A method for partition detection of board quality

Abstract

The application discloses a kind of partition detection methods of board quality, and the application relates to detection technical field.The steps of the method include: dividing board into functional blocks according to functional structure, collecting each block thermal field distribution characteristic data;Standard thermal template is constructed, and the matching degree of functional block thermal imaging is calculated using Euclidean distance algorithm, and the abnormal thermal response area is marked if it exceeds the preset threshold;Collecting resistance and resistance gradient change data of abnormal thermal response area, if the condition is met, it is determined as defect cluster area, and the resistance abnormal amplitude is calculated;Integrate defect cluster area data to form composite characteristic map, input defect sample matching model, output defect classification result and confidence score;Combined with fuzzy logic reasoning and neural network algorithm, the matching degree of thermal imaging, resistance abnormal amplitude and confidence score are integrated to calculate the area quality score;Fusion defect classification result and area quality score, generate quality grade label.Based on quality grade label, complete board quality partition detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the detection engineering field, and particularly to a partition detection method for board card quality. BACKGROUND

[0002] In today's highly integrated electronic devices, the quality of the board card as a key component of the electronic system is directly related to the performance and reliability of the electronic device. Whether it is the motherboard, graphics card, network card of the computer, or the circuit board of the communication equipment and industrial control system, it carries the key signal processing, data transmission and control functions. Any minor defect may cause partial or even overall functional failure, causing equipment failure and inconvenience to users.

[0003] However, the traditional board card quality detection relies on separate thermal imaging detection, which can only find thermal abnormal areas, but cannot determine the specific reasons for the thermal abnormalities, and it is difficult to accurately locate and classify hidden defects. SUMMARY

[0004] In view of the defects of the prior art, the present application provides a partition detection method for board card quality, which solves the problem that the traditional method is difficult to accurately locate and classify hidden defects.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a partition detection method for board card quality, comprising the following steps:

[0006] Step S1: dividing the to-be-tested board card according to the functional structure to obtain a board card functional block; collecting thermal imaging of the board card functional block to obtain thermal field distribution characteristic data;

[0007] Step S2: constructing standard thermal template data, quantifying the difference between the thermal field distribution characteristic data of each board card functional block and the standard thermal template data by using the z-score maximum value method, and calculating the thermal imaging matching degree of each functional block; when the thermal imaging matching degree of the board card functional block is greater than a preset threshold, the board card functional block is marked as an abnormal thermal response area;

[0008] Step S3: collecting resistance data and resistance gradient change data of the abnormal thermal response area; when the resistance value data or the resistance gradient change data meet the preset defect cluster area locking condition, the abnormal thermal response area is further determined as a defect cluster area, and the resistance abnormal amplitude is calculated based on the resistance value and the resistance gradient change data of the abnormal thermal response area;

[0009] Step S4: combine the thermal field distribution characteristic data, resistance data and resistance gradient change data of the defect cluster area to obtain a composite feature spectrum; construct a defect sample matching model, input the composite feature spectrum into the defect sample matching model, and output to obtain a defect classification result and a confidence score;

[0010] Step S5: according to the fuzzy logic reasoning and neural network combination algorithm, the thermal imaging matching degree, the resistance abnormal amplitude and the confidence score are combined, and a region quality score is calculated;

[0011] Step S6: combine the defect classification result and the region quality score to generate a detection region quality level label, and realize the partition detection of the board quality according to the region quality level label.

[0012] Preferably, the step S1 comprises:

[0013] Taking the functional independence and signal flow of the board as the core principle, and combining the modular design in the circuit diagram, the circuit units completing the same type of function are divided into a block, including: a power module, a main control module, an interface module and a power load module;

[0014] Each block corresponds to an independently analyzable thermal imaging unit, and the temperature data of each functional block is collected by scanning the powered board with an infrared thermal imager.

[0015] According to the temperature data of each functional block, a thermal field distribution cloud picture, a temperature rise curve and a thermal spot shape feature are generated.

[0016] Preferably, the construction of the standard thermal template data comprises:

[0017] Standard thermal template data construction: collect the thermal imaging data of the qualified board under the same working condition, and calculate the statistical characteristic data of each block after removing the abnormal values:

[0018] Temperature mean value = average temperature of all qualified board blocks;

[0019] Temperature standard deviation = temperature fluctuation range reflecting the qualified boards;

[0020] Thermal spot shape feature: including the area, shape factor and centroid position of normal thermal spot;

[0021] Template updating mechanism: after detecting 500 boards, new qualified data is automatically included to recalculate μ and σ to adapt to process fluctuations.

[0022] Preferably, the z-score maximum value method is used to quantify the difference between the thermal field distribution characteristic data of each functional block and the standard thermal template data, and the thermal imaging matching degree of each functional block is calculated.

[0023] The difference between the measured data and the standard thermal template data is quantified by the z-score maximum method to obtain the thermal imaging matching degree, and the formula is as follows:

[0024] Calculate the z-score of each point:

[0025]

[0026] Wherein, is the temperature measurement value of the i th point; is the temperature mean value of the i th point; is the temperature standard deviation of the i th point; Among all the calculated z-scores, the maximum value is :

[0027] Define the matching degree: ;

[0028] Define the matching degree:

[0029]

[0030] Wherein, is the thermal imaging matching degree, k is the attenuation coefficient, which determines the speed of attenuation, is the point most deviating from the mean value in the data set, the value range ∈(0,1], when it is automatically close to 0; the higher the value, the closer the measured thermal field to the standard state: ≥0.9: good matching, no significant thermal anomaly; 0.7 <0.9: slight deviation, which needs to be analyzed in combination with other dimensions; ≤0.7: significant anomaly, marked as “abnormal thermal response area”.

[0031] Preferably, the collected resistance data and resistance gradient change data of the abnormal thermal response area include:

[0032] Contact probe: suitable for high-precision detection, through direct contact between the probe needle tip and the surface of the measured object, accurate collection of surface resistance or electrical signal change, and precise mechanical contact to ensure detection stability and reliability;

[0033] Grid division:

[0034] Divide a two-dimensional grid with a 0.5mm spacing in the abnormal thermal response area, for example, a 10mm×10mm area is divided into 20×20=400 grid points;

[0035] Each grid point is marked as (i,j), corresponding to the board card coordinates ​​, );

[0036] Point-by-point scanning procedure:

[0037] The probe positioning servo motor drives the probe array to move directly above the (i,j) point, and the contact probe applies a pressure of 10g;

[0038] Resistance measurement applies a 100mV excitation voltage, measures the current flowing through the probe, and calculates the surface resistance value by Ohm's law:

[0039]

[0040] Wherein, is the resistance value at the (i,j) grid point, V is the input voltage, and I is the resistance value;

[0041] Gradient calculation of the resistance gradient of the current point and the adjacent four points:

[0042]

[0043] Wherein, represents the resistance gradient of the current point ; R represents the resistance value of the current point; , , and R , and R represent the resistance values of the adjacent points above and below the current point, respectively; and R

[0044] represent the resistance values of the adjacent points to the right and left of the current point, respectively. The gradient value reflects the degree of mutation of the resistance distribution, and the larger the gradient, the higher the possibility of defects.

[0045] The resistance anomaly amplitude is calculated based on the resistance value and the resistance gradient change data of the abnormal thermal response area, which includes:

[0046]

[0047] Wherein, is the resistance anomaly amplitude, is the resistance value of the th abnormal grid point in the defect cluster area; is the average resistance value of the region; is the weight coefficient, which is inversely proportional to the distance from the grid point to the cluster center; is the total number of abnormal grid points; the physical meaning of is the weighted square sum of the resistance value of each abnormal grid point in the defect cluster area and the average value of the resistance of the normal area. The physical meaning of the denominator is the sum of the weight coefficients of all abnormal grid points. This calculation method uses the RMS weighted deviation method to calculate the resistance anomaly amplitude of the cluster area, and more accurately evaluates the resistance anomaly situation by considering the weight of each abnormal grid point.

[0048] Preferably, the composite feature spectrum obtained by combining the thermal field distribution characteristic data, resistance data and resistance gradient change data of the defect cluster area includes:

[0049] The following key features are extracted from the defect cluster area to construct a composite feature spectrum:

[0050] The thermal imaging features include:

[0051] The thermal spot centroid coordinates: (x, y), reflecting the position of the thermal spot on the board card; the maximum thermal gradient: , , indicating the most severe degree of change in the thermal field in the region;

[0052] The average temperature: , representing the overall heating level of the region;

[0053] The resistance scanning features include:

[0054] The average resistance value: , measuring the average level of electrical conductivity in the region;

[0055] The average resistance gradient: , describing the average degree of change in the spatial distribution of resistance values;

[0056] The number of impedance reversals: , recording the number of "high-low-high" or "low-high-low" resistance fluctuations in the region;

[0057] The defect cluster area: , indicating the size of the abnormal resistance area;

[0058] The above features are integrated into a composite feature spectrum : ;

[0059] The composite feature spectrum can be represented as a vector in practical applications. This representation is consistent with the relevant definitions and data processing logic. From the definition point of view, the spectrum is essentially an integrated expression of multiple feature information, and the vector is a high-efficiency data structure that can store and operate these features in an ordered numerical form. ​

[0060] Preferably, the output obtains a defect classification result and a confidence score, which comprises:

[0061] The cosine similarity algorithm is used to calculate the similarity between the feature vector of the sample to be detected and the feature vector of the historical sample :

[0062]

[0063] Wherein, is the similarity between the feature vector of the sample to be detected and the feature vector of the historical sample, is the feature vector of the sample to be detected, is the feature vector of the historical sample, The value range of , the value is closer to , indicating that the two vectors are more similar. The numerator is the calculation formula of the dot product of the two vectors, indicating the sum of the products of the corresponding elements in the feature vectors of the sample to be detected and the historical sample, is the jth feature vector in the feature vector of the sample to be detected, is the jth feature vector in the feature vector of the historical sample.

[0064] Defect type output and confidence score:

[0065] Select the top historical samples according to the similarity from high to low, and the predicted defect type is the defect type corresponding to the sample with the highest similarity , and the confidence score is obtained by converting the highest similarity :

[0066] ;

[0067] Wherein, is the confidence score, is the highest similarity.

[0068] Preferably, the area quality score is calculated by combining the thermal imaging matching degree, the resistance anomaly amplitude and the confidence score according to the fuzzy logic reasoning and neural network combination algorithm, which comprises:

[0069] Nonlinear construction of area quality score function:

[0070] The thermal imaging matching degree, the resistance anomaly amplitude and the confidence score are converted into the area quality score by using the fuzzy logic reasoning and neural network combination method;

[0071] Wherein, the fuzzy logic preprocessing is:​​

[0072] Define fuzzy sets: map raw data to "low / medium / high" fuzzy levels:

[0073] Thermal imaging matching degree is:

[0074] Low: Medium: High: ;

[0075] Resistive anomaly amplitude is:

[0076] Low ), medium High: ;

[0077] Confidence score is:

[0078] Low: Medium: High: ;

[0079] Establish fuzzy rule base: train rules through historical data or expert experience, such as:

[0080]

[0081] Among them, the neural network includes:

[0082] Input layer: fuzzy features × 3 dimensions, and original confidence score as input;

[0083] Hidden layer: use 2 layers of fully connected layers, and ReLU as activation function;

[0084] The first layer of fully connected layers includes:

[0085] Construct weight matrix: ;

[0086] Dimension meaning: input feature number × neuron number

[0087] Parameter role: each weight represents the influence strength of the th input feature on the th neuron;

[0088] Its calculation process is: = ReLU( · + );

[0089] Among them, is the first layer output vector, the feature representation after ReLU activation, Y is a 10-dimensional input vector, containing fuzzed features and original confidence, is a 16-dimensional bias vector, used to adjust the neuron activation threshold, ReLU is the activation function: introducing nonlinearity, the formula is ReLU( )=max(0, ), filtering negative signals;

[0090] The second layer is a fully connected layer, and the second layer fully connected layer includes:

[0091] The weight matrix is: ;

[0092] Dimension meaning: 16 first layer output number x 8 second layer neuron number

[0093] Parameter function: compress the high-dimensional features extracted by the first layer into more abstract representations

[0094] The calculation process is: ;

[0095] Among them, The second layer output vector is the final abstract feature representation, is a 16-dimensional vector of the first layer output, is an 8-dimensional bias vector, is the activation function, used to introduce nonlinearity;

[0096] Output layer: single node output area quality score , mapped to the score interval by the Sigmoid function;

[0097] Training method: using historical quality data, using Adam optimizer to minimize the mean square error loss function.

[0098] Preferably, the combination of the defect classification result and the area quality score to generate a detection area quality level label comprises:

[0099] The quality level label is determined by a decision tree, specifically:

[0100] A CART decision tree model is constructed, taking the area quality score and defect type as core features, and outputting quality level label ;

[0101] The decision tree is constructed as follows:

[0102] Splitting condition:

[0103] First, divide according to the area quality score: Stable area; Further determine the defect type; Potential defect area or failure area;

[0104] Secondary refinement according to defect type:

[0105] If the defect type Is "fatal defect", directly determine as failure area; if it is "repairable defect", adjust the level combined with the score;

[0106] Pruning optimization: avoid overfitting by cross-validation, set the minimum sample size to 10 and the maximum depth to 5;

[0107] Decision path:

[0108] Input: , "virtual soldering of soldering points" -> decision tree determines it as a slight abnormal area; input: , "power layer short circuit" -> decision tree determines it as a failure area.

[0109] Beneficial effects:

[0110] The application provides a partition detection method for board quality, which relates to machine learning and deep learning technology, and has the following beneficial effects:

[0111] (1) The partition detection method for board quality combines thermal imaging acquisition and resistance data monitoring, uses Euclidean distance algorithm to quantify thermal field difference, accurately locates abnormal thermal response area, and further determines defect cluster area, effectively improves detection accuracy, can find small thermal abnormalities and resistance changes, and reduces the risk of missing potential defects;

[0112] (2) The partition detection method for board quality innovatively fuses thermal field distribution feature data, resistance data and resistance gradient change data, constructs a composite feature map, fully excavates the information value of multi-source data, provides more abundant and comprehensive basis for board quality evaluation, and makes the detection result more reliable and credible;

[0113] (3) The partition detection method for board quality combines thermal imaging matching degree, resistance abnormal amplitude and confidence score, constructs regional quality score, and generates regional quality level label combined with defect classification result, realizes comprehensive and comprehensive evaluation of board quality, and more accurately reflects the actual quality status of the board. BRIEF DESCRIPTION OF DRAWINGS

[0114] Figure 1 The application provides a flowchart of a partition detection method for board quality.

[0115] Figure 2 This is a hierarchical diagram of a partitioning detection method for board quality proposed in this invention. Detailed Implementation

[0116] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0117] Please see Figure 1 This invention provides a technical solution: a method for partition detection of circuit board quality. Specifically, the following method for partition detection of circuit board quality is provided; please refer to [link / reference]. Figure 1 The method includes the following steps:

[0118] Step S1: Divide the board under test into regions according to its functional structure to obtain board functional blocks; perform thermal imaging on the board functional blocks to obtain thermal field distribution characteristic data.

[0119] Based on the circuit design principles and physical layout characteristics of the circuit board, the complex circuit board system is decomposed into independent blocks with clear functional attributes, establishing a "function-space" mapping relationship to provide structured analysis units for subsequent thermal imaging detection. This process requires deep integration of circuit knowledge, component characteristics, and detection requirements. The specific operations are as follows:

[0120] First, clarify the criteria for division: based on the functional independence and signal flow of the board, and combined with the modular design in the circuit diagram (such as power supply module, main control module, interface module, etc.), circuit units that perform the same type of function are divided into one block:

[0121] Power supply and voltage regulation module: Includes power input interface, filter capacitor, voltage regulator chip, inductor and other components. It is responsible for converting external power into stable voltage required by each module of the board and has an independent power supply circuit.

[0122] Signal processing module: This module includes a central processing unit (CPU), digital signal processor (DSP), memory chips, and a network of peripheral resistors / capacitors. It undertakes core functions such as data processing and logic control, and is the "brain" of the board.

[0123] Communication interface module: including USB interface, Ethernet interface, RF antenna interface, etc., responsible for data transmission or signal interaction with external devices, and usually has independent signal isolation and impedance matching circuits.

[0124] Power load module: contains power transistors, relays, motor drive chips, etc., used to drive large current loads (such as motors, solenoid valves), which have high power consumption and significant heat characteristics different from other modules

[0125] Second, rely on design documents to achieve accurate division:

[0126] Analyze the circuit diagram: by reading the schematic diagram, identify the circuit boundaries of each functional module (such as isolation elements between power modules and signal modules, distribution of ground symbols), and mark the key elements of each module (such as the VCC pin of the power chip, the clock pin of the signal chip).

[0127] Compare BOM list: according to the model and parameters of the components in the bill of materials, judge the functional type of the components (such as electrolytic capacitors are commonly used for power filtering, high-speed operational amplifiers belong to signal processing modules), and record their coordinate positions on the PCB.

[0128] Finally, each functional block needs to meet the following standards:

[0129] Functional singularity: all elements in the block serve the same core function, with no redundant or irrelevant elements;

[0130] Physical divisibility: the block has clear spatial boundaries on the PCB (which can be distinguished by solder masks, isolation grooves, or component arrangement spacing);

[0131] Detection convenience: the block size is moderate (recommended not to exceed 100mm 2 ), which is convenient for thermal imaging instrument focusing scanning, and each block contains at least 1-2 characteristic elements (such as the heat sink of the power module, the main control chip of the signal module) as the reference point for thermal field analysis.

[0132] Through the above steps, the board card is transformed into a "detection map" composed of multiple functional blocks, each block corresponds to an independently analyzable thermal imaging unit, laying the foundation for subsequent defect positioning based on functional semantics. This division method not only improves the detection specificity (such as directly locating the power module overheating problem), but also provides a logical framework for cross-block linkage analysis (such as the impact of power supply abnormalities on signal modules).

[0133] Obtain the thermal field distribution of each block, the core operation: use an infrared thermal imager to scan the powered board card in multiple states, collect temperature data of each functional block, and build a dynamic thermal field distribution model to provide original physical signals for abnormal judgment. The specific process is as follows:

[0134] Device parameter configuration and environmental control

[0135] Thermal imager settings:

[0136] Resolution: ≥640x480 pixels, to ensure that the temperature of a single component (such as a 0402 packaged resistor) can be distinguished; Frame rate: 10 frames / second, to capture rapid changes in thermal response (such as temperature fluctuations under pulsed loading); Wavelength range: 7.5-14 μm (suitable for long-wave infrared detection of objects at room temperature).

[0137] Environmental requirements:

[0138] Detection of the temperature of the chamber: 25°C ± 1°C, humidity: 45%-55% RH, to avoid the temperature and humidity gradient of the environment from interfering with the heat dissipation of the board card; Electromagnetic shielding: turn off the surrounding strong electromagnetic equipment (such as frequency converters, motors), to prevent electromagnetic interference from affecting the accuracy of the thermal imager sensor.

[0139] Multi-condition thermal imaging data collection:

[0140] Idle state (static power consumption detection): the board card is powered on but not connected to the load (such as disconnecting the signal input, running the power module under no load), and is scanned after being stable for 5 minutes;

[0141] Objective: to detect abnormal heating in standby state (such as excessive static current of power supply chip, leakage of energy storage capacitor).

[0142] Full load state (extreme condition detection): the board card runs the maximum load (such as the signal processing module performs full load operation, and the power module drives the rated load), and runs for 30 minutes to reach thermal equilibrium;

[0143] Objective: to expose the defects of heat dissipation design (such as poor contact between chip and heat sink, insufficient current carrying capacity of PCB copper foil).

[0144] Thermal field distribution characteristic data record:

[0145] Each block generates a thermal field distribution cloud map (temperature chromatogram, with the highest / lowest temperature points marked), a temperature rise curve (temperature change trend over time), and thermal spot morphological characteristics (area, centroid position coordinates, etc.).

[0146] Among them, thermal spot is a phenomenon of abnormal temperature rise in a local area of the board card due to excessive current density or poor contact, which is usually manifested as a local high temperature area in the thermal imaging image,

[0147] Method for obtaining thermal spot morphological characteristics:

[0148] Area: convert the actual area by counting the number of pixel points covered by the thermal spot in the thermal imaging image, and combining the pixel-actual size conversion ratio of the infrared equipment ;

[0149] Centroid position coordinates: calculate the average value of each pixel coordinate in the thermal spot area by formula:

[0150]

[0151] in, and These are the coordinates of the centroid on the x-axis and y-axis, respectively. Here are the coordinates of each pixel within the hotspot area. This represents the total number of pixels in the hotspot area.

[0152] This step divides the board into functional blocks according to its functional structure (such as power module and main control module), and collects the thermal field distribution characteristics of each block using an infrared thermal imager. This decomposes the complex board into structured units, reducing the complexity of the analysis. The acquired thermal field data provides the original basis for subsequent comparison with standard templates and for locating abnormal areas.

[0153] Step S2: Construct standard thermal template data, use the z-score maximum value method to quantify the difference between the thermal field distribution characteristic data of each functional block and the standard thermal template data, calculate the thermal imaging matching degree of each functional block, and mark the functional block of the board as an abnormal thermal response area when the thermal imaging matching degree of the functional block of the board is greater than the preset threshold.

[0154] Standard thermal template data construction: Collect thermal imaging data of qualified circuit boards under the same operating conditions, remove outliers, and calculate the statistical characteristic data of each block.

[0155] Average temperature (μ) = Average temperature of this block across all qualified boards;

[0156] Temperature standard deviation (σ) = reflects the temperature fluctuation range between qualified circuit boards;

[0157] Hot spot morphological characteristics: area and centroid location of a normal hot spot.

[0158] Template update mechanism: After every 500 boards are inspected, new qualified data is automatically incorporated and μ and σ are recalculated to adapt to process fluctuations (such as component batch differences and welding process adjustments).

[0159] The difference between the measured data and the standard thermal template data is quantified by the z-score maximum method to obtain the thermal imaging matching degree, as shown in the following formula:

[0160] Calculate the z-score for each point:

[0161]

[0162] in, It is the first Temperature measurements at each point; It is the first The average temperature at each point; It is the first Temperature standard deviation of a point.

[0163] Take the maximum value among all calculated z-scores : ;

[0164] Definition of matching degree:

[0165]

[0166] wherein, is the thermal imaging matching degree, k is the attenuation coefficient, which determines the speed of attenuation, is the point most deviating from the mean in the data set, the value range ∈(0,1], when it automatically approaches 0 (in line with the 3σ principle); the higher the value, the closer the measured thermal field to the standard state: ≥0.9: good match, no significant thermal anomaly; 0.7 <0.9: slight deviation, need to be analyzed in combination with other dimensions; ≤0.7: significant anomaly, marked as "abnormal thermal response area".

[0167] On the basis of matching degree quantitative analysis, the following qualitative judgment rules are superimposed:

[0168] Absolute temperature limit value: if the temperature of any point in the block exceeds the safety threshold (such as the maximum junction temperature limit of the component -10℃), it is directly determined as abnormal, without the need to calculate the matching degree.

[0169] Thermal spot shape anomaly: the thermal spot area exceeds 2 times the standard value (such as the standard thermal spot area ≤50mm 2 , measured >100mm 2 );

[0170] Comprehensive matching degree calculation result and qualitative rule: matching degree ≤0.7 or triggering absolute temperature limit value or thermal spot shape anomaly, mark as "abnormal thermal response area", and mark the abnormal type (such as "overheating" "thermal spot spread anomaly"), and the coordinate range of "abnormal thermal response area".

[0171] This step is based on the thermal imaging data of qualified boards to construct a standard thermal template (including temperature mean, standard deviation, thermal spot shape, etc.), and by calculating the difference between the measured thermal field and the standard template, the thermal imaging matching degree is obtained, which is marked as an abnormal thermal response area if it exceeds the threshold (such as ≤0.7). Through data-driven way, the thermal field abnormal area is accurately located, reducing invalid detection; the matching degree quantization (0-1 interval) provides a comparable index for the abnormal degree.

[0172] Step S3: Collecting resistance data and resistance gradient change data of the abnormal thermal response area, and when the resistance value data or the resistance gradient change data meets the preset defect cluster area locking condition, further determining the abnormal thermal response area as a defect cluster area, and calculating the resistance abnormal amplitude based on the resistance value and the resistance gradient change data of the abnormal thermal response area.

[0173] The labeled "abnormal thermal response area" is subjected to resistance scanning, and whether the thermal anomaly is caused by a real physical defect (such as a virtual weld or a crack) is verified by electrical characteristics, avoiding resource waste of full-area blind scanning, and realizing "thermal-electric" double verification.

[0174] Selection of detection equipment:

[0175] Contact probe: suitable for high-precision detection (such as 0.1 mm pitch solder joints), through direct contact of the probe tip with the surface of the measured object, accurate collection of surface resistance or electrical signal changes, and precise mechanical contact to ensure detection stability and reliability.

[0176] Grid division:

[0177] Divide the abnormal thermal response area into two-dimensional grids with a 0.5 mm pitch, for example, a 10 mm x 10 mm area is divided into 20 x 20 = 400 grid points.

[0178] Each grid point is labeled as (i, j), corresponding to the board coordinates ( , ).

[0179] Point-by-point scanning process:

[0180] The probe positioning servo motor drives the probe array to move directly above the (i, j) point, and the contact probe applies a 10g pressure;

[0181] Resistance measurement applies a 100mV excitation voltage (to avoid damaging the element), measures the current flowing through the probe, and calculates the surface resistance value through Ohm's law:

[0182]

[0183] Wherein, is the resistance value at the (i, j) grid point, V is the input voltage, and I is the resistance value.

[0184] Gradient Calculate the resistance gradient of the current point and the adjacent four points (up, down, left, right):

[0185]

[0186] Wherein, represents the resistance gradient of the current point ; resistance value of the current point; , resistance value of the adjacent point on the left and right of the current point, respectively; , resistance value of the adjacent point above and below the current point, respectively. Gradient value reflects the degree of mutation of resistance distribution, and the greater the gradient, the higher the possibility of defects.

[0187] Locking conditions of defect cluster area:

[0188] When any of the following conditions is detected in a certain area, it is determined to be a "defect cluster area":

[0189] Impedance reversal occurs in more than 5 adjacent grid points: adjacent grid resistance values show "high-low-high" or "low-high-low" fluctuations, and the fluctuation amplitude exceeds 2 times the normal mean value (such as normal resistance value is 50 mΩ, fluctuation interval > 100 mΩ or < 25 mΩ).

[0190] Resistance gradient of more than 5 adjacent grid points is out of limit: resistance gradient > 50 mΩ / mm (normal conductive area gradient < 10 mΩ / mm), and adjacent grid points are continuously distributed in space (such as forming a linear or planar area).

[0191] For the locked "defect cluster area", first calculate the area of the defect cluster area, and then further calculate the overall amplitude level of resistance anomaly in the area to quantify the severity of defects. The specific steps are as follows:

[0192] Calculate the area S of the defect cluster area:

[0193] Single grid area: the physical area of each grid is: 0.5 mm x 0.5 mm = 0.25 m ;

[0194] Abnormal point number statistics: count the total number of abnormal grid points (denoted as q) that meet the locking conditions of the defect cluster area.

[0195] Total area calculation: S = q x 0.25 m ;

[0196] Where S is the area of the defect cluster area, and q is the total number of abnormal grid points that meet the locking conditions of the defect cluster area.

[0197] Area resistance anomaly amplitude calculation: the resistance anomaly amplitude of the cluster area is calculated by the RMS weighted deviation method, and the formula is as follows:

[0198]

[0199] Where, is the resistance anomaly amplitude, It is the first in the defect cluster area Resistance values ​​at anomalous grid points; This is the average resistance of the normal area (taken from a standard template). It is a weighting coefficient, which is set inversely proportional to the distance between the grid point and the cluster center (the closer the distance, the higher the weight, with a maximum value of 1). It is the total number of abnormal grid points ( The numerator represents the weighted sum of squares of the differences between the resistance values ​​of each anomalous grid point within the defect cluster region and the mean normal resistance of that region. The denominator represents the sum of the weight coefficients of all anomalous grid points. This calculation method uses the RMS weighted deviation method to calculate the magnitude of resistance anomalies in the cluster region, and by considering the weight of each anomalous grid point, it more accurately assesses the resistance anomaly.

[0200] This step involves dividing the abnormal thermal response area into a grid. Resistance values ​​and gradients at each grid point are collected via probe scanning. When the resistance or gradient meets the defect locking criteria (e.g., resistance gradients exceeding limits at five or more adjacent points), it is identified as a defect cluster area. The abnormal resistance amplitude is then calculated using the RMS weighted deviation method. This dual "thermal-electrical" verification identifies the true defect area, eliminating misjudgments of thermal anomalies. The abnormal resistance amplitude quantifies the electrical severity of the defect, providing a basis for subsequent quality scoring.

[0201] Step S4: Combine the thermal field distribution feature data, resistance data, and resistance gradient change data of the defect cluster area to obtain a composite feature map; construct a defect sample matching model, input the composite feature map into the defect sample matching model, and output the defect classification result and confidence score.

[0202] Construction of composite feature maps:

[0203] Extract the following key features from the defect cluster region to construct a composite feature map:

[0204] Thermal imaging features include:

[0205] hot spot centroid coordinates: ( , ), reflecting the location of the hot spot on the board; maximum thermal gradient: This indicates the degree of most drastic change in the thermal field within the region, representing the maximum absolute value of the rate of temperature change (based on the slope change in the temperature rise curve).

[0206] Average temperature: This reflects the overall heat level of the region.

[0207] Resistance scanning features include:

[0208] Average resistance value: , the average level of the electrical conductivity in the region;

[0209] resistance gradient mean: , describes the average degree of change in resistance value in spatial distribution;

[0210] impedance inversion times: , records the number of times the "high-low-high" or "low-high-low" resistance fluctuations occur in the region.

[0211] defect cluster area: , indicates the size of the abnormal resistance area.

[0212] Integrate the above features into a real-time feature vector (composite feature map) : ;

[0213] The composite feature map can be represented as a vector in practical applications. This representation is consistent with the relevant definitions and data processing logic. From the definition point of view, the map is essentially an integrated expression of multiple feature information, while the vector is a high-efficiency data structure that can store and operate these features in an ordered numerical form.

[0214] Defect sample matching model construction:

[0215] First, a large number of historical defect sample data need to be collected. For each sample, record its defect type (such as tin bead short circuit, chip virtual welding, etc.), and through the detection process of steps S1 and S3, obtain the thermal imaging features (thermal spot centroid coordinates (x, y) , ), maximum thermal gradient , average temperature , etc.) and resistance scanning features (average resistance value , resistance gradient mean , impedance inversion times , defect cluster area , etc.) corresponding to the sample. Organize these features into a historical feature vector, and perform Z-Score standardization on the historical feature vector to obtain a standardized historical feature vector , where , is the total number of defect samples. At the same time, record the defect type label of each sample, and also perform Z-Score standardization on the detection sample feature vector to keep the dimension consistent.

[0216] Feature weight determination: Different features have different importance in defect type recognition. In order to reflect the importance difference, it is necessary to determine the feature weight vector Based on the expert experience method, experts in the field are invited to score the importance of each feature based on past maintenance and detection experience, and then normalized to obtain the weight;

[0217] Suppose that experts score the importance of features (such as hot spot centroid, resistance gradient, etc.), and each expert gives a score of ( is the expert number, is the feature number), and the score range is usually 1-10 points (10 points are the most important).

[0218] Single feature total score calculation, for each feature , calculate the sum of all expert scores:

[0219] ;

[0220] Among them, is the sum of expert scores of the jth feature, that is, the sum of scores of all experts on this feature, the total number of experts participating in scoring; k is the expert number, the value range is ( ), is the score of the kth expert on the jth feature.

[0221] Feature weight normalization:

[0222] Convert single feature total score to weight , ensure that the sum of all weights is 1:

[0223] ;

[0224] Among them, the weight value of the jth feature reflects the relative importance of the feature in defect recognition; is the sum of expert scores of the jth feature; n is the total dimension number of features (representing 7 features such as hot spot position and resistance response), and j is the feature number.

[0225] Weighted processing is performed on the feature vector to obtain the weighted feature vector and :

[0226] ;

[0227] ;

[0228] Among them, is the detection sample feature vector, is the weighted feature vector of the real-time feature vector X, is the historical sample feature vector, is the historical feature vector weighted feature vector.

[0229] The cosine similarity algorithm is used to calculate the similarity between the detection sample feature vector and the historical sample feature vector :

[0230]

[0231] wherein, is the similarity between the detection sample feature and the historical sample feature vector, i is the sample index, is the detection sample feature vector, is the historical sample feature vector, The value range of , the value is closer to , indicating that the two vectors are more similar. The numerator is the calculation formula of the dot product of the two vectors, indicating the sum of the products of the corresponding elements in the detection sample and the historical sample feature vectors, is the jth feature vector in the detection sample feature vector, is the jth feature vector in the historical sample feature vector.

[0232] Defect type output and confidence score:

[0233] Select the top historical samples (usually take 3) according to the similarity from high to low. The predicted defect type is the defect type corresponding to the sample with the highest similarity , and the confidence score is obtained by converting the highest similarity

[0234]

[0235] wherein, is the confidence score, is the highest similarity. For example, if , the confidence score . If the highest similarity is lower than the set threshold (such as ), it is determined as an unknown defect type, marked as "unrecognizable".

[0236] ​​​The step is to fuse the thermal field features (thermal spot centroid, thermal gradient maximum, etc.) and resistance features (average resistance, resistance gradient mean, etc.) of the defect cluster area to construct a composite feature map, input the defect sample matching model (based on cosine similarity and historical sample comparison), output the defect type (such as false welding, short circuit) and confidence score, and comprehensively improve the defect recognition accuracy by combining thermal and electrical characteristics; the classification result clearly defines the defect type, and the confidence score quantitatively judges the reliability, providing a basis for subsequent decision-making.

[0237] Step S5: According to the fuzzy logic reasoning and neural network combination algorithm, the thermal imaging matching degree, resistance abnormal amplitude and confidence score are combined to calculate the regional quality score.

[0238] Nonlinear construction of regional quality score function:

[0239] The fuzzy logic reasoning and neural network combination method is used to convert the thermal imaging matching degree ( ), resistance abnormal amplitude ( ), confidence score ( ) into regional quality score ( );

[0240] Among them, the fuzzy logic preprocessing is:

[0241] Define fuzzy sets: map the original data to "mild / medium / severe" fuzzy levels. For example:

[0242] The thermal imaging matching degree is:

[0243] Low ( ), medium ( ), high ( );

[0244] The resistance abnormal amplitude is:

[0245] Low ( ), medium ( ), high ( ).

[0246] The confidence score is:

[0247] Low ( %), medium ( ), high ( ).

[0248] Establish a fuzzy rule base: train rules through historical data or develop rules based on expert experience, for example:

[0249]

[0250] Among them, the neural network includes:

[0251] Input layer: 3-dimension of fuzzed features (3 vectors of fuzzy levels) and original confidence score as input, which is the reliability of data based on historical data training or expert experience;

[0252] Hidden layer: 2 layers of fully connected layers (16, 8 neurons respectively), with ReLU as activation function.

[0253] First layer of fully connected layer (16 neurons):

[0254] Construct weight matrix: ;

[0255] Dimension meaning: 10 (input feature number) x 16 (neuron number)

[0256] Parameter role: each weight represents the influence strength of the th input feature on the th neuron;

[0257] Its calculation process is: = ReLU( · + );

[0258] Where, is the first layer output vector, the feature representation after ReLU activation, Y is a 10-dimensional input vector containing fuzzed features and original confidence, is a 16-dimensional bias vector used to adjust the neuron activation threshold, and ReLU is the activation function: introducing nonlinearity, the formula is ReLU( )=max(0, ), filtering negative signals.

[0259] Second layer of fully connected layer (8 neurons):

[0260] Weight matrix: ;

[0261] Dimension meaning: 16 (first layer output number) x 8 (second layer neuron number)

[0262] Parameter role: compress high-dimensional features extracted by the first layer into more abstract representations

[0263] Its calculation process is: ;

[0264] Where, is the second layer output vector, the final abstract feature representation (8-dimensional), is the 16-dimensional vector output by the first layer, is an 8-dimensional bias vector, is an activation function to introduce non-linear characteristics;

[0265] Output layer: single node output area quality score mapped to the score interval by the Sigmoid function;

[0266] Training method: using historical quality data (known grade labels), using Adam optimizer to minimize the mean square error (MSE) loss function.

[0267] The thermal imaging matching degree, resistance anomaly amplitude and confidence score are mapped to "mild / medium / severe" fuzzy levels by fuzzy logic, and then input into the neural network to output the area quality score, which converts multi-dimensional indicators into a single quantitative score, realizes intuitive evaluation of regional quality, and takes into account multiple factors to more objectively reflect the actual quality level of the region.

[0268] Step S6: Combine the defect classification result with the area quality score to generate a detection area quality grade label, and realize partition detection of the board quality according to the area quality grade label.

[0269] Combine the defect classification result with the area quality score to generate a detection area quality grade label:

[0270] The quality grade label is determined by a decision tree, specifically:

[0271] A CART decision tree model is constructed, taking the area quality score ( ) and defect type ( ) as core features, and outputting the quality grade label ( ).

[0272] The decision tree is constructed as follows:

[0273] Splitting condition:

[0274] Firstly, divide according to the area quality score: Stable area; Further judge the defect type; Potential defect area or fault area.

[0275] Secondly, refine according to the defect type:

[0276] If the defect type is "fatal defect" (such as power supply short circuit), it is directly determined as fault area; if it is "repairable defect" (such as virtual welding of welding point), adjust the grade combined with the score (such as score 65 Minor abnormal area).

[0277] Pruning optimization: Avoid overfitting by cross-validation, set the minimum sample size to 10, and the maximum depth to 5.

[0278] Example decision path:

[0279] Input: , "Solder joint virtual soldering" → Decision tree determines minor abnormal area; input: , "Power layer short circuit" → Decision tree determines fault area.

[0280] Detailed quality inspection report generation:

[0281] Report structure design

[0282] Detection overview: detection area coordinates, time consumption, equipment model;

[0283] Key indicator summary: thermal imaging matching degree, resistance abnormal amplitude, quality grade.

[0284] Multi-dimensional analysis: thermal imaging: temperature chart comparison (actual measurement vs. standard), thermal gradient distribution thermograph; resistance scanning: three-dimensional graph of defect cluster area resistance distribution, abnormal amplitude statistics;

[0285] Defect diagnosis: comparison of matched historical samples, analysis of feature contribution degree of confidence score;

[0286] Decision basis: The calculation process of area quality score is: (fuzzy rule + neural network weight visualization); decision tree decision path diagram (showing the logical branches from input to grade label);

[0287] Improvement suggestions: repair plan: repair steps and tools recommended for defect types;

[0288] Process optimization: correlate historical data and propose parameter adjustment strategies (e.g. welding temperature needs to be increased by 10℃). Realize the intelligent upgrade from data collection to quality decision, provide scientific, transparent, traceable quality evaluation system for board production.

[0289] Based on defect classification results (such as fatal defects, repairable defects) and area quality scores, generate quality grade labels (such as stable area, fault area) through CART decision tree, and finally realize board quality partition detection.

[0290] The present application can intuitively reflect the thermal field distribution of each area of the board card through thermal imaging technology, and potential abnormal thermal areas can be found by comparison with a standard template. Resistance scanning can detect the conductive performance of the board card, and the defect cluster area can be determined by analyzing the resistance data and the gradient change thereof. Through data processing and machine learning algorithm, comprehensive analysis can be performed on the collected data, and accurate evaluation of the quality of the board card can be realized.

[0291] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between such entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover a non-exclusive inclusion, so that a process, method, article or apparatus that includes a list of elements does not only include those elements, but also includes other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the statement "comprising a limited element" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element.

[0292] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives and variations can be made thereto without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for partitioning and detecting the quality of a circuit board, characterized in that, Includes the following steps: Step S1: Divide the board under test into regions according to its functional structure to obtain the board's functional blocks; perform thermal imaging on the board's functional blocks to obtain thermal field distribution characteristic data; Step S2: Construct standard thermal template data, and use the z-score maximum method to quantify the difference between the thermal field distribution characteristic data of each board functional block and the standard thermal template data. Calculate the thermal imaging matching degree of each functional block. When the thermal imaging matching degree of a board functional block is greater than a preset threshold, the board functional block is marked as an abnormal thermal response area; specifically: The difference between the measured data and the standard thermal template data is quantified by the z-score maximum method to obtain the thermal imaging matching degree, as shown in the following formula: Calculate the z-score for each point: ； in, It is the first Temperature measurements at each point; It is the first The average temperature at each point; It is the first Temperature standard deviation at each point; Take the maximum value among all calculated z-scores. : ; Define the degree of matching: ； in, It refers to the thermal imaging matching degree, and k is the attenuation coefficient, which determines the rate of attenuation. It is the point in the dataset that deviates most from the mean. The range is (0,1], The value automatically approaches 0; a higher value indicates that the measured thermal field is closer to the standard state. ≥0.9: Good match, no significant thermal anomalies; 0.7< <0.9: Slight deviation, requires analysis in conjunction with other dimensions; ≤0.7: Significantly abnormal, marked as "abnormal thermal response zone"; Step S3: Collect resistance data and resistance gradient change data of the abnormal thermal response area. When the resistance value data or resistance gradient change data meets the preset defect cluster area locking conditions, the abnormal thermal response area is further determined as a defect cluster area. Simultaneously, based on the resistance value and resistance gradient change data of the abnormal thermal response area, the abnormal resistance amplitude is calculated. Specifically: The abnormal resistance amplitude of the cluster area is calculated using the RMS weighted deviation method, as shown in the following formula: ； in, Abnormal resistance amplitude It is the first in the defect cluster area Resistance values ​​at anomalous grid points; This is the average resistance value of the normal range in this region; It is a weighting coefficient, set inversely proportional to the distance between the grid point and the cluster center; The numerator represents the total number of abnormal grid points; the numerator is the weighted sum of the squares of the differences between the resistance values ​​of each abnormal grid point within the defect cluster region and the average normal resistance value of that region; the denominator is the sum of the weight coefficients of all abnormal grid points. Step S4: Combine the thermal field distribution feature data, resistance data, and resistance gradient change data of the defect cluster area to obtain a composite feature map; construct a defect sample matching model, input the composite feature map into the defect sample matching model, and output the defect classification result and confidence score; Step S5: Based on the algorithm combining fuzzy logic reasoning and neural networks, the thermal imaging matching degree, resistance anomaly amplitude, and confidence score are combined to calculate the area quality score; Step S6: Combine the defect classification results with the regional quality score to generate a quality level label for the detection area. Based on the regional quality level label, perform zoned detection of the board quality.

2. The method for partitioning and detecting the quality of a circuit board according to claim 1, characterized in that, Step S1 includes: Based on the core principles of functional independence and signal flow of the board, and combined with the modular design in the circuit diagram, the circuit units that perform the same type of function are divided into a block, including: power supply module, main control module, interface module and power load module. Each block corresponds to an independently analyzable thermal imaging unit. The power-on board is scanned in multiple states using an infrared thermal imager to collect temperature data of each functional block. Based on the temperature data of each functional area, a thermal field distribution cloud map, temperature rise curve, and hot spot morphology characteristics are generated.

3. The method for partitioning and detecting the quality of a circuit board according to claim 2, characterized in that, The standard hot template data for construction includes: Standard thermal template data construction: Collect thermal imaging data of qualified circuit boards under the same operating conditions, remove outliers, and calculate the statistical characteristic data of each block. Average temperature = the average temperature of this block across all qualified boards; Temperature standard deviation = reflects the temperature fluctuation range between qualified circuit boards; Hot spot morphological characteristics: including the area, shape factor, and centroid location of a normal hot spot; Template update mechanism: After every 500 boards are inspected, new qualified data is automatically incorporated and μ and σ are recalculated to adapt to process fluctuations.

4. The method for partitioning and detecting the quality of a circuit board according to claim 3, characterized in that, The acquisition of resistance data and resistance gradient change data in the abnormal thermal response zone includes: Contact probes: suitable for high-precision detection. The probe tip makes direct contact with the surface of the object being measured to accurately collect changes in surface resistance or electrical signals. Precision mechanical contact ensures the stability and reliability of the detection. Grid generation: Within the abnormal thermal response zone, a two-dimensional grid is divided at a spacing of 0.5 mm, that is, a 10 mm × 10 mm area is divided into 20 × 20 = 400 grid points; Each grid point is labeled (i,j), corresponding to the board coordinates (i,j). , ); Point-by-point scanning process: The probe positioning servo motor drives the probe array to move directly above point (i,j), and the contact probe applies a pressure of 10g. Resistance measurement involves applying an excitation voltage of 100mV, measuring the current flowing through the probe, and calculating the surface resistance using Ohm's law. ； in, I is the resistance value at grid point (i,j), where V is the input voltage and I is the resistance value. Gradient calculation: The resistance gradient between the current point and its four adjacent points. ； in, Indicates the current point The resistance gradient; This represents the resistance value at the current point; , These represent the resistance values ​​of the adjacent points above and below the current point, respectively. , These represent the resistance values ​​of the adjacent points to the right and left of the current point, respectively. The gradient value reflects the degree of abrupt change in the resistance distribution, and the larger the gradient, the higher the probability of a defect.

5. The method for partitioning and detecting the quality of a circuit board according to claim 4, characterized in that, The composite feature map is obtained by combining the thermal field distribution characteristics, resistance data, and resistance gradient change data of the defect cluster region, including: Extract the following key features from the defect cluster region to construct a composite feature map: Thermal imaging features include: hot spot centroid coordinates: ( , ), reflecting the location of the hot spot on the board; maximum thermal gradient: This indicates the degree of most drastic change in the thermal field within the region; Average temperature: This reflects the overall heat level of the region; Resistance scanning features include: Average resistance value: It measures the average level of electrical conductivity within a region; Mean resistance gradient: This describes the average degree of variation in resistance values ​​across spatial distribution; Number of impedance reversals: Record the number of times "high-low-high" or "low-high-low" resistance fluctuations occur within the region; Defect cluster area: , indicating the size of the abnormal resistance region; The above features are integrated into a composite feature map. : .

6. The method for partitioning and detecting the quality of a circuit board according to claim 5, characterized in that, The output yields defect classification results and confidence scores, including: The cosine similarity algorithm is used to calculate the feature vector of the sample to be detected. With historical sample feature vectors similarity : ； in, It detects the similarity between the features of a sample and the feature vectors of historical samples. It is the feature vector of the detected sample. It is a feature vector of historical samples. The range of values ​​is The closer the value is The more similar two vectors are, the higher the numerator is. This is the formula for calculating the dot product of two vectors, which represents the sum of the products of corresponding elements in the feature vectors of the sample to be detected and the historical samples. It is the j-th feature vector in the detected sample feature vector. It is the j-th feature vector in the historical sample feature vector; Defect type output and confidence score: Select the one with the highest similarity Based on historical samples, sorted from highest to lowest similarity, the predicted defect type is the defect type corresponding to the sample with the highest similarity. Confidence score From the highest similarity The result of the conversion is: ； in, It is a confidence score. It represents the highest similarity.

7. The method for partitioning and detecting the quality of a circuit board according to claim 6, characterized in that, The method, based on a combination of fuzzy logic reasoning and neural network algorithms, combines the thermal imaging matching degree, resistivity anomaly amplitude, and confidence score to calculate the region quality score, including: Nonlinear construction of the regional quality scoring function: By combining fuzzy logic reasoning with neural networks, the thermal imaging matching degree, resistance anomaly amplitude, and confidence score are transformed into a regional quality score. The fuzzy logic preprocessing is as follows: Define a fuzzy set: map the original data to "low / medium / high" fuzzy levels. The thermal imaging matching degree is: Low: ,middle: ,high: ; The abnormal resistance amplitude is: Low ,middle ,high: ; The confidence score is: Low: ,middle: ,high: ; Neural networks include: Input layer: The blurred features are multiplied by 3 and the original confidence scores are used as input; Hidden layers: Two fully connected layers are used, with ReLU activation function; The first fully connected layer includes: Construct the weight matrix: ; Dimension meaning: Number of input features × Number of neurons Parameter purpose: Each weight Indicates the first The input feature for the th _th The influence strength of each neuron; The calculation process is as follows: =ReLU( · + ); in, Y is the first layer output vector, representing the features after ReLU activation. Y is a 10-dimensional input vector containing fuzzy features and the original confidence score. It is a 16-dimensional bias vector used to adjust the neuron activation threshold. ReLU is the activation function: introducing non-linearity, the formula is ReLU( )=max(0, ), filtering negative signals; The second fully connected layer includes: Weight matrix: ; Dimension meaning: 16 (number of outputs in the first layer) × 8 (number of neurons in the second layer) The purpose of the parameters is to compress the high-dimensional features extracted in the first layer into a more abstract representation. The calculation process is as follows: ; in, The second layer output vector is the final abstract feature representation. It is the 16-dimensional vector output from the first layer. It is an 8-dimensional bias vector. It is an activation function used to introduce nonlinear characteristics; Output layer: Quality score of single-node output area The score range is mapped to the score range using the Sigmoid function. Training method: Using historical quality data, the Adam optimizer is used to minimize the mean squared error loss function.

8. The method for partitioning and detecting the quality of a circuit board according to claim 7, characterized in that, The step of combining the defect classification results with the regional quality score to generate a quality level label for the detection area includes: The quality level label is determined by the decision tree, specifically: Construct a CART decision tree model for regional quality scoring. and defect type As the core feature, output quality level labels. ; The decision tree is constructed as follows: Split conditions: Priority will be given to classifying regions based on their quality scores: Stable region; Further determine the type of defect; Potential defect area or fault area; The secondary level is refined based on the defect type: If the defect type If it is a "fatal defect", it is directly identified as a fault area; if it is a "repairable defect", the level is adjusted based on the score. Pruning optimization: Overfitting is avoided through cross-validation, with a minimum sample size of 10 and a maximum depth of 5. Decision-making path: enter: , "Poor solder joint" → Decision tree identifies it as a minor anomaly area; Input: , "Power layer short circuit" → Decision tree identifies it as a fault area.

Images