Circuit board conductivity detection method and system based on deep learning

Abstract

The invention relates to the technical field of circuit board conductivity detection, and discloses a circuit board conductivity detection method and system based on deep learning, and the method comprises the steps: obtaining high-resolution image data and circuit design parameters of a to-be-detected circuit board, and generating a circuit board feature data set; calculating theoretical impedance distribution of each area of the circuit board by using a physical constraint neural network based on the circuit board feature data set, and outputting an impedance prediction parameter set; inputting the impedance prediction parameter set and the circuit symbol rule into a neural symbol inference device to generate a conductance fault diagnosis result; and obtaining diagnosis results of a plurality of detection sites, executing cross-site diagnosis fusion, and outputting a unified circuit board conductivity detection report. The method overcomes the limitation that a traditional visual detection method cannot identify electrical characteristic defects, and solves the detection problem of hidden faults such as discontinuous impedance of the high-frequency circuit board.

Description

Technical Field

[0001] This invention relates to the field of circuit board conductivity testing technology, and more specifically, to a circuit board conductivity testing method and system based on deep learning. Background Technology

[0002] In modern electronics manufacturing, circuit boards (PCBs) are core components of electronic products, and their conductivity directly affects the reliability and lifespan of the products. Especially in high-frequency applications, impedance matching and conductivity continuity of PCBs become key factors determining product performance. Currently, electronics manufacturers typically need to conduct quality inspections on high-frequency PCBs across multiple product lines, including consumer-grade, industrial-grade, and automotive-grade PCBs. While PCBs from different product lines exhibit significant differences in impedance characteristics, failure modes, and testing standards, they all adhere to common circuit physics principles.

[0003] However, existing circuit board conductivity testing technologies have the following technical problems:

[0004] First, electrical defects in high-frequency circuit boards, such as impedance discontinuities, minute open circuits, and signal reflections, are difficult to identify using traditional visual inspection methods. These defects may manifest as extremely subtle changes in physical form, even difficult to detect in high-resolution images, but they can severely affect the transmission quality of high-frequency signals. Existing purely data-driven inspection methods lack an understanding of the physical constraints of the circuit and cannot effectively deduce potential electrical characteristic problems.

[0005] Secondly, circuit boards from different product lines vary significantly in design specifications, material properties, and process requirements, resulting in different failure modes and testing standards. Existing testing systems are typically developed independently for a single product line, lacking an effective cross-product line knowledge transfer mechanism. This fails to fully utilize common knowledge across different product lines, leading to the duplication of testing experience and wasted resources.

[0006] Finally, in real-world production environments, companies typically set up multiple testing sites for distributed testing. However, these sites often operate independently, lacking effective coordination mechanisms. This fragmented testing model fails to establish a unified diagnostic logic, and different sites may produce inconsistent judgments for the same type of fault. Furthermore, existing purely data-driven methods neglect symbolic constraints in circuit design, such as impedance matching principles and signal integrity requirements, potentially leading to diagnostic results that violate basic circuit logic. Summary of the Invention

[0007] This invention provides a deep learning-based method and system for detecting the conductivity of circuit boards, which overcomes the limitations of traditional visual inspection methods in identifying electrical characteristic defects and solves the problem of detecting latent faults such as impedance discontinuities in high-frequency circuit boards.

[0008] This invention provides a deep learning-based method for detecting circuit board conductivity, comprising: acquiring high-resolution image data and circuit design parameters of the circuit board to be tested, generating a circuit board feature dataset; calculating the theoretical impedance distribution of each region of the circuit board based on the circuit board feature dataset using a physical constraint neural network, and outputting an impedance prediction parameter set; inputting the impedance prediction parameter set and circuit symbol rules into a neural symbol inference engine, performing physical constraint-based diagnostic inference, and generating conductivity fault diagnosis results; acquiring diagnostic results from multiple testing sites, performing cross-site diagnostic fusion, and outputting a unified circuit board conductivity detection report; wherein, the loss function of the physical constraint neural network includes two parts: data fitting loss and physical constraint loss. The data fitting loss is used to minimize the difference between the predicted impedance value and the true impedance value, and the physical constraint loss is used to constrain the continuity of impedance in adjacent regions.

[0009] Furthermore, high-resolution image data and circuit design parameters of the circuit board under test are acquired, including:

[0010] The process involves: acquiring visible light images of the circuit board using optical imaging equipment to extract geometric feature data of the traces; acquiring temperature distribution data of the circuit board under power-on conditions using infrared thermal imaging equipment to identify abnormal heating areas; processing the acquired image data based on image segmentation technology to generate structured feature vectors containing trace width, spacing, and via positions; reading the CAD design file of the circuit board to extract the theoretical impedance values ​​and dielectric constant parameters of each layer of traces; parsing the netlist file of the circuit board to obtain the connection relationships of signal paths and impedance matching requirements; and checking the design rule file to extract the electrical constraints and tolerance ranges of the circuit board.

[0011] Furthermore, the physically constrained neural network specifically includes:

[0012] The circuit board is divided into multiple detection areas, and a unique position code is assigned to each area. The characteristic impedance of each area is calculated based on the transmission line theory, and the characteristic impedance is calculated based on the ratio of inductance to capacitance per unit length. Image features are extracted using a convolutional neural network and trained in combination with a physical constraint loss function. The physical constraint loss function constrains the continuity of impedance by calculating the difference between the maximum value of the impedance gradient of adjacent areas and a preset threshold. The preset threshold is determined based on the process precision of the circuit board.

[0013] Furthermore, the calculation of inductance and capacitance per unit length is based on a microstrip line transmission model, with frequency-dependent corrections considered:

[0014] For high-frequency applications, the skin effect correction needs to be considered for the inductance per unit length. The skin depth is related to the operating frequency, material permeability, and conductivity. The effective dielectric constant needs to be frequency-corrected, including the frequency characteristics of dielectric loss. The signal propagation delay needs to meet the time-domain constraints, and the maximum allowable delay is related to the clock period. The phase difference between adjacent transmission lines needs to meet the phase matching constraints.

[0015] Furthermore, neural symbolic reasoners include:

[0016] The circuit symbol rules are encoded into first-order logic expressions to form a rule knowledge base; based on the forward chain reasoning method, the impedance prediction results are matched with the symbol rules to generate initial diagnostic hypotheses; the confidence propagation algorithm is used to calculate the confidence of each diagnostic hypothesis, and the fault diagnosis results with confidence exceeding the threshold are selected; wherein, the rule matching process is achieved by calculating the matching degree between the rule antecedent and the current state, and the matching degree is the product of the membership function values ​​of each condition.

[0017] Furthermore, after obtaining the circuit board feature dataset, the process also includes:

[0018] Based on task embedding network analysis, the conductivity detection task characteristics of different product lines are analyzed, and the task similarity matrix between product lines is calculated. According to the task similarity matrix, the detection model parameters of the product lines with similarity higher than a preset threshold are selected, and transfer learning is performed. The transferred model parameters are adjusted based on the domain adaptation algorithm to adapt to the specific detection requirements of the current product line. The task similarity is obtained by calculating the cosine similarity of the task embedding vectors, and the domain adaptation algorithm uses gradient inversion layer technology to achieve domain-invariant feature learning.

[0019] Furthermore, prior to performing cross-site diagnostic fusion, the following steps are also included:

[0020] The collaborative strategy of each detection station is trained based on multi-agent reinforcement learning to learn the optimal information sharing mechanism; detection tasks and computing resources are dynamically allocated according to the detection load and expertise of each station; a distributed consensus method is used to ensure that each station reaches a consensus on the diagnosis results of key faults; wherein, the multi-agent reinforcement learning adopts an Actor-Critic architecture, and the reward function is defined as a weighted combination of detection accuracy, task completion efficiency and response latency.

[0021] Furthermore, the distributed consensus method employs a Byzantine fault-tolerant protocol, including:

[0022] Proposal Phase: Each site packages its local diagnostic results and confidence level into a proposal message and broadcasts it to all other sites; Voting Phase: After receiving proposals from other sites, each site calculates a weighted diagnostic result and decides to vote based on the difference from its local diagnosis; Confirmation Phase: The number of votes for each diagnostic result is counted, and the diagnosis that receives more than two-thirds of the votes is selected as the final consensus result; The protocol can tolerate fewer than one-third of the total number of sites with Byzantine faults.

[0023] Furthermore, it also includes:

[0024] The knowledge base of the neural symbolic inference engine is updated based on an incremental learning method, and newly discovered fault modes are encoded as symbolic rules. The model updates of each detection site are aggregated using a federated learning framework to generate a globally optimized detection model. The performance indicators of the detection model are evaluated regularly, and model retraining and knowledge base updates are automatically triggered. The incremental learning method adopts elastic weight consolidation technology, which prevents catastrophic forgetting by adding a regularization term based on the Fisher information matrix to the loss function.

[0025] This invention provides a circuit board conductivity detection system based on deep learning, comprising: a data acquisition module for acquiring high-resolution image data and circuit design parameters of the circuit board; a physical constraint neural network module for calculating the theoretical impedance distribution of each region of the circuit board based on transmission line theory; a neural symbol inference module for performing diagnostic inference by combining impedance prediction results with circuit symbol rules; and a cross-site fusion module for integrating diagnostic results from multiple detection sites to generate a unified detection report.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention overcomes the limitations of traditional visual inspection methods in identifying electrical characteristic defects by integrating physical constraint neural networks and neural symbol inference engines, and solves the problem of detecting latent faults such as impedance discontinuities in high-frequency circuit boards.

[0028] Specifically, physically constrained neural networks embed the physical laws of transmission line theory into deep learning models, enabling the detection system to infer theoretical impedance distributions based on the geometric features and material parameters of the circuit board. This introduction of physical constraints transforms the model from a purely black box into one capable of understanding the physical characteristics of the circuit. By combining the physical constraint loss function with the data fitting loss function, the model output is ensured to conform to the fundamental laws of electromagnetic field propagation, thus enabling the identification of impedance anomalies that are visually imperceptible but affect signal transmission.

[0029] The introduction of neural symbolic inference addresses the problem of purely data-driven methods neglecting circuit symbol rules. By encoding knowledge in areas such as circuit design rules and impedance matching principles into first-order logic expressions, neural symbolic inference can simultaneously consider data characteristics and logical constraints during the diagnostic process. The forward chain inference algorithm ensures that the generation of diagnostic hypotheses follows the causal relationships of the circuit, while the confidence propagation algorithm quantifies the reliability of each diagnostic result. This combination of symbolic reasoning and neural networks ensures that diagnostic results are both data-supported and conform to circuit logic, avoiding erroneous diagnoses that violate basic circuit principles.

[0030] The cross-site diagnostic fusion mechanism solves the problem of multiple detection sites operating independently. By executing the diagnostic fusion algorithm, the detection results from different sites can mutually verify and complement each other, forming a unified diagnostic logic. The distributed consensus algorithm ensures that all sites reach a consistent judgment on key faults, eliminating diagnostic inconsistencies caused by site differences. This collaborative mechanism fully utilizes the detection capabilities of multiple sites, improving the overall reliability of the diagnosis.

[0031] Furthermore, task similarity analysis and transfer learning mechanisms address the lack of knowledge sharing across different product lines. By calculating task similarity between product lines, the system can identify detection tasks with commonalities and selectively transfer relevant model parameters and diagnostic experience. The domain adaptation algorithm further adjusts the transferred parameters to suit the specific requirements of the target product line, avoiding the negative impacts of blind transfer. This selective knowledge transfer mechanism enables the effective reuse of detection experience across product lines, reducing the resource consumption of repetitive learning.

[0032] Therefore, this invention comprehensively solves the key technical problems in high-frequency circuit board conductivity testing by deeply integrating physical constraints and symbolic reasoning, cross-site collaborative diagnosis, and cross-product line knowledge transfer, and achieves accurate, reliable, and efficient automated testing. Attached Figure Description

[0033] Figure 1 This is a flowchart of a circuit board conductivity detection method based on deep learning according to the present invention;

[0034] Figure 2 This is a comparative analysis chart of the impedance prediction results of the present invention;

[0035] Figure 3 This is a correlation analysis diagram of impedance deviation and temperature anomaly in this invention;

[0036] Figure 4 This is a flowchart of the fault diagnosis reasoning process of the present invention;

[0037] Figure 5 This is a thermal diagram of the impedance distribution in the circuit board area according to the present invention; Detailed Implementation

[0038] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0039] First, the following explanations are provided for the terms that may be used in this article:

[0040] Physically constrained neural networks (PCNs) are deep learning models that embed physical laws as constraints into the network structure or loss function during the training and inference processes of a neural network. In this embodiment, PCNs integrate physical knowledge such as transmission line theory and electromagnetic field propagation laws into the circuit board impedance prediction model, ensuring that the model output conforms to the physical characteristics of the circuit.

[0041] Neural symbolic inferencers are hybrid intelligent systems that combine the data processing capabilities of neural networks with symbolic logic reasoning capabilities. They combine numerical feature data with symbolic logical rules to perform rule-based reasoning tasks. In this embodiment, the neural symbolic inferencer combines impedance prediction results with circuit design rules to generate diagnostic results that conform to logical constraints.

[0042] Task similarity matrix: This is a numerical matrix that quantifies the correlation between different tasks. The elements in the matrix represent the degree of similarity between corresponding task pairs. In this embodiment, the task similarity matrix is ​​used to measure the correlation between conductivity detection tasks from different product lines. The similarity value ranges from 0 to 1, with a larger value indicating greater task similarity.

[0043] Cross-site diagnostic fusion refers to the process of integrating and unifying the diagnostic results of multiple testing sites located in different physical locations. The cross-site diagnostic fusion process uses specific fusion algorithms to comprehensively consider the testing results, confidence levels, and areas of expertise of each site to generate a consistent final diagnostic conclusion.

[0044] Physical constraint loss function: In neural network training, this refers to a penalty term used to ensure that the model output satisfies physical laws. The physical constraint loss function guides the model to learn a mapping relationship that conforms to physical constraints by quantifying the degree of deviation between the model's predictions and physical laws. In this implementation, the physical constraint loss function ensures that the impedance prediction results conform to transmission line theory.

[0045] Forward chain reasoning algorithm: This refers to a symbolic reasoning method that starts from known facts and gradually derives new conclusions by applying reasoning rules. The forward chain reasoning algorithm applies applicable rules sequentially according to the triggering conditions of the rules until no new conclusions can be derived or the target state is reached.

[0046] Confidence propagation algorithm: This refers to an inference algorithm in a probabilistic graphical model that calculates the marginal probability distribution of each node through a message passing mechanism. In this embodiment, the confidence propagation algorithm is used to calculate the confidence of each diagnostic hypothesis. By propagating confidence information in the inference graph, the final failure probability distribution is obtained.

[0047] Domain adaptation algorithms are techniques used to adjust model parameters or feature representations to adapt to the data distribution of a target domain when transferring a model trained in a source domain to a target domain. In this implementation, a domain adaptation algorithm is used to adjust the parameters of a detection model transferred from a similar product line to adapt it to the specific detection requirements of the current product line.

[0048] At least one embodiment of the present invention discloses a circuit board conductivity detection method based on deep learning, such as... Figures 1-5 As shown, it includes the following steps:

[0049] Step 100: Obtain high-resolution image data and circuit design parameters of the circuit board to be inspected, and generate a circuit board feature dataset.

[0050] Step 200: Based on the circuit board feature dataset, use a physical constraint neural network to calculate the theoretical impedance distribution of each region of the circuit board and output the impedance prediction parameter set.

[0051] Step 300: Input the impedance prediction parameter set and circuit symbol rules into the neural symbol inference engine, perform physical constraint-based diagnostic inference, and generate conductance fault diagnosis results.

[0052] Step 400: Obtain the diagnostic results from multiple testing sites, perform cross-site diagnostic fusion, and output a unified circuit board conductivity test report.

[0053] It should be noted that the high-resolution image data in step 100 includes:

[0054] Step 101: Use optical imaging equipment to acquire visible light images of the circuit board and extract geometric feature data of the circuit traces;

[0055] Step 102: Use infrared thermal imaging equipment to collect temperature distribution data of the circuit board under power-on conditions and identify abnormal heat generation areas;

[0056] Step 103: Process the acquired image data based on image segmentation technology to generate a structured feature vector containing line width, spacing, and via position; since line width and spacing are measured in distance units, while via position is represented in coordinate form, perform standardized preprocessing on the structured feature vector to normalize all features to the [0,1] interval.

[0057] It should be noted that the circuit design parameters in step 100 include:

[0058] Step 104: Read the CAD design file of the circuit board and extract the theoretical impedance value and dielectric constant parameter of each layer of the circuit;

[0059] Step 105: Parse the netlist file of the circuit board to obtain the connection relationship of the signal path and the impedance matching requirements;

[0060] Step 106: Based on the design rule check file, extract the electrical constraints and tolerance range of the circuit board; since the circuit design parameters include parameters with different dimensions such as impedance value (ohms), dielectric constant (dimensionless), and tolerance range (percentage), these parameters are standardized and preprocessed to ensure the consistency of dimensions between parameters.

[0061] It should be noted that the physical constraint neural network in step 200 specifically includes:

[0062] Step 201: Divide the circuit board into multiple detection areas and assign a unique location code to each area;

[0063] Step 202: Calculate the characteristic impedance of each region based on transmission line theory. The calculation formula is as follows:

[0064]

[0065] in, Indicates characteristic impedance, Inductance per unit length Indicates capacitance per unit length;

[0066] Furthermore, inductance per unit length and capacitor The calculation method is based on the microstrip line transmission model. For a rectangular cross-section microstrip line, the formula for calculating the inductance per unit length is: ,in H / m is the permeability of free space. For the thickness of the medium, For line width, For effective width, edge correction item , Where is the conductor thickness; the formula for calculating capacitance per unit length is: ,in F / m is the vacuum dielectric constant;

[0067] For the effective relative permittivity, is the relative permittivity of the substrate. The physical parameters of the microstrip line transmission model were obtained by analyzing the circuit board CAD file and material parameter database.

[0068] Furthermore, the frequency dependence and time-dimensional constraints of transmission line parameters include: First, frequency correlation correction, for high-frequency applications ( (GHz), skin effect correction needs to be considered for inductance per unit length. ,in For DC inductors, skin depth , S / m represents the conductivity of copper; the second step is to determine the frequency characteristics of dielectric loss and correct the effective dielectric constant for its frequency. ,in , For reference loss tangent, GHz is the reference frequency. The third step is to define the frequency exponent and the time-domain constraints, including the signal propagation delay. ,in For transmission line length, the maximum permissible delay , For clock cycles; fourth step, frequency effective range constraint. ,in MHz, GHz represents the effective frequency range of the model; Step 5: Phase matching constraints. ,in For the first The phase of a transmission line, the maximum phase difference .

[0069] Step 203: Extract image features using a convolutional neural network and train it using a physical constraint loss function.

[0070] The aforementioned physically constrained neural network uses a convolutional neural network as its feature extractor. Its input layer receives high-resolution image data of the circuit board, with an image size of [missing information]. ,in Indicates the image height. The width represents the image, and 3 represents the three RGB color channels. The network contains multiple convolutional layers, and each convolutional layer outputs a feature map. The calculation formula is:

[0071]

[0072] in, Indicates the input feature map, Indicates the convolution kernel weights, This represents the convolution operation, used in convolutional neural networks. Indicates the bias term. This represents the ReLU activation function.

[0073] The output layer of the aforementioned physically constrained neural network is a fully connected layer with an output dimension of . ,in This indicates the number of detection areas divided on the circuit board, and each output value represents the predicted impedance value of the corresponding area.

[0074] The aforementioned training method for the physically constrained neural network employs a supervised learning model, uses the Adam optimizer as the optimization strategy, and sets the learning rate to [value missing]. Loss function Defined as:

[0075]

[0076] in, This represents the number of detection regions from i=1 to... The summation operation is used to calculate the data fitting loss. Indicates the first Predicted impedance values ​​for each region Indicates the first The true impedance value of each region; This represents the number of pairs from j=1 to the adjacent region. The summation operation is used to calculate the physical constraint loss, which is used to constrain the continuity of impedance between adjacent regions. Indicates the number of adjacent region pairs. This represents the maximum allowable impedance gradient. This represents the weighting coefficient.

[0077] Furthermore, the impedance gradient in the physical constraint loss function The calculation method and adjacent region definition include: First, the circuit board region is meshed, dividing the circuit board according to fixed intervals. Millimeters are divided into A rectangular grid, with the detection area at the center of each grid. ,in Indicates the range of row indexes. The second step is to define the adjacent range of the column index; for each range... Its 4-adjacent neighbor set is The adjacent sets in the boundary region are reduced accordingly; the third step is to calculate the impedance gradient using the central difference method. ,in , The fourth step is boundary processing, which involves using forward or backward differencing for the boundary regions. or Fifth step: Constraint verification, ensuring that all adjacent regions are compatible. Satisfying continuity constraints ,in Distance from the center of the region The set of all adjacent pairs of regions. .

[0078] Furthermore, the parameters in the physical constraint loss function The value is determined based on the manufacturing precision of the circuit board, ranging from [0.01, 0.1] ohms / mm. The specific value is obtained by analyzing the statistical characteristics of impedance variation gradients across different product lines. For high-precision circuit boards... For standard process circuit boards For low-precision circuit boards Weighting coefficients It is used to balance the importance of data fitting and physical constraints. Its value range is [0.1, 10]. It is usually set to a small value of 0.1 in the early stage of training to prioritize data fitting. In the later stage of training, it is gradually increased to 1.0 to strengthen physical constraints. When the physical constraints are violated to a large extent, it can be set to the maximum value of 10.

[0079] The aforementioned physically constrained neural network performs the following steps during the inference phase:

[0080] First, high-resolution image data of the circuit board is received as input, and the image is preprocessed to normalize the pixel values ​​to the [0,1] range;

[0081] Then, multi-scale feature maps are extracted through convolutional layers, and the convolutional kernels automatically learn the geometric features of the circuit routing, such as edges, corners, and vias;

[0082] Next, the extracted feature maps are downsampled through a pooling layer to reduce computational complexity while retaining key feature information;

[0083] Subsequently, the feature vector is input into the fully connected layer, and the impedance value of each detection region is predicted by combining the position encoding information;

[0084] Finally, physical constraint post-processing is applied to ensure that the impedance changes in adjacent regions meet the continuity constraint conditions.

[0085] The aforementioned impedance prediction results need to be decoded into specific detection conclusions. By setting impedance threshold ranges, continuous prediction values ​​are converted into discrete detection states: normal state (impedance value within ±5% of the design range), abnormal state (impedance value deviates from the design range by more than ±10%), and fault state (impedance value deviates from the design range by more than ±20% or an open circuit / short circuit occurs).

[0086] Furthermore, the impedance threshold range is set based on the IPC-2141 standard of the International Electron Industries Connection Association and actual production experience. The ±5% threshold under normal conditions corresponds to the signal integrity requirements of high-frequency circuit boards. Impedance deviations within the ±5% threshold range under normal conditions will not significantly affect signal transmission. The ±10% threshold under abnormal conditions is calculated based on the reflection coefficient in transmission line theory. When the impedance deviation exceeds 10%, the reflection coefficient will increase significantly, affecting signal quality. The ±20% threshold under fault conditions is determined based on statistical analysis of a large number of fault samples. The ±20% threshold range under fault conditions usually corresponds to obvious physical defects such as open circuits, short circuits, or severe geometric deformation.

[0087] It should be noted that the neural symbolic reasoning unit in step 300 includes:

[0088] Step 301: Encode the circuit symbol rules into first-order logic expressions to form a rule knowledge base;

[0089] Step 302: Based on the forward chain reasoning method, match the impedance prediction results with the sign rule to generate initial diagnostic hypotheses;

[0090] Step 303: Calculate the confidence level of each diagnostic hypothesis using the confidence propagation algorithm, and filter out fault diagnosis results with confidence levels exceeding the threshold.

[0091] Furthermore, the confidence threshold in the confidence propagation algorithm is set to 0.8, determined based on ROC curve analysis and the false positive rate requirements in practical applications. When the confidence threshold is set to 0.8, the system accuracy can reach over 95%, while keeping the false alarm rate below 5%. For critical fault types such as open circuits and short circuits, the confidence threshold can be appropriately reduced to 0.7 to improve the detection rate; for minor faults such as slight impedance deviations, the confidence threshold can be increased to 0.9 to reduce false alarms.

[0092] The diagnostic results output by the aforementioned neural symbol inference engine need to be converted into specific, executable test reports and maintenance recommendations. The conversion process includes: generating corresponding problem descriptions based on the fault type, generating precise coordinate information based on the fault location, determining processing priorities based on the severity of the fault, and matching corresponding maintenance plan recommendations based on the fault characteristics.

[0093] The aforementioned neural symbolic inference engine employs a hybrid inference architecture, comprising two parts: a symbolic processing module and a numerical computation module. The symbolic processing module manages and executes the logical rules, while the numerical computation module handles the impedance data for continuous values ​​and calculates confidence levels. The core steps of the inference engine include:

[0094] First, the impedance prediction results are converted into a discretized symbolic representation. Based on a preset threshold range, continuous impedance values ​​are mapped to symbolic states such as {normal, high, low, abnormal}, representing the discretized state set of impedance values.

[0095] Then, a reasoning graph structure is constructed, where nodes represent the state of the detection region, edges represent the constraint relationships between regions, and the weight of the edges is determined by physical adjacency and circuit connection relationships.

[0096] Next, the rule matching process is executed, traversing each rule in the rule knowledge base, calculating the matching degree between the rule's antecedent and the current state, and the matching degree... The calculation formula is:

[0097]

[0098] in, Indicates the number of conditions in the antecedent of the rule. Indicates the first The membership function value for each condition. Indicates the condition number;

[0099] Furthermore, the membership function The trapezoidal fuzzy membership function is adopted, and its mathematical expression is as follows:

[0100]

[0101] in, These are the four key parameter points of the trapezoidal function. For the impedance normal state, the parameters are set to... , , , ,in For designing impedance values, the trapezoidal fuzzy membership function can effectively handle fuzzy boundaries of impedance values, avoiding instability caused by hard threshold judgment.

[0102] Subsequently, based on the matching degree, corresponding rules are triggered to generate new diagnostic hypotheses, each of which includes attributes such as fault type, location, and severity.

[0103] Finally, the confidence of each hypothesis is updated through confidence propagation, and the propagation is iterated until convergence or the maximum number of iterations is reached.

[0104] The aforementioned forward chain reasoning method takes as input a set of impedance prediction parameters and a rule knowledge base, and outputs a set of diagnostic hypotheses. The reasoning process adopts a data-driven model, starting from the known impedance state and gradually deriving possible fault conclusions.

[0105] The aforementioned confidence propagation algorithm takes the inference graph structure and initial confidence distribution as input, and outputs the final confidence value of each diagnostic hypothesis. Based on the message-passing principle of Bayesian networks, the algorithm updates the global confidence level through local computation.

[0106] In this embodiment of the application, in order to improve the detection adaptability across product lines, the method further includes the following after step 100:

[0107] Step 110: Analyze the characteristics of conductivity detection tasks in different product lines based on task embedding network analysis, and calculate the task similarity matrix between product lines;

[0108] Step 111: Based on the task similarity matrix, select product line detection model parameters with similarity higher than the preset threshold, and perform transfer learning;

[0109] Step 112: Adjust the migration model parameters based on the domain adaptation algorithm to adapt them to the specific testing requirements of the current product line.

[0110] The aforementioned task embedding network uses a multilayer perceptron (MLP) as its basic architecture. Its input layer receives the feature vectors of the product line's detection tasks, with a dimension of [missing information]. This includes characteristics such as circuit board size, number of layers, material type, impedance range, and detection accuracy requirements. The network contains three hidden layers, and the calculation formula for each hidden layer is as follows:

[0111]

[0112] in, Indicates the first The hidden state of the layer Indicates the first The weight matrix of the layer, This represents the bias vector. Represents the ReLU activation function. This represents the network layer index.

[0113] The output layer of the aforementioned task embedding network is a linear mapping layer with an output dimension of . This generates task embedding vectors for the product line. Task similarity is obtained by calculating the cosine similarity between the two task embedding vectors:

[0114]

[0115] in, and They represent the first The and the first Task embedding vectors for each product line This indicates the similarity between them.

[0116] Furthermore, cosine similarity The value of is strictly limited to the interval [-1, 1], where -1 indicates that the two task embedding vectors are completely opposite, 0 indicates that the two vectors are orthogonal (i.e., completely unrelated), and 1 indicates that the two vectors are exactly the same. Since the task embedding vectors are processed by the ReLU activation function, all components are non-negative. Therefore, the actual calculated cosine similarity value range is the interval [0, 1], which ensures that the physical meaning of the similarity is clear and the value is stable.

[0117] The aforementioned task embedding network is trained using a contrastive learning model, with the SGD optimizer used as the optimization strategy and a learning rate of 0.01. The loss function is the triplet loss. :

[0118]

[0119] in, This indicates the embedding of anchor samples. Embeddings representing positive samples (similar tasks), Embeddings representing negative samples (dissimilar tasks) This represents the interval parameter.

[0120] Furthermore, the interval parameter The value range is [0.1, 2.0], and the interval parameter α is used to control the minimum distance boundary between positive and negative sample pairs. A value that is too small will cause the boundaries between positive and negative samples in the embedding space to become blurred, affecting the task's ability to distinguish between them. An excessively large value will make the training process difficult to converge. In practical applications, the value is dynamically adjusted according to the complexity of the product line tasks: for product lines with significant differences (such as consumer-grade and automotive-grade), the value is set... For product lines with minor differences (such as industrial-grade products in different frequency bands), set For highly similar product lines To enhance the learning of subtle differences.

[0121] The aforementioned task embedding network performs the following steps when calculating task similarity:

[0122] First, the task feature vector of the product line is extracted, including the physical characteristics of the circuit board (number of layers, board thickness, material dielectric constant), electrical characteristics (operating frequency range, impedance range, signal rate), process characteristics (minimum line width, via diameter, copper foil thickness), and quality requirements (inspection accuracy, fault tolerance). Since each feature has different dimensions and numerical ranges, the feature vector is preprocessed using standardization, employing min-max normalization to scale all features to the [0,1] interval. ,in These are the standardized eigenvalues;

[0123] Then, the feature vector is input into a multilayer perceptron, and a low-dimensional task embedding representation is generated through nonlinear transformation. The embedding dimension is usually set to 128 or 256.

[0124] Next, the embedding vectors are L2 normalized to ensure that all task embeddings lie on the unit hypersphere.

[0125] Subsequently, the cosine similarity between task pairs is calculated to generate a symmetric similarity matrix;

[0126] Finally, transferable source tasks are filtered based on a similarity threshold, which is typically set between 0.7 and 0.8.

[0127] Furthermore, the similarity threshold was determined based on statistical analysis of numerous cross-product line transfer learning experiments. Evaluation of transfer learning performance across 100 different product line pairs revealed that when the similarity threshold was set to 0.7, the success rate of transfer learning was 85%, meaning 85% of transfers improved the performance of the target task. When the threshold was increased to 0.8, the success rate rose to 92%, but the number of available source tasks decreased by approximately 40%. When the threshold was further increased to 0.9, the success rate reached 98%, but the number of available source tasks was extremely limited. Therefore, in practical applications, the threshold is dynamically adjusted based on the number of available source tasks and the required transfer quality: 0.8 is chosen when source tasks are plentiful to ensure transfer quality, and 0.7 is lowered when source tasks are scarce to expand the range of options.

[0128] The aforementioned domain adaptation algorithm takes source domain model parameters and target domain data samples as input and outputs adapted model parameters. The algorithm employs a gradient inversion layer technique, using adversarial training to enable the feature extractor to learn domain-invariant feature representations.

[0129] Furthermore, the specific implementation steps of the domain adaptation algorithm include: First, constructing a domain classifier network, whose input is the output features of the feature extractor, and whose output is the binary classification probability. ,in For feature vectors, and For domain classifier parameters, The second step involves inserting a gradient inversion layer between the feature extractor and the domain classifier, whose output during forward propagation is the sigmoid function. Keeping features unchanged, gradients are reversed during backpropagation. ,in The gradient reversal coefficient has a value range of [0.1, 1.0]; the third step is to define the adversarial loss function. :

[0130]

[0131]

[0132] in For the domain labels of the sample, Indicates the sample index. The fourth step involves performing a summation operation; the feature extractor minimizes the task loss. Simultaneously maximize the domain classification loss To learn domain-invariant features, the total loss is ,in This is the balance coefficient, which is usually set to 0.1.

[0133] The aforementioned convolutional neural network serves as a feature extractor, taking circuit board image data as input and outputting multi-scale feature maps. The network structure comprises multiple convolutional blocks, each consisting of convolutional layers, batch normalization layers, and activation layers. The parameters for the convolutional operation include kernel size (typically 3×3 or 5×5), stride, and padding.

[0134] Furthermore, the specific parameter configuration and initialization methods for the convolutional neural network include: First, network structure design, using the ResNet-50 architecture, containing 5 convolutional blocks. The first convolutional block uses a 7×7 kernel to capture large-scale features, and the subsequent 4 convolutional blocks use 3×3 kernels to extract fine-grained features. The number of channels in each convolutional block are 64, 128, 256, 512, and 1024, respectively. Second, convolutional kernel parameter initialization, using the Xavier initialization method. ,in and These represent the number of input and output channels, respectively, with the bias term initialized to 0. The third step represents a normal distribution; the batch normalization parameter setting and scaling parameter settings are also included. Initialized to 1, offset parameter Initialized to 0, moving average coefficient The fourth step is to configure the convolution operation parameters. For the feature extraction layer, use stride=1 to maintain the feature map size, and for the downsampling layer, use stride=2 to reduce the size. Set the dropout rate to 0.5 to prevent overfitting.

[0135] The fifth step involves a learning rate scheduling strategy, employing cosine annealing.

[0136] ;

[0137] in , , The total number of training rounds. Indicates a time step. This represents the cosine function.

[0138] The aforementioned Adam optimizer is used to update network parameters. Its inputs are gradient information and historical momentum, and its output is the parameter update. The algorithm maintains first-order momentum. and second momentum The updated formula is:

[0139]

[0140]

[0141] in, Indicates the current gradient. , The momentum coefficient, Indicates a time step.

[0142] Furthermore, the complete parameter update process of the Adam optimizer includes specific manifestations of bias correction and the time dimension: the first step is to calculate the momentum estimate after bias correction. and ,in For the current time step, and express and of Power of 1 This represents the corrected first-order momentum. This represents the corrected second-order momentum; the second step is to perform a parameter update. ,in express The parameters at time, For learning rate, It is the numerical stability constant. The third step is to represent the square root function; apply gradient clipping constraints. When the gradient norm exceeds the threshold At that time, the gradient is normalized. To prevent gradient explosion; the fourth step is to enforce parameter range constraints. The parameter boundaries are set according to the network layer type, and the weight parameters... , Bias parameters , .

[0143] In this embodiment of the application, in order to improve the efficiency of multi-site collaborative detection, the method further includes the following step before step 400:

[0144] Step 310: Train the collaborative strategy of each detection station based on multi-agent reinforcement learning to learn the optimal information sharing mechanism;

[0145] Step 311: Dynamically allocate detection tasks and computing resources based on the detection load and expertise of each site;

[0146] Step 312: Use a distributed consensus method to ensure that all sites reach a consensus on the diagnosis results of critical faults.

[0147] The aforementioned multi-agent reinforcement learning takes as input the state information of each detection station (including detection load, historical accuracy, and hardware configuration) and the environmental state (queue of tasks to be detected, urgency level) and outputs the cooperative action strategy of each station. Since the items in the state information have different dimensions and numerical ranges, data preprocessing is performed before inputting into the network: the detection load and historical accuracy are normalized to the [0,1] interval, the hardware configuration parameters are standardized, and the urgency level is encoded as a discrete value. The algorithm adopts an Actor-Critic architecture, where each detection station acts as an agent, learning the optimal strategy through interaction with the environment.

[0148] The aforementioned Actor-Critic architecture comprises two parts: an Actor network and a Critic network. The output layer of the Actor network is a softmax layer, with the output dimension being the size of the action space, outputting the probability distribution of each action selection.

[0149]

[0150] in, Indicates the state Select action The probability, This represents the hidden layer output of the Actor network. and These represent the weight matrix and bias vector of the output layer, respectively.

[0151] The aforementioned Actor network output needs to be decoded into specific cooperative actions. The action space includes: task allocation actions (assigning detection tasks to specific sites), resource scheduling actions (adjusting the allocation ratio of computational resources), and information sharing actions (determining the type and frequency of information to be shared with other sites). This can be achieved through a greedy strategy or... - A greedy strategy selects a specific action from a probability distribution: .

[0152] The output layer of the aforementioned Critic network is a linear layer with an output dimension of 1, used to evaluate state value:

[0153]

[0154] in, Representing state The value estimate, This represents the hidden layer output of the Critic network. and These represent the weights and biases of the output layer, respectively.

[0155] The loss function of the aforementioned Actor-Critic architecture It includes two parts: strategy loss and value loss.

[0156]

[0157] in, This represents the policy loss of the Actor network. Represents the dominance function; This represents the value loss of the Critic network. Indicates the target value; The weighting coefficient represents the value loss.

[0158] Furthermore, the specific calculation methods and constraints for each parameter in the Actor-Critic loss function include: First, the action value function. Computation through temporal difference learning ,in For instant rewards, As a discount factor, The next state; the second step, target value. Using target network computing Target network parameters per Step software update once The third step is to define the policy gradient constraints. Gradient clipping threshold Fourth step: Policy entropy regularization constraint The lower bound of policy entropy By adding an entropy penalty term to the loss function Implementation, entropy weight Fifth step: Bounded constraints on the value function ,in , This is achieved through the tanh activation function and linear transformation. .

[0159] Furthermore, the weighting coefficient for value loss The value ranges from [0.1, 1.0] and is used to balance the importance of strategy optimization and value estimation. If the value is too small, the Critic network will not be trained sufficiently, the value estimation will be inaccurate, and this will affect the policy update of the Actor network. If the value is too large, it will overemphasize the accuracy of value estimation and neglect policy improvement. It is usually set at the beginning of training. To ensure balanced training of the Actor and Critic, adjustments can be made in the later stages of training: increase to 0.8 when the system converges slowly to enhance value learning, and decrease to 0.3 when overfitting occurs to reduce the influence of the Critic.

[0160] The aforementioned reward function Defined as a weighted combination of detection efficiency and accuracy:

[0161]

[0162] in, , , These represent the standardized detection accuracy, task completion efficiency, and response latency, respectively. To eliminate the influence of dimensions, each indicator is standardized using the Z-score: ,in This is the historical average. The standard deviation is the historical value. , , These represent the weight coefficients of the corresponding items, typically set to... , , .

[0163] Furthermore, weighting coefficients , , The settings are based on the actual needs of the PCB testing industry and multi-objective optimization theory. Weighting coefficients This reflects the core importance of testing accuracy in quality control, as false positives and false negatives directly impact product quality and customer satisfaction; weighting coefficient This highlights the importance of task completion efficiency; increasing testing throughput while ensuring quality is key to a company's competitiveness; weighting coefficient This indicates a relatively low weight for response latency, as a small amount of latency is acceptable in actual production. This weighting scheme has been verified by multiple manufacturing companies and can balance inspection quality with production efficiency. In special application scenarios, the weights can be adjusted according to specific needs: for scenarios with high-quality requirements, the weights can be increased. Increased to 0.8; for high-throughput scenarios, it can be... Increased to 0.5.

[0164] The aforementioned distributed consensus method takes as input the local diagnostic results and confidence levels of each site, and outputs a globally consistent diagnostic decision. The algorithm employs a Byzantine fault-tolerant protocol, enabling it to reach correct consensus even when some sites experience failures or errors. The consensus process includes three phases: proposal, voting, and confirmation.

[0165] Furthermore, the specific implementation steps of the Byzantine Fault Tolerance Protocol include: First, the proposal phase, each site... Its local diagnostic results and confidence level Packaged as a proposal message And broadcast to all other sites; the second step, the voting phase, is when each site receives... After considering suggestions from other sites, the weighted diagnostic results are calculated. ,in For the total number of stations, if With local diagnosis The difference is less than the threshold Then vote in favor; otherwise vote against. The third step is the confirmation phase, where the number of votes for each diagnostic result is tallied, and the result receiving more than [a certain number of votes] is selected. The diagnosis of the ticket was taken as the final consensus result, in which The fourth step is a floor function; if no diagnostic results receive enough votes, a faulty site detection is performed. Potential faulty sites are identified by comparing the consistency of diagnoses across sites. After excluding faulty sites, the consensus process is re-executed. The protocol can tolerate a maximum of [number missing] votes. A Byzantine fault site ensures that correct consensus can still be reached even if some sites experience arbitrary failures.

[0166] In this embodiment of the application, to enhance the interpretability of the diagnostic results, the method further includes the following after step 400:

[0167] Step 410: Analyze the formation path of the fault based on causal reasoning and trace the root cause of the fault;

[0168] Step 411: Generate a tiered diagnostic report and repair recommendations based on the fault type and severity;

[0169] Step 412: Use visualization technology to mark the fault location and affected area on the circuit board image.

[0170] The aforementioned causal inference takes fault diagnosis results and historical fault data as input and outputs a fault causal link graph. The algorithm constructs a directed acyclic graph (DAG) to represent the causal relationships between variables and uses structural equation modeling to quantify the strength of causal effects. The inference process includes two stages: causal discovery and causal effect estimation.

[0171] Furthermore, the specific implementation methods of causal reasoning include: First, the causal discovery stage, using the PC algorithm to construct a causal graph and passing the conditional independence test. Identify causal relationships between variables, where , For fault variables, The condition set is used; conditional independence is achieved through the chi-square test, and the test statistic is... ,in For the frequency of observation, The expected frequency is set to a significance level of 1. The second step, the causal effect estimation stage, involves using structural equation modeling to assess the identified causal relationships. Quantifying the strength of causal effects, among which The causal effect coefficient is estimated using the least squares method. The third step is to construct the fault propagation path. This involves traversing the causal graph using a depth-first search algorithm to find all possible paths from the root cause fault to the final fault. The path weights are calculated by multiplying the causal effect coefficients. ,in The fourth step is to sort the paths by their weights and select the top three paths with the highest weights as the main fault-causing paths, generating a causal link diagram for fault analysis and prevention.

[0172] In this embodiment of the application, in order to achieve continuous improvement in detection capability, it further includes:

[0173] Step 500: Update the knowledge base of the neural symbolic inference engine based on the incremental learning method, and encode the newly discovered fault patterns into symbolic rules;

[0174] Step 501: Use the federated learning framework to aggregate the model updates from each detection site and generate a globally optimized detection model;

[0175] Step 502: Periodically evaluate the performance metrics of the detection model and automatically trigger model retraining and knowledge base updates.

[0176] The aforementioned incremental learning method takes new fault samples and existing model parameters as input and outputs updated model parameters and an expanded knowledge base. The algorithm employs Elastic Weight Consolidation (EWC) to prevent catastrophic forgetting by adding a regularization term to the loss function. The regularization term calculates the importance of the parameters based on the Fisher information matrix.

[0177] Furthermore, the specific implementation of the Elastic Weight Consolidation (EWC) technique includes: First, before learning a new task, calculating the Fisher information matrix of the old task. ,in For the first The first step is to calculate the expected value of the squared gradient of the old task dataset; the second step is to record the optimal parameters of the old task. As a reference point; the third step is to define the EWC loss function. :

[0178]

[0179] in Losses due to new missions, The importance weight coefficients range from [1, 1000]. The fourth step involves optimizing the EWC loss function using gradient descent, with the gradient calculated as follows: This ensures that important parameters from the old task are retained while learning new tasks. The Fisher information matrix is ​​specifically calculated using the second derivative of the log-likelihood function: ,in This represents the number of samples.

[0180] The aforementioned federated learning framework takes the local model updates from each detection site as input and outputs the aggregated global model. The algorithm employs a federated averaging (FedAvg) strategy, where a central server periodically collects model parameters from each site and generates a global model through weighted averaging. The weights are dynamically adjusted based on the amount and quality of data from each site.

[0181] Furthermore, the specific implementation steps of the federated learning framework include: First, each detection station executes on its local dataset. In the local training round, the local parameter update formula is: ,in For the site In the The wheel's local parameters, For learning rate, For the site The first step is to determine the local loss function; the second step is for each site to update its local parameters. The data is uploaded to the central server; the third step is for the central server to calculate the weighted aggregate weights. ,in For the site The amount of data, The data quality is scored (based on annotation accuracy and data diversity). The fourth step is to perform global parameter aggregation, which represents the total number of stations. Fifth step, update the global parameters. The data was distributed to all sites to begin the next round of training. Data quality scoring. The calculation formula is ,in For the site The accuracy of the annotation, This is a data diversity indicator (calculated using category distribution entropy). , These are the weighting coefficients.

[0182] Furthermore, the time constraints and convergence criteria of the federated learning framework include: First, communication time constraints. ,in The maximum allowed communication latency is defined in seconds; when the network latency exceeds the threshold, the participation of the corresponding site is suspended. The second step involves training time budget constraints. ,in For a single round of local training time, The third step is to determine the total time budget; the convergence criteria are as follows: The convergence threshold Fourth step, performance stability constraints ,in For the first Accuracy and stability threshold of the global model on the validation set Step 5: Maximum number of training rounds limit ,in To prevent endless training; the sixth step is the early stop strategy, when continuous... Training will automatically terminate if performance does not improve; performance improvement is defined as... ,in .

[0183] A large electronics manufacturing company has circuit board testing requirements for three product lines: consumer-grade smartphone motherboards (operating frequency 2.4GHz), industrial-grade 5G base station PCBs (operating frequency 3.5GHz), and automotive-grade electronic control units (ECUs) (operating frequency 1.8GHz). The company has established three testing sites (A, B, and C) to perform conductivity testing on a batch of automotive-grade ECU control boards. This batch of circuit boards uses a 4-layer design, with FR-4 substrate material, a design impedance of 50±2.5 ohms, and a minimum trace width of 0.1 mm.

[0184] Taking the testing process of an automotive-grade ECU control board as an example, the specific implementation of the method of the present invention is demonstrated:

[0185] Step 100 Data Acquisition Example:

[0186] The system acquires high-resolution image data and circuit design parameters of the circuit board under test. An optical imaging device acquires a visible light image with a resolution of 4096×4096 pixels, extracting the geometric features of the circuit traces. An infrared thermal imaging device acquires the temperature distribution after applying a 1.5V test voltage, detecting abnormal heating in region R7. After image segmentation, a structured feature vector containing trace width, spacing, and via positions is generated. See Table 1.

[0187] Table 1. Feature extraction data from circuit board images

[0188]

[0189] Simultaneously, the system reads the CAD design file to obtain the theoretical impedance values ​​and dielectric constant parameters of each layer, and parses the netlist file to obtain the signal path connection relationships. Through the above processing, a circuit board feature dataset is finally generated, providing input for the subsequent physical constraint neural network. As shown in Table 2:

[0190] Table 2 Circuit Design Parameter Data

[0191]

[0192] Step 200: Example of physical constraint neural network impedance prediction:

[0193] The circuit board was divided into 64 detection areas, each assigned a unique location code. Based on transmission line theory, the characteristic impedance of area R7 was calculated using the inductance per unit length. H / m, capacitance per unit length F / m, characteristic impedance Ω. Considering high-frequency correction, the skin depth at 3.5GHz. m, corrected inductance H / m.

[0194] The loss function calculation result of the physically constrained neural network is: data fitting loss. Physical constraint loss Total loss .

[0195] After network training, the output impedance prediction parameter set includes predicted impedance values, confidence scores, and detection status labels for 64 detection regions. See Table 3.

[0196] Table 3 Impedance prediction results data

[0197]

[0198] Step 300: Example of a Neural Symbolic Reasoning Engine Diagnosis

[0199] The neural symbolic inference engine receives the impedance prediction parameter set and circuit symbol rules output from step 200 as input. The impedance prediction results are converted into symbolic representations, and the impedance deviation of -13.6% in region R7 is mapped to an {abnormal} state. After constructing the inference graph, circuit rules are applied for matching. For region R7, the trigger rule "IF impedance deviation > 10% AND temperature abnormality THEN possible micro-open circuit" is applied, and the matching degree is calculated as follows:

[0200] .

[0201] After iterative calculation using the confidence propagation algorithm, the final confidence levels for each diagnostic hypothesis are: micro-open circuit 0.84, material aging 0.23, and design defect 0.15. The final conductivity fault diagnosis results are generated, including fault type, affected area, confidence level, fault location coordinates, severity, and maintenance recommendations. See Table 4 for details.

[0202] Table 4 Fault Diagnosis Results Data

[0203]

[0204] Steps 310-312: Examples of Multi-Agent Cooperative Strategies

[0205] The collaborative strategy for training based on multi-agent reinforcement learning is as follows:

[0206] Station A State Vector (Detection load, historical accuracy, hardware configuration, number of urgent tasks);

[0207] Station B's state vector ;

[0208] C site state vector .

[0209] The action probability distribution output by the Actor network is: Site A This corresponds to the selection probability of three types of actions: task allocation, resource scheduling, and information sharing. Through... - Greedy strategy selects actions, site A selects task assignment actions ( Site B selects the resource scheduling action, and Site C selects the information sharing action.

[0210] Collaborative strategy execution results: Site A was assigned to handle the high-complexity ECU control board detection task, with computing resources adjusted to 80%; Site B optimized the detection algorithm parameters, improving processing efficiency by 15%; Site C shared the fault mode feature library, achieving cross-site knowledge synchronization. Critic network-assessed state value. , , This indicates that the collaborative strategy effectively improved the overall detection performance.

[0211] Step 400 Cross-site diagnostic fusion example:

[0212] The system obtains diagnostic results from multiple testing sites as input. After optimization based on the collaborative strategy, the testing results of the three testing sites for the same batch of circuit boards are as follows: Site A has a fault detection rate of 12.3% (confidence level 0.86), Site B has a fault detection rate of 11.8% (confidence level 0.82), and Site C has a fault detection rate of 12.1% (confidence level 0.84).

[0213] Cross-site diagnostic fusion is performed using a distributed consensus algorithm to calculate weighted diagnostic results.

[0214] .

[0215] The voting results from all three sites showed a failure rate of approximately 12%, receiving unanimous approval with three votes, reaching a final consensus, and resulting in the output of a unified circuit board conductivity test report. See Table 5 for details.

[0216] Table 5 Cross-site diagnostic fusion results

[0217]

[0218] Steps 110-112: Cross-product line knowledge transfer example:

[0219] Step 110: Analyze the characteristics of conductivity detection tasks across different product lines using a task embedding network, and calculate the task similarity matrix between product lines. The system extracts feature vectors through the task embedding network:

[0220] Automotive-grade task embedding ;

[0221] Industrial-grade task embedding Calculate cosine similarity. The migration conditions are met.

[0222] Step 111: Based on the task similarity matrix, select product line detection model parameters with similarity higher than a preset threshold and perform transfer learning. The system selects the detection model parameters of industrial-grade 5G base station PCBs as source domain parameters and transfers them to the automotive-grade ECU control board detection task.

[0223] Step 112: Adjust the model parameters for the transition based on the domain adaptation algorithm to adapt them to the specific detection requirements of the current product line. The domain adaptation algorithm is used to adjust the model parameters, specifically the inversion coefficients of the gradient inversion layer. The accuracy of the domain classifier decreased from 85.2% to 52.1%, indicating successful learning of domain-invariant features. After transfer learning, the model's accuracy on automotive-grade inspection tasks improved from 73.4% to 91.2%. (See Table 6.)

[0224] Table 6 Cross-product line task similarity analysis

[0225]

[0226] As can be seen from the above application examples, the method of the present invention can effectively detect electrical conduction faults on circuit boards, especially hidden defects such as micro-open circuits that are difficult to identify by traditional visual methods. It realizes multi-site collaborative diagnosis and cross-product line knowledge reuse, and greatly improves detection efficiency and accuracy.

[0227] It is understood that data preprocessing methods known to those skilled in the art include data cleaning, data transformation, and data reduction. Data transformation includes type conversion and normalization and standardization. Although the dimensions and types of data were omitted in the description of the preceding embodiments, data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite step in data processing. Therefore, the previously described well-known data preprocessing steps were not described independently.

[0228] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A deep learning-based circuit board conductance detection method, characterized in that, The method comprises the following steps: obtaining high-resolution image data and circuit design parameters of a circuit board to be detected, and generating circuit board feature data set; based on the circuit board feature data set, using a physically constrained neural network to calculate the theoretical impedance distribution of each region of the circuit board, and outputting impedance prediction parameter set; inputting the impedance prediction parameter set and circuit symbol rules into a neural symbol reasoner, performing physically constrained diagnostic reasoning, and generating conductance fault diagnosis results; obtaining diagnosis results of multiple detection sites, performing cross-site diagnosis fusion, and outputting a unified circuit board conductance detection report; wherein the loss function of the physically constrained neural network includes two parts of data fitting loss and physical constraint loss, the data fitting loss is used to minimize the difference between the predicted impedance value and the real impedance value, and the physical constraint loss is used to constrain the continuity of adjacent region impedance.

2. The deep learning-based circuit board electrical conductivity detection method according to claim 1, characterized by, The method comprises the following steps: using an optical imaging device to collect visible light images of the circuit board and extract geometric feature data of the circuit traces; using an infrared thermal imaging device to collect temperature distribution data of the circuit board under the condition of being powered on, and identifying abnormal heating areas; processing the collected image data based on image segmentation technology to generate a structured feature vector containing circuit width, spacing, and via position; reading the CAD design file of the circuit board, and extracting the theoretical impedance value and dielectric constant parameter of each layer of circuit; analyzing the netlist file of the circuit board to obtain the connection relationship and impedance matching requirements of the signal path; based on the design rule check file, extracting the electrical constraint conditions and tolerance range of the circuit board. 3.The deep learning-based circuit board electric conductivity detection method of claim 1, wherein, The physically constrained neural network specifically comprises: dividing the circuit board into multiple detection regions, and assigning a unique position code to each region; based on the transmission line theory, calculating the characteristic impedance of each region, wherein the characteristic impedance is calculated according to the ratio of unit length inductance and unit length capacitance; using a convolutional neural network to extract image features and combining a physical constraint loss function for training; wherein the physical constraint loss function constrains the continuity of impedance by calculating the difference between the maximum value of the impedance gradient of adjacent regions and a preset threshold, and the preset threshold is determined according to the process accuracy of the circuit board.

4. The deep learning-based circuit board electrical conductivity detection method according to claim 3, characterized by, The calculation of unit length inductance and capacitance is based on the microstrip line transmission model and considers frequency-dependent correction: for high-frequency applications, the unit length inductance needs to consider the skin effect correction, and the skin depth is related to the working frequency, material permeability and conductivity; the effective dielectric constant needs to consider the frequency correction, which includes the frequency characteristics of dielectric loss; the signal propagation delay needs to meet the time domain constraint condition, and the maximum allowed delay is related to the clock period; the phase difference between adjacent transmission lines needs to meet the phase matching constraint.

5. The deep learning-based circuit board electrical conductivity detection method according to claim 1, characterized by, The neural symbol reasoner comprises: encoding the circuit symbol rules into first-order logic expressions to form a rule knowledge base; based on the forward chain reasoning method, matching the impedance prediction results with the symbol rules to generate initial diagnostic hypotheses; using a belief propagation algorithm to calculate the credibility of each diagnostic hypothesis, and screening out fault diagnosis results with credibility exceeding a threshold value; The rule matching process is achieved by calculating the matching degree of the rule antecedent and the current state, and the matching degree is the product of the membership degree function values of each condition.

6. The deep learning-based circuit board electrical conductivity detection method according to claim 1, characterized by, After obtaining the circuit board feature data set, further comprising: Based on the task embedding network, the electrical conductivity detection task characteristics of different product lines are analyzed, and the task similarity matrix between product lines is calculated; According to the task similarity matrix, the product line detection model parameters with similarity higher than the preset threshold are selected, and the transfer learning is performed; Based on the domain adaptation algorithm, the transferred model parameters are adjusted to adapt to the specific detection requirements of the current product line; The task similarity is obtained by calculating the cosine similarity of the task embedding vector, and the domain adaptation algorithm uses the gradient reversal layer technology to realize domain invariant feature learning.

7. The deep learning-based circuit board electrical conductivity detection method according to claim 1, characterized by, Before performing cross-site diagnosis fusion, further comprising: Based on multi-agent reinforcement learning, the cooperative strategy of each detection site is trained, and the optimal information sharing mechanism is learned; According to the detection load and the field of expertise of each site, the detection task and the computing resource are dynamically allocated; Using a distributed consensus method to ensure that each site agrees on the diagnosis result of the key fault; The multi-agent reinforcement learning adopts the Actor-Critic architecture, and the reward function is defined as the weighted combination of detection accuracy, task completion efficiency and response delay. 8.The deep learning-based circuit board electric conductivity detection method of claim 7, wherein, The distributed consensus method adopts the Byzantine fault tolerance protocol, including: Proposal stage: each site packages its local diagnosis result and confidence as a proposal message and broadcasts it to all other sites; Voting stage: after receiving the proposals from other sites, each site calculates the weighted diagnosis result and decides the vote according to the difference with the local diagnosis; Confirmation stage: count the number of votes for each diagnosis result, and select the diagnosis with more than two-thirds of the votes as the final consensus result; The protocol can tolerate less than one-third of the total number of Byzantine fault sites. 9.The deep learning-based circuit board electrical conductivity detection method of claim 1, wherein, Further comprising: Based on the incremental learning method, update the knowledge base of the neural symbolic reasoner, and encode the newly discovered fault mode as a symbolic rule; Using the federated learning framework to aggregate the model updates of each detection site to generate a globally optimized detection model; Periodically evaluate the performance indicators of the detection model to automatically trigger model retraining and knowledge base updating; The incremental learning method uses the elastic weight consolidation technology to prevent catastrophic forgetting by adding a regularization term based on the Fisher information matrix to the loss function. 10.A deep learning-based circuit board conductance detection system configured to perform the deep learning-based circuit board conductance detection method according to any one of claims 1 to 9. Comprising: A data acquisition module for acquiring high-resolution image data and circuit design parameters of the circuit board; A physical constraint neural network module for calculating the theoretical impedance distribution of each region of the circuit board based on transmission line theory; A neural symbolic reasoning module for combining impedance prediction results with circuit symbolic rules to perform diagnostic reasoning; A cross-site fusion module for integrating the diagnosis results of multiple detection sites to generate a unified detection report.

Images