PCBA fault detection system and method

Abstract

The invention relates to the technical field of PCBA fault detection, and discloses a PCBA fault detection system and method, and the method comprises the following steps: collecting the multi-modal data of a PCBA circuit image: obtaining an element side image through a side image collection module, and obtaining the current and voltage data through an electrical test module; preprocessing and feature extraction are carried out on the multi-modal data, a solder paste area is intercepted from the element side image, solder paste area and contour features are calculated, an open circuit suspected element and a short circuit suspected element are identified based on a predefined area threshold, and an image feature vector is generated. According to the method, PCBA board quality evaluation is carried out through the comprehensive feature fault detection model, deep fusion and collaborative analysis of multi-source heterogeneous features can be realized, and the model is used for evaluating the quality of the PCBA board by mining hidden feature-quality association and feature-feature interaction in historical data. Therefore, the problems of single-index view angle limitation and one-sided judgment basis are overcome, and the comprehensiveness and scientificity of quality evaluation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of PCBA fault detection, and in particular to a PCBA fault detection system and method. BACKGROUND

[0002] Current PCBA fault detection relies on single modal data or manual experience for judgment, such as visual inspection of solder joints or electrical test analysis of waveforms. Such methods are difficult to comprehensively capture complex faults caused by the combined effects of welding defects, electrical parameter abnormalities, etc., resulting in limited detection accuracy, high misjudgment rate, and lack of early warning and comprehensive assessment of potential quality risks.

[0003] An existing patent discloses a PCBA circuit board fault detection method and device (publication number CN117169246A). The existing patent discloses a technology that has the problems of single detection dimension, insufficient comprehensive judgment ability, and lack of early quality warning. That is, the existing method only relies on visual solder paste detection to screen suspicious components and then performs probe electrical testing, without integrating multi-source features for collaborative analysis, making it difficult to comprehensively capture complex faults caused by the combined effects of welding defects and electrical parameter abnormalities, resulting in limited detection accuracy, high misjudgment rate, and inability to effectively predict and comprehensively assess potential quality risks. SUMMARY

[0004] The present application provides a PCBA fault detection system and method to solve the existing technical problems, solving the problems of limited detection accuracy and high misjudgment rate.

[0005] To solve the above technical problems, according to one aspect of the present application, more specifically, a PCBA fault detection method, comprising the following steps:

[0006] S1, collecting multi-modal data of PCBA circuit images, including obtaining component side image through side image acquisition module, and obtaining current and voltage data through electrical test module;

[0007] S2, pre-processing and feature extraction of the multi-modal data, wherein:

[0008] The tin paste area and contour features are calculated by intercepting the tin paste area from the component side image, the open circuit suspicious components and short circuit suspicious components are identified based on the pre-defined area threshold, and the image feature vector is generated;

[0009] The current and voltage data are divided into cycles, the waveform factor, effective value, and jitter distribution random feature coefficient are extracted, and the electrical feature vector is generated;

[0010] S3. Multimodal fusion of at least two key indicators, image feature vector and electrical feature vector, from PCBA circuit images in the historical database to generate a comprehensive feature fault detection model.

[0011] S4. Input the comprehensive feature vector into the preset comprehensive feature fault detection model and output the estimated quality status of the PCBA circuit image.

[0012] S5. Based on the estimated quality status of the PCBA circuit image, make visual markings on the PCBA circuit image and automatically generate a diagnostic report.

[0013] Furthermore, in step S2, identifying suspected open-circuit components and suspected short-circuit components includes:

[0014] If the outline feature of the solder paste at the component pin position is a first image outline, and the first image area of ​​the first image outline is smaller than a first preset area, it is determined to be an open circuit suspected component.

[0015] If the outline feature of the solder paste on the bottom surface of the component is a second image outline, and the second image area of ​​the second image outline is larger than the second preset area, and the solder paste at the component pin position is connected to the solder paste outline at the bottom surface of the component, it is determined to be a short-circuit suspected component.

[0016] Furthermore, the suspected open circuit element obtains an open circuit suspicion score based on the difference between the area of ​​the first image and the first preset area;

[0017] The short-circuit suspected element obtains a short-circuit suspected score based on the difference between the area of ​​the second image and the second preset area.

[0018] Furthermore, in step S2, the main indicators of the electrical characteristic vector are obtained as follows:

[0019] The current data is divided into multiple cycles using a clustering algorithm;

[0020] Calculate the smoothness index of the current waveform within each cycle, the steady-state coefficient of the cycle jitter, and the random characteristic coefficient of the jitter distribution.

[0021] Furthermore, in step S3, the multimodal fusion adopts a feature-level fusion method, including:

[0022] Normalize the image feature vector, electrical feature vector, and thermal feature vector;

[0023] A comprehensive feature vector is generated by weighted summation based on an attention mechanism.

[0024] Furthermore, in step S3, generating the comprehensive feature-based fault detection model specifically includes:

[0025] S301. Collect at least two main indicators from the image feature vector and electrical feature vector of the PCBA circuit image in the historical database, and also obtain the actual quality of the PCBA circuit image in subsequent tests.

[0026] S302. Use the controlled variable method to analyze the characteristic-mass relationship between each characteristic vector and the actual mass, as well as the characteristic-characteristic relationship between each characteristic vector;

[0027] S303. Generate a comprehensive feature-based fault detection model based on feature-quality relationship and feature-feature relationship fitting.

[0028] Furthermore, the comprehensive feature fault detection model is constructed by multimodal feature-level fusion of the main indicators from image feature vectors and electrical feature vectors, and by comprehensive fitting based on feature-quality relationship and feature-feature relationship, which can output a unified estimated quality assessment result that reflects the overall quality status of PCBA.

[0029] Furthermore, step S5 includes:

[0030] Mark the fault location with a highlighted box on the circuit diagram;

[0031] Generate maintenance suggestion text based on the fault type;

[0032] The labeled images and text are combined into a diagnostic report, which is then transmitted to the quality inspection platform via the Internet of Things.

[0033] A PCBA fault detection system includes: a data acquisition module, a feature extraction module, a multimodal fusion module, a fault detection module, a report generation module, and a database module.

[0034] The present invention provides a PCBA fault detection system and method, which, compared with the prior art, achieves the following advantages:

[0035] 1. This invention uses a comprehensive feature fault detection model to assess the quality of PCBA boards, enabling deep fusion and collaborative analysis of multi-source heterogeneous features. By mining the hidden feature-quality correlation and the interaction between features in historical data, this model constructs a unified and comprehensive quality prediction coefficient, thereby overcoming the limitations of a single indicator perspective and the one-sided judgment basis, and significantly improving the comprehensiveness and scientific nature of quality assessment.

[0036] 2. This invention employs a data-driven modeling approach. The comprehensive model adaptively learns and quantifies the weights of different fault characteristics on the final quality, rather than relying on fixed thresholds set by human experience. This dynamic weighting mechanism allows the quality assessment results to more accurately reflect the complex operating conditions of PCBA boards. Especially when faced with contradictory or critically oriented characteristics, the model can make more accurate and reliable comprehensive judgments than independent standards, effectively reducing the probability of missed and false alarms.

[0037] 3. This invention uses the estimated quality coefficient output by the model as a scalarized and intuitive comprehensive indicator, greatly simplifying the quality decision-making process. Production and quality inspection personnel no longer need to cross-compare and interpret multiple independent parameters with different dimensions. They can directly use this coefficient to quickly classify PCBA boards (e.g., good, questionable, defective) and set unified quality control thresholds accordingly. This achieves the unification of quality judgment standards and the standardization of quality inspection processes, improving the efficiency and consistency of quality control on the production line.

[0038] 4. This invention, based on a comprehensive model for quality prediction, enables early and proactive detection of potential PCBA faults, rather than merely confirming them after they have become apparent. This predictive maintenance capability allows companies to identify boards at risk of defects early in the production cycle, enabling timely intervention or repair. This prevents defective products from flowing into subsequent stages, thereby helping to reduce overall maintenance costs, decrease scrap rates, and ultimately improve product reliability and long-term service performance. Attached Figure Description

[0039] Figure 1 This is a flowchart of the present invention;

[0040] Figure 2 This is a graph showing the relationship between the open-circuit suspicion score and the quality coefficient in this invention;

[0041] Figure 3 This is a graph showing the relationship between the random characteristic coefficients and the quality coefficient of the jitter distribution in this invention;

[0042] Figure 4 This is a graph showing the relationship between the open-circuit suspicion score and the random characteristic coefficient of the jitter distribution in this invention. Detailed Implementation

[0043] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1:

[0045] like Figure 1As shown, the PCBA fault detection method includes acquiring multimodal data of PCBA circuit images, including obtaining component side images through a side image acquisition module and obtaining current and voltage data through an electrical test module; preprocessing and feature extraction of the multimodal data, wherein: the solder paste area is extracted from the component side image to calculate the solder paste area and contour features, and suspected open circuit components and suspected short circuit components are identified based on a predefined area threshold to generate image feature vectors;

[0046] Identifying suspected open-circuit and short-circuit components includes:

[0047] If the outline feature of the solder paste at the component pin position is a first image outline, and the first image area of ​​the first image outline is smaller than a first preset area, it is determined to be an open circuit suspected component.

[0048] If the outline feature of the solder paste on the bottom surface of the component is a second image outline, and the second image area of ​​the second image outline is larger than the second preset area, and the solder paste at the component pin position is connected to the solder paste outline at the bottom surface of the component, it is determined to be a short-circuit suspected component.

[0049] The method for determining the connectivity between the lead solder paste and the bottom solder paste contours includes: extracting the edges of the first and second image contours; calculating the edge similarity; and determining connectivity if the similarity is greater than a preset threshold. The image feature vectors are normalized, and a weighted summation based on an attention mechanism is performed to generate a comprehensive feature vector. Specific key indicators include:

[0050] Suspicion rating for road clearing :

[0051] ;

[0052] In the above formula, This indicates the area of ​​the solder paste on the leads (the area of ​​the first image, calculated by cropping from the side image of the component). This represents the first preset area (open circuit threshold, set based on historical data or experience). This represents the sensitivity factor (used to adjust the steepness of the open-circuit suspicion scoring curve).

[0053] This rating indicates the degree of suspicion of an open-circuit fault in the component. When the solder paste area of ​​the pin ( Less than the preset minimum requirement ( When the open circuit suspicion score is close to 1, it indicates a high risk of open circuit; conversely, it approaches 0, indicating a low risk. Therefore, the higher the open circuit suspicion score, the greater the possibility of open circuit failure, and the lower the safety performance of the PCBA.

[0054] Short circuit suspicion score :

[0055] ;

[0056] In the above formula, This indicates the area of ​​the solder paste on the bottom surface (the area of ​​the second image, calculated by cropping from the side image of the component). This indicates the second preset area (short-circuit threshold, set based on historical data or experience). This represents the edge similarity (calculated by extracting the edge lines of the first and second image contours, with a value between 0 and 1). This represents the sensitivity factor (used to adjust the steepness of the short-circuit suspicion scoring curve).

[0057] This rating indicates the degree of suspicion of a short circuit fault in the component, taking into account factors such as excessive solder paste (area of ​​solder paste on the bottom surface). Exceeding the maximum allowed value ) and contour connectivity (edge ​​similarity) The higher the score, the higher the risk of short circuit. Therefore, the higher the short circuit suspicion score, the greater the likelihood of a short circuit failure, and the lower the safety performance of the PCBA.

[0058] Furthermore, the current and voltage data are periodically divided, and waveform factors, RMS values, and jitter distribution random characteristic coefficients are extracted to generate an electrical feature vector. The generation of the electrical feature vector includes: dividing the current data into multiple periods using a clustering algorithm; calculating the current waveform smoothing index, periodic jitter steady-state coefficient, and jitter distribution random characteristic coefficient within each period; analyzing the jitter distribution random characteristic coefficient based on an anomaly detection algorithm to identify abnormal current periods. The electrical feature vector is then normalized; a comprehensive feature vector is generated through weighted summation based on an attention mechanism. Specific key indicators include:

[0059] Current waveform smoothness index :

[0060] ;

[0061] In the above formula, The waveform smoothing index is represented by the waveform smoothing index (obtained by periodic division and calculation of current data, such as a combination of waveform factor and RMS value). The threshold representing the waveform smoothing index (set based on historical normal data). This represents the sensitivity factor (used to adjust the steepness of the current waveform smoothing exponential curve).

[0062] This index represents the probability of an anomaly in the smoothness of the current waveform. Original waveform smoothness index ( A higher value indicates a smoother waveform (safer), but after normalization, when... Below the threshold ( )hour, Approaching 0 (normal); when When it is above the threshold, Approaching 1 (abnormal). In fact, in fault detection, The current waveform smoothness index is designed to represent the probability of anomalies; a higher value indicates a more severe waveform anomaly. Therefore, the higher the value of the current waveform smoothness index, the greater the probability of current waveform anomalies, and the lower the safety performance of the PCBA.

[0063] Periodic jitter steady-state coefficient :

[0064] ;

[0065] In the above formula, This represents the steady-state coefficient of periodic jitter (obtained by calculating the drift and diffusion coefficients from current data). The threshold representing the jitter steady-state coefficient (set based on historical normal data); This represents the sensitivity factor (used to adjust the steepness of the steady-state coefficient curve of the periodic jitter).

[0066] This coefficient represents the probability of anomalies in the steady-state characteristics of current jitter. Original jitter steady-state coefficient ( A higher value indicates more stable (safe) jitter, but after normalization, when... Below the threshold ( ) hour, Approaching 0 (normal); when When it is above the threshold, Approaching 1 (abnormal). Similarly, Designed as an anomaly probability, a higher value indicates greater instability in the jitter. Therefore, the higher the value of the periodic jitter steady-state coefficient, the greater the probability of current jitter anomalies, and the lower the safety performance of the PCBA.

[0067] jitter distribution random characteristic coefficients :

[0068] ;

[0069] In the above formula, Represents the random characteristic coefficients of the jitter distribution (calculated based on the distribution characteristics of the time interval error sequence); The threshold representing the random characteristic coefficient of jitter (set based on historical normal data); This represents the sensitivity factor (used to adjust the steepness of the random characteristic coefficient curve of the jitter distribution).

[0070] This coefficient represents the probability of anomalies in the random characteristics of current jitter. Original jitter random characteristic coefficient ( A higher value indicates more severe random jitter (abnormality). After normalization, when... Above the threshold ( )hour, Approaching 1 (abnormal); when When below the threshold, The value should be close to 0 (normal). A higher value indicates a more severe random jitter anomaly. Therefore, the higher the value of the random characteristic coefficient of the jitter distribution, the greater the probability of random current jitter anomaly, and the lower the safety performance of the PCBA.

[0071] Example 2:

[0072] like Figures 1-4 As shown, the image feature vectors and electrical feature vectors of PCBA circuit images in the historical database are fused using a multimodal method to generate a comprehensive feature-based fault detection model. Specifically, the generation of the comprehensive feature-based fault detection model includes:

[0073] S301. Collect at least two key indicators from the image feature vector and electrical feature vector of the PCBA circuit image in the historical database, and also obtain the actual quality of the PCBA circuit image in subsequent tests.

[0074] For example, image feature vectors and electrical feature vectors are obtained from 1000 identically soldered radio PCBA boards (such as open circuit suspicion score in the image feature vector and jitter distribution random feature coefficient in the electrical feature vector). Then, stress tests are performed on these 1000 identically soldered radio PCBA boards. The stress test specifically involves:

[0075] Under the same conditions, 1000 radio PCBA boards were simultaneously subjected to overload operation with a power supply exceeding 50% of the voltage or current.

[0076] 1) After a specified time (e.g., 72 hours), conduct temperature tests on each of the 1000 radio PCBA boards again. Rank the 1000 radio PCBA boards according to their temperatures. If the temperature of a radio PCBA board is lower than that of the other 500 radio PCBA boards, its quality coefficient can be marked as 50.0%.

[0077] 2) Without a specified time limit, continuously connect 1000 radio PCBA boards to power and operate them under overload conditions until the radio PCBA boards fail. Rank the radio PCBA boards according to the duration of operation. If the duration of operation of a radio PCBA board exceeds that of the other 500 radio PCBA boards, its quality coefficient can be marked as 50.0%.

[0078] Furthermore, the quality coefficients obtained above are used as a metric to measure the actual quality of subsequent tests on radio PCBA circuit images.

[0079] S302. The control variable method is used to analyze the feature-quality relationship between each feature vector and the actual quality, as well as the feature-quality relationship between each feature vector. The specific steps to obtain the feature-quality relationship between each feature vector and the actual quality (between the open-path suspicion score and the quality coefficient, and between the jitter distribution random feature coefficient and the quality coefficient) are as follows:

[0080] 1) Among the above 1000 radio PCBA boards, extract 100 radio PCBA boards with a jitter distribution random characteristic coefficient of 0.3, and denote the quality coefficient of these 100 radio PCBA boards as X.

[0081] Therefore, a feature-quality relationship is established based on the relationship between the quality coefficient X and the open-circuit suspicion score. (The red dots in the diagram represent the distribution of the 100 collected samples), therefore:

[0082] (Formula 1);

[0083] In Formula 1 above, , Used for control A constant that approximates the mass coefficient X. And by... Figure 2 The data in the middle can be determined , When, in Formula 1 It approximates the quality coefficient X (where, express Figure 2 The curve in the figure shows the red dot representing the quality coefficient X.

[0084] 2) Among the above 1000 radio PCBA boards, 100 radio PCBA boards with an open circuit suspicion score of 0.4 are extracted, and the quality coefficient of these 100 radio PCBA boards is denoted as Y.

[0085] Therefore, a feature-quality relationship is established based on the relationship between the quality coefficient Y and the random characteristic coefficients of the jitter distribution. (The red dots in the diagram represent the distribution of the 100 collected samples), therefore:

[0086] (Formula 2);

[0087] In Formula 1 above, , Used for control A constant that approximates the mass coefficient Y. And by... Figure 3 The data in the middle can be determined , When, in Formula 1 It approximates the mass coefficient Y (where, express Figure 3 The curve in the figure shows the red dot representing the quality coefficient Y.

[0088] S303. Generate a comprehensive feature-based fault detection model based on feature-quality relationships and feature-feature relationships. The specific steps for obtaining the feature-feature relationships between each feature vector (between the open-circuit suspicion score and the random feature coefficients of the jitter distribution) are as follows:

[0089] From the aforementioned 1000 radio PCBA boards, 100 radio PCBA board samples are randomly selected. The open-circuit suspicion score and jitter distribution random characteristic coefficients of these 100 radio PCBA boards are obtained, along with their quality coefficient W. Based on the deterministic feature-feature relationship between the open-circuit suspicion score and the jitter distribution random characteristic coefficients, the expression for the comprehensive feature fault detection model is as follows:

[0090] G= × (Formula 1) × (Formula 2);

[0091] In the above, The constant used to control the approximation of the formula G and the quality coefficient W. And by... Figure 4 The data in the middle can be determined At this point, formula G and the quality coefficient W tend to approximate each other. The mathematical expression for the comprehensive feature fault detection model generated by the fitting at this time is:

[0092] ;

[0093] In the above formula, G represents the estimated quality coefficient output by the comprehensive feature fault detection model of the radio PCBA board. The larger the estimated quality coefficient, the more reliable the quality of the radio PCBA board. A threshold can even be set based on the estimated quality coefficient output by this comprehensive feature fault detection model to control the yield rate of the radio PCBA board. For example:

[0094] Table 1. Data from some radio PCBA boards and abnormal conditions encountered during subsequent stress testing.

[0095]

[0096] This model organically combines multiple features (such as image features and electrical features) through multimodal fusion technology. It is trained using historical data and considers the interactions between features (such as feature-quality relationships and feature-feature relationships), thus providing a more comprehensive, accurate, and reliable quality assessment. This not only reduces the risk of misjudgment due to the limitations of single features but also intuitively reflects the overall quality status of PCBA boards through unified quantitative indicators (estimated quality coefficients). This facilitates setting thresholds to control yield rates, improves the accuracy and efficiency of fault detection, and ultimately enhances the automation level of product quality control.

[0097] Next, the comprehensive feature vector is input into a preset comprehensive feature fault detection model, which outputs the estimated quality status of the PCBA circuit image. Based on the estimated quality status of the PCBA circuit image, visual markings are made on the PCBA circuit image, and a diagnostic report is automatically generated. This step includes: marking a highlighted box on the circuit image according to the fault location; generating maintenance suggestion text according to the fault type; combining the marked image and text into a diagnostic report, and transmitting it to the quality inspection platform via the Internet of Things.

[0098] Example 3:

[0099] A PCBA fault detection system includes: a data acquisition module, a feature extraction module, a multimodal fusion module, a fault detection module, a report generation module, and a database module.

[0100] The data acquisition module is used to acquire multimodal data of PCBA circuit images, including a side image acquisition unit and an electrical test unit;

[0101] The side image acquisition unit is configured to acquire side images of PCBA components from the side of the components;

[0102] The electrical test unit is configured to measure the current and voltage data of the PCBA;

[0103] The feature extraction module, connected to the data acquisition module, is used for preprocessing and feature extraction of multimodal data, including:

[0104] The image feature extraction unit is configured to extract the solder paste area from the side image of the component, calculate the solder paste area and contour features, and identify suspected open circuit components and suspected short circuit components based on a predefined area threshold, and generate an image feature vector.

[0105] The electrical feature extraction unit is configured to periodically divide current and voltage data, extract waveform factors, effective values ​​and jitter distribution random feature coefficients, and generate electrical feature vectors.

[0106] The multimodal fusion module, connected to the feature extraction module, is configured to perform feature-level fusion of image feature vectors and electrical feature vectors to generate a comprehensive feature vector;

[0107] The fault detection module, connected to the multimodal fusion module, includes a pre-trained comprehensive feature fault detection model, configured to receive comprehensive feature vectors and output fault location and confidence level;

[0108] The report generation module, connected to the fault detection module, is configured to perform visual marking on the PCBA circuit image based on the fault location and confidence level, and automatically generate a diagnostic report.

[0109] The database module, connected to the fault detection module and the feature extraction module, is used to store historical multimodal data, feature vectors, actual quality data, and comprehensive feature fault detection model parameters.

[0110] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A PCBA fault detection method, characterized in that, Includes the following steps: S1. Acquire multimodal data of PCBA circuit images, including obtaining component side images through the side image acquisition module and obtaining current and voltage data through the electrical test module; S2. Preprocess and extract features from the multimodal data, wherein: Solder paste area and contour features are calculated by cropping solder paste area from the side image of the component, and suspected open circuit and suspected short circuit components are identified based on predefined area thresholds to generate image feature vectors. The current and voltage data are periodically divided, and waveform factors, effective values, and jitter distribution random characteristic coefficients are extracted to generate an electrical feature vector. S3. Multimodal fusion of at least two key indicators, image feature vector and electrical feature vector, from PCBA circuit images in the historical database to generate a comprehensive feature fault detection model. S4. Input the comprehensive feature vector into the preset comprehensive feature fault detection model and output the estimated quality status of the PCBA circuit image. S5. Based on the estimated quality status of the PCBA circuit image, make visual markings on the PCBA circuit image and automatically generate a diagnostic report.

2. The PCBA fault detection method according to claim 1, characterized in that: In step S2, identifying suspected open-circuit components and suspected short-circuit components includes: If the outline feature of the solder paste at the component pin position is a first image outline, and the first image area of ​​the first image outline is smaller than a first preset area, it is determined to be an open circuit suspected component. If the outline feature of the solder paste on the bottom surface of the component is a second image outline, and the second image area of ​​the second image outline is larger than the second preset area, and the solder paste at the component pin position is connected to the solder paste outline at the bottom surface of the component, it is determined to be a short-circuit suspected component.

3. The PCBA fault detection method according to claim 2, characterized in that: The suspected open circuit element obtains an open circuit suspicion score based on the difference between the area of ​​the first image and the first preset area; The short-circuit suspected element obtains a short-circuit suspected score based on the difference between the area of ​​the second image and the second preset area.

4. The PCBA fault detection method according to claim 1, characterized in that: In step S2, the main indices of the electrical characteristic vector are obtained as follows: The current data is divided into multiple cycles using a clustering algorithm; Calculate the smoothness index of the current waveform within each cycle, the steady-state coefficient of the cycle jitter, and the random characteristic coefficient of the jitter distribution.

5. The PCBA fault detection method according to claim 1, characterized in that: In step S3, multimodal fusion adopts a feature-level fusion method, including: Normalize the image feature vector, electrical feature vector, and thermal feature vector; A comprehensive feature vector is generated by weighted summation based on an attention mechanism.

6. The PCBA fault detection method according to claim 1, characterized in that: In step S3, the specific steps for generating the comprehensive feature-based fault detection model include: S301. Collect at least two main indicators from the image feature vector and electrical feature vector of the PCBA circuit image in the historical database, and also obtain the actual quality of the PCBA circuit image in subsequent tests. S302. Use the controlled variable method to analyze the characteristic-mass relationship between each characteristic vector and the actual mass, as well as the characteristic-characteristic relationship between each characteristic vector; S303. Generate a comprehensive feature-based fault detection model based on feature-quality relationship and feature-feature relationship fitting.

7. The PCBA fault detection method according to claim 1, characterized in that: The comprehensive feature fault detection model is constructed by multimodal feature-level fusion of the main indicators from image feature vectors and electrical feature vectors, and by comprehensive fitting based on feature-quality relationship and feature-feature relationship. It can output a unified estimated quality assessment result that reflects the overall quality status of PCBA.

8. The PCBA fault detection method according to claim 1, characterized in that: Step S5 includes: Mark the fault location with a highlighted box on the circuit diagram; Generate maintenance suggestion text based on the fault type; The labeled images and text are combined into a diagnostic report, which is then transmitted to the quality inspection platform via the Internet of Things.

9. A PCBA fault detection system, characterized in that, The PCBA fault detection system, applicable to any one of claims 1-9, comprises: a data acquisition module, a feature extraction module, a multimodal fusion module, a fault detection module, a report generation module, and a database module.

Images