to the configuration unit; sends the data acquisition device configuration parameters to the smart factory according to the client ID; receives raw inspection data collected by the smart factory based on the data acquisition device configuration parameters, and sends the raw inspection data to the data processing unit; sends the defect identification results to the smart factory.

[0032] Configuration Unit: Receives production quality inspection requests, matches and determines the configuration parameters of the acquisition device according to the preset category and parameter mapping table, and sends the acquisition device configuration parameters to the transceiver unit.

[0033] Data processing unit: processes the raw detection data to obtain a set of feature vectors; and performs weighted fusion of the feature vector set through dynamic weights to form a fused feature vector.

[0034] Defect identification unit: It is equipped with a pre-built AI-based defect identification model; the fused feature vector is input into the pre-built AI-based defect identification model, the AI-based defect identification model outputs the defect identification result and sends it to the transceiver unit.

[0035] like Figure 2 As shown, this application provides a smart factory production quality inspection method integrating artificial intelligence, including the following steps:

[0036] S1: Receives a production quality inspection request from the smart factory containing the client ID and the current product category, matches and determines the configuration parameters of the acquisition device according to the preset category and parameter mapping table, and sends the configuration parameters of the acquisition device to the smart factory according to the client ID.

[0037] Furthermore, the sub-steps for receiving production quality inspection requests from the smart factory containing the client ID and the current product category, and matching and determining the configuration parameters of the acquisition device based on a preset mapping table of categories and parameters are as follows:

[0038] S11: Using the current product category in the production quality inspection request as the query keyword, traverse the preset category and parameter mapping table; wherein, the preset category and parameter mapping table includes: multiple standard product categories, each standard product category corresponds to a set of standard acquisition configuration parameters, and the standard acquisition configuration parameters include at least: electrical acquisition configuration parameters, size acquisition configuration parameters and image acquisition configuration parameters.

[0039] Specifically, the products corresponding to the standard product categories in the preset category and parameter mapping table are all electronic components. The specific number and content of the standard product categories in the preset category and parameter mapping table are set according to the actual situation. The specific content of the electrical acquisition configuration parameters, dimension acquisition configuration parameters, and image acquisition configuration parameters are set according to the actual situation.

[0040] For example, the standard product category is "chip (electrically sensitive type)". The electrical acquisition configuration parameters for the chip (electrically sensitive type) include: the accuracy and acquisition frequency of the current sensor. Specifically, the accuracy of the current sensor is: an error range not exceeding 0.001% of the measured value, and the acquisition frequency is 5kHz. The size acquisition configuration parameters for the chip (electrically sensitive type) include: the accuracy of the displacement sensor and the acquisition object. Specifically, the accuracy of the displacement sensor is: the error between the actual and measured dimensions is within the range of [-1μm, +1μm], and the acquisition object is: chip thickness and pin spacing. The image acquisition configuration parameters for the chip (electrically sensitive type) include: pixels and magnification. Specifically, the pixels are 20 million, and the magnification is 500x.

[0041] The standard product category is: Precision Capacitors (Structure Sensitive Type). The electrical acquisition configuration parameters for Precision Capacitors (Structure Sensitive Type) include: voltage sensor accuracy and acquisition frequency. The voltage sensor accuracy is: error range not exceeding 0.01% of the measured value; the acquisition frequency is: 1kHz. The dimensional acquisition configuration parameters for Precision Capacitors (Structure Sensitive Type) include: laser sensor accuracy and acquisition object. The laser sensor accuracy is: the error between the actual and measured dimensions is within the range of [-0.5μm, +0.5μm]; the acquisition object is the capacitor diameter and pin perpendicularity. The image acquisition configuration parameters for Precision Capacitors (Structure Sensitive Type) include: pixels and magnification. The pixels are: 10 million; the magnification is: 1000x.

[0042] S12: Determine the standard collection configuration parameters corresponding to the standard product categories that directly match the query keywords as the collection device configuration parameters.

[0043] Specifically, the standard product category that directly matches the query keywords refers to a standard product category that is exactly the same as the current product category.

[0044] Traditional data acquisition devices have fixed configuration parameters, which can easily lead to data redundancy or missing data when dealing with different product categories. However, the category-parameter mapping table in this application allows for dynamic adaptation of the data acquisition device configuration parameters based on the current product category, thereby effectively improving acquisition accuracy and data utilization.

[0045] S2: Receives raw detection data collected by the smart factory according to the configuration parameters of the acquisition equipment, processes the raw detection data, and obtains a set of feature vectors; wherein, the raw detection data includes at least: raw electrical signals, raw dimensional data, and raw image data; the set of feature vectors includes: instantaneous feature vectors, dimensional feature vectors, and image feature vectors.

[0046] Furthermore, the sub-steps for receiving raw detection data collected by the smart factory according to the configuration parameters of the acquisition equipment and processing the raw detection data to obtain a set of feature vectors are as follows:

[0047] S21: Extract features from the original electrical signal to obtain an instantaneous feature vector; wherein, the original electrical signal includes: time-domain current signal, time-domain voltage signal and time-domain resistance signal; instantaneous feature vector = [transient fluctuation features of current signal, transient drop / spiking features of voltage signal, transient jump features of resistance signal].

[0048] Furthermore, the sub-steps for extracting features from the original electrical signal to obtain the instantaneous feature vector are as follows:

[0049] S211: Input the time-domain current signal into a pre-built current feature extraction model, and output the transient fluctuation characteristics of the current signal from the current feature extraction model.

[0050] Furthermore, the formula for the pre-constructed current feature extraction model is as follows:

[0051]

[0052] Among them, WT I (a I ,b I ) represents the transient fluctuation characteristics of the current signal; I(t) represents the time-domain current signal at time t; Let be the wavelet basis function of the current signal at time t; dt is the time derivative; a I b is the scale parameter of the current signal; I This is the translation parameter for the current signal.

[0053] Specifically, the wavelet basis function for the current signal can be implemented using existing mature wavelet basis functions, such as the Haar wavelet or the Daubechies wavelet. I Used to adjust the width of the wavelet basis functions, a I The specific value is set according to the frequency characteristics of the current anomaly. For example, during a short circuit, the sudden current change is a high-frequency signal, a I Take [0.1, 1]; normal fluctuations in current are low-frequency signals, a I Take [5, 10]. b Ib is used to control the position of the wavelet basis functions on the time axis. I The specific value is set according to the time resolution of the current signal.

[0054] S212: Input the time-domain voltage signal into a pre-built voltage feature extraction model, and output the transient drop / spiking features of the voltage signal from the voltage feature extraction model.

[0055] Furthermore, the formula for the pre-built voltage feature extraction model is as follows:

[0056]

[0057] Among them, WT U (a U ,b U ) represents the transient drop / spiking characteristics of the voltage signal; U(t) is the time-domain voltage signal at time t; Let be the wavelet basis function of the voltage signal at time t; dt is the time derivative; a U b is the scale parameter of the voltage signal; U This refers to the translation parameter of the voltage signal.

[0058] Specifically, the wavelet basis function for the voltage signal can be implemented using existing mature wavelet basis functions, such as the Daubechies wavelet. U The specific value is adapted to the frequency setting of the voltage anomaly signal. For example, in the case of a cold solder joint, the voltage drop is an intermediate frequency signal, a U Take [1, 3]; normal voltage fluctuation is a low-frequency signal, a U Take [6, 8]. b U The specific value is set according to the time resolution of the voltage signal.

[0059] S213: Input the time-domain resistance signal into the pre-constructed resistance feature extraction model, and output the transient jump characteristics of the resistance signal from the resistance feature extraction model.

[0060] Furthermore, the formula for the pre-built resistance feature extraction model is as follows:

[0061]

[0062] Among them, WT R (a R ,b R R(t) represents the transient jump characteristic of the resistance signal; R(t) is the time-domain resistance signal at time t. Let be the wavelet basis function of the resistance signal at time t; dt is the time derivative; a R b is the scale parameter of the resistance signal; R This is the translation parameter for the resistance signal.

[0063] Specifically, the wavelet basis function for the resistive signal can be implemented using existing mature wavelet basis functions, such as the smoothed Morlet wavelet. R The specific value is adapted to the frequency setting of the resistor abnormality signal. For example, the slow switching of the resistor due to component oxidation is a low-frequency signal. R Take [8, 10]; resistance fluctuations caused by poor contact are intermediate frequency signals, a R Take [3, 5]. b R The specific value is set according to the time resolution of the resistance signal.

[0064] S214: A transient feature vector is formed based on the transient fluctuation characteristics of the current signal, the transient drop / spiking characteristics of the voltage signal, and the transient jump characteristics of the resistance signal.

[0065] Furthermore, based on machine learning or deep learning techniques, WTs are arranged in a preset order. I (a I ,b I ), WT U (a U ,b U ) and WT R (a R ,b R ), which can be combined into an instantaneous feature vector (F elec ).

[0066] S22: Extract features from the original size data to obtain a size feature vector; where the original size data includes: original length, original spacing and original thickness; size feature vector = [length deviation rate, spacing deviation rate, thickness deviation rate].

[0067] Furthermore, the sub-steps for extracting features from the original size data to obtain size feature vectors are as follows:

[0068] S221: Input the original length into the pre-built length feature extraction model, and the length feature extraction model outputs the length deviation rate.

[0069] Furthermore, the formula for the pre-built length feature extraction model is as follows:

[0070]

[0071] Where, φ cd Zcd represents the length deviation rate. sc The original length; Zcd bz This is the standard value for length.

[0072] Specifically, Zcd bzThe specific value is set according to the manufacturing standards of electronic components. The original length is a critical physical dimension of the electronic component, such as the pin length of a chip and the body length of a resistor.

[0073] S222: Input the original spacing into the pre-built spacing feature extraction model, and the spacing feature extraction model outputs the spacing deviation rate.

[0074] Furthermore, the formula for the pre-built spacing feature extraction model is:

[0075]

[0076] Where, φ jj For spacing deviation rate; Zjj sc This represents the original spacing; Zjj bz This is the standard value for the spacing.

[0077] Specifically, Zjj bz The specific value is set according to the manufacturing standards of electronic components. The original spacing is a critical physical dimension of the electronic component, such as the distance between adjacent pins of a chip and the spacing between capacitor pins.

[0078] S223: Input the original thickness into the pre-built thickness feature extraction model, and the thickness feature extraction model outputs the thickness deviation rate.

[0079] Furthermore, the formula for the pre-built thickness feature extraction model is as follows:

[0080]

[0081] Where, φ hd For thickness deviation rate; Zhd sc Original thickness; Zhd bz This is the standard value for thickness.

[0082] Specifically, Zhd bz The specific value is set according to the manufacturing standards of electronic components. The original thickness is a critical physical dimension of the electronic component, such as the thickness of the component body and the thickness of the leads.

[0083] S224: Form a dimensional feature vector based on the length deviation rate, spacing deviation rate, and thickness deviation rate.

[0084] Specifically, based on existing machine learning or deep learning techniques, the length deviation rate, spacing deviation rate, and thickness deviation rate can be arranged in a preset order to form a size feature vector (F). size ).

[0085] S23: Extract features from the original image data to obtain image feature vectors.

[0086] Furthermore, feature vectors are obtained by extracting features from the original image data using a pre-built attention-based CNN (convolutional neural network).

[0087] Furthermore, the pre-built expression for an attention-based CNN (Convolutional Neural Network) is as follows:

[0088]

[0089] Among them, F img A is the image feature vector; s It is the spatial attention weight matrix of the electronic components corresponding to the original image data; A c It is the channel attention weight matrix of the electronic component corresponding to the original image data; For element-wise multiplication; Wqz is the convolution kernel weight; I mg The original image data; b pz For bias.

[0090] Specifically, Used to multiply corresponding elements of different matrices or vectors. A s By weighting the raw image data at different spatial locations, the characteristics of key locations of electronic components (e.g., pin edges, package edges, and chip surfaces) are highlighted. pz Used to adjust the output of convolution operations. A c This is used to highlight channel features (e.g., grayscale variations) in the solder joint area. The attention-based CNN (convolutional neural network) pre-built in this application improves the feature response to minute defects (e.g., 10µm scratches) and the feature extraction accuracy through dual attention.

[0091] S24: Form a feature vector set based on the instantaneous feature vector, size feature vector, and image feature vector.

[0092] S3: The feature vector set is weighted and fused using dynamic weights to form a fused feature vector.

[0093] Furthermore, the sub-steps for weighted fusion of the feature vector set using dynamic weights to form a fused feature vector are as follows:

[0094] S31: Input the current product category corresponding to the feature vector set into the pre-built AI-based dynamic weight model, and output the corresponding dynamic weights from the AI-based dynamic weight model. The dynamic weights include at least: instantaneous feature weights, size feature weights, and image feature weights.

[0095] Specifically, the core training data is based on actual production and testing data of electronic components (e.g., characteristic data of historical qualified / defective samples). Combined with industry research experience, and according to the set testing accuracy target and specific testing requirements (e.g., the electrical performance of a certain type of component needs to be focused on), the AI-based dynamic weight model is optimized and trained. The AI-based dynamic weight model is logistic regression or a lightweight neural network, but is not limited to logistic regression or lightweight neural networks.

[0096] Furthermore, new sample detection data is acquired according to the preset parameter update time, and the pre-built AI-based dynamic weight model is optimized based on the new sample detection data using the gradient descent method.

[0097] Specifically, the preset parameter update time is set according to the actual situation; in this application, it is preferably 24 hours. Gradient descent is an existing machine learning and optimization algorithm, so it will not be described in detail here.

[0098] S32: Construct a fused feature vector based on dynamic weights and feature vector sets.

[0099] Furthermore, the expression for the fused feature vectors is:

[0100] Frh=η ss ·F elec +η tx ·F img +η cc ·F size ;

[0101] Where Frh is the fused feature vector; η ss For instantaneous feature weights, η cc For the size feature weights, η tx For image feature weights, η ss +η tx +η cc =1; F elec F is the instantaneous feature vector; size F is the size feature vector; img This is the image feature vector.

[0102] S4: Input the fused feature vector into the pre-built AI-based defect identification model, which then outputs the defect identification result and sends it to the smart factory.

[0103] Furthermore, the sub-steps for inputting the fused feature vector into a pre-built AI-based defect recognition model, and for the AI-based defect recognition model to output the defect recognition result, are as follows:

[0104] S41: Input the fused feature vector into a pre-built AI-based defect identification model, and output the current defect probability of multiple defect types through the AI-based defect identification model.

[0105] Furthermore, the expression for the pre-built AI-based defect identification model is:

[0106]

[0107] Among them, Pqx k Let f be the current defect probability of the k-th defect type; σ(·) is the sigmoid function; f k (Frh) represents the degree of matching between the features in the fused feature vector and the defect features corresponding to the k-th defect type; e is the base.

[0108] S42: Analyze the current defect probability of the corresponding defect type using the defect probability threshold for each defect type, and generate a sub-defect result; if the current defect probability is greater than or equal to the defect probability threshold, the generated sub-defect result is defective, and if the current defect probability is less than the defect probability threshold, the generated sub-defect result is defect-free.

[0109] Specifically, the details of each defect type are set according to the specific detection scenario (e.g., chip manufacturing quality inspection). The defect probability threshold for each defect type is set based on test set verification (e.g., statistical analysis of false negative and false positive rates under different defect probability thresholds) and the importance of the object being detected (e.g., raising the threshold for critical defects in safety components).

[0110] S43: Summarize all sub-defect results and generate defect identification results; if the sub-defect results for each defect type are all defect-free, the generated defect identification results will only include: defect-free; if there are one or more defect categories whose sub-defect results are defective, the generated defect identification results will include not only defective but also the corresponding defect type.

[0111] The beneficial effects achieved by this application are as follows:

[0112] (1) The intelligent factory production quality inspection method and system integrating artificial intelligence of this application can effectively process and integrate multiple types of raw inspection data, construct a more comprehensive and accurate integrated feature vector, and improve the comprehensiveness and accuracy of defect identification.

[0113] (2) The intelligent factory production quality inspection method and system integrating artificial intelligence of this application adapts the configuration parameters of the acquisition equipment according to the client ID and the current product category, which enhances the adaptability of the inspection process to different scenarios and products, improves the intelligence and personalization level of intelligent factory production quality inspection, ensures the efficient transmission and application of defect identification results, and helps intelligent factories to control product quality in a timely manner.

[0114] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the scope of protection of this application is intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application. Obviously, those skilled in the art can make various alterations and variations to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of protection of this application and its equivalents, this application also intends to include these modifications and variations.

Claims

1. A smart factory production quality inspection method integrating artificial intelligence, characterized in that, Includes the following steps: S1: Receives a production quality inspection request from the smart factory containing the client ID and the current product category, matches and determines the configuration parameters of the acquisition device according to the preset category and parameter mapping table, and sends the configuration parameters of the acquisition device to the smart factory according to the client ID; S2: Receives raw detection data collected by the smart factory according to the configuration parameters of the acquisition equipment, processes the raw detection data, and obtains a set of feature vectors; wherein, the raw detection data includes at least: raw electrical signals, raw dimensional data, and raw image data; the set of feature vectors includes: instantaneous feature vectors, dimensional feature vectors, and image feature vectors; S3: The feature vector set is weighted and fused using dynamic weights to form a fused feature vector; S4: Input the fused feature vector into the pre-built AI-based defect identification model, which then outputs the defect identification result and sends it to the smart factory.

2. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 1, characterized in that, The sub-steps for receiving a production quality inspection request from a smart factory containing the client ID and the current product category, and matching and determining the configuration parameters of the data acquisition device based on a preset mapping table of categories and parameters are as follows: S11: Using the current product category in the production quality inspection request as the query keyword, traverse the preset category and parameter mapping table; wherein, the preset category and parameter mapping table includes: multiple standard product categories, each standard product category corresponds to a set of standard acquisition configuration parameters, and the standard acquisition configuration parameters include at least: electrical acquisition configuration parameters, size acquisition configuration parameters and image acquisition configuration parameters. S12: Determine the standard collection configuration parameters corresponding to the standard product categories that directly match the query keywords as the collection device configuration parameters.

3. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 1, characterized in that, The sub-steps for receiving raw detection data collected by the smart factory according to the configuration parameters of the acquisition equipment and processing the raw detection data to obtain the feature vector set are as follows: S21: Extract features from the original electrical signal to obtain the instantaneous feature vector; wherein, the original electrical signal includes: time-domain current signal, time-domain voltage signal and time-domain resistance signal; instantaneous feature vector = [transient fluctuation characteristics of current signal, transient drop / spiking characteristics of voltage signal, transient jump characteristics of resistance signal]; S22: Extract features from the original size data to obtain a size feature vector; where the original size data includes: original length, original spacing, and original thickness; size feature vector = [length deviation rate, spacing deviation rate, thickness deviation rate]; S23: Extract features from the original image data to obtain image feature vectors; S24: Form a feature vector set based on the instantaneous feature vector, size feature vector, and image feature vector.

4. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 3, characterized in that, The sub-steps for extracting features from the original electrical signal to obtain the instantaneous feature vector are as follows: S211: Input the time-domain current signal into a pre-built current feature extraction model, and output the transient fluctuation characteristics of the current signal from the current feature extraction model; S212: Input the time-domain voltage signal into a pre-built voltage feature extraction model, and output the transient drop / spiking features of the voltage signal from the voltage feature extraction model; S213: Input the time-domain resistance signal into a pre-built resistance feature extraction model, and output the transient jump characteristics of the resistance signal from the resistance feature extraction model; S214: A transient feature vector is formed based on the transient fluctuation characteristics of the current signal, the transient drop / spiking characteristics of the voltage signal, and the transient jump characteristics of the resistance signal.

5. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 4, characterized in that, Based on machine learning or deep learning techniques, the transient fluctuation characteristics of current signals, the transient drop / spiking characteristics of voltage signals, and the transient jump characteristics of resistance signals can be arranged in a preset order to form an instantaneous feature vector.

6. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 3, characterized in that, The sub-steps for extracting features from the original size data to obtain the size feature vector are as follows: S221: Input the original length into the pre-built length feature extraction model, and the length feature extraction model outputs the length deviation rate; S222: Input the original spacing into the pre-built spacing feature extraction model, and the spacing feature extraction model outputs the spacing deviation rate; S223: Input the original thickness into the pre-built thickness feature extraction model, and the thickness feature extraction model outputs the thickness deviation rate; S224: Form a dimensional feature vector based on the length deviation rate, spacing deviation rate, and thickness deviation rate.

7. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 3, characterized in that, Image feature vectors are obtained by extracting features from the original image data using a pre-built convolutional neural network based on an attention mechanism.

8. The smart factory production quality inspection method integrating artificial intelligence according to claim 1, characterized in that, The sub-steps for weighted fusion of feature vector sets using dynamic weights to form fused feature vectors are as follows: S31: Input the current product category corresponding to the feature vector set into the pre-built AI-based dynamic weight model, and output the corresponding dynamic weights from the AI-based dynamic weight model. The dynamic weights include at least: instantaneous feature weights, size feature weights, and image feature weights. S32: Constructing a fused feature vector based on dynamic weights and feature vector sets; The expression for the fused feature vector is: Frh=η ss ·F elec +n tx ·F img +n cc ·F size ; Where Frh is the fused feature vector; η ss For instantaneous feature weights, η cc For the size feature weights, η tx For image feature weights, η ss +η tx +η cc =1; F elec F is the instantaneous feature vector; size F is the size feature vector; img This is the image feature vector.

9. The intelligent factory production quality inspection method integrating artificial intelligence according to claim 1, characterized in that, The sub-steps for inputting the fused feature vector into a pre-built AI-based defect recognition model, and for the AI-based defect recognition model to output the defect recognition result, are as follows: S41: Input the fused feature vector into a pre-built AI-based defect identification model, and output the current defect probability of multiple defect types through the AI-based defect identification model; S42: Analyze the current defect probability of the corresponding defect type using the defect probability threshold for each defect type, and generate a sub-defect result; if the current defect probability is greater than or equal to the defect probability threshold, the generated sub-defect result is defective; if the current defect probability is less than the defect probability threshold, the generated sub-defect result is defect-free. S43: Summarize all sub-defect results and generate defect identification results; If the sub-defect result for each defect type is "no defect", the generated defect identification result will only include "no defect". If there are one or more sub-defect results for defect categories that are "defective", the generated defect identification result will include not only "defective" but also the corresponding defect type.

10. A smart factory production quality inspection system integrating artificial intelligence, characterized in that, include: At least one smart factory and AI testing center for the production of electronic components; Each smart factory is equipped with a smart factory client and a data acquisition subsystem. Smart Factory Client: Used to send production quality inspection requests; receive configuration parameters of data acquisition equipment; send raw inspection data; and receive defect identification results. The data acquisition subsystem is used to complete the configuration according to the configuration instructions sent by the smart factory client based on the configuration parameters of the data acquisition device, acquire the raw detection data, and send the raw detection data to the smart factory client. Artificial Intelligence Testing Center: Used to execute the smart factory production quality testing method integrating artificial intelligence as described in any one of claims 1-9.

Images