High-frequency connector contact reliability evaluation method based on multi-source vibration signals

Abstract

The application discloses a high-frequency connector contact reliability evaluation method based on multi-source vibration signals and relates to the field of artificial intelligence.The method comprises the following steps: adopting a multi-scale transient impact and slow modulation separation enhancement strategy to extract and nonlinearly reconstruct two types of physical features;an adaptive multi-scale feature extraction network composed of a deformable receptive field module, a multi-period gate cycle module, a three-axis coupling interaction module and a deep feature fusion module guided by period information is constructed;and a degradation order constraint and a category center adaptive adjustment mechanism are introduced in the classifier training process.The method can accurately capture weak vibration features in the degradation process of the connector and effectively improve the contact state recognition precision and physical consistency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for evaluating the contact reliability of high-frequency connectors based on multi-source vibration signals. Background Technology

[0002] High-frequency connectors, as key components for signal interconnection and transmission in various electronic devices, are widely used in fields with stringent reliability requirements, such as aerospace, rail transportation, communication base stations, and new energy vehicles. During service, connectors are often subjected to complex environmental stresses, including continuous mechanical vibration, temperature cycling, and impact loads. These stresses lead to progressive degradation phenomena at the connector's internal contact interfaces, such as fretting wear, contact pressure relaxation, material fatigue, and surface oxidation. This manifests as fluctuations in contact resistance, deterioration in signal transmission quality, and may ultimately lead to contact failure, causing malfunctions or even catastrophic accidents in the entire system. Therefore, accurately assessing the contact state of high-frequency connectors under vibration environments is crucial for ensuring equipment reliability and implementing predictive maintenance.

[0003] The main drawbacks of the existing technology are as follows: Existing methods for assessing the reliability of high-frequency connector contacts typically employ filters with fixed parameters to preprocess vibration signals. This makes it difficult to simultaneously account for both microsecond-level transient impacts and millisecond-level or longer-term slow-varying modulation characteristics. Consequently, the spike impacts corresponding to early contact anomalies are easily filtered out, while the slow-varying envelope reflecting changes in contact pressure cannot be effectively preserved.

[0004] Existing deep learning-based vibration signal analysis methods mostly use conventional convolutional kernels with fixed sampling positions, which makes it difficult to accurately capture sparsely distributed transient impact positions. At the same time, single-scale recurrent neural networks or convolutional networks cannot simultaneously model the slowly varying modulation patterns at different time scales, and they usually use independent processing methods for triaxial vibration signals, ignoring the synchronous response information generated by structural coupling between multi-axis channels.

[0005] Existing classification models typically use only cross-entropy as the loss function, requiring only that the classification result be correct. They do not consider the monotonic degradation sequence relationship that exists in the process of connector contact state from normal to failure, which may lead to unreasonable prediction results where the order of "severe degradation" and "slight degradation" is reversed, which contradicts the physical mechanism.

[0006] Existing methods, when the training sample class distribution is unbalanced, have the majority class samples dominate the loss function, resulting in the minority class samples being scattered in the feature space and the classification boundary being blurred. This leads to a significantly low accuracy in identifying degraded state samples, making it difficult to meet the needs of high reliability monitoring of failure states in engineering practice. Summary of the Invention

[0007] To address the technical problems in the prior art, this invention provides a method for evaluating the contact reliability of high-frequency connectors based on multi-source vibration signals.

[0008] This invention is achieved through the following technical solution: A method for evaluating the contact reliability of high-frequency connectors based on multi-source vibration signals, comprising: S1. Data acquisition; including synchronously acquiring vibration response signals of the connector in the X, Y and Z axes using a triaxial accelerometer; S2, High-frequency connector vibration signal enhancement; including multi-scale transient impact and slowly modulated separation enhancement, specifically including transient impact component extraction, slowly modulated component extraction, channel correlation-guided differential enhancement, and nonlinear reconstruction based on contact degradation mechanism; S3. Construct an adaptive multi-scale feature extraction network, including a deformable receptive field extraction module, a multi-period gated cyclic modeling module, a three-axis coupled interactive modeling module, and a periodic information-guided deep feature fusion module. S4. Classifier training based on physical constraints of connector contact state; training is performed by introducing degenerate order constraint training and class center adaptive adjustment mechanism; S5. High-frequency connector contact reliability assessment: The trained improved deep neural network model is deployed in the high-frequency connector contact reliability assessment system to perform online or offline assessment of newly acquired unlabeled vibration signals.

[0009] Furthermore, the channel correlation-guided differential enhancement includes performing differential enhancement on the transient impact component and the slowly modulated component respectively using the cross-correlation relationship between axial channels.

[0010] Furthermore, the differential enhancement specifically includes calculating normalized cross-correlation coefficients among the three axial channels for both transient impact components and slowly modulated components, generating transient impact enhancement weights and slowly modulated enhancement weights for each axial channel based on the cross-correlation results, and dynamically adjusting the contribution of the channel feature based on the correlation between the current axial channel and other channels.

[0011] Furthermore, the transient impact enhancement weights and the slowly varying modulation enhancement weights are calculated in the following form: ; ; in, Indicates the first Transient impact enhancement weights for each axial channel; Indicates the first Gradual modulation enhancement weights for each axial channel; Indicates the transient impact strength coefficient; Indicates the enhancement intensity coefficient of the gradual modulation. Indicates the first The first axial channel and the first The correlation of each axial channel in transient impact components Characterizing the first The first axial channel and the first The correlation of each axial channel with the slowly modulated component.

[0012] Furthermore, the nonlinear reconstruction based on the contact degradation mechanism includes performing amplitude compression and range constraints on the enhanced slowly modulated component, and using the enhanced slowly modulated component to perform dynamic modulation on the enhanced transient impact component, as expressed below: ; in, Characterizing the first The enhanced feature signal is ultimately output by each axial channel; Indicates the first Transient impact components enhanced by axial channels; This indicates point-by-point multiplication; Indicates the fusion strength coefficient; Indicates the first The gradually modulated component is enhanced by axial channel enhancement.

[0013] Furthermore, the three-axis coupled interactive modeling module includes stitching the deformable receptive field feature maps of the three axial channels along the feature channel dimensions to obtain a deformable receptive field joint feature map, constructing three parallel projection branches for the deformable receptive field joint feature map, each projection branch performing linear projection on the channel vector at each time position, and performing point-by-point multiplication on the outputs of the three projection branches at the same time position and the same channel index to obtain a three-axis coupled interactive feature map.

[0014] Furthermore, the periodic information-guided deep feature fusion includes performing a linear mapping on the periodic feature direction of each axial channel, inputting the periodically modulated features into the channel mapping layer to obtain the depth feature map of each axial channel, and stitching the depth feature maps of each axial channel along the feature channel dimension to obtain the depth feature map.

[0015] Furthermore, the classifier training includes inputting a global statistical feature vector into the contact state classification layer to obtain a contact state prediction score.

[0016] Furthermore, the degenerate order constraint training introduces a degenerate ranking score, and only applies order constraints to this degenerate ranking score. When the degenerate ranking scores of mild and severe samples already satisfy the order relationship and the interval is large enough, the sample pair no longer generates loss; when the order is reversed or the interval is insufficient, the sample pair generates a positive loss value, thereby prompting the model to readjust the degenerate ranking score.

[0017] Furthermore, the adaptive adjustment mechanism for category centers calculates the mean of the global statistical feature vectors of each category sample in each training batch and updates the corresponding category center vector using an exponential moving average method. When a category does not exist in the current training batch, the category center vector remains unchanged. The update method for the category center vector is expressed as follows: ; in, Indicates the updated number Each category center vector; Indicates the number before the update Each category center vector; Indicates the momentum coefficient; Indicates the number of times the current training batch is executed. The mean of the global statistical feature vectors of samples of each category.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In response to the simultaneous presence of two physical characteristics in the vibration signal of high-frequency connectors—microsecond-level transient impact and millisecond-level or longer-term gradual modulation—this invention proposes a separation and enhancement strategy that combines adaptive morphological filtering with Hilbert envelope low-pass filtering. Furthermore, the transient impact component is nonlinearly dynamically modulated by the gradually modulated component, making the enhanced characteristic signal more closely resemble the actual physical degradation mechanism of "contact pressure change modulating fretting wear impact."

[0019] 2. To address the three technical challenges of capturing the sparse distribution of transient impacts in vibration signals, the existence of multiple periodic scales in slowly varying modulations, and the coupling relationship between the three-axis channels, this invention constructs an adaptive multi-scale feature extraction network consisting of a deformable receptive field module, a multi-period gated loop module, and a three-axis coupling interaction module. This enables the convolutional kernel sampling points to actively focus on the impact location, while simultaneously capturing the periodic patterns at different time scales and highlighting the coupling mode of multiple axes being active at the same time.

[0020] 3. To address the issue that the physical constraints of connector contact states have a clear degradation sequence that conventional classification losses cannot reflect, this invention introduces degradation ranking scores and ranking constraint losses during classifier training. This forces the ranking scores output by the model to be consistent with the monotonic degradation sequence of normal state, slight degradation, severe degradation, and contact failure, making the classification results not only accurate but also in line with the irreversible physical development law.

[0021] 4. To address the class imbalance problem caused by the fact that there are far more normal state samples than degraded state samples in actual engineering, this invention maintains a dynamically updated class center vector for each contact state category in the global statistical feature space, and combines intra-class compactness and inter-class separation constraints to enable minority class samples to form compact clusters in the feature space, effectively improving the model's ability to identify degraded states. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of a high-frequency connector contact reliability assessment method based on multi-source vibration signals according to an embodiment of this application; Figure 2 shows the original vibration signals according to an embodiment of this application, wherein Figure 2(a) is the original vibration signal along the X-axis, Figure 2(b) is the original vibration signal along the Y-axis, and Figure 2(c) is the original vibration signal along the Z-axis. Figure 3(a) shows the comparison between the original X-axis signal and the extracted transient impact component, and Figure 3(b) shows the extracted transient impact component along the X-axis. Figure 4(a) shows the slowly modulated components of the original X-axis signal and the Hilbert envelope low-pass filter, and Figure 4(b) shows the slowly modulated envelope of the X-axis. Figure 5 shows the enhanced feature signal after nonlinear reconstruction according to an embodiment of this application, where 5(a) is the transient impulse component after X-axis enhancement, Figure 5(b) is the variation of the modulation coefficient, and Figure 5(c) is the enhanced feature signal after X-axis reconstruction. Detailed Implementation

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] See Figure 1 A method for evaluating the contact reliability of high-frequency connectors based on multi-source vibration signals includes the following steps: S1, Data Acquisition; A high-frequency connector vibration test platform was built, and the high-frequency connector under test was installed on the vibration table. A vibration load spectrum that can simulate actual working conditions was set. During the vibration test, the vibration response signals of the connector in the X-axis, Y-axis and Z-axis directions were collected synchronously by a triaxial accelerometer. The sampling frequency was set to 100 kHz to fully capture the transient impact components caused by fretting wear and instantaneous poor contact at the contact interface of the high-frequency connector, as well as the slowly modulated components caused by changes in contact pressure and material fatigue.

[0027] During the data acquisition process, the vibration stress level applied to the vibration table is gradually increased according to the pre-designed accelerated test profile, causing the connector to undergo a complete degradation process from normal state to contact failure.

[0028] To construct a labeled sample set for subsequent model training, this invention employs a high-frequency connector dynamic contact resistance monitoring system as an auxiliary measurement method. During vibration testing, the system synchronously records the contact resistance fluctuations of the connector. Based on the trend of contact resistance changes and failure criteria, each 0.1-second raw triaxial vibration signal sample is divided into four contact state categories: normal state, where the contact resistance is stable at an initial low value with no significant fluctuations; slight contact degradation, where the contact resistance exhibits intermittent small fluctuations but does not exceed the threshold; severe contact degradation, where the contact resistance frequently exhibits large fluctuations and approaches the failure threshold; and contact failure, where the contact resistance continuously exceeds the failure threshold or the signal is completely interrupted.

[0029] During the labeling process, the experimental system automatically assigns a corresponding state label to each vibration sample based on the synchronously acquired contact resistance data, ensuring that the sample label is strictly aligned with the physical degradation process of the contact. For each type of contact state, a sufficient number of samples are collected to cover different vibration stress levels, different degradation stages, and differences between individual connectors, ultimately forming a labeled dataset for subsequent model training. Each sample in this dataset contains a triaxial raw vibration signal matrix and its corresponding contact state category label.

[0030] S2, Enhanced separation of multi-scale transient impact and slow-varying modulation of high-frequency connector vibration signals; The degradation of contact reliability of high-frequency connectors under vibration environment usually manifests as both short-term sudden impact response and slow modulation response over a longer time range. The former mainly corresponds to fretting wear, instantaneous poor contact or local jump at the contact interface, while the latter mainly corresponds to periodic envelope changes caused by contact pressure changes, material fatigue and structural coupling. Conventional fixed parameter filtering methods can only highlight one type of feature and cannot simultaneously take into account microsecond-level impact and millisecond-level, or even longer-term slow-change features.

[0031] This invention performs transient impact component extraction, slowly modulated component extraction, channel correlation-guided enhancement, and nonlinear reconstruction on the original triaxial vibration signal samples to obtain enhanced feature signals that are strongly correlated with connector contact reliability. These enhanced feature signals retain both the spike impact information from the early stages of contact failure and the low-frequency envelope information caused by changes in contact pressure. The specific steps are as follows: S201. Construction and Axial Decomposition of Original Triaxial Vibration Signal Samples 1> Acquire triaxial vibration sampling data of the high-frequency connector during vibration testing to form raw triaxial vibration signal samples. , The original triaxial vibration signal matrix representing a single sample has a size of [size missing]. ,in, The first dimension represents the axial channel, and the second dimension represents the number of time sampling points; In one implementation, the sampling frequency is 100kHz and the duration of a single sample is 0.1s, so each axial channel contains 10,000 sampling points.

[0032] Furthermore, the original triaxial vibration signal samples The axial channel is split into three one-dimensional original vibration signals, denoted as follows: , Characterizing the first The original one-dimensional vibration signal of each axial channel has a length of 10000, where... This represents the axial channel index, with values ​​of 1, 2, and 3, corresponding to the X-axis, Y-axis, and Z-axis, respectively.

[0033] In one embodiment, as shown in Figures 2(a), 2(b), and 2(c), the original vibration acceleration signals of the X, Y, and Z axes within a time window of 0 to 0.02 seconds are displayed respectively. The vertical axis represents the vibration acceleration amplitude in units of gravitational acceleration (g), and the horizontal axis represents time in units of seconds (s). In the figures, it can be observed that there are obvious transient impact spikes and slow amplitude fluctuations caused by contact pressure fluctuations in the X-axis signal. The Y-axis is somewhat similar to the X-axis, while the Z-axis is relatively weaker.

[0034] 2> Perform DC component removal and amplitude normalization processing on the one-dimensional raw vibration signal of each axial channel to reduce the impact of sensor zero drift and dimensional differences between different test batches on subsequent feature extraction; In practical implementation, for the one-dimensional original vibration signal of each axial channel, the mean of the axial channel signal can be subtracted first, and then divided by the standard deviation. When the standard deviation is too small, a small stability constant can be set to avoid division by zero.

[0035] S202, Extraction of transient impact components In the early stages of contact degradation in high-frequency connectors, the first manifestation is usually a sudden jump in local contact points. This is reflected in the vibration signal as a transient impact component with a very short duration and a prominent peak. This type of component often only occupies a small number of sampling points and is easily averaged out if conventional smoothing methods are used directly.

[0036] This invention utilizes adaptive morphological filtering to separate transient impact components from each axial channel. The specific steps are as follows: 1> One-dimensional original vibration signal for each axial channel Construct multiple sets of one-dimensional structuring elements at different scales. The structuring elements are used to describe the local comparison range of morphological filtering. In one implementation, the length of the structuring element can be 3, 5, 7, 9, 11, or 15 sampling points, and the shape of the structuring element can be a one-dimensional flat-top linear structuring element.

[0037] Furthermore, for each candidate structuring element, respectively... Perform open-closed combined filtering and closed-open combined filtering. In open-closed combined filtering, the opening operation is performed first to suppress isolated spikes, and then the closing operation is performed to restore the overall baseline. In closed-open combined filtering, the closing operation is performed first to suppress isolated depressions, and then the opening operation is performed to eliminate local spikes.

[0038] Furthermore, the two filtering results are averaged to obtain a transient response estimate that is relatively stable for both forward and reverse shocks.

[0039] 2> Calculate the impulse pulse index for the filtering results corresponding to each candidate structural element, and select the structural element with the largest impulse pulse index as the optimal structural element for the current axial channel. The impulse pulse index measures whether the peak value is sufficiently prominent while the overall energy is not excessively dispersed, and is defined as the ratio of the peak amplitude to the root mean square value. The larger this indicator is, the better the structural element can highlight short-term shocks without excessively preserving background fluctuations; in, Indicates the first The impact pulse index of each axial channel under the current structural element. This represents the absolute peak value in the filtering result. This represents the root mean square value of the filtering result.

[0040] 3> Re-engineer using the selected optimal structuring element Perform open-closed combined filtering and closed-open combined filtering, and average the two results to obtain the first result. Transient impact components of each axial channel , Characterizing the first The transient impact component signal extracted from each axial channel has a length of 10,000, reflecting the short-term abrupt changes caused by fretting wear, contact jump, and local poor contact at the connector contact interface.

[0041] It should be noted that the impact duration varies for different connectors, operating conditions, and degradation stages. Therefore, this invention does not use fixed-scale structural elements, but rather adaptively selects the optimal structural elements based on the impact pulse index in the current sample. This makes it easier for the transient impact components to match the actual contact anomaly duration, thereby improving the sensitivity of feature extraction to early failure signs.

[0042] In one embodiment, as shown in Figure 3(a), the original X-axis signal and the transient impact component extracted by adaptive morphological filtering are compared (magnified 2 times). As shown in Figure 3(b), the extracted impact component is given separately, with the horizontal axis representing time (s) and the vertical axis representing amplitude. The transient impact response with extremely short duration and prominent peak value can be effectively separated by adaptive scale morphological filtering, which reflects the sensitivity of this method to early contact anomalies (such as fretting wear and instantaneous poor contact), while the background vibration and slow-motion components in the original signal are significantly suppressed.

[0043] S203, Extraction of Slow-Modifying Components During continuous vibration, connectors will also exhibit slow-varying envelope characteristics caused by contact pressure fluctuations, elastic recovery, structural coupling, and material fatigue. These characteristics usually do not manifest as sharp peaks, but rather as amplitudes that fluctuate slowly over time. If only transient impact components are considered, the trend of continuous deterioration of the connector's condition may be easily missed.

[0044] This invention extracts slowly varying modulation components from each axial channel, thereby capturing long-period trends in slowly varying degradation processes such as connector contact pressure changes, material fatigue, and structural coupling. This provides time-scale-independent envelope information for subsequent multi-scale feature fusion. The specific steps are as follows: 1> Using the one-dimensional original vibration signal of each axial channel As input, an analytical signal is constructed to extract the instantaneous amplitude envelope; In one implementation, first... Perform a Hilbert transform to obtain components orthogonal to the original signal. Then combine the original components and the orthogonal components to calculate the envelope amplitude at each sampling point.

[0045] Furthermore, a low-pass filter is applied to the obtained envelope amplitude to remove high-frequency jitter and retain only the low-frequency modulation information related to contact pressure changes, contact material fatigue, and slow-changing structural response. In one implementation, the low-pass filter is a fourth-order Butterworth low-pass filter with a cutoff frequency of 500 Hz. This frequency range can cover typical contact pressure variations and low-frequency modulation envelopes of the connector, while suppressing high-frequency impact residue.

[0046] 2> The result after low-pass filtering is denoted as , Characterizing the first The slowly modulated component signal extracted from each axial channel has a length of 10,000, reflecting the slowly varying envelope and periodic modulation trend of the connector contact interface over a longer time scale.

[0047] Furthermore, in order to maintain numerical stability during subsequent enhancement processes, for Amplitude normalization is performed, which can be done by scaling according to the maximum absolute value of the axial channel, or by using quantile scaling to suppress the influence of extreme values, so that the slowly modulated components enter a stable numerical range.

[0048] In one embodiment, taking the X-axis signal as an example, Figure 4(a) shows the original X-axis signal and the slowly varying modulation components after Hilbert envelope low-pass filtering superimposed, and Figure 4(b) shows the slowly varying modulation envelope separately. The horizontal axis is time (s) and the vertical axis is the envelope amplitude (g). Through Hilbert transform and low-pass filtering, the slowly varying trend caused by contact pressure fluctuations, material fatigue, etc. can be successfully extracted, which reflects the ability of the slowly varying modulation components to capture the continuous deterioration process of the connector state.

[0049] S204, Channel Correlation-Guided Differentiation Enhancement In triaxial vibration signals, a certain type of contact anomaly often does not only appear in a single axial channel, but also exhibits a certain degree of synchronicity or coupling in multiple axial channels. Random noise usually lacks this cross-axis consistency.

[0050] This invention utilizes the cross-correlation relationship between axial channels to perform differentiated enhancement on transient impact components and slowly modulated components, with the specific steps as follows: 1> Specifically targeting transient impact components and gradual modulation components The normalized cross-correlation coefficient is calculated among the three axial channels. The normalized cross-correlation coefficient is used to measure the degree of synchronization of different axial channels within a certain time delay range. In practice, the mean value of each axial channel signal is subtracted and divided by the standard deviation. Then, the cross-correlation value is calculated by sliding within a given delay window, and the maximum value is taken as the cross-correlation result between the two axial channels (i.e., the normalized cross-correlation coefficient).

[0051] Furthermore, the cross-correlation results of the transient impact components are denoted as... The cross-correlation results of the slowly modulated components are denoted as ,in, Characterizing the first The first axial channel and the first The correlation of each axial channel in transient impact components Characterizing the first The first axial channel and the first Correlation of individual axial channels on slowly modulated components; In one implementation, the delay search range of the transient impact component can be taken as... For each sampling point, the delay search range of the slowly modulated component can be taken as... 10 sampling points were used to match the different time scales of the two types of components.

[0052] 2> Based on the cross-correlation results, transient impact enhancement weights and slowly varying modulation enhancement weights are generated for each axial channel, which are used to dynamically adjust the contribution of the channel's features according to the correlation between the current axial channel and other channels; In one implementation, it can take the following form: ; ; in, Indicates the first Transient impact enhancement weights for each axial channel; Indicates the first Gradual modulation enhancement weights for each axial channel; This represents the transient impact enhancement strength coefficient, with a preferred value of 0.3; This represents the intensity coefficient of the gradual modulation, with a preferred value of 0.2.

[0053] It should be noted that when the features of an axial channel (transient impact or gradual modulation) are highly correlated with those of other channels, it indicates that the features of that channel are more likely to be caused by multiaxial coupling response due to connector contact abnormalities rather than local random noise, and therefore are given a greater enhancement weight; conversely, if the correlation is low, a smaller weight is given. Through this differentiated weighting, isolated noise can be effectively suppressed, multiaxial consistency response strongly correlated with contact reliability can be highlighted, and the robustness of subsequent feature extraction can be improved.

[0054] Furthermore, and Multiplying them together yields the enhanced transient impact component. Then, and Multiplying them yields the enhanced slowly modulated components. ; in, Characterizing the first Transient impact components enhanced by axial channels; Characterizing the first The gradually modulated component is enhanced by axial channel enhancement.

[0055] It should be noted that this invention does not simply average the three axial channels, but rather utilizes the physical fact that "multiple simultaneous occurrences are more likely to be real contact anomalies, while isolated occurrences on a single axis are more likely to be local noise" to differentiate the weighting of different axial channels, thereby highlighting the real contact anomaly response in the model input and suppressing isolated noise.

[0056] S205, Nonlinear reconstruction based on contact degradation mechanism 1> Enhanced slowly modulated components Perform amplitude compression and range constraints to make it suitable for subsequent nonlinear modulation; In the specific implementation, a normalization process is adopted, firstly... Normalization to The range is then determined by inputting the hyperbolic tangent function to avoid excessive local amplitude leading to overly strong modulation.

[0057] Furthermore, the enhanced transient impulse component is dynamically modulated using the enhanced slowly varying modulation component (which has already been normalized) to obtain the first... Enhanced feature signals of each axial channel ; In one implementation, it can take the following form: ; in, Characterizing the first The enhanced feature signal output by each axial channel has a length of 10000. This indicates point-by-point multiplication; This represents the fusion strength coefficient, with a preferred value of 2; due to the hyperbolic tangent function The output range is ,therefore The output range is This allows the slowly modulated components to amplify or suppress transient impact components in a controlled manner.

[0058] 2> Enhance the characteristic signals of the three axial channels , and The enhanced feature signal is obtained by splicing along the axial channel dimension. , Characterizing the preprocessed triaxial enhancement feature matrix, with size [missing information]. It contains both transient impact information and slowly modulated information, and enhances the physical response that is closely related to the reliability of connector contacts.

[0059] It should be noted that the fretting wear and transient contact anomalies at the connector contact interface do not occur uniformly over the entire time axis. Instead, they often exhibit the phenomenon that "impact is more easily triggered in certain envelope stages" as the contact pressure changes. This invention modulates the transient impact component by using a slowly varying modulation component, which can more closely approximate the actual physical process of connector contact degradation, thereby improving the ability of subsequent models to express the degradation law.

[0060] In one embodiment, taking the X-axis as an example, the enhanced feature signal after nonlinear reconstruction is analyzed. Figure 5(a) shows the enhanced transient impact component (after weighting), Figure 5(b) shows the change of the modulation coefficient, and Figure 5(c) shows the reconstructed signal obtained by multiplying the two point by point. The horizontal axis is time (s), and the vertical axis is the impact amplitude (g), the modulation coefficient (dimensionless), and the reconstructed amplitude (g) in sequence. The experimental results show that by dynamically modulating the transient impact with the slowly varying modulation component, the impact is amplified in some stages of the envelope and suppressed in other stages.

[0061] S3. Adaptive multi-scale feature extraction network for vibration signals oriented towards connector contact conditions. High-frequency connector vibration signals present three challenges: First, transient impact features are extremely sparse, and convolution kernels with fixed sampling positions are prone to missing the impact location; second, the slowly varying modulation features have multiple periodic scales, and it is difficult for a fixed receptive field to cover different periods simultaneously; third, there is a clear coupling relationship between the three axial channels, and performing independent feature extraction will result in the loss of cross-axis information.

[0062] This invention constructs an adaptive multi-scale feature extraction network, which includes deformable receptive field extraction, multi-period gated cyclic modeling, triaxial coupling interaction, and periodic information guided fusion. This network can simultaneously complete sparse impact localization, multi-scale periodic modeling, and triaxial coupling relationship modeling, and output a deep feature map with strong discriminative ability. The adaptive multi-scale feature extraction network constructed in this invention comprises four core modules: First, through the deformable receptive field extraction module, the enhanced feature signal of each axial channel learns an adaptive sampling offset, enabling the convolutional kernel to actively focus on sparse transient impact positions, thus solving the problem of missing impacts at fixed sampling positions. Secondly, through the multi-period gated recurrent modeling module, multiple periodic branches of different lengths are set in parallel. Each branch first performs segmented compression of the feature map at the corresponding scale, then uses the gated recurrent unit to learn the long-range temporal dependency at that scale, and finally uses the attention mechanism to adaptively fuse the outputs of each branch, thereby capturing the slow-varying modulation patterns of multiple periods at the same time. Then, through the three-axis coupled interactive modeling module, the deformable receptive field feature maps of the three axial channels are stitched together in the channel dimension. After passing through three independent projection branches, the projection vectors at the same time position are multiplied point by point to highlight the coupling mode of multiple axes being active at the same time and enhance the model's ability to identify real contact anomalies. Finally, the periodic feature fusion module guided by periodic information maps the periodic feature vector of each axial channel as the modulation coefficient, adjusts the intensity of the triaxial coupled interactive feature map and retains the residual connection, so that the periodic information can guide the feature expression, and finally outputs a deep feature map with strong discriminative ability.

[0063] The specific steps are as follows: S301, Deformable receptive field extraction of sparse transient impact 1> Enhance the feature signal The input is split into three paths based on the axial channel, with each path corresponding to the enhanced feature signal of one axial channel. , The enhanced feature signal, with a length of 10000, is obtained after the a-th axial channel is separated, enhanced, and reconstructed by transient impact and slow modulation.

[0064] Furthermore, an offset prediction branch is established for each axial channel to generate adaptive sampling positions for subsequent deformable convolutions; In one implementation, the offset prediction branch can adopt a two-layer one-dimensional convolution structure. The first layer of one-dimensional convolution can have 16 output channels, a kernel length of 5, and a stride of 1, followed by a nonlinear activation function. The second layer of one-dimensional convolution has the same number of output channels as the number of sampling points, a kernel length of 3, and a stride of 1, and is used to output the offset corresponding to each sampling point. To avoid the sampling position from deviating from the effective signal area due to excessive offset, a hyperbolic tangent constraint can be added at the output end and multiplied by the maximum offset. Preferably, the maximum offset can be 4 sampling points.

[0065] 3> Construct a one-dimensional deformable convolution, since the input is the enhanced feature signal of the current axial channel. Since it is a one-dimensional time-series signal, the convolution kernel also adopts a one-dimensional sampling method; In one implementation, the deformable convolution uses 9 regular sampling points, corresponding to fixed offset positions relative to the current time step. Furthermore, for the first The current time position of each axial channel is first taken as 9 regular sampling positions, and then the 9 adaptive offsets output by the offset prediction branch are superimposed to obtain 9 continuous sampling positions.

[0066] It should be noted that when the continuous sampling positions are not integer sampling points, one-dimensional linear interpolation is used to obtain the corresponding signal value. Specifically, the interpolation result at the position is obtained by weighting the two adjacent integer sampling points on the left and right sides of the continuous sampling position according to the distance. Then, the interpolation results of the nine sampling positions are weighted and summed to obtain the deformable convolution output.

[0067] Furthermore, the first The result of deformable convolution on each axial channel is denoted as , Characterizing the first Deformable receptive field feature map of each axial channel, with a size of ,in, The number of feature channels can be 32 or 64. The purpose of the deformable receptive field feature map is to adaptively adjust the position of the receptive field according to the local morphology of the input signal, so as to capture sparse transient impacts more accurately.

[0068] It should be noted that this invention does not simply increase the number of convolutional layers, but rather makes the sampling position of the convolutional kernel change with the local features of the input signal. When there is a local spike impact in a certain time series, the sampling points will actively gather near the impact. When the signal is relatively flat, the sampling points remain relatively stable, thereby significantly improving the network's ability to locate short-term abrupt changes.

[0069] S302, Multi-period Slow Modulation Modeling 1> Deformable receptive field feature map for each axial channel Construct multiple parallel periodic branches, each corresponding to a preset period length, and let the number of periodic branches be . ;

[0070] In one implementation, Four sampling points were selected for four cycle lengths: 50, 100, 200, and 500, respectively, corresponding to the modulation patterns at different time scales.

[0071] Furthermore, regarding the first Each periodic branch will Along the time dimension by length The window is segmented, and the feature values ​​of the same feature channel are averaged within each window to obtain the resampled compressed sequence. , Characterizing the first The axial channel in the first A periodically compressed sequence at a period scale, approximately [length missing] If the last window is not a complete window, you can either use zero padding and then calculate the average, or you can simply discard the last incomplete window.

[0072] 2> Compress the sequence for each period Input the corresponding gated recurrent unit, learn the temporal dependencies under that periodic scale, the gated recurrent unit reads each time step in the compressed sequence step by step and updates the hidden state, thereby extracting the slow modulation pattern under that periodic scale; In one implementation, each periodic branch corresponds to a single-level gated loop unit with independent parameters. The hidden state dimension can be 32. The gated loop unit outputs a hidden state at each time step, and these hidden states constitute a hidden state sequence.

[0073] 3> Perform convergence on the hidden state sequence output by each periodic branch to obtain the periodic sub-features of that branch; In one implementation, the hidden state of the last time step can be directly taken as the periodic sub-feature of the branch, or average pooling or attention pooling can be performed on all hidden states to enhance the stable representation capability of long sequences.

[0074] 4> The periodic sub-features of the four periodic branches are concatenated and then input into the attention fusion layer to obtain the first... Periodic eigenvectors of each axial channel , Characterizing the first The comprehensive periodic feature vector of each axial channel at multiple periodic scales, with a dimension of up to 64; The role of the attention fusion layer is to automatically assign weights to different periodic branches based on the actual modulation patterns of the current sample, enabling the network to adaptively select a more discriminative periodic scale. In practice, the attention fusion layer calculates a scalar score for each periodic sub-feature, then normalizes it using Softmax to obtain the weights of the four branches. Finally, it performs a weighted summation on the four periodic sub-features and maps them through a fully connected layer to obtain the periodic feature vector. .

[0075] In the specific implementation, the periodic sub-feature vector output by each periodic branch is mapped to a real value through a learnable fully connected layer (or a linear projection layer). This real value reflects the importance of the periodic branch in the current sample. Then, these real values ​​are input into the Softmax function for normalization to obtain four weight coefficients that sum to 1, thereby achieving adaptive weighting of different periodic branches.

[0076] S303, Triaxial Coupled Interactive Modeling 1> Deformable receptive field feature maps of the three axial channels , and By concatenating along the feature channel dimension, a deformable receptive field joint feature map is obtained. , The joint feature matrix, representing the fusion of local features from three axial channels, has a size of [size missing]. .

[0077] Furthermore, regarding the joint feature map of deformable receptive fields... Three parallel projection branches are constructed, and each projection branch performs a linear projection on the channel vector at each time position; In one implementation, three independent One-dimensional convolutional layers implement three linear projections, and the number of output channels for each projection branch can be chosen as follows: For example, 64 can be used, and each projection branch output is followed by a ReLU activation function.

[0078] 2> Perform point-by-point multiplication of the outputs of the three projection branches at the same time position and the same channel index to obtain the three-axis coupled interactive feature map. , The feature map, representing the nonlinear coupling relationship of the three axial channels, has a size of [size missing]. Since point-by-point multiplication amplifies the position where "all three features are simultaneously significant", it is more conducive to highlighting the contact anomaly pattern that occurs consistently across axes.

[0079] In practical implementation, to avoid excessively large feature values ​​after point-by-point multiplication, a three-axis coupled interactive feature map can be obtained. Batch normalization, layer normalization, or amplitude clipping operations are then added to ensure numerical stability during subsequent fusion processes.

[0080] It should be noted that high-frequency connector contact anomalies are not necessarily released in only one direction. Many real anomalies are transmitted to multiple axial channels through the connector structure. If each axial channel is treated independently, the model will find it difficult to capture this cross-axial coupling relationship. By projecting three paths and then multiplying them point by point, the coupling mode of multiple axes being active at the same time can be highlighted, thereby enhancing the model's ability to identify real contact anomalies.

[0081] S304, Deep Feature Fusion Guided by Periodic Information 1> Periodic eigenvectors for each axial channel Perform a linear mapping to obtain an interactive feature map coupled with the three axes. Modulation vectors with the same number of channels; In one implementation, a single fully connected layer is used to... Mapped to a length of The vector is then compressed to a finite range using the hyperbolic tangent function.

[0082] Furthermore, the mapped modulation vector is broadcast and copied along the time dimension to 10,000 time positions to form a periodic modulation matrix. Then, the periodic modulation matrix is ​​coupled with the triaxial interactive feature map. Perform pointwise multiplication and preserve residual paths, so that periodic information modulates the intensity of triaxial coupling interaction features rather than completely covering them; In one implementation, the modulation coefficients can be formed using the method of "1 + hyperbolic tangent mapping result", that is, the mapped modulation vector is passed through the hyperbolic tangent function. Compress it to limit its value range to... Within the range, then add a constant 1, so that the range of the modulation coefficient becomes Its function is to generate a modulation coefficient centered at 1 and fluctuating between 0 and 2, thereby allowing the intensity of periodic information to be adjusted while preserving the original characteristics, rather than being completely covered.

[0083] 2> Map the periodically modulated feature input channels to a layer to obtain the first... Depth feature map of each axial channel , Characterizing the first The final depth feature map output by each axial channel has a size of ,in, This represents the final number of feature channels, for example, a value of 64; and the channel mapping layer can use... One-dimensional convolution is implemented.

[0084] 3> Depth feature maps of the three axial channels , and By concatenating along the feature channel dimensions, a depth feature map is obtained. , The final depth feature matrix, representing the contact state classification, has a size of [size missing]. ;when At that time, depth feature map The size is .

[0085] S4. Classifier training based on physical constraints of connector contact states Connector contact reliability assessment is not a typical unordered classification problem, but an ordered classification problem with a clear degradation sequence. There is a physically monotonic degradation relationship between normal state, slight contact degradation, severe contact degradation and contact failure. At the same time, in actual samples, there are usually more samples in the normal state and fewer samples in the degraded state, which can easily lead to class imbalance.

[0086] In the classifier training phase, this invention introduces a degradation order constraint loss and a class center adaptive adjustment mechanism, enabling the model to possess both conventional classification accuracy and the ability to learn ordered discriminative boundaries that conform to the contact degradation mechanism. The specific steps are as follows: S401, Global Statistical Feature Extraction from Deep Feature Maps For deep feature maps Process each feature channel individually; specifically, set up a depth feature map. The total number of feature channels is For each feature channel, four statistics were calculated at 10,000 time sampling points: mean, standard deviation, peak-to-peak value, and kurtosis. The mean is used to characterize the average activation level of the feature channel over the entire sample time range; the standard deviation is used to characterize the fluctuation strength of the feature channel; the peak-to-peak value is used to characterize the maximum change amplitude; and the kurtosis is used to characterize the sharpness of the distribution and tail characteristics, and is particularly sensitive to whether short-term shocks are prominent. In practical implementation, kurtosis can be calculated by dividing the fourth central moment by the fourth power of the standard deviation; when the standard deviation is too small, a stability constant can be set to avoid numerical anomalies.

[0087] Furthermore, the four statistical measures corresponding to each feature channel are concatenated sequentially to form a global statistical feature vector. , The representation is the global statistical feature vector extracted from the deep feature map, with dimension . ,when hour, The dimension is 768.

[0088] It should be noted that the differences in connector contact status are reflected not only in whether the local transient response is sharp, but also in whether the overall energy, dispersion and peak structure of the entire signal have undergone systematic changes. By combining four types of statistical measures, namely mean, standard deviation, peak-to-peak value and kurtosis, we can simultaneously reflect the overall trend and local extreme value characteristics.

[0089] S402, Calculation of Contact State Prediction Score global statistical feature vector Input the contact state classification layer to obtain the contact state prediction score. , The vector representing the prediction results of the current sample for each contact state category, with dimension 1. ,in, This indicates the total number of contact status categories. For example, if the value is 4, the four contact status categories are defined as normal state, slight contact degradation, severe contact degradation, and contact failure, respectively. In one implementation, the contact state classification layer is implemented using a single fully connected layer, whose input is a global statistical feature vector. The output consists of four real-valued scores, which represent the relative tendency of the current sample to be in a normal state, slightly degraded, severely degraded, or failed.

[0090] It should be noted that during the reasoning phase, scores can be predicted for the contact states. Perform Softmax normalization and select the category with the highest probability as the final contact state classification result.

[0091] S403, Degeneracy Sequence Constraint Training The conventional cross-entropy loss only requires correct classification and does not guarantee the physical order of "mild degradation should be ranked before severe degradation". Directly imposing global monotonic constraints on the scores of the four mutually exclusive categories can easily lead to training instability.

[0092] This invention further introduces a degenerate sorting score and applies order constraints only to this degenerate sorting score. The specific steps are as follows: 1> In addition to the contact state classification layer, add a degenerate sorting output node to process the global statistical feature vector. Mapping to scalar , The degeneration ranking score represents the current sample; the larger the value, the more severe the degeneration.

[0093] In one implementation, an additional independent fully connected layer is added after the fully connected layer of the model. The input of this fully connected layer is the global statistical feature vector. (Dimension 768), output dimension 1 (i.e., scalar) This fully connected layer does not share weights and works in parallel with the contact state classification layer.

[0094] 2> Within each training batch, construct sample pairs with different true class labels. Specifically, Let the sample The true category is lighter than the sample Then the sample is required Degenerate sorting score Smaller than the sample Degenerate sorting score To achieve this constraint while maintaining a certain interval, the following loss calculation method can be used: ; in, This represents the loss due to degradation order constraints. This represents the set of sample pairs in the current training batch where "samples with lighter true class values ​​come first, followed by samples with heavier true class values". Indicates the number of sample pairs; and Representing samples respectively and samples Degenerate sorting score; and Representing samples respectively and samples The true category label; This represents the basic sorting interval coefficient, with a preferred value of 0.1.

[0095] It should be noted that when the degradation ranking scores of mild and severe samples already satisfy the order relationship and the interval is large enough, the sample pair no longer generates loss; when the order is reversed or the interval is insufficient, the sample pair generates a positive loss value, thereby prompting the model to readjust the degradation ranking scores.

[0096] It should also be noted that by using degradation ranking scores for order constraints, the model learns not only "which category it belongs to" but also "where the degree of degradation is located overall," thus better conforming to the physical law of the gradual deterioration and irreversible development of the connector contact state.

[0097] S404, Category Center Adaptive Adjustment 1> In the global statistical feature space, maintain a class center vector for each contact state category. ,in, This represents the category index, with values ​​of 1, 2, 3, and 4, corresponding to normal condition, slight contact degradation, severe contact degradation, and contact failure, respectively. Indicates the first The category center vector of each category in the global statistical feature space, and its dimension are related to the global statistical feature vector. Consistency (i.e.) ).

[0098] 2> Within each training batch, the mean of the global statistical feature vectors of each class of samples in the current batch is calculated, and the corresponding class center vector is updated using an exponential moving average method. When a class does not exist in the current training batch, the class center vector remains unchanged. The update method for the class center vector is expressed as follows: ; in, Indicates the updated number Each category center vector; Indicates the number before the update Each category center vector; This represents the momentum coefficient, with a preferred value of 0.95. Indicates the number of times the current training batch is executed. The mean of the global statistical feature vectors of samples of each category.

[0099] 3> Define the class center constraint loss, which consists of two parts: intra-class compaction term and inter-class separation term. The intra-class compaction term is used to push the samples closer to the class center vector of their own class; the inter-class separation term is used to keep the class center vectors of different classes sufficiently separated, which can improve the clustering of minority class samples and avoid excessive overlap of different classes in the statistical feature space. In practical implementation, the squared Euclidean distance between each sample and the class center vector of the true class can be calculated, and a minimum distance threshold can be set between the class center vectors of different classes. Preferably, the inter-class separation strength coefficient can be 0.1, and the minimum allowable squared distance between the class center vectors can be 1.0.

[0100] It should be noted that when there are more normal state samples and fewer degenerate state samples, minority class samples are prone to being scattered, leading to unstable classification boundaries. By dynamically maintaining the class center vector and constraining intra-class contraction and inter-class separation, the separability of degenerate state samples can be enhanced, and the model's ability to identify minority classes can be improved.

[0101] S405, Joint Loss Optimization and Model Acquisition 1> The overall loss function is composed of cross-entropy loss, degenerate order constraint loss, and class center constraint loss. The overall loss function is expressed as: ; in, Represents the overall loss function; This represents the cross-entropy loss, used to ensure basic classification accuracy; This represents the degeneracy order constraint loss, used to guarantee the degeneracy ordering relationship; This represents the class center constraint loss, used to improve intra-class compactness and inter-class separation. The weighting coefficient representing the degradation order constraint loss is preferably set to 0.5; The weighting coefficient representing the class center constraint loss is preferably set to 0.2.

[0102] 2> The Adam optimizer is used to jointly train all trainable parameters in the adaptive multi-scale feature extraction network (S3), the contact state classification layer (S402), and the degenerate ranking output node (S403); In one implementation, the initial learning rate can be chosen as... The batch size can be 32 or 64, and the learning rate can be decayed during training based on the performance of the validation set.

[0103] 3> In each round of training, the original triaxial vibration signal samples are first... The enhanced feature signal is obtained after S2. Then, the depth feature map is obtained through S3. Then, through S4, the global statistical feature vector is obtained. Contact state prediction score Degenerate sorting score Then, based on the overall loss function Perform backpropagation to update network parameters and class center vectors.

[0104] 4. Classification accuracy and degradation order consistency on the validation set are used as criteria for model convergence. Degradation order consistency measures whether the degradation ranking scores of samples from different true classes satisfy the correct order relationship. Training is stopped when classification accuracy no longer improves and degradation order consistency tends to stabilize.

[0105] 5> Retaining the optimal model parameters from the validation set, an improved deep neural network model for high-frequency connector contact reliability assessment is obtained. The input to this model is the original triaxial vibration signal samples. The output is a classification of four contact states: normal state, slight contact degradation, severe contact degradation, and contact failure.

[0106] S5, High-Frequency Connector Contact Reliability Assessment After completing the joint training of the adaptive multi-scale feature extraction network based on multi-source vibration signals and the classifier with physical constraints, the trained improved deep neural network model is deployed in the high-frequency connector contact reliability assessment system for online or offline evaluation of newly acquired unlabeled vibration signals.

[0107] During the evaluation phase, triaxial vibration signals of the high-frequency connector to be evaluated under vibration conditions were first collected in a manner completely consistent with the training data. The sampling frequency was also set to 100 kHz, and each sample was truncated to an evaluation sample with a duration of 0.1 seconds, resulting in the original triaxial vibration signal matrix. Then, the evaluation sample was input into the trained model. The model first performed transient impact and slowly varying modulation separation and enhancement processing on the evaluation sample according to the method described in step S2. Specifically, this included splitting the triaxial signal along the axis, extracting the transient impact component using adaptive morphological filtering, extracting the slowly varying modulation component based on Hilbert transform and low-pass filtering, performing differential enhancement using the cross-correlation relationship between axial channels, and nonlinear reconstruction based on the contact degradation mechanism, ultimately obtaining an enhanced feature signal that can simultaneously retain transient impact information and slowly varying modulation information. Then, the model performs deep feature extraction on the enhanced feature signal using the adaptive multi-scale feature extraction network constructed in step S3. It sequentially uses a deformable receptive field extraction module to actively focus on sparse transient impact locations, a multi-period gated cyclic modeling module to capture slowly varying modulation patterns at different time scales, and a three-axis coupled interactive modeling module to highlight the simultaneous active coupling modes across multiple axes. Finally, a deep feature fusion module guided by periodic information outputs a deep feature map with strong discriminative capabilities. Next, following the method in step S4, the model calculates four statistical measures along the time dimension for the deep feature map: mean, standard deviation, peak-to-peak value, and kurtosis. These are concatenated to form a global statistical feature vector, which is then input into the contact state classification layer and the degradation ranking output node to obtain predicted scores for the four contact state categories and a ranking score representing the degree of degradation. Finally, the system performs Softmax normalization on the four predicted scores, using the category with the highest probability as the contact state evaluation result for the current sample, outputting one of the following: normal state, slight contact degradation, severe contact degradation, or contact failure. Furthermore, the degradation ranking score can be used to further assist in determining the relative position of the current sample in the degradation process.

[0108] Through the above methods, the present invention can achieve automated and high-precision assessment of the contact reliability of high-frequency connectors under vibration environment, providing a quantitative basis for connector condition monitoring and life prediction.

[0109] In this embodiment, for the two physical characteristics of microsecond-level transient impact and millisecond-level or longer-term slow modulation that coexist in the vibration signal of high-frequency connectors, this invention proposes a separation and enhancement strategy that combines adaptive morphological filtering and Hilbert envelope low-pass filtering. Furthermore, the transient impact component is nonlinearly dynamically modulated by the slow modulation component, so that the enhanced feature signal is more in line with the actual physical degradation mechanism of "contact pressure change modulating micro-motion wear impact". To address the three technical challenges of capturing the sparse distribution of transient impacts in vibration signals, the existence of multiple periodic scales in slowly varying modulations, and the coupling relationship between the three-axis channels, this invention constructs an adaptive multi-scale feature extraction network consisting of a deformable receptive field module, a multi-period gated loop module, and a three-axis coupling interaction module. This network enables the convolutional kernel sampling points to actively focus on the impact location, while simultaneously capturing the periodic patterns at different time scales and highlighting the coupling mode of simultaneous activity across multiple axes. To address the issue that the physical constraints of connector contact states, which exhibit a clear degradation sequence, cannot be reflected by conventional classification losses, this invention introduces degradation ranking scores and ranking constraint losses during classifier training. This forces the ranking scores output by the model to maintain consistency with the monotonic degradation sequence of normal state, slight degradation, severe degradation, and contact failure, ensuring that the classification results are not only accurate but also conform to the irreversible physical development laws. To address the class imbalance problem caused by the fact that there are far more normal state samples than degraded state samples in actual engineering, this invention maintains a dynamically updated class center vector for each contact state category in the global statistical feature space, and combines intra-class compactness and inter-class separation constraints to enable minority class samples to form compact clusters in the feature space, effectively improving the model's ability to identify degraded states.

[0110] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for evaluating high-frequency connector contact reliability based on multi-source vibration signals, characterized in that, include: S1, Data Acquisition; This includes synchronously acquiring vibration response signals of the connector in the X, Y, and Z axes using a triaxial accelerometer; S2, High-frequency connector vibration signal enhancement; including multi-scale transient impact and slowly modulated separation enhancement, specifically including transient impact component extraction, slowly modulated component extraction, channel correlation-guided differential enhancement, and nonlinear reconstruction based on contact degradation mechanism; S3. Construct an adaptive multi-scale feature extraction network; This includes a deformable receptive field extraction module, a multi-period gated cyclic modeling module, a three-axis coupled interactive modeling module, and a deep feature fusion module guided by periodic information. S4. Classifier training based on physical constraints of connector contact state; training is performed by introducing degenerate order constraint training and class center adaptive adjustment mechanism; S5. High-frequency connector contact reliability assessment: The trained improved deep neural network model is deployed in the high-frequency connector contact reliability assessment system to perform online or offline assessment of newly acquired unlabeled vibration signals.

2. The method for evaluating the contact reliability of a high-frequency connector based on multi-source vibration signals according to claim 1, characterized in that, The channel correlation-guided differential enhancement includes performing differential enhancement on transient impact components and slowly modulated components respectively by utilizing the cross-correlation relationship between axial channels.

3. The method for high-frequency connector contact reliability evaluation based on multi-source vibration signals according to claim 2, characterized in that, The differential enhancement specifically includes calculating normalized cross-correlation coefficients among the three axial channels for both transient impact components and slowly modulated components. Based on the cross-correlation results, transient impact enhancement weights and slowly modulated enhancement weights are generated for each axial channel. The contribution of the channel feature is dynamically adjusted according to the correlation between the current axial channel and other channels.

4. The method for evaluating the contact reliability of a high-frequency connector based on multi-source vibration signals according to claim 3, characterized in that, The transient impact enhancement weight and the slowly varying modulation enhancement weight are calculated in the following form: ； ； wherein, represents a transient impulse enhancement weight for the represents a ramp modulation enhancement weight for the represents a transient impulse enhancement intensity coefficient, represents a ramp modulation enhancement intensity coefficient, represents a correlation between the represents a correlation between the ​​​​​​ 5. The method for high-frequency connector contact reliability evaluation based on multi-source vibration signals according to claim 4, characterized in that, The nonlinear reconstruction based on the contact degradation mechanism includes amplitude compression and range constraint on the enhanced slowly modulated component, and dynamic modulation of the enhanced transient impact component using the enhanced slowly modulated component, as expressed below: ； in, Characterizing the first The enhanced feature signal is ultimately output by each axial channel; Indicates the first Transient impact components enhanced by axial channels; This indicates point-by-point multiplication; Indicates the fusion strength coefficient; Indicates the first The gradually modulated component is enhanced by axial channel enhancement.

6. The high-frequency connector contact reliability assessment method based on multi-source vibration signals according to claim 1, characterized in that, The three-axis coupled interactive modeling module includes stitching the deformable receptive field feature maps of the three axial channels along the feature channel dimensions to obtain a deformable receptive field joint feature map. Three parallel projection branches are constructed for the deformable receptive field joint feature map. Each projection branch performs linear projection on the channel vector at each time position. The outputs of the three projection branches are multiplied point by point at the same time position and the same channel index to obtain the three-axis coupled interactive feature map.

7. The high-frequency connector contact reliability assessment method based on multi-source vibration signals according to claim 1, characterized in that, The periodic information-guided deep feature fusion includes performing a linear mapping on the periodic features of each axial channel, inputting the periodically modulated features into the channel mapping layer to obtain the depth feature map of each axial channel, and then stitching the depth feature maps of each axial channel along the feature channel dimension to obtain the depth feature map.

8. The method for evaluating the contact reliability of high-frequency connectors based on multi-source vibration signals according to claim 1, characterized in that, The classifier training involves inputting a global statistical feature vector into the contact state classification layer to obtain a contact state prediction score.

9. The method for evaluating the contact reliability of a high-frequency connector based on multi-source vibration signals according to claim 8, characterized in that, The degenerate order constraint training introduces a degenerate ranking score and applies order constraints only to this score. When the degenerate ranking scores of mild and severe samples already satisfy the order relationship and the interval is large enough, the sample pair no longer incurs loss. When the order is reversed or the interval is insufficient, the sample pair generates a positive loss value, thereby prompting the model to readjust the degenerate ranking score.

10. The method for evaluating the contact reliability of a high-frequency connector based on multi-source vibration signals according to claim 9, characterized in that, The adaptive adjustment mechanism for class centers calculates the mean of the global statistical feature vectors of each class sample in each training batch and updates the corresponding class center vector using an exponential moving average. When a class does not exist in the current training batch, the class center vector remains unchanged. The update method for the category center vector is expressed as follows: ； in, Indicates the updated number Each category center vector; Indicates the number before the update Each category center vector; Indicates the momentum coefficient; Indicates the number of times the current training batch is executed. The mean of the global statistical feature vectors of samples in each category.

Images