CN-122023257-A - Russian circuit board failure analysis method

CN122023257ACN 122023257 ACN122023257 ACN 122023257ACN-122023257-A

Abstract

The invention belongs to the technical field of electronic device detection and intelligent repair, and particularly relates to a Russian circuit board failure analysis method. The method comprises the steps of carrying out N-type synchronous scanning on a circuit board coating to be tested, marking an acquisition time stamp and space coordinates, aligning a plurality of types of scanning images according to the acquisition time stamp and the space coordinates to obtain a multi-mode data set, extracting key material characteristic parameters of the russian circuit board as priori knowledge, constructing N-channel Transformer encoders, designing different encoder structures for each channel respectively, constructing a failure feature library, determining failure modes of the circuit board and key features corresponding to each failure mode, carrying out similarity calculation on the multi-mode data set, the priori knowledge and the failure feature library by utilizing the N-channel Transformer encoders, outputting confidence of each failure mode after Softmax normalization to form a 32-dimensional confidence vector, and judging the main failure mode of the current coating according to the maximum value in the confidence vector.

Inventors

GAO KEJIN
ZHANG FENGTAO
LIN XINGWEI
HAO WEI
SHI SHENGZHI
Ma Pengsen

Assignees

西安益翔航电科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251227

Claims (9)

1. A Russian circuit board failure analysis method is characterized by comprising the following steps: step one, synchronously scanning the plating layer of the circuit board to be tested in N forms, and marking an acquisition time stamp and a space coordinate; Aligning the scanning images in various forms according to the acquisition time stamp and the space coordinates to obtain a multi-mode data set; step three, extracting key material characteristic parameters of the Russian circuit board as priori knowledge; Step four, constructing N-channel transducer encoders, and respectively designing different encoder structures for each channel; Step five, constructing an invalidation feature library, and determining a circuit board invalidation mode and key features corresponding to each invalidation mode; Step six, performing similarity calculation on the multi-mode data set, the priori knowledge and the failure feature library by using an N-channel transducer encoder to obtain an original output vector with the length of M, wherein each element in the vector represents the matching strength of an input sample and a certain failure mode; carrying out Softmax normalization processing on the original output vector, and converting the original output vector into an M-dimensional confidence distribution vector which represents the matching probability between the current sample and each failure mode; judging a main failure mode of the current coating according to the maximum value in the confidence coefficient distribution vector, and judging that the coating is in composite failure if the confidence coefficient of a plurality of modes is higher; and step nine, matching corresponding repairing modes according to the judging result of the failure mode.
2. The method of claim 1, wherein in the first step, the four forms are performed by an atomic force microscope AFM, an energy dispersive X-ray spectroscopy EDX, an X-ray diffraction XRD device and a high frame rate thermal infrared imager, respectively; the AFM is responsible for acquiring a high-resolution image of the surface morphology and is used for assisting in positioning the microstructure area; Element distribution in the EDX synchronous acquisition area, especially concentration information of Br element, is used for judging the brittle characteristics of the material; the XRD device collects the grain orientation angle sequence at a fixed frequency to form a theta-t dynamic characteristic; the thermal infrared imager captures the thermal field change in the working state and is used as a macroscopic thermal stress analysis basis.
3. The method of claim 2, wherein in the second step, the alignment process is as follows: a) The AFM and EDX data use a sigma 1 coordinate system, and the thermal infrared imager uses a sigma 2 coordinate system; b) Performing image-morphology matching through preset feature mark points in the acquisition cabin to generate a coordinate transformation matrix; c) Performing iterative optimization on the coordinate transformation matrix by using a Levenberg-Marquardt optimization algorithm; d) Setting an initial guess value, and aligning based on the characteristic points of the overlapping areas of the AFM image and the infrared heat map; e) After 500 iterations, controlling the coordinate error to be within 0.3 micron; f) Performing Dynamic Time Warping (DTW) processing on XRD and infrared time series data; g) Setting an alignment window of + -15 frames; h) The slope of the constrained path change is between 0.8 and 1.2; i) A multi-modal dataset in a unified space-time coordinate system is generated.
4. The method of claim 3, wherein in the third step, key material characteristic parameters related to the coating of the circuit board of the II series of the product are extracted from Russian gamma-oC standard document, wherein the key material characteristic parameters comprise the orientation characteristics of a metal crystal face in the coating, the concentration threshold value of a brittle element Br and the typical range of the grain size of the coating, and the parameters are used as material priori knowledge and are reflected in the input processing and characteristic extraction process of a model; before starting to process input data, injecting the preloaded material parameters into an initial processing layer of an encoder as a contextual characteristic, and participating in a characteristic extraction process together with the input multi-mode data; The injected contextual features participate with the multi-modal dataset in a subsequent multi-scale feature extraction process.
5. The method of claim 4, wherein in step four, the transducer encoder comprises: EDX channel, processing three-dimensional voxel data and identifying local element abnormality; XRD channel, namely processing orientation angle sequence and capturing variation trend of crystal grain structure; An infrared channel for processing the multi-scale temperature image and analyzing abnormal distribution of the thermal field; and a material parameter channel is used for processing priori knowledge extracted from the gamma-O-T standard and taking the priori knowledge as an auxiliary judgment basis.
6. The method of claim 5, wherein in step four, the encoder structure of each channel is as follows: EDX extracts spatial distribution features using 33 convolution kernels; XRD processes θ -t sequences using one-dimensional convolution (5×1); the infrared image is subjected to multi-scale coding by using a multi-sigma Gaussian kernel; the material parameter channels map them into vector representations using an embedding layer.
7. The method of claim 6, wherein the failure feature library comprises 32 failure modes, key features of each failure mode are converted into feature vector forms to serve as model matching targets, and the key features comprise Br concentration, theta angle offset and thermal gradient.
8. The method of claim 7, wherein in step nine, if the brittle fracture is occurred, the pulse plating repair process is triggered; if the crystal grain is deflected, triggering a laser annealing adjustment process; If the thermal stress concentration is significant, a local heat dissipation enhancement scheme is started.
9. The method of claim 8, wherein the method further comprises incremental optimization of: If the confidence coefficient of all failure modes in a certain identification is lower than a set threshold value, which indicates that the characteristics of the current sample are not matched with the existing failure modes, the system enters an incremental learning mode, a StyleGAN model is called to perform transformation enhancement on the spatial distribution, time sequence and element concentration dimension of the low-confidence sample, and the number of generated enhancement samples is 5 times that of the original samples and is used for model updating.

Description

Russian circuit board failure analysis method Technical Field The invention belongs to the technical field of electronic device detection and intelligent repair, and particularly relates to a Russian circuit board failure analysis method. The method is particularly suitable for the fault identification, life prediction and repair suggestion generation of the Russian circuit board in high-reliability application scenes such as aerospace, military equipment and the like. Background With the increase of electronic devices in extreme environments (such as low temperature, high temperature, radiation), the reliability of circuit boards becomes a key factor affecting the stable operation of the devices. The Russian circuit board has obvious differences with international general IEC standards due to unique material systems, manufacturing processes and design specifications (such as GOST standards), so that the traditional detection means has low identification precision, high omission factor and long analysis period in the application process, and is difficult to meet the maintenance requirements in a high-reliability scene. The existing failure recognition technology has three major limitations: 1. the traditional scheme adopts a mixed architecture of CNN processing images, LSTM processing time sequence and random forest processing structured data, so that the data needs multipath processing and the recognition delay is more than 2s; 2. the Russian material adaptation is lacking, the general model is not optimized for the specific crystal face orientation (such as 32% -38% of Ni (200) face) and Br segregation characteristic (2.1 at% of brittleness threshold) of the coating of the Pi series, and the misjudgment rate is up to 22%; 3. The cross-scale characteristic fracture is that the existing method separates microscopic (atomic level), mesoscopic (micron level) and macroscopic (millimeter level) data, so that the association miss rate of interface segregation and welding spot fracture is more than 35%. Disclosure of Invention The invention provides a failure analysis method of a Russian circuit board, which aims to solve at least one of the following problems: 1. How to realize the joint identification of the Russian circuit board three-level failure mode through a single model architecture; 2. how to optimize the model structure to adapt to the crystal plane parameters and element distribution characteristics of the coating material of the second class; 3. How to complete the whole process of feature extraction, classification decision and restoration linkage through a single model; 4. How to reduce the delay (existing scheme >1.5 s) and the false positive rate (conventional scheme > 20%) caused by multi-model data conversion. The technical scheme is as follows: A Russian circuit board failure analysis method comprises the following steps: step one, synchronously scanning the plating layer of the circuit board to be tested in N forms, and marking an acquisition time stamp and a space coordinate; Aligning the scanning images in various forms according to the acquisition time stamp and the space coordinates to obtain a multi-mode data set; step three, extracting key material characteristic parameters of the Russian circuit board as priori knowledge; Step four, constructing N-channel transducer encoders, and respectively designing different encoder structures for each channel; Step five, constructing an invalidation feature library, and determining a circuit board invalidation mode and key features corresponding to each invalidation mode; And step six, performing similarity calculation on the multi-mode data set, the priori knowledge and the failure feature library by using an N-channel transducer encoder, wherein each failure mode corresponds to a group of feature templates, and the system obtains an original output vector with the length of 32 by calculating similarity scores between input features and the templates, and each element in the vector represents the matching strength of an input sample and a certain failure mode. And step seven, after feature matching is completed, the system carries out Softmax normalization processing on the original matching score output by the multi-head attention mechanism, and converts the original matching score into a 32-dimensional confidence distribution vector, so as to represent the matching probability between the current sample and each failure mode. The normalization operation ensures comparability among different failure characteristics and provides a quantitative basis for subsequent pattern recognition and repair decisions. Judging a main failure mode of the current coating according to the maximum value in the confidence coefficient vector, and judging that the coating is in composite failure if the confidence coefficient of a plurality of modes is higher; and step nine, matching corresponding repairing modes according to the judging result of the failure mode. In