A Charger PCBA Fault Detection Method Based on Multimodal Feature Fusion

Abstract

This invention discloses a charger PCBA fault detection method based on multimodal feature fusion, belonging to the field of electronic circuit fault detection technology. The method includes: applying a dynamic load to the charger PCBA under test, and simultaneously acquiring voltage ripple signals, electromagnetic radiation signals, and thermal distribution signals; performing time-frequency domain decomposition on the voltage ripple signal to obtain a ripple energy spectrum sequence, performing spectral broadening analysis on the electromagnetic radiation signal to obtain a radiation feature vector, and performing gradient calculation on the thermal distribution signal to obtain a thermal diffusion trajectory; inputting the above three modal data into a cross-feature mapping layer, generating a first interactive feature map through point-by-point multiplication and integration, and generating a second interactive feature map through bidirectional convolution; after concatenation, outputting a multimodal fusion feature tensor through a gated recurrent fusion unit, and then outputting the fault type and fault location of the charger PCBA by a fault discrimination network.

Description

Technical Field

[0001] This invention relates to the field of electronic circuit fault detection technology, specifically to a charger PCBA fault detection method based on multimodal feature fusion. Background Technology

[0002] During the manufacturing and use of charger PCBAs, faults such as poor soldering of components, capacitor aging, and chip damage can cause multi-physical field anomalies, including output voltage fluctuations, abnormal electromagnetic radiation, and localized overheating. Existing fault detection methods mainly rely on manual visual inspection or single electrical parameter testing, which cannot capture the transient evolution of multi-physical field signals under dynamic operating conditions. Some technologies attempt to introduce infrared thermography or near-field electromagnetic scanning, but they often analyze the characteristics of each mode independently, processing the temperature matrix, electromagnetic spectrum, and ripple waveform separately and then directly stitching them together for classification. This approach ignores the inherent coupling relationship between the voltage ripple energy spectrum, the broadening characteristics of the electromagnetic radiation spectrum, and the spatiotemporal trajectory of heat diffusion under dynamic load conditions, failing to establish a joint characterization of different physical fields under the same fault excitation. The energy mutations in specific frequency bands of voltage ripple and the power fluctuations of electromagnetic radiation in the corresponding frequency bands exhibit a nonlinear product interaction, while the direction and timing of the movement of heat diffusion hotspots are closely related to the transient changes in voltage ripple. Existing stitching-based fusion methods cannot extract such cross-modal cross-correlation features. Meanwhile, the temporal context information of the heat diffusion trajectory fails to achieve bidirectional alignment and deep interaction with electrical signal characteristics. This makes fault location determination rely solely on static hotspot regions, making it difficult to distinguish between multi-point heat diffusion caused by a single fault source and the superimposed temperature rise caused by multiple independent faults, resulting in insufficient fault location accuracy. Therefore, it is urgent to address how to perform fine-grained cross-modal interactive mapping of voltage ripple, electromagnetic radiation, and heat distribution synchronously acquired under dynamic loads, and how to achieve bidirectional recursive fusion of the spatiotemporal evolution law of heat diffusion with voltage ripple characteristics, in order to improve the accuracy of charger PCBA fault type identification and location. Summary of the Invention

[0003] This paper presents a fault detection method for charger PCBA based on multimodal feature fusion. It constructs cross-modal interactive features from voltage ripple, electromagnetic radiation and thermal distribution signals synchronously acquired under dynamic load, and realizes bidirectional time-series fusion of thermal diffusion trajectory and ripple features, thereby obtaining the fault type and precise fault location of the charger PCBA.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a charger PCBA fault detection method based on multimodal feature fusion. By synchronously collecting voltage ripple signal, electromagnetic radiation signal and thermal distribution signal of charger PCBA under dynamic load excitation, feature extraction and cross-fusion of different modal signals are performed respectively, and finally the fault type and fault location are accurately identified.

[0005] In this invention, a dynamic load is applied to the charger PCBA under test. During the application of the dynamic load, the voltage ripple signal, electromagnetic radiation signal, and thermal distribution signal of the charger PCBA are simultaneously acquired. The voltage ripple signal is decomposed in the time-frequency domain to obtain a ripple energy spectrum sequence. The electromagnetic radiation signal is subjected to spectral broadening analysis to obtain a radiation feature vector. The thermal distribution signal is calculated using gradients to obtain a thermal diffusion trajectory. The ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory are input into a cross-feature mapping layer. The cross-feature mapping layer generates a first interactive feature map between voltage ripple and electromagnetic radiation through point-by-point multiplication and integration operations, and simultaneously generates a second interactive feature map between thermal diffusion and voltage ripple through bidirectional convolution operations. The first and second interactive feature maps are concatenated channel-by-channel and then input into a gated cyclic fusion unit. The gated cyclic fusion unit outputs a multimodal fusion feature tensor. The multimodal fusion feature tensor is input into a fault discrimination network, which outputs the fault type and fault location of the charger PCBA.

[0006] In the above scheme, dynamic load can excite latent defects in PCBA that are difficult to expose under real-world operating conditions; voltage ripple directly reflects the stability of the power conversion circuit; electromagnetic radiation captures switching noise and near-field coupling anomalies; and thermal distribution characterizes local power consumption abrupt changes and heat accumulation effects. These three modal signals complementarily depict fault characteristics from electrical, electromagnetic, and thermodynamic dimensions. By performing time-frequency domain decomposition on the voltage ripple to obtain the ripple energy spectrum sequence, the energy migration patterns of different frequency bands can be accurately identified, exhibiting high sensitivity to transient disturbances. Spectral broadening analysis of the electromagnetic radiation signal extracts the average power, peak power, and autocorrelation attenuation coefficient of sub-bands, effectively describing the concentration and persistence characteristics of broadband radiation and suppressing environmental noise interference. The thermal diffusion trajectory obtained by gradient calculation of the thermal distribution signal retains the spatial movement information of hotspots in the time dimension, expanding the thermal anomaly from a static temperature rise to a dynamic behavioral characterization.

[0007] The cross-feature mapping layer generates the first interactive feature map through pointwise multiplication and integration operations. This process multiplies the ripple energy spectrum vector and the radiation feature vector element-wise, strengthening the nonlinear correlation between the voltage ripple frequency domain features and the electromagnetic radiation frequency domain features, thus amplifying coupling anomalies related to switching stress. Simultaneously, bidirectional convolution operations generate the second interactive feature map. Utilizing the systematic encoding of the thermal diffusion trajectory and bidirectional temporal convolution, it achieves deep fusion of thermal diffusion behavior and voltage ripple features, capturing the forward and backward dependencies of thermal-electrical parameters over time, and highlighting the temporal synergistic changes between local hotspots and ripple fluctuations. The gated cyclic fusion unit performs adaptive gated fusion on the stitched interactive feature map, dynamically selecting the cross-modal combination with the most informational value for fault discrimination, thereby forming a compact and highly discriminative multimodal fusion feature tensor.

[0008] Based on the multimodal fusion feature tensor, the fault discrimination network adopts a dual-branch structure. The first branch outputs the probability distribution of fault types through convolution and classification layers, while the second branch outputs the two-dimensional coordinates of the fault location through convolution and regression layers. This simultaneously completes fault classification and localization, avoiding the accumulation of errors caused by step-by-step prediction, and improving fault detection from single type identification to accurate spatial localization.

[0009] As a technical solution of this invention, when a dynamic load is applied to the charger PCBA under test, the output terminal of the charger PCBA is connected to a programmable electronic load. The programmable electronic load outputs multiple load current values ​​sequentially according to a preset current step curve. The duration of each load current value is a preset duration, and a transient load transition is inserted between two adjacent load current values. The rise time and fall time of the transient load transition are both controlled at the microsecond level. By superimposing the step current with the microsecond-level transient transition, a wide range of operating points from light load to heavy load can be covered in a short time, and key components such as switching transistors and transformers can generate rapid di / dt changes, fully stimulating abnormal ripple, electromagnetic emission, and heat dissipation at the fault point, thereby improving the detectability of fault characteristics.

[0010] Preferably, during the acquisition phase, a differential probe is connected to the positive and negative terminals of the charger PCBA's output. The output of the differential probe is connected to the first channel of a high-speed data acquisition card, which records the voltage ripple signal at a preset sampling rate. A near-field electromagnetic probe is placed at a preset distance above the charger PCBA, and its output is connected to a spectrum analyzer. The spectrum analyzer outputs electromagnetic radiation signals in real time via an Ethernet interface. An infrared thermal imager is used to target the component surfaces of the charger PCBA. The infrared thermal imager continuously outputs a temperature matrix sequence of heat distribution signals at a preset frame rate, with the spatial resolution of the temperature matrix sequence corresponding to the component layout of the charger PCBA. The near-field electromagnetic probe can be a magnetic field probe or an electric field probe, with a preset distance of 1 mm to 5 mm. This acquisition architecture ensures the time synchronization and spatial consistency of the voltage, electromagnetic, and thermal signals. The near-field measurement method effectively captures localized areas of strong radiation, and the matching of the infrared frame rate and spatial resolution allows the heat diffusion trajectory to accurately reflect the component-level thermal response.

[0011] For processing voltage ripple signals, preferably, the voltage ripple signal is segmented into non-overlapping segments according to a preset time window length. Wavelet packet decomposition is performed on the voltage ripple signal within each time window, with a preset number of decomposition levels. Wavelet packet coefficients are extracted from each terminal node obtained from the wavelet packet decomposition. The sum of the squares of the wavelet packet coefficients of each terminal node is calculated as the frequency band energy value corresponding to that terminal node. The frequency band energy values ​​are arranged into a vector form according to the frequency band order of the terminal nodes to obtain the ripple energy spectrum vector corresponding to each time window. The ripple energy spectrum vectors of all time windows are stacked in chronological order to form a ripple energy spectrum sequence. The wavelet packet decomposition preferably uses the Daubechies wavelet basis function, with 3 or 4 decomposition levels. Wavelet packet transform can finely divide the high-frequency band of the signal, non-overlapping segmentation ensures time resolution, and energy value calculation highlights the frequency band energy redistribution caused by faults, enabling quantitative expression of abnormal ripple and subharmonic oscillations near integer multiples of the switching frequency.

[0012] For spectral broadening analysis of electromagnetic radiation signals, preferably, the electromagnetic radiation signal is divided into multiple sub-bands according to a preset frequency range. The average and peak radiated power values ​​within each sub-band are calculated, and an autocorrelation attenuation coefficient is obtained by performing autocorrelation calculations on the radiated power value sequence within each sub-band. The average radiated power value, peak radiated power value, and autocorrelation attenuation coefficient of each sub-band are concatenated to form a sub-band feature vector. All sub-band feature vectors are then concatenated in frequency band order to form a radiation feature vector. This method comprehensively utilizes statistical measures and time-varying characteristics, describing not only radiation intensity but also the periodicity and randomness of radiation. It is sensitive to intermittent discharge and radiation broadening effects, and helps distinguish normal switching noise from abnormal radiation caused by solder joint cracks or filter capacitor degradation.

[0013] For processing the heat distribution signal, preferably, spatial gradient calculation is performed on the temperature matrix of each frame of the heat distribution signal to obtain the gradient field matrix corresponding to each frame of the temperature matrix. The gradient field matrix contains horizontal gradient components and vertical gradient components. A predetermined number of pixels with the largest temperature gradient amplitude are located in the gradient field matrix of each frame, and the coordinate positions and gradient direction angles of these pixels are recorded. The coordinate positions of the pixels recorded in all frames are connected in chronological order to form a set of heat diffusion trajectories. Each heat diffusion trajectory corresponds to the movement path of a temperature hotspot on the surface of the charger PCBA. By extracting the point with the largest gradient amplitude in the gradient field and tracking its movement, the influence of the background temperature field can be eliminated, focusing only on the area with the most active heat transport. Even weak hot spots can be effectively identified through their movement trajectories, which is particularly suitable for intermittent short circuits or high-resistance connection point faults.

[0014] To achieve accurate cross-modal interaction, a preferred approach is to perform time alignment before inputting the ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory into the cross-feature mapping layer: Using the time sampling points of the ripple energy spectrum sequence as the reference time axis, the acquisition time corresponding to each ripple energy spectrum vector is located; the radiation feature sub-vector closest to the acquisition time of the ripple energy spectrum vector is extracted from the acquisition timestamp of the radiation feature vector, and the extracted radiation feature sub-vector is paired with the ripple energy spectrum vector to form a time-aligned pair; the thermal diffusion trajectory segment of the same frame as the acquisition time of the ripple energy spectrum vector is extracted from the frame timestamp of the thermal diffusion trajectory, and this thermal diffusion trajectory segment is paired with the ripple energy spectrum vector to form a time-aligned triplet. This time alignment operation eliminates differences in sampling rate and response latency among multiple sensors, ensuring that subsequent interaction feature mapping is based on strictly time-corresponding data, avoiding feature ambiguity caused by misaligned fusion, and improving the timing accuracy of fault diagnosis.

[0015] In generating the first interactive feature map, preferably, each ripple energy spectrum vector in the ripple energy spectrum sequence is multiplied element-wise with its corresponding radiation feature vector to obtain a first product vector. This first product vector is then input into a first fully connected layer, where the number of output nodes is the same as the dimension of the ripple energy spectrum vector. The first fully connected layer outputs a first intermediate vector. The first intermediate vector is then reconstructed into a two-dimensional matrix by performing element-wise addition with the ripple energy spectrum vector, resulting in the first interactive feature map. In this design, element-wise multiplication establishes a one-to-one cross-modal intensity mapping. The fully connected layer performs a nonlinear transformation on the multiplicative interaction result, and then connects the original ripple energy spectrum vector through residual addition. This captures the modulation effect between electromagnetic and voltage signals while preserving the time-frequency structure of the voltage ripple itself, enabling the first interactive feature map to possess both cross-modal interactive information and single-modal integrity.

[0016] In generating the second interactive feature map, preferably, the coordinate position and gradient direction angle of each trajectory point in the heat diffusion trajectory are encoded as local thermal feature vectors. These local thermal feature vectors are then arranged according to the order of the trajectory points in the heat diffusion trajectory to form a thermal feature sequence. A one-dimensional convolution kernel is used to perform a convolution operation on the thermal feature sequence in the forward direction to obtain a forward convolution result. Simultaneously, the same one-dimensional convolution kernel is used to perform a convolution operation on the thermal feature sequence in the reverse direction to obtain a backward convolution result. The forward and backward convolution results are then element-wise averaged and input into a second fully connected layer. The number of output nodes of the second fully connected layer is the same as the dimension of the ripple energy spectrum vector. The output of the second fully connected layer is reshaped into a two-dimensional matrix to obtain the second interactive feature map. Bidirectional convolution fully utilizes the temporal context of the heat diffusion trajectory, capturing the correlation between past and future states and the current voltage ripple in the heat movement path. The averaging and fusion of forward and backward features makes the temporal features smoother and more robust. Ultimately, the second interactive feature map can effectively express the conduction direction and intensity of heat-induced voltage fluctuations.

[0017] The gated cyclic fusion unit performs temporal modeling and gating selection on the spliced ​​interactive feature map, adaptively enhances the modal interactive channels that contribute significantly to fault identification, suppresses redundant or conflicting channel responses, and outputs a multimodal fusion feature tensor with low noise level and high fault identification accuracy.

[0018] The fault discrimination network is specifically structured as follows: The first branch receives a multimodal fusion feature tensor and contains a first convolutional layer, a first pooling layer, and a first classification layer connected in sequence. The first classification layer outputs a fault type probability distribution vector, where each element corresponds to a preset charger PCBA fault type. The second branch receives the multimodal fusion feature tensor and contains a second convolutional layer, a second pooling layer, and a location regression layer connected in sequence. The location regression layer outputs a fault location coordinate vector, which includes the x-coordinate and y-coordinate values ​​of the fault point on the charger PCBA. The fault type corresponding to the highest probability value in the fault type probability distribution vector is determined as the fault type of the charger PCBA, and the coordinate point determined by the fault location coordinate vector is determined as the fault location of the charger PCBA. The two branches share the same fusion feature, making them computationally efficient and compact. Simultaneously outputting both type and location allows fault localization to assist in interpreting the classification results, improving the reliability of the detection results.

[0019] This invention achieves full-dimensional characterization and precise localization of charger PCBA faults through dynamic load excitation, multi-modal signal synchronous acquisition, feature extraction, cross-feature mapping, gating fusion, and dual-branch discrimination, significantly improving the detection capability and diagnostic accuracy of weak faults, intermittent faults, and complex faults.

[0020] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The first interaction feature map between voltage ripple and electromagnetic radiation is generated through point-by-point multiplication and integration. In this operation, the ripple energy spectrum vector is multiplied element-by-element by the radiation feature vector at the corresponding time. The product directly reflects the joint response mode of energy and radiation intensity increasing in the same direction or decreasing in opposite directions in each frequency band. This point-by-point multiplication interaction enhances the squared anomalous components of common-mode faults caused by the same fault, while random noise or non-homogeneous fluctuations are effectively suppressed due to the lack of product consistency. The product vector is nonlinearly transformed by the first fully connected layer to extract the high-order interaction structure across frequency bands. Then, it is added element-by-element to the original ripple energy spectrum vector to form a residual connection. This introduces the modulation information of electromagnetic radiation on ripple features while maintaining the basic distribution of the ripple energy spectrum, avoiding the submersion of effective information due to deep mapping. Finally, the first interaction feature map is reconstructed into a two-dimensional matrix form, preserving the topological relationship between the ripple frequency band dimension and the time window. This allows subsequent gated fusion units to perceive the local connectivity of cross-modal interaction modes in the time-frequency plane through convolution, enhancing the separability of different fault types in the interaction feature space.

[0021] A second interactive feature map between thermal diffusion and voltage ripple is generated through bidirectional convolution. The coordinates and gradient direction angles of each trajectory point are extracted from the thermal diffusion trajectory and encoded into local thermal feature vectors. These vectors are then arranged in temporal order to form a thermal feature sequence, preserving the direction and rate information of hotspot movement in sequence. The same one-dimensional convolution kernel is used to convolve the thermal feature sequence in the forward direction to capture the forward propagation context of thermal diffusion, while convolution is performed in the reverse direction to capture the backward dependencies. The results of the forward and reverse convolutions are averaged element-wise to integrate the bidirectional temporal perspective. This bidirectional convolution mechanism makes balanced use of the causal and retrospective features of the thermal diffusion trajectory, ensuring that even if there are local loops or abrupt velocity changes in the thermal diffusion path, the alignment and correlation between hotspot movement patterns and voltage ripple events on the time axis can be fully extracted. The averaged features are mapped to a dimension of equal length to the ripple energy spectrum vector through a second fully connected layer, reshaping it into a two-dimensional matrix form of the second interactive feature map. This directly embeds the spatiotemporal dynamics of thermal diffusion into the feature space of the ripple energy spectrum. This interactive method ensures that the voltage ripple state at a given moment not only corresponds to the hotspot location at that moment but also contains evolutionary information in the direction before and after heat diffusion. It tightly couples the causal chain of thermal fault propagation with the transient characteristics of the voltage signal, significantly improving the accuracy of regressing the spatial location of the fault source on the PCBA. The gated cyclic fusion unit performs channel-level adaptive fusion on the first and second interactive feature maps, generating a multimodal fusion feature tensor. This retains the fault type-sensitive features encoded by the voltage-electromagnetic interaction and the location discrimination features implicit in the thermal-electrical bidirectional interaction, suppressing redundant information. Ultimately, this enables the fault discrimination network to simultaneously parse the fault category and fault coordinates of the charger PCBA from the same fusion tensor. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart of a charger PCBA fault detection method based on multimodal feature fusion; Figure 2 This is a schematic diagram of a synchronous acquisition system for charger PCBA voltage ripple, electromagnetic radiation, and heat distribution under dynamic load. Figure 3 This is a flowchart of the ripple energy spectrum sequence generation process; Figure 4 This is a flowchart of electromagnetic radiation feature vector extraction and thermal diffusion trajectory generation; Figure 5 This is a flowchart of the cross-feature mapping layer processing; Figure 6 This is a schematic diagram of the forward and reverse convolution of the thermal feature sequence of the thermal diffusion trajectory and the element-wise averaging result. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0025] See Figure 1This invention provides a charger PCBA fault detection method based on multimodal feature fusion. The method applies a dynamic load to the charger PCBA under test, simultaneously acquiring voltage ripple signals, electromagnetic radiation signals, and thermal distribution signals during the application of the dynamic load. Time-frequency domain decomposition is performed on the voltage ripple signal to obtain a ripple energy spectrum sequence; spectral broadening analysis is performed on the electromagnetic radiation signal to obtain a radiation feature vector; and gradient calculation is performed on the thermal distribution signal to obtain a thermal diffusion trajectory. The ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory are input into a cross-feature mapping layer. This layer generates a first interactive feature map between voltage ripple and electromagnetic radiation through point-by-point multiplication and integration operations, and simultaneously generates a second interactive feature map between thermal diffusion and voltage ripple through bidirectional convolution operations. The first and second interactive feature maps are concatenated channel-by-channel and then input into a gated cyclic fusion unit, which outputs a multimodal fusion feature tensor. The multimodal fusion feature tensor is input into a fault discrimination network, which outputs the fault type and fault location of the charger PCBA.

[0026] Example 1:

[0027] In specific implementation, please refer to Figure 2 A dynamic load is applied to the charger PCBA under test. The output terminal of the charger PCBA is connected to the input terminal of the programmable electronic load via wires. The programmable electronic load has a pre-stored current step curve. The current step curve contains multiple load current values ​​arranged in chronological order, and each load current value is maintained in the programmable electronic load for a preset duration. The preset duration is set to 500 milliseconds based on the loop response characteristics of the charger PCBA. The load current value changes gradually from 10% to 110% of the rated output current of the charger PCBA, with the current increase between every two adjacent load current values ​​being 20% ​​of the rated output current. A transient load jump is inserted between two adjacent load current values ​​output by the programmable electronic load. The transient load jump refers to the process of the load current switching from one load current value to the next. The rise time of a transient load transition is defined as the time required for the load current to rise from 10% of the lower load current value to 90% of the higher load current value. The fall time of a transient load transition is defined as the time required for the load current to fall from 90% of the higher load current value to 10% of the lower load current value. Both rise and fall times are determined by the switching speed of the field-effect transistors inside the programmable electronic load. The control circuit of the programmable electronic load controls both rise and fall times to within 80 microseconds. Throughout the application of the dynamic load, the programmable electronic load continuously records the correspondence between timestamps and load current values.

[0028] During the application of dynamic load, the voltage ripple signal of the charger PCBA is synchronously acquired. The pair of test tips of the differential probe are contacted to the positive and negative output terminals of the charger PCBA, respectively, and the ground terminal of the differential probe is connected to the reference ground of the charger PCBA. The output port of the differential probe is connected to the first channel of a high-speed data acquisition card via a coaxial cable. The high-speed data acquisition card is installed in an industrial computer. The preset sampling rate of the high-speed data acquisition card is set to 100 Mbps, and the input coupling mode is set to AC coupling to remove DC bias voltage and retain only the voltage ripple signal. The high-speed data acquisition card starts acquisition simultaneously with the start of dynamic load output from the programmable electronic load, continuously recording the waveform data of the voltage ripple signal, and storing the waveform data along with the acquisition timestamp as a voltage ripple signal file.

[0029] During dynamic load application, electromagnetic radiation signals from the charger PCBA are simultaneously acquired. A magnetic field probe is fixed above the charger PCBA using a clamp, with the probe's induction coil plane parallel to the component surface of the PCBA. The vertical distance between the bottom of the probe and the top surface of the PCBA is set to 3 mm. The output of the magnetic field probe is connected to the RF input port of the spectrum analyzer via an RF coaxial cable. The spectrum analyzer's center frequency is set to 150 MHz, sweep width to 20 MHz, resolution bandwidth to 10 kHz, video bandwidth to 10 kHz, and scan time to 50 ms. The spectrum analyzer establishes a TCP / IP communication connection with the industrial control computer via the Ethernet interface on the front panel. Each time the industrial control computer receives a frame of spectrum data (containing power values ​​at various frequency points on the frequency axis), the spectrum analyzer simultaneously transmits the acquisition timestamp corresponding to each frame of spectrum data to the industrial control computer via the Ethernet interface, forming a time-continuous sequence of electromagnetic radiation signals.

[0030] During dynamic load application, the thermal distribution signal of the charger PCBA is simultaneously acquired. An infrared thermal imager is mounted directly above the charger PCBA, with its infrared lens aimed at the component surface of the PCBA. The imager's focal length is adjusted to fill the imager's field of view. The preset frame rate of the infrared imager is set to 60 frames per second, the detector array resolution is 640 pixels by 480 pixels, and the internal non-uniformity correction function is enabled. The imager continuously captures infrared thermal images of the charger PCBA. Each frame is converted from analog to digital and then used to generate a temperature matrix. Each element in the temperature matrix corresponds to the temperature value at a spatial location on the charger PCBA surface. The entire temperature matrix sequence is transmitted in real-time to an industrial control computer via the imager's gigabit Ethernet interface. The spatial resolution of the temperature matrix sequence generated by the infrared thermal imager corresponds to the component layout of the charger PCBA. That is, the surface area of ​​the charger PCBA covered by each temperature matrix pixel is no larger than the size of the smallest surface mount component on the charger PCBA. By adjusting the object distance between the infrared thermal imager and the charger PCBA and the optical magnification of the infrared thermal imager, the spatial size corresponding to a single pixel is 0.2 mm by 0.2 mm.

[0031] Example 2:

[0032] In specific implementation, please refer to Figure 3 The voltage ripple signal is decomposed in the time-frequency domain to obtain the ripple energy spectrum sequence. The industrial control computer reads all voltage ripple signal waveform data from the voltage ripple signal file recorded by the high-speed data acquisition card. The voltage ripple signal waveform data is a one-dimensional discrete time series, and the sequence length is the product of the total acquisition time and the preset sampling rate. The industrial control computer segments the voltage ripple signal waveform data into non-overlapping segments according to a preset time window length. The preset time window length is set to 10 milliseconds, and the number of sampling points in each time window is the product of the preset time window length and the preset sampling rate, i.e., 1,000,000 sampling points. The segmentation operation starts from the beginning of the voltage ripple signal waveform data and sequentially extracts multiple continuous and non-overlapping time windows. The voltage ripple signal segment within each time window is treated as an independent processing unit.

[0033] Wavelet packet decomposition is performed on the voltage ripple signal segment within each time window. The wavelet packet decomposition uses the Daubechies wavelet basis function, with the vanishing moment order set to 4 (i.e., the db4 wavelet basis function is used). The decomposition level is set to 3 levels. The basis for this 3-level decomposition is that each additional level doubles the frequency resolution. The 3-level decomposition divides the frequency range of the voltage ripple signal into 8 equal-bandwidth terminal node frequency bands. The fault-related characteristic frequency components in the voltage ripple signal of the charger PCBA are mainly concentrated in several specific frequency bands in the low-frequency and mid-low-frequency ranges. The frequency band division accuracy provided by the 3-level decomposition can cover the ripple frequency distribution range corresponding to common fault modes of the charger PCBA. The wavelet packet decomposition process is as follows: The voltage ripple signal segment is taken as the level 0 node. A low-pass decomposition filter and a high-pass decomposition filter are simultaneously applied to the level 0 node to obtain the low-frequency and high-frequency node coefficients of level 1, respectively. The same low-pass and high-pass decomposition filters are then applied again to the low-frequency and high-frequency node coefficients of level 1 to obtain the four node coefficients of level 2. This process continues, with low-pass and high-pass decomposition filters applied to each node coefficient of level 2 to obtain the eight terminal node coefficients of level 3. The coefficients of the low-pass and high-pass decomposition filters are uniquely determined by the db4 wavelet basis functions.

[0034] Wavelet packet coefficients are extracted from each terminal node obtained by wavelet packet decomposition. The eight terminal nodes are numbered sequentially from terminal node 1 to terminal node 8 in ascending order of frequency. For terminal nodes... ,in The value can be an integer from 1 to 8, and the terminal node It contains a set of wavelet packet coefficients, the number of which depends on the number of decomposition levels and the number of sampling points for the voltage ripple signal segment. The sum of the squares of the wavelet packet coefficients for each terminal node is calculated as the frequency band energy value corresponding to the terminal node. The formula for calculating the frequency band energy value is: in, Indicates terminal node The corresponding frequency band energy value, This is the sequence number of the terminal node. The value of is an integer from 1 to 8; Indicates terminal node The first in Wavelet packet coefficients, For wavelet packet coefficients at terminal nodes The index number within, The value range is from 1 to Integers; Indicates terminal node The total number of wavelet packet coefficients included. The value depends on the length of the voltage ripple signal segment and the number of wavelet packet decomposition levels. Under the condition of 3-level wavelet packet decomposition... It equals the number of sampling points of the voltage ripple signal segment divided by 8.

[0035] The energy values ​​of each frequency band are arranged into a vector according to the frequency band order of the terminal nodes. Specifically, the energy values ​​of terminal node 1, terminal node 2, and so on, up to terminal node 8, are arranged into a vector containing 8 elements, each element representing a frequency band energy value. This yields the ripple energy spectrum vector corresponding to the current time window. The ripple energy spectrum vector is 8-dimensional. Each element corresponds to a terminal node. Bandwidth energy value Repeat the wavelet packet decomposition and frequency band energy value extraction operations for the voltage ripple signal segments of all time windows. Arrange the ripple energy spectrum vectors of all time windows from top to bottom according to the order in which the time windows appear, and stack them to form a two-dimensional matrix. The number of rows in the two-dimensional matrix is ​​the total number of time windows, and the number of columns in the two-dimensional matrix is ​​8. Each row of the two-dimensional matrix represents the ripple energy spectrum vector of a time window. The two-dimensional matrix as a whole constitutes a ripple energy spectrum sequence.

[0036] Example 3:

[0037] In specific implementation, please refer to Figure 4 The electromagnetic radiation signal is subjected to spectral broadening analysis to obtain the radiation characteristic vector. The industrial control computer reads all electromagnetic radiation signal frames from the electromagnetic radiation signal sequence received by the spectrum analyzer via the Ethernet interface. Each frame of the electromagnetic radiation signal consists of a sequence of power values ​​at each frequency point within a set sweep width of the spectrum analyzer. The industrial control computer divides the frequency coverage of the electromagnetic radiation signal into multiple sub-bands according to a preset frequency band range. The preset frequency band range is consistent with the sweep width of the spectrum analyzer, set to 20 MHz, with a center frequency of 150 MHz, and a frequency coverage range from 140 MHz to 160 MHz. The 140 MHz to 160 MHz frequency band is evenly divided into 10 sub-bands, each with a bandwidth of 2 MHz. The 10 sub-bands are numbered sequentially from low to high frequency as sub-band 1 to sub-band 10. The division is based on the fact that the switching frequency and harmonic components of the internal switching power supply of the charger PCBA present as discrete narrow-band peaks in the electromagnetic radiation spectrum. The 2 MHz sub-band bandwidth can completely cover the spectral broadening range of a single switching harmonic peak, while avoiding adjacent harmonic components falling into the same sub-band and causing aliasing.

[0038] For each frame of electromagnetic radiation signal, in the sub-band Internal calculation sub-band The average radiated power value within, of which This is the sub-band number. The value ranges from 1 to 10. The average radiated power value is calculated by extracting sub-bands. The sum of the radiated power values ​​at all frequencies within the frequency range, divided by the sub-band. The total number of internal frequency points. Simultaneously calculate sub-bands. The peak radiated power value within the sub-band is calculated as follows: The radiated power values ​​at all frequency points within the frequency range are taken as the peak radiated power value.

[0039] Autocorrelation calculations are performed on the radiated power value sequence within each sub-band to obtain the autocorrelation attenuation coefficient for each sub-band. The radiated power value sequence is derived from the same sub-band of electromagnetic radiation signals across multiple consecutive frames. The average radiated power values ​​are arranged in chronological order of the frames. For sub-bands The sequence of radiated power values ​​is represented as a vector. ,in The number of consecutive frames participating in the autocorrelation operation. The value is set to 50 frames. Indicates sub-band In the The average radiated power value of the frame, The value range is from 1 to The integer. The formula for calculating the autocorrelation decay coefficient is: in, Indicates sub-band The autocorrelation decay coefficient, The value of is a real number between -1 and 1. The closer the value is to 1, the more likely it is to be a sub-band. The stronger the time-domain correlation of the radiated power value sequence, The closer the value is to 0, the more likely it is to be a sub-band. The weaker the time-domain correlation of the radiation power value sequence; Indicates sub-band In the The average radiated power value of the frame; Indicates sub-band In the The average radiated power value of the frame; Indicates sub-band All of the radiation power values ​​in the sequence The arithmetic mean of the average radiant power values; This represents the total number of consecutive frames contained in the radiated power value sequence.

[0040] Each sub-band Average radiant power, peak radiant power, and autocorrelation attenuation coefficient The elements are concatenated to form a sub-frequency band feature vector. The sub-frequency band feature vector is a vector containing three elements, the first element being the sub-frequency band. The average radiated power value, the second element is the sub-band. The peak radiated power value, the third element is the sub-band autocorrelation attenuation coefficient The sub-band feature vectors of all 10 sub-bands are concatenated in the order of sub-band 1 to sub-band 10 to form a vector containing 30 elements. This vector is the radiation feature vector corresponding to the electromagnetic radiation signal of the current frame.

[0041] The heat diffusion trajectory is obtained by gradient calculation on the heat distribution signal. The industrial control computer reads the temperature matrix of all frames from the temperature matrix sequence transmitted by the infrared thermal imager. Each frame's temperature matrix is ​​a two-dimensional matrix with dimensions of 640 rows by 480 columns, and each matrix element stores the temperature value of the corresponding pixel location. Spatial gradient calculation is performed on the temperature matrix of each frame of the heat distribution signal using the central difference operator. For the temperature matrix located at the... Line number The number of pixels in the column, where The value of is an integer from 2 to 639. The value of is an integer from 2 to 479. The horizontal gradient component is calculated as follows: [The text abruptly ends here, so the translation stops.] Line number The temperature value of the column minus the first Line number The temperature values ​​in each column are calculated by dividing the difference by twice the pixel pitch, which is 0.2 mm. The vertical gradient component is calculated as follows: [The text abruptly ends here, so the translation stops as well.] Line number The temperature value of the column minus the first Line number The temperature values ​​in each column are calculated by dividing the difference by twice the pixel spacing. Pixels at the edges of the temperature matrix are processed using either forward or backward differencing. The vertical gradient component of the first row of pixels uses forward differencing, the vertical gradient component of the 640th row uses backward differencing, the horizontal gradient component of the first column uses forward differencing, and the horizontal gradient component of the 480th column uses backward differencing. Each frame's temperature matrix is ​​processed through spatial gradient operations to obtain a corresponding gradient field matrix. Each pixel in the gradient field matrix stores a pair of values: the horizontal gradient component and the vertical gradient component.

[0042] In each frame of the gradient field matrix, a preset number of pixels with the largest temperature gradient amplitude are located. The temperature gradient amplitude is calculated as follows: for each pixel in the gradient field matrix, the square root of the sum of the squares of the horizontal and vertical gradient components is taken to obtain the temperature gradient amplitude. The preset number is set to 15 pixels. This is based on the fact that the number of main heat-generating components on a charger PCBA is typically no more than 10, such as power switches, transformers, and rectifier diodes. 15 pixels can cover all key heat-generating components while reserving some redundant pixels for tracking the heat dissipation path. From all pixels in the gradient field matrix, the 15 pixels with the largest temperature gradient amplitude are selected, and their coordinate positions and gradient direction angles are recorded. The coordinate positions are represented in pixel coordinates, with the horizontal coordinate ranging from 1 to 480 and the vertical coordinate ranging from 1 to 640. The gradient direction angle is calculated as follows: for each selected pixel, the vertical gradient component is divided by the horizontal gradient component, and the arctangent of the quotient is taken to obtain the gradient direction angle in radians. The gradient direction angle has a negative range. arrive .

[0043] The pixel coordinates recorded in all frames are concatenated in chronological order to form a set of heat diffusion trajectories. For each frame's temperature matrix, a set of records containing the coordinates and gradient direction angles of 15 pixels is obtained. This is for continuously captured images by an infrared thermal imager. Frame temperature matrix, The value of is determined by the product of the total duration of dynamic load application and the preset frame rate of the infrared thermal imager, generating... Group pixel records. For pixels with adjacent coordinates and continuous gradient direction angles across all frames, perform inter-frame association. The inter-frame association method is as follows: in the first... The 15 pixels of the frame and the first Between the 15 pixels in a frame, the Euclidean distance of all pixel pairs is calculated. Pixel pairs with the smallest Euclidean distance (less than 5 pixels) are identified as the same temperature hotspot in adjacent frames. The pixel coordinates of the same temperature hotspot in all frames are then connected sequentially according to the frame's time order to form a heat diffusion trajectory. Each frame contains a maximum of 15 temperature hotspots, and the heat diffusion trajectory set contains a maximum of 15 heat diffusion trajectories. Each heat diffusion trajectory corresponds to the movement path of a temperature hotspot on the charger PCBA surface. The data structure of the heat diffusion trajectory is a coordinate sequence arranged in time order. Each element in the coordinate sequence contains the pixel coordinate position and gradient direction angle of the temperature hotspot in the temperature matrix of a frame.

[0044] Example 4:

[0045] In specific implementation, please refer to Figure 5The cross-feature mapping layer generates the first interactive feature map between voltage ripple and electromagnetic radiation through pointwise multiplication and integration operations. The cross-feature mapping layer reads the ripple energy spectrum vector and radiation feature vector from the time-aligned triplet. The ripple energy spectrum vector is an 8-dimensional vector, and the radiation feature vector is a 30-dimensional vector. Element-wise multiplication is performed on the ripple energy spectrum vector and the radiation feature vector. The condition for performing element-wise multiplication is that the dimensions of the ripple energy spectrum vector and the radiation feature vector are different. A dimension-aligned broadcasting method is used: the 8-dimensional ripple energy spectrum vector is expanded into a 30-dimensional vector by repeating each element of the ripple energy spectrum vector a certain number of times. The number of repetitions is allocated according to the grouping structure of the sub-frequency band feature vectors in the radiation feature vector. The radiation feature vector is composed of 10 sub-frequency band feature vectors, each containing 3 elements. For the Each sub-band feature vector, where The value of is an integer from 1 to 10, and the in the ripple energy spectrum vector is... The value of the nth element is assigned to the nth element. The three positions corresponding to the sub-frequency band feature vectors, among which The calculation method is as follows , This represents the floor operation, which, after expansion, yields a 30-dimensional extended ripple energy spectrum vector. Multiplying the extended ripple energy spectrum vector element-wise with the corresponding elements of the radiation eigenvector yields a 30-dimensional first product vector.

[0046] The first product vector is input into the first fully connected layer. The network structure of the first fully connected layer consists of an input layer, a hidden layer, and an output layer. The input layer contains 30 nodes, each receiving one element of the first product vector. The hidden layer contains 64 nodes, and its activation function is a modified linear unit function. The output layer contains 8 nodes, and its activation function is linear activation; the number of output nodes is set to 8, the same as the dimension of the ripple energy spectrum vector. The weight matrix parameters and bias term parameters of the first fully connected layer are learned through backpropagation during the overall training of the fault detection model. The training data is a multi-modal signal sample set of a charger PCBA labeled with fault type and fault location. After performing matrix multiplication and bias addition operations on the first product vector, the first fully connected layer outputs an 8-dimensional first intermediate vector.

[0047] The first intermediate vector is added element-wise to the original 8-dimensional ripple energy spectrum vector. The element-wise addition is performed by adding the first intermediate vector to the first intermediate vector in a specific order. The element and the ripple energy spectrum vector of the first element Adding the elements together, where The value of is an integer from 1 to 8. After element-wise addition, an 8-dimensional fusion vector is obtained. The 8-dimensional fusion vector is then reshaped into a two-dimensional matrix with 2 rows and 4 columns. The reshaping method is to fill the first to fourth elements of the fusion vector into the first row of the two-dimensional matrix in row-major order, and fill the fifth to eighth elements of the fusion vector into the second row of the two-dimensional matrix in row-major order, thus obtaining the first interactive feature map.

[0048] The cross-feature mapping layer generates a second interactive feature map between thermal diffusion and voltage ripple through bidirectional convolution operations. The cross-feature mapping layer reads thermal diffusion trajectory segments from time-aligned triples. Each thermal diffusion trajectory segment is a set of coordinate positions and gradient direction angles corresponding to all thermal diffusion trajectories within the same time window. For each trajectory point in the thermal diffusion trajectory segment, the coordinate position and gradient direction angle are encoded into a local thermal feature vector. The coordinate position includes an x-coordinate value and a y-coordinate value, where the x-coordinate value ranges from 1 to 480, and the y-coordinate value ranges from 1 to 640. The gradient direction angle ranges from... arrive The encoding method for the local thermal feature vector is as follows: The horizontal coordinate value is normalized to a normalized horizontal coordinate value by dividing the horizontal coordinate value by 480; the vertical coordinate value is normalized to a normalized vertical coordinate value by dividing the vertical coordinate value by 640; the gradient direction angle is normalized to a normalized direction angle value by dividing the gradient direction angle by 480. The normalized values ​​of the horizontal axis, the vertical axis, and the orientation angle are combined to form a 3-dimensional vector, which yields the local thermal feature vector.

[0049] The local thermal feature vectors of all trajectory points in a heat diffusion trajectory segment are arranged in order of their position within the heat diffusion trajectory to form a thermal feature sequence. The order of the trajectory points in the heat diffusion trajectory is the order of their corresponding frame timestamps from earliest to latest. When the heat diffusion trajectory set contains multiple heat diffusion trajectories, the local thermal feature vector sequences of each heat diffusion trajectory are concatenated end-to-end according to the trajectory numbering order to form a complete thermal feature sequence. The thermal feature sequence is a two-dimensional tensor, with the first dimension being the sequence length and the second dimension being 3.

[0050] The forward convolution result is obtained by performing a convolution operation on the hot feature sequence using a one-dimensional convolution kernel along the forward direction. The length of the one-dimensional convolution kernel is set to 5, the number of one-dimensional convolution kernels is set to 8, and the stride of the one-dimensional convolution kernel along the sequence length direction is set to 1. Zero padding is used to ensure that the length of the convolution output sequence is consistent with the length of the input sequence. The forward direction is defined as the direction of movement from the first local hot feature vector to the last local hot feature vector in the hot feature sequence. The forward convolution operation is as follows: the one-dimensional convolution kernel starts from the beginning of the hot feature sequence, covering 5 consecutive local hot feature vectors each time. A weighted summation operation is performed on all elements within the covered area, with the weighting coefficients being the weight parameters of the one-dimensional convolution kernel. After the weighted summation, a bias term is added to obtain the output value of the forward convolution result at the current position. Simultaneously, the same one-dimensional convolution kernel is used to perform a convolution operation on the hot feature sequence in the reverse direction to obtain the backward convolution result. The reverse direction is defined as the direction of movement from the last local hot feature vector to the first local hot feature vector in the hot feature sequence. The process of reverse convolution is the same as that of forward convolution, except that the starting position of the one-dimensional convolution kernel on the hot feature sequence is the end of the sequence, and the direction of movement is opposite to that of forward convolution.

[0051] Both the forward and backward convolution results are two-dimensional tensors of the same size. The first dimension of the two-dimensional tensor is the sequence length, and the second dimension is 8. The forward and backward convolution results are then averaged element-wise. The element-wise averaging method is as follows: for position indices... ,in The index value is the sequence length. The index value is the channel dimension index value, which is used to store the results of the forward convolution. The value at that point is the same as the result of the inverse convolution. Add the values ​​at each position and divide by 2 to get the element-wise average. The value at that location.

[0052] The element-wise averaging result is input into the second fully connected layer. The second fully connected layer's network structure consists of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer equals the total number of elements after flattening the element-wise averaging result; the flattening operation expands the two-dimensional tensor into a one-dimensional vector in row-major order. The hidden layer contains 64 nodes, and its activation function is a modified linear unit function. The output layer contains 8 nodes, and its activation function is linear activation; the number of output nodes is set to 8, the same as the dimension of the ripple energy spectrum vector. The weight matrix parameters and bias term parameters of the second fully connected layer are learned through backpropagation during the overall training of the fault detection model. The training data is the same labeled dataset used by the first fully connected layer. The second fully connected layer outputs an 8-dimensional second intermediate vector. This 8-dimensional second intermediate vector is reshaped into a two-dimensional matrix, using the same reshaping method as the first interactive feature map, with 2 rows and 4 columns, resulting in the second interactive feature map.

[0053] See Figure 6 The figure shows the trends of the forward convolution result, the backward convolution result, and the element-wise averaging result obtained by one-dimensional convolution operation between the heat diffusion trajectory and the voltage ripple signal, as a function of the sequence index. The horizontal axis represents the sequence index, ranging from 0 to 500, and the vertical axis represents the feature value, ranging from 0 to 1.1. The three curves represent the forward convolution result (solid line), the backward convolution result (dashed line), and the element-wise averaging result (dotted line), respectively.

[0054] From the trend of the curves, within the sequence index range of 0 to 190, the eigenvalues ​​of all three curves remain at a low level close to zero, indicating that the interaction between the thermal characteristic sequence and the voltage ripple signal is weak during this period. Between approximately 190 and 390, the eigenvalues ​​rapidly rise to a high level close to 1.0 and remain stable within this range, indicating that the convolutional interaction between thermal diffusion and the voltage ripple signal is significantly enhanced in this range, and the fusion feature exhibits a strong response. Between approximately 390 and 490, the eigenvalues ​​show a significant decrease to a moderate level of approximately 0.5, indicating that the interaction feature strength is weaker than in the previous stage. Between approximately 490 and 500, the eigenvalues ​​further decrease, gradually falling back to a low level close to zero, indicating that the interaction feature gradually disappears.

[0055] The curves of the forward and backward convolution results are basically the same, and the fluctuation range of their feature values ​​in each interval is similar. This indicates that bidirectional convolution effectively captures information in both directions of the hot feature sequence. The element-wise averaging result smooths out the subtle differences between the forward and backward convolution results and provides a stable interactive feature representation.

[0056] The changing trend of the convolutional interactive feature map reflects the strength of the interaction between the heat diffusion trajectory and the voltage ripple signal in the time series, which is consistent with the processing flow and feature expression method of generating the second interactive feature map by bidirectional convolution operation of the cross feature mapping layer in Example 4.

[0057] Example 5:

[0058] In practice, before inputting the ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory into the cross feature mapping layer, a time alignment operation is performed on the ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory.

[0059] After the industrial control computer generates the ripple energy spectrum sequence, each ripple energy spectrum vector in the sequence corresponds to a time window. Each time window has a unique start and end time for data acquisition, with the start time serving as the acquisition time corresponding to the ripple energy spectrum vector. Using the time sampling points of the ripple energy spectrum sequence as the reference time axis, each time sampling point is defined as a sequence formed by arranging the acquisition times of all ripple energy spectrum vectors in the sequence in chronological order. The method for finding the acquisition time corresponding to each ripple energy spectrum vector is as follows: For the ripple energy spectrum vector in the sequence... ripple energy spectrum vectors, where The value ranges from 1 to the total number of time windows. integers, The value is the total duration of the voltage ripple signal divided by the preset time window length and then rounded down. The reading is the first... The first ripple energy spectrum vector corresponding to the first The start time of the acquisition in the first time window will be the first... The start time of the acquisition in the first time window is taken as the first... The acquisition time corresponding to each ripple energy spectrum vector is denoted as . .

[0060] During the extraction of radiation feature vectors, the radiation feature vectors are generated frame-by-frame from the multi-frame electromagnetic radiation signal output by the spectrum analyzer. Each frame of the radiation feature vector has a corresponding acquisition timestamp, which is recorded by the spectrum analyzer when it outputs the electromagnetic radiation signal frame. The acquisition timestamps of all radiation feature vectors are arranged in chronological order to form a sequence of acquisition timestamps for the radiation feature vectors. The acquisition timestamps corresponding to the ripple energy spectrum vector acquisition time are then extracted from this sequence. The nearest neighbor radiative feature vector is extracted by sequentially comparing each acquisition timestamp in the acquisition timestamp sequence of the radiative feature vector with... Find the absolute value of the difference, identify the acquisition timestamp with the smallest absolute value of the difference, and determine the radiation feature vector corresponding to that acquisition timestamp as the one with the smallest absolute value of the difference. The time-aligned radiative feature vector of the ripple energy spectrum vector is obtained. The extracted radiative feature vector is then compared with the first... A time-aligned pair is formed by ripple energy spectrum vectors. The data structure of the time-aligned pair is a tuple containing an 8-dimensional ripple energy spectrum vector and a 30-dimensional radiation eigenvector.

[0061] When extracting thermal diffusion trajectory segments, the frame timestamps of the thermal diffusion trajectory are recorded by the infrared thermal imager when generating each frame of the temperature matrix. Each frame of the temperature matrix corresponds to a frame timestamp, and each trajectory point in the thermal diffusion trajectory belongs to a frame of the temperature matrix. Therefore, each trajectory point inherits the frame timestamp of its respective temperature matrix. The acquisition time of the ripple energy spectrum vector is extracted from the frame timestamps of the thermal diffusion trajectory. The extraction method for thermal diffusion trajectory segments of the same frame is as follows: sequentially compare the timestamps of all frames from the infrared thermal imager with... Compare when the frame timestamp equals At that time, the frame timestamp is equal to The set of all trajectory points corresponding to the temperature matrix frame in the thermal diffusion trajectory set is determined as the set of all trajectory points corresponding to the first temperature matrix frame. A time-aligned thermal diffusion trajectory segment of a ripple energy spectrum vector. When there is no frame timestamp exactly equal to... When the frame is in time, look up the frame timestamp in the temperature matrix sequence that matches... The frame timestamp with the smallest absolute difference is used to determine the set of all trajectory points corresponding to the temperature matrix frame in the thermal diffusion trajectory set, which is then used as the thermal diffusion trajectory segment. This thermal diffusion trajectory segment, along with the ripple energy spectrum vector and the radiation characteristic sub-vector, forms a time-aligned triplet. The data structure of the time-aligned triplet is a triplet, where the first element is the... The ripple energy spectrum vector, the second element is related to the ripple energy spectrum vector. The third element is the radiation eigenvector paired with the ripple energy spectrum vector. The thermal diffusion trajectory segment is a pair of ripple energy spectrum vectors.

[0062] For all of the ripple energy spectrum sequences Perform the above time alignment operation on each ripple energy spectrum vector to obtain All time-aligned triples are arranged in chronological order of the ripple energy spectrum vector and used as input data for the cross-feature mapping layer.

[0063] During the application phase of the fault diagnosis network, the multimodal fusion feature tensor is input into the network. The multimodal fusion feature tensor is a three-dimensional tensor, where the first dimension is the index dimension of the time-aligned triples, the second dimension is the height dimension, and the third dimension is the width dimension. The fault diagnosis network contains a first branch and a second branch connected in parallel, each receiving the exact same multimodal fusion feature tensor independently.

[0064] The first branch of the fault detection network comprises a first convolutional layer, a first pooling layer, and a first classification layer connected in sequence. The first convolutional layer contains 32 convolutional kernels, each 3x3 in size. The kernels have a stride of 1 in both height and width dimensions, and are zero-padding to ensure the output size matches the input size in both dimensions. The activation function of the first convolutional layer is a modified linear unit function (MRJ). The first pooling layer uses max pooling with a 2x2 pooling window, and the stride of the pooling window is set to 2 in both height and width dimensions. The weights of the first convolutional and pooling layers are learned through backpropagation during the overall training of the fault detection model. The output feature map of the first pooling layer is flattened into a one-dimensional vector and input to the first classification layer. The first classification layer contains a fully connected hidden layer and a softmax output layer. The fully connected hidden layer contains 128 nodes, and its activation function is a modified linear unit function (MRJ). The softmax output layer contains... 1 node The value is equal to the total number of preset charger PCBA fault types. The value is 6, and the 6 preset charger PCBA fault types are: switching transistor short circuit fault, switching transistor open circuit fault, transformer inter-turn short circuit fault, output capacitor aging fault, rectifier diode breakdown fault, and PCBA trace solder joint fault. The Softmax output layer outputs a fault type probability distribution vector, which is a... The fault type probability distribution vector is a 3D vector. Each element in the vector corresponds to a preset charger PCBA fault type, and the sum of all elements in the vector is 1. The fault type corresponding to the highest probability value in the vector is determined as the fault type of the charger PCBA.

[0065] The second branch of the fault detection network consists of a second convolutional layer, a second pooling layer, and a location regression layer connected sequentially. The second convolutional layer contains 64 convolutional kernels, each 5x5 in size. The kernels have a stride of 1 in both height and width dimensions, and are zero-padding to ensure the output size matches the input size in both dimensions. The activation function for the second convolutional layer is a modified linear unit function (MRJ). The second pooling layer uses average pooling with a 2x2 pooling window. The pooling window has a stride of 2 in both height and width dimensions. The weights of the second convolutional and pooling layers are learned through backpropagation during the overall training of the fault detection model. The output feature map of the second pooling layer is flattened into a one-dimensional vector and input to the location regression layer. The location regression layer contains two fully connected hidden layers: the first with 256 nodes and the second with 64 nodes. Both fully connected hidden layers use the MRJ. The output layer of the location regression layer contains two nodes and uses linear activation. The location regression layer outputs a fault location coordinate vector, a 2D vector. The first element of this vector is the x-coordinate of the fault point on the charger PCBA, in pixels. This x-coordinate is denormalized to correspond to the actual x-coordinate position on the charger PCBA surface by multiplying the x-coordinate by the ratio of the actual horizontal dimension of the charger PCBA to the horizontal pixel count of the infrared thermal imager's temperature matrix. The second element is the y-coordinate of the fault point on the charger PCBA, also in pixels. This y-coordinate is denormalized to correspond to the actual vertical dimension of the charger PCBA surface by multiplying the vertical pixel count of the infrared thermal imager's temperature matrix. The coordinates determined by this fault location vector are then identified as the fault location on the charger PCBA.

[0066] The weights of the first and second branches in the fault detection network are obtained through joint training using a labeled dataset. Each sample in the labeled dataset contains a set of time-aligned triples and a corresponding label. The label includes a fault type label and a fault location coordinate label, with the fault type label being one-hot encoded. The fault location coordinates are labeled as normalized 2D coordinate vectors. The joint training loss function is defined as a weighted sum of the classification cross-entropy loss of the first branch and the Euclidean distance loss of the second branch. The weight coefficient for the classification cross-entropy loss is set to 1.0, and the weight coefficient for the Euclidean distance loss is set to 0.5. The weight coefficient of 0.5 is determined by trying different values ​​in the range of 0.1 to 2.0 with a step size of 0.1 during training using a grid search to select the weight coefficient value that minimizes the average regression error of the fault location on the validation set. The training optimizer uses the Adam algorithm, with an initial learning rate of 0.001, which decays to one-tenth of its current value every 50 training epochs. The batch size is set to 32, and the training epochs are set to 200. During training, the total loss value on the validation set is monitored. Early stopping is triggered when the total loss value on the validation set does not decrease for 15 consecutive training epochs. The model parameters with the minimum total loss value on the validation set are saved as the weight parameters of the final fault discrimination network.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A charger PCBA fault detection method based on multimodal feature fusion, characterized in that, include: A dynamic load is applied to the charger PCBA under test, and the voltage ripple signal, electromagnetic radiation signal and thermal distribution signal of the charger PCBA are collected simultaneously during the application of the dynamic load. The voltage ripple signal is decomposed in the time-frequency domain to obtain the ripple energy spectrum sequence; the electromagnetic radiation signal is subjected to spectral broadening analysis to obtain the radiation feature vector; and the thermal distribution signal is calculated by gradient to obtain the thermal diffusion trajectory. The ripple energy spectrum sequence, radiation feature vector, and thermal diffusion trajectory are input into the cross feature mapping layer. The cross feature mapping layer generates the first interactive feature map between voltage ripple and electromagnetic radiation through point-by-point multiplication and integration operations, and at the same time generates the second interactive feature map between thermal diffusion and voltage ripple through bidirectional convolution operations. After performing channel-by-channel concatenation on the first and second interactive feature maps, the data is input into the gated cyclic fusion unit, which outputs a multimodal fusion feature tensor. The multimodal fusion feature tensor is input into the fault discrimination network, which outputs the fault type and fault location of the charger PCBA.

2. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of applying a dynamic load to the charger PCBA to be tested includes: Connect the output terminal of the charger PCBA to the programmable electronic load. The programmable electronic load outputs multiple load current values ​​in sequence according to the preset current step curve. The duration of each load current value is a preset duration. A transient load transition is inserted between two adjacent load current values, and the rise and fall times of the transient load transition are both controlled in the microsecond range.

3. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of simultaneously acquiring the voltage ripple signal, electromagnetic radiation signal, and thermal distribution signal of the charger PCBA during the application of dynamic load includes: The differential probe is connected to the positive and negative output terminals of the charger PCBA. The output terminal of the differential probe is connected to the first channel of the high-speed data acquisition card. The high-speed data acquisition card records the voltage ripple signal at a preset sampling rate. Place the near-field electromagnetic probe at a preset distance above the charger PCBA. Connect the output of the near-field electromagnetic probe to the spectrum analyzer. The spectrum analyzer outputs the electromagnetic radiation signal in real time through the Ethernet interface. An infrared thermal imager is used to target the component surface of the charger PCBA. The infrared thermal imager continuously outputs a temperature matrix sequence of heat distribution signals at a preset frame rate. The spatial resolution of the temperature matrix sequence corresponds to the component layout of the charger PCBA.

4. The charger PCBA fault detection method based on multimodal feature fusion according to claim 3, characterized in that, The near-field electromagnetic probe is a magnetic field probe or an electric field probe, and the preset distance is 1 mm to 5 mm.

5. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of performing time-frequency domain decomposition on the voltage ripple signal to obtain the ripple energy spectrum sequence includes: The voltage ripple signal is segmented into non-overlapping segments according to a preset time window length. Wavelet packet decomposition is performed on the voltage ripple signal within each time window. The number of decomposition levels for wavelet packet decomposition is a preset number. Wavelet packet coefficients are extracted from each terminal node obtained by wavelet packet decomposition, and the sum of the squares of the wavelet packet coefficients of each terminal node is calculated as the frequency band energy value corresponding to that terminal node. Arrange the energy values ​​of each frequency band into a vector form according to the frequency band order of the terminal nodes to obtain the ripple energy spectrum vector corresponding to each time window. Stack the ripple energy spectrum vectors of all time windows in chronological order to form a ripple energy spectrum sequence.

6. The charger PCBA fault detection method based on multimodal feature fusion according to claim 5, characterized in that, The wavelet packet decomposition uses the Daubechies wavelet basis function, and the decomposition level is 3 or 4.

7. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The steps for obtaining the radiation feature vector by performing spectral broadening analysis on the electromagnetic radiation signal include: The electromagnetic radiation signal is divided into multiple sub-bands according to a preset frequency range, and the average radiation power value and peak radiation power value in each sub-band are calculated. Autocorrelation calculation is performed on the radiated power value sequence within each sub-band to obtain the autocorrelation attenuation coefficient of each sub-band; The average radiated power, peak radiated power, and autocorrelation attenuation coefficient of each sub-band are concatenated to form a sub-band feature vector. All the feature vectors of the sub-bands are then concatenated in band order to form a radiative feature vector.

8. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of calculating the gradient of the heat distribution signal to obtain the heat diffusion trajectory includes: Spatial gradient operation is performed on the temperature matrix of each frame of the heat distribution signal to obtain the gradient field matrix corresponding to each frame of the temperature matrix. The gradient field matrix contains horizontal gradient components and vertical gradient components. In each frame's gradient field matrix, locate a preset number of pixels with the largest temperature gradient magnitude, and record the coordinates and gradient direction angles of these pixels. The pixel coordinates recorded in all frames are connected in chronological order to form a set of heat diffusion trajectories. Each heat diffusion trajectory corresponds to the movement path of a temperature hotspot on the surface of the charger PCBA.

9. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of generating the first interactive feature map between voltage ripple and electromagnetic radiation through point-by-point multiplication and integration operations in the cross-feature mapping layer includes: Perform element-wise multiplication between each ripple energy spectrum vector in the ripple energy spectrum sequence and the corresponding radiation eigenvector to obtain the first product vector; The first product vector is input into the first fully connected layer. The number of output nodes of the first fully connected layer is the same as the dimension of the ripple energy spectrum vector. The first fully connected layer outputs the first intermediate vector. After performing element-wise addition on the first intermediate vector and the ripple energy spectrum vector, it is reshaped into a two-dimensional matrix form to obtain the first interactive feature map.

10. The charger PCBA fault detection method based on multimodal feature fusion according to claim 1, characterized in that, The step of generating the second interactive feature map between thermal diffusion and voltage ripple through bidirectional convolution operation includes: The coordinate position and gradient direction angle of each trajectory point in the heat diffusion trajectory are encoded as local thermal feature vectors, and the local thermal feature vectors are arranged in the order of the trajectory points in the heat diffusion trajectory to form a thermal feature sequence. A forward convolution result is obtained by performing a convolution operation on the thermal feature sequence using a one-dimensional convolution kernel in the forward direction, and a reverse convolution result is obtained by performing a convolution operation on the thermal feature sequence using the same one-dimensional convolution kernel in the reverse direction. The results of forward and backward convolutions are averaged element-wise and then input into the second fully connected layer. The number of output nodes of the second fully connected layer is the same as the dimension of the ripple energy spectrum vector. The output of the second fully connected layer is reshaped into a two-dimensional matrix to obtain the second interactive feature map.

Images