Liquid crystal display screen binding conduction detection method and system

Abstract

The application relates to the technical field of fault detection, and discloses a liquid crystal display screen binding conduction detection method and system, wherein the method comprises the following steps: acquiring resistance, current, temperature, alternating current impedance spectrum, gray scale image, micro-strain and stress time series data of a binding area, forming a multi-dimensional state characteristic sequence through temperature compensation, equivalent circuit fitting, abnormal density analysis and non-elastic strain energy accumulation; obtaining a fusion degradation index sequence by using bidirectional time sequence coding and entropy adjustment attention fusion; generating a candidate degradation trajectory group in parallel through a data-driven branch and a physical constraint branch; performing learnable gate fusion according to trajectory deviation; and determining a predicted failure time and a remaining conduction life by combining a failure threshold value. The application solves the problem that a single feature is difficult to reflect thermal, electrical and force coupling degradation, and makes multi-source state representation, double-branch prediction, trajectory correction and threshold value determination continuously connected, so that the stability and interpretability of the binding conduction life detection result are improved.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a method and system for detecting the bonding continuity of a liquid crystal display screen. Background Technology

[0002] Currently, in LCD manufacturing, the driver IC and glass substrate are bonded together using anisotropic conductive films, and the conductivity reliability of the bonding area directly determines the display lifespan and image quality. Under accelerated aging test conditions, current, temperature, and mechanical stress are typically applied to the bonding area simultaneously, while DC resistance values ​​are collected concurrently. Multi-frequency AC impedance spectroscopy, grayscale image sequence of surface temperature distribution, and micro-strain values Multi-source sensor data is used. A common processing approach in existing technologies is to extract degradation features separately for each type of sensor data. For example, the relative rate of change of DC resistance is calculated to characterize the overall degradation trend of the conductive circuit, or the impedance modulus drift at a fixed frequency is analyzed to determine the degree of interface corrosion. Then, degradation curves based on a single feature are extrapolated and a fixed threshold is set to determine failure. This partitioned and independent analysis method severs the coupling relationship between multiple physical fields and fails to incorporate synergistic degradation information such as local anomalies in the temperature field, accumulation of inelastic strain energy, and electrochemical interface reactions into a unified consideration, leading to a one-sided judgment of the true health status.

[0003] For example, under the harsh aging conditions of Joule heating, multiple degradation modes may simultaneously occur within the bonding area, including fretting fatigue of the metal conductors, oxidation of conductive particles, and propagation of microcracks in the glass substrate. DC resistance changes gradually in the early stages of degradation, only exhibiting abrupt changes when the contact area is significantly reduced. In contrast, the charge transfer resistance and dispersion index of the AC impedance spectrum show identifiable changes in the early stages of damage. Abnormal clustering indices at local hotspots in the grayscale image of surface temperature distribution can reveal micro-current congestion, while the nonlinear accumulation of micro-strain energy directly reflects the degree of material fatigue damage. Relying solely on the rate of change of DC resistance for degradation trend analysis will miss these subtle early changes, failing to provide timely warnings before the failure threshold. Especially under the combined effects of thermo-electrical-mechanical fields, degradation modes dynamically switch, and a single indicator can easily misjudge complex, progressive failures as stable operation, severely underestimating the true remaining lifespan. This ultimately leads to sudden bonding continuity failures, causing display abnormalities or even complete system failure.

[0004] Therefore, there is an urgent need for a method that can effectively integrate the above-mentioned multimodal degradation characteristics when obtaining multi-dimensional sensing data including DC resistance, AC impedance, temperature distribution, and micro-strain under accelerated aging conditions. This method would allow for the comprehensive utilization of early and weak degradation information across physical fields, even when degradation modes are diverse and data dimensions are high. This would enable accurate early identification of bonding conduction degradation and prediction of remaining lifetime, thereby significantly improving the accuracy and robustness of lifetime prediction and providing key support for online assessment and health management of the reliability of LCD bonding processes. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to propose a method for detecting the bonding continuity of a liquid crystal display screen. This method addresses the problem that existing technologies cannot fully reflect the bonding failure process through single-feature degradation trend analysis, especially when multiple modes of degradation, such as DC resistance, AC impedance, temperature, and micro-strain, coexist under accelerated aging conditions. This makes it impossible to integrate multiple physical features to accurately predict the remaining lifetime of bonding continuity failure.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for detecting the bonding conductivity of a liquid crystal display screen.

[0007] The method for detecting the bonding continuity of a liquid crystal display screen includes: Step S10: Obtain multi-source time-series detection data of the bonding area of ​​the liquid crystal display screen, and perform state feature extraction task based on the multi-source time-series detection data using temperature compensation and multi-source state feature construction mechanism, and output multi-dimensional state feature sequence; Step S20: Based on the multidimensional state feature sequence, a bidirectional temporal coding and entropy-modulated attention fusion mechanism is used to perform the degradation representation task, and output the fused degradation index sequence; Step S30: Based on the fused degradation index sequence, a data-driven branch and physical constraint branch parallel prediction mechanism is used to perform the degradation trajectory prediction task and output the candidate degradation trajectory group; Step S40: Based on the candidate degenerate trajectory group, a learnable gating fusion mechanism is used to perform the degenerate trajectory correction task and output the final predicted degenerate trajectory; Step S50: Determine the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculate the remaining conduction lifetime based on the predicted failure time, and output the remaining conduction lifetime detection result; Preferably, step S10, which involves acquiring multi-source temporal detection data of the liquid crystal display bonding area, performing state feature extraction based on the multi-source temporal detection data using a temperature compensation and multi-source state feature construction mechanism, and outputting a multi-dimensional state feature sequence, specifically includes: Step S101: According to the preset sampling time interval, acquire the DC resistance sequence, conduction current sequence, ambient temperature sequence, surface average temperature sequence, multi-frequency AC impedance spectrum sequence, surface grayscale image sequence, micro-strain sequence and stress sequence corresponding to the micro-strain sequence of the liquid crystal display bonding area to form a multi-source time-series detection dataset. Step S102: Based on the multi-source time-series detection dataset, temperature compensation is performed on the DC resistance sequence according to a preset reference temperature and a preset conductor resistance temperature coefficient to obtain a temperature-compensated resistance sequence. The temperature-compensated resistance at the first sampling moment in the temperature-compensated resistance sequence is determined as the initial temperature-compensated resistance, and the change ratio of the temperature-compensated resistance at the remaining sampling moments relative to the initial temperature-compensated resistance is calculated to obtain a resistance relative change rate sequence. Equivalent circuit fitting is performed on the multi-frequency AC impedance spectrum sequence to obtain a charge transfer resistance sequence, a constant phase element parameter sequence, and a dispersion index sequence. The characteristic relaxation time is determined based on the charge transfer resistance, constant phase element parameters, and dispersion index at the same sampling moment. The characteristic relaxation time at the first sampling moment is determined as the initial characteristic relaxation time, and the logarithmic difference between the characteristic relaxation time at the remaining sampling moments and the initial characteristic relaxation time is calculated to obtain an impedance relaxation offset sequence. Step S103: Based on the multi-source time-series detection dataset and the impedance relaxation offset sequence, determine the image abnormal density sequence according to the deviation of the local reachable density of each pixel in the surface grayscale image sequence from the maximum local reachable density in the corresponding image; fit the elastic strain range according to the correspondence between the micro-strain sequence and the stress sequence, determine the micro-strain increments that exceed the elastic strain range as inelastic strain increments, and accumulate the product of each inelastic strain increment and the corresponding stress to obtain the inelastic strain energy density sequence; perform sliding window standardization on the relative resistance change rate sequence, the impedance relaxation offset sequence, the image abnormal density sequence, and the inelastic strain energy density sequence respectively, and stitch the standardization results according to the preset feature order to output the multidimensional state feature sequence.

[0008] Preferably, step S20, which involves performing a degradation representation task based on the multidimensional state feature sequence using a bidirectional temporal coding and entropy-modulated attention fusion mechanism, and outputting a fused degradation index sequence, specifically includes: Step S201: The multidimensional state feature sequence is continuously truncated according to the preset window length and preset sliding step size, and each truncated result is arranged according to the sampling time and feature dimension to form a window feature matrix sequence; Step S202: Input the window feature matrix sequence into the bidirectional long short-term memory network encoder, extract the forward temporal correlation features and backward temporal correlation features at each sampling time, and concatenate the forward temporal correlation features and backward temporal correlation features at the same sampling time to output the hidden state sequence. Step S203: Based on the hidden state sequence, calculate the absolute difference between adjacent hidden states in each feature dimension, and normalize each absolute difference to obtain the degradation mode probability corresponding to each feature dimension; calculate the degradation mode entropy according to the degradation mode probability, convert the degradation mode entropy into a temperature regulation quantity through a hyperbolic tangent function, and determine the dynamic attention temperature parameter according to the preset baseline attention temperature parameter and the temperature regulation quantity; scale and normalize the attention score calculated by multi-head self-attention using the dynamic attention temperature parameter, perform weighted fusion on the hidden state sequence according to the normalized attention weight, and map the weighted fusion result into a single-dimensional degradation index, outputting the fused degradation index sequence.

[0009] Preferably, step S202, which involves inputting the window feature matrix sequence into a bidirectional long short-term memory network encoder, extracting forward temporal correlation features and backward temporal correlation features at each sampling time, and concatenating the forward temporal correlation features and backward temporal correlation features at the same sampling time to output the hidden state sequence, specifically includes: The feature matrices of each window are sequentially input into the feedforward long short-term memory network layer in the order of increasing sampling time to obtain the forward hidden state corresponding to each sampling time. The same window feature matrix is ​​input into the backward long short-term memory network layer in decreasing order of sampling time to obtain the backward hidden state corresponding to each sampling time. The forward hidden state and the backward hidden state are concatenated at the same sampling time, and the concatenation result is normalized to form the hidden state sequence.

[0010] Preferably, step S30, which involves performing the degradation trajectory prediction task based on the fused degradation index sequence using a data-driven branching and physical constraint branching parallel prediction mechanism, and outputting a candidate degradation trajectory group, specifically includes: Step S301: Input the fused degradation index sequence into the time-series decoder, and recursively decode the fused degradation index sequence according to the increasing direction of the prediction time to obtain the data-driven prediction trajectory; Step S302: Determine the prediction interval according to the preset prediction step number, and perform linear trend extrapolation on the conduction current sequence, the ambient temperature sequence, the DC resistance sequence, and the micro-strain sequence respectively to obtain the predicted conduction current sequence, the predicted ambient temperature sequence, the predicted DC resistance sequence, and the predicted micro-strain sequence, and calculate the equivalent junction temperature of the binding region at each prediction time according to the following formula:

[0011] The normalized rate of change of damage state at each prediction time is calculated using the following formula:

[0012] In the formula, For the predicted time; For the predicted time The equivalent junction temperature of the binding region; For the predicted time Predicted ambient temperature; For the predicted time The predicted on-current; For the predicted time Predicted DC resistance; The preset equivalent heat transfer coefficient; The preset effective heat exchange area; For the predicted time Normalized damage state; Preset damage evolution pre-factor; Preset activation energy; Boltzmann's constant; For the predicted time Predicting microstrain; The preset micro-strain influence index; Using a preset initial damage state as the initial value for integration, the normalized damage state change rate is recursively integrated to obtain a normalized damage state sequence within the prediction interval. The normalized damage state sequence is then converted into a physical constraint prediction trajectory based on the monotonic mapping relationship between the preset damage state and the degradation index. Step S303: Align the data-driven prediction trajectory and the physical constraint prediction trajectory according to the same prediction time, and form the candidate degradation trajectory group by aligning the data-driven prediction trajectory and the physical constraint prediction trajectory.

[0013] Preferably, step S40, which involves performing a degradation trajectory correction task based on the candidate degradation trajectory group using a learnable gating fusion mechanism and outputting the final predicted degradation trajectory, specifically includes: Step S401: Read the data-driven prediction value and the physical constraint prediction value at the same prediction time from the candidate degraded trajectory group, and calculate the absolute difference between the data-driven prediction value and the physical constraint prediction value to obtain the trajectory deviation value; Step S402: Calculate the gating coefficients for each prediction time according to the following formula:

[0014] In the formula, For the predicted time The gating coefficient; For the Sigmoid function; For learnable gated weight matrix; For learnable gated bias; For the predicted time Data-driven forecasts; For the predicted time The predicted values ​​of physical constraints; For the predicted time The trajectory deviation value; the semicolon in square brackets indicates vector concatenation; Step S403: Gated fusion of the data-driven prediction value and the physical constraint prediction value according to the following formula:

[0015] In the formula, For the predicted time The correction degradation index is calculated; the correction degradation indexes are arranged in ascending order according to the prediction time, and the final predicted degradation trajectory is output.

[0016] Preferably, step S50, which involves determining the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculating the remaining conduction lifetime based on the predicted failure time, and outputting the remaining conduction lifetime detection result, specifically includes: Step S501: Read the current detection time, and compare each corrected degradation index in the final predicted degradation trajectory with a preset failure threshold in the order of increasing predicted time; when the corrected degradation index of the first predicted time is greater than or equal to the preset failure threshold, generate a threshold arrival flag for the first time; when there is an adjacent corrected degradation index after the first predicted time that changes from less than the preset failure threshold to greater than or equal to the preset failure threshold, determine the predicted time corresponding to the previous corrected degradation index as the first adjacent predicted time, and determine the predicted time corresponding to the next corrected degradation index as the second adjacent predicted time; when all corrected degradation indices in the prediction interval are less than the preset failure threshold, generate a threshold not reached flag. Step S502: When the first time threshold is reached, the first predicted time is determined as the predicted failure time; when the first adjacent predicted time and the second adjacent predicted time are obtained, linear interpolation is performed based on the correction degradation index corresponding to the two adjacent predicted times and the preset failure threshold to obtain the predicted failure time; when the threshold is not reached, the end time of the prediction interval is determined as the lifetime lower limit calculation time. Step S503: Calculate the time difference between the predicted failure time and the current detection time to obtain the remaining conduction lifetime prediction value; or, calculate the time difference between the lifetime lower limit calculation time and the current detection time to obtain the remaining conduction lifetime lower limit; combine the remaining conduction lifetime prediction value or the remaining conduction lifetime lower limit with the corresponding threshold arrival flag to output the remaining conduction lifetime detection result.

[0017] The present invention also provides a liquid crystal display screen bonding continuity detection system, the system comprising: The state feature construction module is used to acquire multi-source temporal detection data of the bonding area of ​​the liquid crystal display screen, and to perform state feature extraction tasks based on the multi-source temporal detection data using a temperature compensation and multi-source state feature construction mechanism, and output a multi-dimensional state feature sequence. The degradation index construction module is used to perform degradation representation tasks based on the multidimensional state feature sequence using a bidirectional temporal coding and entropy-modulated attention fusion mechanism, and output a fused degradation index sequence. The degradation trajectory prediction module is used to perform degradation trajectory prediction tasks based on the fused degradation index sequence using a parallel prediction mechanism of data-driven branch and physical constraint branch, and output candidate degradation trajectory groups. The degradation trajectory correction module is used to perform degradation trajectory correction tasks based on the candidate degradation trajectory group using a learnable gating fusion mechanism, and output the final predicted degradation trajectory. The remaining lifetime detection module is used to determine the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculate the remaining conduction lifetime based on the predicted failure time, and output the remaining conduction lifetime detection result.

[0018] The present invention also provides a liquid crystal display screen bonding continuity testing device, the liquid crystal display screen bonding continuity testing device comprising: a memory, a processor, and a liquid crystal display screen bonding continuity testing program stored in the memory and executable on the processor, wherein the liquid crystal display screen bonding continuity testing program implements the above method when executed by the processor.

[0019] The present invention also provides a computer program product, the computer program product including a liquid crystal display screen bonding continuity detection program, which implements the above method when executed by a processor.

[0020] The beneficial effects of this invention are as follows: This invention incorporates DC resistance, conduction current, ambient temperature, average surface temperature, AC impedance spectrum, surface grayscale image, microstrain, and stress into the same sampling time reference. By using temperature compensation, impedance relaxation offset, image anomaly density, and inelastic strain energy density, a multidimensional state feature sequence is constructed, enabling degradation information of different dimensions to have a clear physical source and comparable temporal expression.

[0021] This invention utilizes a bidirectional long short-term memory network to extract forward and backward temporal associations within a window, then calculates the degradation pattern entropy based on the feature dimension changes of adjacent hidden states, and adjusts the attention temperature with the entropy value, so that multi-head self-attention maintains a relatively smooth weight when degradation patterns are scattered, and enhances the effect of corresponding feature changes when the dominant pattern is clear, forming a continuous fusion degradation index sequence.

[0022] On the one hand, the present invention recursively generates data-driven prediction trajectories using a time-series decoder. On the other hand, it recursively normalizes the damage state based on extrapolation results of conduction current, ambient temperature, DC resistance, and micro-strain, and maps it to a physically constrained prediction trajectory, so that statistical time-series changes and damage evolution related to Joule heating and strain can both enter the candidate degradation trajectory group.

[0023] This invention uses the deviation between two candidate trajectories at the same prediction time as one of the gating inputs, combines data-driven prediction values ​​and physical constraint prediction values ​​to calculate the gating coefficient, and completes the fusion according to the gating coefficient and its complementary coefficient, so that the trajectory correction can adjust the contribution ratio according to the consistency of the two branches.

[0024] This invention sequentially processes three cases: the first prediction time has reached the threshold, the adjacent prediction time crosses the threshold, and the prediction interval has not yet reached the threshold. It outputs the remaining conduction lifetime prediction value or the lower limit of the remaining conduction lifetime respectively, and retains the corresponding threshold arrival indicator, so that the judgment boundary and result nature of the lifetime value can be distinguished. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the first embodiment of a liquid crystal display screen bonding continuity detection method according to the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] Example 1: As Figure 1 The diagram shown is a flowchart of a first embodiment of a liquid crystal display screen bonding continuity testing method according to the present invention, which presents a first embodiment of a liquid crystal display screen bonding continuity testing method according to the present invention.

[0029] In the first embodiment, the liquid crystal display screen bonding continuity detection method includes: Step S10: Obtain multi-source time-series detection data of the bonding area of ​​the liquid crystal display screen, and perform state feature extraction task based on the multi-source time-series detection data using temperature compensation and multi-source state feature construction mechanism, and output multi-dimensional state feature sequence; The bonding area of ​​a liquid crystal display (LCD) screen refers to the interface where the driver chip or flexible circuit board is electrically connected to the glass substrate via bonding materials, and its adjacent conductive paths. Multi-source time-series testing data are acquired synchronously at preset sampling time intervals, including DC resistance sequences, conduction current sequences, ambient temperature sequences, surface average temperature sequences, multi-frequency AC impedance spectrum sequences, surface grayscale image sequences, micro-strain sequences, and stress sequences corresponding to micro-strain time-by-time. Each sequence retains the same sampling time identifier, enabling the resistance, impedance, thermal image, and mechanical state at a given moment to be interpreted collectively, without misaligning and splicing data from different aging stages.

[0030] The DC resistance is first compensated for based on a preset reference temperature and the conductor's temperature coefficient of resistance. During processing, the difference between the average surface temperature and the reference temperature determines the proportion of resistance change caused by the conductor's temperature. The measured DC resistance is then converted to the reference temperature, forming a temperature-compensated resistance sequence. The system uses the temperature-compensated resistance at the first sampling moment as the initial reference, and calculates the ratio of change relative to this reference at each subsequent sampling moment, obtaining a sequence of relative resistance change rates. Thus, the reversible resistance change caused by ambient temperature rise and the continuous resistance shift caused by changes in the bonding contact state are distinguished at the characteristic level.

[0031] Multi-frequency AC impedance spectra are fitted hourly according to the same equivalent circuit structure to obtain charge transfer resistance, constant-phase element parameters, and dispersion index. The system determines the characteristic relaxation time based on the three fitting results at the same sampling time, and uses the relaxation time of the first sampling time as the initial characteristic relaxation time; for each subsequent time, the logarithmic difference between the current relaxation time and the initial relaxation time is calculated to form an impedance relaxation offset sequence. By using the logarithmic difference, relaxation variations spanning orders of magnitude are compressed to a comparable scale, while preserving the directional changes in interface charge transfer and non-ideal capacitive response during the aging process.

[0032] Surface grayscale image sequences are used to describe spatial anomalies in the thermal distribution or surface condition of the bound region. For each image, the system calculates the local reachability density of each pixel and forms the image anomaly density based on the deviation of this pixel density from the maximum local reachability density within the image; the deviation results of all pixels are summarized as the image anomaly density value at the corresponding sampling time. This processing focuses on the differences between local pixel distributions and their neighborhoods, and does not directly identify uniform grayscale changes in the entire image caused by exposure or overall temperature rise as local anomalies, thus allowing minute hotspots or local structural anomalies to be included in the temporal features.

[0033] The microstrain and stress sequences are correlated one-to-one at each sampling time. The system fits an elastic strain range based on this correspondence, separating strain changes within this range that can recover with unloading from irreversible increments outside this range; the excess is identified as inelastic strain increments. Each inelastic strain increment is multiplied by the stress at the same moment and accumulated over time to form an inelastic strain energy density sequence. This sequence is used to characterize the irreversible dissipation that continuously accumulates in the bonded material and adjacent structures under cyclic loading, rather than judging fatigue level solely by the microstrain amplitude at a single moment.

[0034] The relative rate of change of resistance, impedance relaxation offset, image anomaly density, and inelastic strain energy density have different dimensions and ranges of variation. The system standardizes each of these parameters within a sliding window and then concatenates them into a multidimensional state feature sequence according to a preset feature order. Each dimension only uses the statistical benchmark of its own window to complete the scaling transformation, preserving its temporal order and feature identity. Compared with directly concatenating the original data, this process reduces the impact of temperature, dimension, and fixed baseline differences on subsequent weight allocation, while unifying electrical connections, interface impedance, spatial anomalies, and irreversible mechanical damage into a temporal input that can be encoded in step S20.

[0035] In a series of high-temperature and high-humidity aging records, the ambient temperature slowly increased from 24°C to 28°C, and the measured DC resistance increased synchronously. However, after compensation with a reference temperature, the relative rate of change of resistance remained close to the baseline for the first 20 hours. By the 45th hour, the compensated resistance began to deviate continuously from the initial value. Simultaneously, the characteristic relaxation time obtained from multi-frequency impedance spectroscopy fitting showed a stable logarithmic shift, local density anomalies formed in the surface grayscale image at the bonding edges, and the inelastic strain increment corresponding to cyclic stress also began to accumulate. The system standardized four types of results within their respective sliding windows and stitched them together in a fixed order, so that this period was represented as multiple slight but directional state changes, rather than being dominated by a single resistance increase caused by ambient temperature.

[0036] Step S20: Based on the multidimensional state feature sequence, a bidirectional temporal coding and entropy-modulated attention fusion mechanism is used to perform the degradation representation task and output the fused degradation index sequence; The window feature matrix sequence is obtained by continuously extracting from the multidimensional state feature sequence. A preset window length determines the historical sampling range covered by each matrix, and a preset sliding step size determines the advancement distance of adjacent matrices on the time axis. The rows of the matrices are arranged according to the sampling time, and the columns are arranged according to the feature order fixed in step S10. Adjacent windows are allowed to retain common samples, making the degradation process continuous at the window boundaries. When the sampling sequence length is less than a complete window, subsequent samples can be waited for to enter, rather than constructing a degradation representation with unsourced filler values.

[0037] Each window's feature matrix is ​​input into the forward Long Short-Term Memory (LSTM) layer in the increasing direction of sampling time and into the backward LSM layer in the decreasing direction of sampling time. The forward hidden state reflects the correlation between the current sampling time and the previous state, while the backward hidden state reflects the correlation between the current sampling time and subsequent states within the window. The system concatenates two hidden states at the same sampling time and performs normalization to form a hidden state sequence. Bidirectional encoding does not utilize future measured data within the prediction interval, but instead simultaneously describes the entry process and the continuation result of changes within the acquired historical window.

[0038] The degradation pattern probability is generated by the absolute difference between adjacent hidden states across each feature dimension. The system calculates the change amplitude of the hidden states at adjacent time steps dimension by dimension, and then normalizes the absolute differences of each dimension to make the probabilities of all feature dimensions at the same time step form a comparable distribution. If multiple dimensions change approximately, the probability distribution is relatively dispersed; if the change in one dimension is significantly higher than that of the other dimensions, the probability is concentrated in the corresponding degradation pattern. The system calculates the degradation pattern entropy based on this probability distribution, using the entropy value to describe whether the degradation change of the current window is driven by multiple states or dominated by a few states.

[0039] The degradation pattern entropy is converted into a temperature regulation value using a hyperbolic tangent function, and this value, together with a preset baseline attention temperature parameter, determines the dynamic attention temperature parameter. This parameter is used to scale the attention score calculated by multi-head self-attention. The scaled result is normalized to form attention weights, which are then used to weight and fuse the hidden state sequence. The temperature regulation corresponding to the entropy value prevents the weights from concentrating too early when the pattern is dispersed, and allows attention to highlight continuously changing dimensions when the pattern is clear. After one-dimensional mapping, the weighted fusion result is arranged in chronological order to form a fused degradation index sequence, which serves as the common input to the two prediction branches in step S30.

[0040] Fixed-weight combinations maintain the same feature contribution throughout the aging cycle, even if the dominant degradation mode shifts from interface impedance change to contact resistance increase, the fusion result may still use early weights; simply selecting the dimension with the largest change can easily amplify occasional noise. This step first uses bidirectional encoding to confirm the connection between changes within the window, and then uses changes in adjacent hidden states to form probabilities and entropies, allowing the attention temperature to adjust according to the degradation mode distribution. Thus, early multi-source weak changes can be included in the degradation index, and features that continue to dominate in the later stages can receive higher attention without losing information from other dimensions.

[0041] Taking a window with a length of 24 sampling points and a sliding step of 4 sampling points as an example, in the early stage of aging, the changes in the hidden states of the four feature dimensions are similar, the normalized degradation mode probabilities are relatively uniform, the degradation mode entropy is at a high level, and the dynamic attention temperature makes the multi-head weight distribution relatively smooth. After entering the stage of rapid increase in contact resistance, the dimension corresponding to the relative rate of change of resistance maintains a large difference in hidden states in multiple consecutive adjacent time moments, the probability distribution begins to concentrate, the entropy value decreases, and the attention weights then tilt towards this continuously changing dimension. The final fused degradation index retains the joint shift of impedance, image, and strain in the early stage, and can also follow the accelerated change dominated by resistance in the later stage.

[0042] Step S30: Based on the fused degradation index sequence, a data-driven branch and physical constraint branch parallel prediction mechanism is used to perform the degradation trajectory prediction task and output the candidate degradation trajectory group; The candidate degradation trajectory group consists of a data-driven predicted trajectory and a physically constrained predicted trajectory, with both trajectories using the same prediction time and a preset number of prediction steps. The data-driven branch inputs the fused degradation index sequence into the temporal decoder, which recursively outputs subsequent degradation indices in the direction of increasing prediction time. Each predicted value is used as part of the generated temporal information during subsequent decoding, allowing historical fused degradation trends, recent rate of change, and continuity to jointly influence the data-driven predicted trajectory. This branch is used to preserve the statistical change patterns already exhibited by a specific sample.

[0043] The physical constraint branch first performs linear trend extrapolation on the conduction current, ambient temperature, DC resistance, and micro-strain, respectively, to obtain four prediction sequences covering the same prediction interval. Linear extrapolation is based on valid observations prior to the detection time, maintaining the sampling interval of each sequence consistent with the prediction time. If an isolated missing measurement point exists in a channel, the continuous trend can be restored based on its adjacent valid samples before extrapolation, avoiding directly incorporating the sudden drop caused by missing measurements into the physical damage calculation. The four prediction sequences do not replace the fusion degradation index but are used to construct physical constraint trajectories with thermal, electrical, and mechanical implications.

[0044] The equivalent junction temperature of the binding region at each prediction time is determined by the predicted ambient temperature and the Joule thermal rise. The Joule thermal rise increases with the sum of the squares of the predicted conduction current and the predicted DC resistance, and is constrained by the preset equivalent heat transfer coefficient and the preset effective heat transfer area; the stronger the heat transfer capacity or the larger the effective heat dissipation area, the smaller the junction temperature increment corresponding to the same current and resistance. This textual relationship is consistent with the equivalent junction temperature expression in claim 5, enabling the current, resistance, and heat dissipation conditions to form a thermal state input at the same prediction time.

[0045] The normalized damage state change rate is jointly determined by the preset damage evolution pre-factor, activation energy, Boltzmann constant, equivalent junction temperature, predicted microstrain, and preset microstrain influence index. The temperature term decays exponentially according to the negative ratio of the activation energy to the product of the Boltzmann constant and the equivalent junction temperature, while the microstrain exhibits a power-law influence according to the preset influence index. Together with the damage evolution pre-factor, these factors give the damage growth rate at the predicted time. The system uses the preset initial damage state as the initial value for integration and recursively calculates the cumulative change rate according to the predicted time to obtain the normalized damage state sequence.

[0046] The normalized damage state sequence is converted into a physical constraint prediction trajectory through a preset monotonic mapping relationship between damage states and degradation indices. The monotonic mapping ensures that a higher damage state does not correspond to a lower degradation indices, but does not require the data-driven trajectories to strictly conform to the same analytical shape. The system aligns the data-driven prediction trajectories and the physical constraint prediction trajectories at the same prediction time. If the time steps of the two trajectories are different, the common prediction time defined in the claims is used to complete the correspondence, and misaligned prediction values ​​are not sent to step S40. The two aligned trajectories together form a candidate degradation trajectory group.

[0047] When relying solely on historical degradation indices for recursive extrapolation, high-temperature or high-current conditions not covered by the training samples may cause long-range trajectories to remain approximately linear, failing to reflect the accelerated damage after Joule heating. Conversely, when relying solely on physical branches, preset heat transfer and damage parameters may not adequately reflect individual sample differences. Parallel prediction retains both information sources, enabling step S40 to adjust the contributions of both based on the actual trajectory deviation at the same time, rather than pre-fixing a particular branch to always be dominant.

[0048] Assume 12 prediction steps are set after the current detection time. The data-driven branch outputs a stable trajectory based on the slow increase of recent fusion degradation indicators. The physical constraint branch extrapolates the conduction current, ambient temperature, DC resistance, and microstrain, finding that the current remains constant but the DC resistance and ambient temperature increase simultaneously, leading to a corresponding increase in the equivalent junction temperature and a gradually accelerating rate of increase in the recursive damage state. After monotonic mapping, the physical constraint trajectory exhibits a more pronounced upward curve in the later prediction times. The two trajectories are aligned and output as a group according to prediction times 1 to 12, providing direct data for the next step to identify when they are consistent and when significant deviations occur.

[0049] Step S40: Based on the candidate degenerate trajectory group, a learnable gating fusion mechanism is used to perform the degenerate trajectory correction task, and the final predicted degenerate trajectory is output; The learnable gating fusion reads the data-driven prediction value and the physical constraint prediction value at each prediction time, and calculates the absolute difference between the two as the trajectory deviation value. The absolute difference only indicates the degree of divergence between the branches and does not change the sign due to which branch has a larger value. The data-driven prediction value, the physical constraint prediction value, and the trajectory deviation value are arranged in a fixed order to form the gating input, so that the gating unit can see both the position of the two trajectories and their current degree of separation.

[0050] The gating coefficients are obtained by applying the learnable gating weight matrix and learnable gating bias to the concatenated vector, followed by transformation using the Sigmoid function. The Sigmoid function restricts the gating result to between 0 and 1, making it easier to interpret as the participation ratio of the data-driven branch at the current prediction time. The gating weight matrix and bias are learned during the training phase based on the trajectory correction target and remain determined during the detection phase; the gating input strictly adopts the two predicted values ​​and their trajectory deviation as defined in claim 6, without introducing any additional state variables.

[0051] The correction degradation index consists of two parts: the data-driven predicted value multiplied by the current gating coefficient, and the physical constraint predicted value multiplied by 1 minus the current gating coefficient. The two parts are then added together. When the gating coefficient is close to 1, the correction result is closer to the data-driven predicted value; when the gating coefficient is close to 0, the correction result is closer to the physical constraint predicted value; in the intermediate range, the two trajectories participate together in a complementary proportion. The system arranges all correction degradation indices along the increasing direction of the prediction time to form the final predicted degradation trajectory.

[0052] Simple averaging always contributes the same amount to both candidate trajectories, and cannot be adjusted even if there is a significant divergence between them at some point. While fixed weights can reflect overall preference, they cannot change with the prediction stage or trajectory deviation. Gating mechanisms directly incorporate the degree of divergence into the calculation, allowing the training process to learn to use different fusion ratios when candidate trajectories are close, separated, or in different numerical ranges. This approach does not require directly determining whether a branch is correct or incorrect, but rather forms continuous weight changes, reducing discontinuous inflections caused by hard switching on the final trajectory.

[0053] Within a prediction interval, the data-driven values ​​and physical constraint values ​​are close in the first four prediction times, resulting in a small trajectory deviation. The gating coefficient can remain in the middle range, and the correction result lies between the two trajectories. At the seventh prediction time, the data-driven trajectory experiences a short-term spike due to recent observation fluctuations, while the physical constraint trajectory continues to grow, increasing the trajectory deviation. The gating unit reduces the participation ratio of the data-driven branch at this time based on the three inputs, preventing the correction value from completely following the spike. If the data-driven values ​​continue to rise and approach the physical trajectory again in subsequent times, the gating coefficient can be readjusted according to the input changes, ultimately maintaining a continuous trajectory.

[0054] Step S50: Determine the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculate the remaining conduction lifetime based on the predicted failure time, and output the remaining conduction lifetime detection result; The remaining conduction lifetime detection takes the current detection time, the final predicted degradation trajectory, and a preset failure threshold as input. The preset failure threshold represents the degradation index level corresponding to when the bound conduction state reaches the failure judgment boundary. The system compares and corrects the degradation index with the threshold item by item along the increasing direction of the predicted time, not only finding the position where the threshold is first reached, but also distinguishing between three types of boundary states: the first predicted time has been reached, a crossing has occurred between adjacent times, and the entire prediction interval has not yet been reached.

[0055] If the correction degradation index at the first predicted time is greater than or equal to the preset failure threshold, the system generates a threshold arrival flag for the first time and directly determines the first predicted time as the predicted failure time. This rule applies when the current detection state is close to the failure boundary and the predicted trajectory is above the threshold from the starting point, avoiding continuing to search later and mistakenly identifying later times as the first failure. The corresponding remaining conduction lifetime is obtained from the time difference between the first predicted time and the current detection time.

[0056] If the first predicted time has not yet reached the threshold, but a pair of adjacent correction degradation indicators subsequently change from less than the threshold to greater than or equal to the threshold, the system records the time corresponding to the former indicator as the first adjacent predicted time and the time corresponding to the latter indicator as the second adjacent predicted time. The predicted failure time is linearly interpolated based on the two correction degradation indicators, the two predicted times, and the preset failure threshold, so that the threshold arrival position can fall between two discrete predicted times, rather than uniformly taking the latter time. Then, the current detection time is subtracted from the interpolated failure time to form the predicted value of the remaining conduction lifetime.

[0057] If all corrected degradation indices within the prediction interval are less than the preset failure threshold, the system generates a threshold not reached flag and determines the end time of the prediction interval as the calculation time for the lower limit of conduction lifetime. Since a definite threshold arrival time cannot be given within the current prediction range, the time difference between the end time and the current detection time is output as the lower limit of the remaining conduction lifetime. The combination of the lower limit of conduction lifetime and the threshold not reached flag allows users to recognize that this value indicates at least a corresponding remaining lifetime, rather than mistakenly taking the prediction interval length as a definite failure time.

[0058] When using only a simple intersection of trajectory and threshold, output rules are easily missing for boundary cases where the first point has exceeded the threshold and the entire interval has not. Furthermore, taking only the discrete sampling time of the first threshold exceedance limits the remaining lifetime accuracy to the prediction step size. This step uses a three-branch decision to cover the entire prediction interval, performs linear interpolation when the interval crosses the threshold, and outputs the lifetime prediction value or lifetime lower limit along with the corresponding identifier. This ensures that the result of step S50 includes both the time quantity and the decisional properties of that time quantity.

[0059] The current time of a certain detection is hour 150, the prediction step size is 10 hours, and the failure threshold is 1.0. If the corrected degradation indexes at hour 220 and hour 230 are 0.94 and 1.04 respectively, the system interpolates between the two times according to the proportion of the threshold to obtain a predicted failure time of approximately hour 226, and outputs a predicted remaining conduction lifetime of approximately 76 hours and a threshold arrival flag. If all predicted values ​​up to hour 270 are still below 1.0, no out-of-interval failure point is fabricated, and instead, a lower limit of remaining conduction lifetime of no less than 120 hours and a threshold not reached flag are output; if the first predicted value at hour 160 has reached 1.0, then hour 160 is directly used as the predicted failure time.

[0060] Example 2: Furthermore, the liquid crystal display (LCD) bonding continuity testing system provided by this invention employs the LCD bonding continuity testing method described in the above embodiments, and can solve a technical problem related to LCD bonding continuity testing. The beneficial effects of the LCD bonding continuity testing system provided by this invention are the same as those of the LCD bonding continuity testing method described in the above embodiments, and other technical features of the LCD bonding continuity testing system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0061] Example 3: This invention provides a liquid crystal display (LCD) bonding continuity testing device. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which, when executed, enable the processor to perform the LCD bonding continuity testing method described in Example 1. The LCD bonding continuity testing device in this invention may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This LCD bonding continuity testing device is merely an example and should not limit the functionality or scope of the invention. The device may also include a processing unit (e.g., a central processing unit, a graphics processing unit), which can perform various appropriate actions and processes based on a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of a liquid crystal display bonding continuity testing device. The processing unit, read-only memory, and random access memory are interconnected via a bus. The I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the liquid crystal display bonding continuity testing device to communicate wirelessly or wiredly with other devices to exchange data. While a liquid crystal display bonding continuity testing device with various systems has been described, it should be understood that it is not required to implement or possess all of the described systems. Alternatively, more or fewer systems can be implemented.

[0062] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the liquid crystal display screen bonding continuity detection method described above. The computer program product provided by this invention can solve a technical problem related to liquid crystal display screen bonding continuity detection. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the liquid crystal display screen bonding continuity detection method provided in the above embodiments, and will not be repeated here.

[0063] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0064] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for detecting the bonding continuity of a liquid crystal display screen, characterized in that, The method includes: Step S10: Obtain multi-source time-series detection data of the bonding area of ​​the liquid crystal display screen, and perform state feature extraction task based on the multi-source time-series detection data using temperature compensation and multi-source state feature construction mechanism, and output multi-dimensional state feature sequence; Step S20: Based on the multidimensional state feature sequence, a bidirectional temporal coding and entropy-modulated attention fusion mechanism is used to perform the degradation representation task and output the fused degradation index sequence; Step S30: Based on the fused degradation index sequence, a data-driven branch and physical constraint branch parallel prediction mechanism is used to perform the degradation trajectory prediction task and output the candidate degradation trajectory group; Step S40: Based on the candidate degenerate trajectory group, a learnable gating fusion mechanism is used to perform the degenerate trajectory correction task, and the final predicted degenerate trajectory is output; Step S50: Determine the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculate the remaining conduction lifetime based on the predicted failure time, and output the remaining conduction lifetime detection result.

2. The liquid crystal display screen bonding continuity detection method as described in claim 1, characterized in that, Step S10 involves acquiring multi-source timing detection data of the liquid crystal display bonding area, and performing a state feature extraction task based on the multi-source timing detection data using a temperature compensation and multi-source state feature construction mechanism to output a multi-dimensional state feature sequence. Specifically, this includes: Step S101: According to the preset sampling time interval, acquire the DC resistance sequence, conduction current sequence, ambient temperature sequence, surface average temperature sequence, multi-frequency AC impedance spectrum sequence, surface grayscale image sequence, micro-strain sequence and stress sequence corresponding to the micro-strain sequence of the liquid crystal display bonding area to form a multi-source time-series detection dataset. Step S102: Based on the multi-source time-series detection dataset, temperature compensation is performed on the DC resistance sequence according to a preset reference temperature and a preset conductor resistance temperature coefficient to obtain a temperature-compensated resistance sequence. The temperature-compensated resistance at the first sampling moment in the temperature-compensated resistance sequence is determined as the initial temperature-compensated resistance, and the change ratio of the temperature-compensated resistance at the remaining sampling moments relative to the initial temperature-compensated resistance is calculated to obtain a resistance relative change rate sequence. Equivalent circuit fitting is performed on the multi-frequency AC impedance spectrum sequence to obtain a charge transfer resistance sequence, a constant phase element parameter sequence, and a dispersion index sequence. The characteristic relaxation time is determined based on the charge transfer resistance, constant phase element parameters, and dispersion index at the same sampling moment. The characteristic relaxation time at the first sampling moment is determined as the initial characteristic relaxation time, and the logarithmic difference between the characteristic relaxation time at the remaining sampling moments and the initial characteristic relaxation time is calculated to obtain an impedance relaxation offset sequence. Step S103: Based on the multi-source time-series detection dataset and the impedance relaxation offset sequence, determine the image abnormal density sequence according to the deviation of the local reachable density of each pixel in the surface grayscale image sequence from the maximum local reachable density in the corresponding image; fit the elastic strain range according to the correspondence between the micro-strain sequence and the stress sequence, determine the micro-strain increments that exceed the elastic strain range as inelastic strain increments, and accumulate the product of each inelastic strain increment and the corresponding stress to obtain the inelastic strain energy density sequence; perform sliding window standardization on the relative resistance change rate sequence, the impedance relaxation offset sequence, the image abnormal density sequence, and the inelastic strain energy density sequence respectively, and stitch the standardization results according to the preset feature order to output the multidimensional state feature sequence.

3. The liquid crystal display screen bonding continuity detection method as described in claim 1, characterized in that, Step S20, which involves performing a degradation representation task based on the multidimensional state feature sequence using a bidirectional temporal coding and entropy-modulated attention fusion mechanism, and outputting a fused degradation index sequence, specifically includes: Step S201: The multidimensional state feature sequence is continuously truncated according to the preset window length and preset sliding step size, and each truncated result is arranged according to the sampling time and feature dimension to form a window feature matrix sequence; Step S202: Input the window feature matrix sequence into the bidirectional long short-term memory network encoder, extract the forward temporal correlation features and backward temporal correlation features at each sampling time, and concatenate the forward temporal correlation features and backward temporal correlation features at the same sampling time to output the hidden state sequence. Step S203: Based on the hidden state sequence, calculate the absolute difference between adjacent hidden states in each feature dimension, and normalize each absolute difference to obtain the degradation mode probability corresponding to each feature dimension; calculate the degradation mode entropy according to the degradation mode probability, convert the degradation mode entropy into a temperature regulation quantity through a hyperbolic tangent function, and determine the dynamic attention temperature parameter according to the preset baseline attention temperature parameter and the temperature regulation quantity; scale and normalize the attention score calculated by multi-head self-attention using the dynamic attention temperature parameter, perform weighted fusion on the hidden state sequence according to the normalized attention weight, and map the weighted fusion result into a single-dimensional degradation index, outputting the fused degradation index sequence.

4. The liquid crystal display screen bonding continuity detection method as described in claim 3, characterized in that, Step S202, which involves inputting the window feature matrix sequence into a bidirectional long short-term memory network encoder to extract forward temporal correlation features and backward temporal correlation features at each sampling time, and concatenating the forward and backward temporal correlation features at the same sampling time to output the hidden state sequence, specifically includes: The feature matrices of each window are sequentially input into the feedforward long short-term memory network layer in the order of increasing sampling time to obtain the forward hidden state corresponding to each sampling time. The same window feature matrix is ​​input into the backward long short-term memory network layer in decreasing order of sampling time to obtain the backward hidden state corresponding to each sampling time. The forward hidden state and the backward hidden state are concatenated at the same sampling time, and the concatenation result is normalized to form the hidden state sequence.

5. The liquid crystal display screen bonding continuity detection method as described in claim 2, characterized in that, Step S30, which involves performing the degradation trajectory prediction task based on the fused degradation index sequence using a parallel prediction mechanism of data-driven branching and physical constraint branching, and outputting a candidate degradation trajectory group, specifically includes: Step S301: Input the fused degradation index sequence into the time-series decoder, and recursively decode the fused degradation index sequence according to the increasing direction of the prediction time to obtain the data-driven prediction trajectory; Step S302: Determine the prediction interval according to the preset prediction step number, and perform linear trend extrapolation on the conduction current sequence, the ambient temperature sequence, the DC resistance sequence, and the micro-strain sequence respectively to obtain the predicted conduction current sequence, the predicted ambient temperature sequence, the predicted DC resistance sequence, and the predicted micro-strain sequence, and calculate the equivalent junction temperature of the binding region at each prediction time according to the following formula: The normalized rate of change of damage state at each prediction time is calculated using the following formula: In the formula, For the predicted time; For the predicted time The equivalent junction temperature of the binding region; For the predicted time Predicted ambient temperature; For the predicted time The predicted on-current; For the predicted time Predicted DC resistance; The preset equivalent heat transfer coefficient; The preset effective heat exchange area; For the predicted time Normalized damage state; Preset damage evolution pre-factor; Preset activation energy; Boltzmann's constant; For the predicted time Predicting microstrain; The preset micro-strain influence index; Using a preset initial damage state as the initial value for integration, the normalized damage state change rate is recursively integrated to obtain a normalized damage state sequence within the prediction interval. The normalized damage state sequence is then converted into a physical constraint prediction trajectory based on the monotonic mapping relationship between the preset damage state and the degradation index. Step S303: Align the data-driven prediction trajectory and the physical constraint prediction trajectory according to the same prediction time, and form the candidate degradation trajectory group by aligning the data-driven prediction trajectory and the physical constraint prediction trajectory.

6. The liquid crystal display screen bonding continuity detection method as described in claim 5, characterized in that, Step S40, which involves performing a degradation trajectory correction task based on the candidate degradation trajectory group using a learnable gating fusion mechanism and outputting the final predicted degradation trajectory, specifically includes: Step S401: Read the data-driven prediction value and the physical constraint prediction value at the same prediction time from the candidate degraded trajectory group, and calculate the absolute difference between the data-driven prediction value and the physical constraint prediction value to obtain the trajectory deviation value; Step S402: Calculate the gating coefficients for each prediction time according to the following formula: In the formula, For the predicted time The gating coefficient; For the Sigmoid function; For learnable gated weight matrix; For learnable gated bias; For the predicted time Data-driven forecasts; For the predicted time The predicted values ​​of physical constraints; For the predicted time The trajectory deviation value; the semicolon in square brackets indicates vector concatenation; Step S403: Gated fusion of the data-driven prediction value and the physical constraint prediction value according to the following formula: In the formula, For the predicted time The correction degradation index is calculated; the correction degradation indexes are arranged in ascending order according to the prediction time, and the final predicted degradation trajectory is output.

7. The liquid crystal display screen bonding continuity detection method as described in claim 6, characterized in that, Step S50, which involves determining the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculating the remaining conduction lifetime based on the predicted failure time, and outputting the remaining conduction lifetime detection result, specifically includes: Step S501: Read the current detection time, and compare each corrected degradation index in the final predicted degradation trajectory with a preset failure threshold in the order of increasing predicted time; when the corrected degradation index of the first predicted time is greater than or equal to the preset failure threshold, generate a threshold arrival flag for the first time; when there is an adjacent corrected degradation index after the first predicted time that changes from less than the preset failure threshold to greater than or equal to the preset failure threshold, determine the predicted time corresponding to the previous corrected degradation index as the first adjacent predicted time, and determine the predicted time corresponding to the next corrected degradation index as the second adjacent predicted time; when all corrected degradation indices in the prediction interval are less than the preset failure threshold, generate a threshold not reached flag. Step S502: When the first time threshold is reached, the first predicted time is determined as the predicted failure time; when the first adjacent predicted time and the second adjacent predicted time are obtained, linear interpolation is performed based on the correction degradation index corresponding to the two adjacent predicted times and the preset failure threshold to obtain the predicted failure time; when the threshold is not reached, the end time of the prediction interval is determined as the lifetime lower limit calculation time. Step S503: Calculate the time difference between the predicted failure time and the current detection time to obtain the remaining conduction lifetime prediction value; or, calculate the time difference between the lifetime lower limit calculation time and the current detection time to obtain the remaining conduction lifetime lower limit; combine the remaining conduction lifetime prediction value or the remaining conduction lifetime lower limit with the corresponding threshold arrival flag to output the remaining conduction lifetime detection result.

8. A liquid crystal display screen bonding continuity testing system, applied to the liquid crystal display screen bonding continuity testing method according to any one of claims 1 to 7, characterized in that, The liquid crystal display screen bonding continuity detection system includes: The state feature construction module is used to acquire multi-source temporal detection data of the bonding area of ​​the liquid crystal display screen, and to perform state feature extraction tasks based on the multi-source temporal detection data using a temperature compensation and multi-source state feature construction mechanism, and output a multi-dimensional state feature sequence. The degradation index construction module is used to perform degradation representation tasks based on the multidimensional state feature sequence using a bidirectional temporal coding and entropy-modulated attention fusion mechanism, and output a fused degradation index sequence. The degradation trajectory prediction module is used to perform degradation trajectory prediction tasks based on the fused degradation index sequence using a parallel prediction mechanism of data-driven branch and physical constraint branch, and output candidate degradation trajectory groups. The degradation trajectory correction module is used to perform degradation trajectory correction tasks based on the candidate degradation trajectory group using a learnable gating fusion mechanism, and output the final predicted degradation trajectory. The remaining lifetime detection module is used to determine the predicted failure time based on the final predicted degradation trajectory and the preset failure threshold, calculate the remaining conduction lifetime based on the predicted failure time, and output the remaining conduction lifetime detection result.

9. A liquid crystal display screen bonding continuity testing device, characterized in that, The liquid crystal display bonding continuity testing device includes: a memory, a processor, and a liquid crystal display bonding continuity testing program stored in the memory and executable on the processor. When the liquid crystal display bonding continuity testing program is executed by the processor, it implements a liquid crystal display bonding continuity testing method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a liquid crystal display screen bonding continuity detection program, which, when executed by a processor, implements a liquid crystal display screen bonding continuity detection method according to any one of claims 1 to 7.

Images