Logistics roadside equipment bionic self-repairing test method and system

Abstract

The invention discloses a bionic self-repairing testing method for logistics roadside equipment, and belongs to the technical field of logistics infrastructure intelligent operation and maintenance and material performance testing. The method aims at solving the technical problems that an existing self-repairing testing technology is low in efficiency and depends on artificial experience, and active management of the testing process cannot be achieved. According to the technical scheme, the method comprises the following steps: acquiring multi-source time sequence monitoring data including a visual image sequence, an acoustic emission signal and an environmental parameter data stream in real time; performing feature extraction on the data to obtain a time sequence feature vector; inputting the feature vector into a pre-trained time sequence deep learning model, and dynamically outputting a normalized repair completion degree score and predicted residual repair time; and finally, automatically judging and triggering a test termination instruction based on the score, and simultaneously generating a test report containing the repair process curve and time comparison information. The method is mainly used for realizing rapid, accurate and automatic testing and performance evaluation of the bionic self-repairing process of the logistics roadside equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and material performance testing technology for logistics infrastructure. More specifically, this invention relates to a biomimetic self-healing testing method and system for logistics roadside equipment. Background Technology

[0002] Roadside equipment in logistics, such as monitoring poles, traffic light covers, and sensor housings, are exposed to complex outdoor environments for extended periods, inevitably suffering damage such as microcracks and scratches due to external impacts, material aging, and environmental corrosion. To extend equipment lifespan and reduce maintenance costs, biomimetic self-healing materials are increasingly being applied to such equipment. However, effectively and reliably testing and evaluating the self-healing performance of these materials has become a key factor restricting their engineering application.

[0003] Currently, testing methods for self-healing materials used in logistics roadside equipment have several limitations. First, testing efficiency is generally low. Common testing methods rely on manual, timed observation or destructive sampling followed by offline laboratory analysis. For example, multiple images of the damaged area are taken at different time points using a microscope for comparison, or the mechanical property recovery rate is measured at specific time intervals. The entire process is time-consuming, ranging from hours to days, and cannot meet the need for rapid screening and evaluation of the material's repair performance. This problem stems from the lack of continuous, automated monitoring of the repair process in traditional methods, and the determination of whether the repair is complete largely depends on the operator's experience, making it highly subjective and difficult to determine the optimal testing termination point.

[0004] Secondly, the accuracy and objectivity of the test results need improvement. Existing methods often focus on acquiring a single type of signal, such as relying solely on visual image analysis to determine crack closure. However, visual assessment is easily affected by factors such as lighting conditions and viewing angle, and is insensitive to repair activities within the material or at the microscopic scale, potentially leading to misjudgments of the repair status. Simultaneously, environmental factors such as temperature and humidity significantly impact the repair rate and effectiveness, but conventional tests are usually conducted in a constant laboratory environment, failing to consider the dynamic changes in real roadside environmental parameters. This results in a disconnect between test conditions and actual working conditions, limiting the guiding value of the test results. Attempts to integrate data from multiple sensor types have faced difficulties, primarily due to the lack of effective automated processing solutions for temporal alignment and feature fusion of data from different sources (such as images, acoustic signals, and environmental parameters), making it difficult to objectively determine the contribution weight of each data source to the repair status.

[0005] Furthermore, the standardization and repeatability of the testing process are insufficient. The initial state of the test, namely the initial damage artificially created on the material surface, has significant human error in terms of shape, size, and depth. This inconsistency in initial conditions directly leads to large fluctuations in test results between different batches, and even between different samples within the same batch, making effective cross-sectional comparisons impossible and posing difficulties for the objective evaluation and classification of material properties.

[0006] Furthermore, the pursuit of intelligent testing systems also presents new challenges. For instance, if the testing system is to self-optimize through accumulated data, incremental learning is required. However, ensuring the quality of new data and preventing performance fluctuations or regressions in the model during the learning process are issues that need careful handling in practical applications. Simultaneously, to ensure the stability of long-term automated testing, the system's fault tolerance, such as ensuring uninterrupted testing in the event of critical sensor failures, is also a crucial aspect that needs to be considered in the design.

[0007] In summary, existing technologies for testing biomimetic self-healing materials for logistics roadside equipment have room for improvement in terms of testing efficiency, result accuracy, environmental simulation realism, testing standardization, and the stability and self-optimization capabilities of the system during long-term operation. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0009] Another objective of this invention is to provide a biomimetic self-healing testing method and system for logistics roadside equipment, which solves the technical problems of low efficiency, reliance on human experience, and inability to proactively manage the testing process in existing self-healing testing technologies. This invention enables rapid, accurate, and automated testing and performance evaluation of the biomimetic self-healing process of logistics roadside equipment.

[0010] To achieve these objectives and other advantages according to the present invention, a biomimetic self-healing testing method for roadside equipment is provided, comprising the following steps: S1. Real-time acquisition of multi-source time-series monitoring data of the roadside equipment under test during the self-repair process. The multi-source time-series monitoring data includes at least a sequence of visual images characterizing changes in its surface morphology, acoustic emission signals characterizing micro-deformation of the equipment housing and connectors, and environmental parameter data streams containing roadside high temperature and ultraviolet radiation characteristics. S2: Extract features from multi-source time-series monitoring data to obtain a time-series feature vector representing the self-healing process; S3: Input the time-series feature vector into the pre-trained repair status prediction model, and output the current self-repair process status assessment and the estimated remaining repair time required to reach the predetermined repair completion standard; wherein, the pre-trained repair status prediction model is a time-series deep learning model trained based on historical self-repair data; the current self-repair process status assessment output is a normalized repair completion score between 0 and 1 that integrates multi-source time-series monitoring data; the estimated remaining repair time is the time span required for the pre-trained repair status prediction model to dynamically extrapolate and predict the repair completion score to reach the preset threshold based on the historical sequence of the time-series feature vector and the current state; S4: Based on the output status assessment and estimated remaining repair time, automatically determine the test progress and automatically generate a test termination command and test report when the predetermined repair completion criteria are met. The automatic judgment logic is as follows: monitor the repair completion score in real time, and trigger the test termination command immediately when the repair completion score reaches or exceeds the preset threshold for the first time. The test report should include at least the repair completion rate change curve over time, the final repair completion score, and the comparison information between the estimated remaining repair time and the actual repair time.

[0011] Preferably, in step S2, feature extraction of the multi-source time-series monitoring data specifically includes: S21: Extract primary features from visual image sequences, acoustic emission signals, and environmental parameter data streams respectively. From the visual image sequence, extract the area and perimeter of the damaged area as visual primary features using an image segmentation algorithm. From the acoustic emission signal, extract event counts, signal amplitude, and energy as acoustic primary features. From the environmental parameter data stream, extract the instantaneous values ​​and changing trends of temperature and humidity as environmental primary features. S22: Weighted fusion of visual primary features, acoustic primary features, and environmental primary features to generate a temporal feature vector; The weight coefficients of each primary feature are learned through backpropagation during the training process of the pre-trained repair state prediction model, and can be dynamically adjusted according to different environmental primary features.

[0012] Preferably, in step S1, the multi-source time-series monitoring data is collected in a programmable multi-factor environmental simulation test chamber; wherein the test chamber is configured to dynamically and synchronously apply at least two of the following environmental stresses according to a preset environmental stress spectrum during the test cycle of the self-healing process: Cyclic temperature stress: It cycles between high temperature limit and low temperature limit according to a preset time program; Mechanical vibration stress: Simulates the vibration frequency and amplitude of the roadside environment near logistics transportation vehicles or equipment; Ultraviolet radiation stress: Simulates the ultraviolet components in outdoor sunlight and irradiates at a preset irradiance. The environmental parameter data stream originates from real-time monitoring data of the aforementioned environmental stresses applied within the test chamber.

[0013] Preferably, step S4 is followed by: S5. After generating the test report, the multi-source time-series monitoring data with complete self-healing process obtained in this test and the final measured repair time are used as a new set of labeled samples. The pre-trained repair state prediction model is incrementally learned using new labeled samples to update its parameters for prediction in subsequent test tasks.

[0014] Preferably, in step S5, a data credibility verification step is added before incremental learning, specifically as follows: Based on a preset set of data quality rules, the multi-source time-series monitoring data that constitute the new labeled samples are verified. The data quality rule set includes at least: whether the sharpness of the visual image sequence is higher than a set sharpness threshold, whether the signal-to-noise ratio of the acoustic emission signal is higher than a set signal-to-noise ratio threshold, and whether the values ​​of the environmental parameter data stream are within a preset effective range. New labeled samples are only used for incremental learning of the pre-trained repair state prediction model if they pass the data credibility verification.

[0015] Preferably, a standardized initial damage preparation step is included before step S1, specifically: Using a programmable laser etching device integrated into the testing system, initial damage with consistent geometry and physical dimensions is prepared on the surface of the self-healing material of the logistics roadside equipment under test, based on pre-stored standardized damage patterns and depth parameters.

[0016] The initial damage is the starting target monitored by the visual image sequence and acoustic emission signal in step S1.

[0017] Preferably, in step S1, the acquisition of the visual image sequence is jointly ensured by a primary camera and a backup camera. Step S1 includes a fault diagnosis step while simultaneously acquiring multi-source time-series monitoring data of the roadside equipment under test during its self-repair process: real-time monitoring of the operating status of the main camera and the quality of the visual image sequence. When the primary camera is determined to be faulty or the image quality is consistently unsatisfactory, the data acquisition source is automatically switched to the backup camera to ensure the continuity of multi-source time-series monitoring data.

[0018] Preferably, step S5, which involves incrementally learning the pre-trained repair state prediction model using new labeled samples, specifically includes: S51. Use new labeled samples to incrementally update a copy of the pre-trained repair state prediction model. Only update the fully connected layer parameters of the pre-trained repair state prediction model, without retraining the feature extraction layer, to obtain a candidate model. S52: Using historical data corresponding to the new labeled samples, evaluate the prediction accuracy of the candidate model and the current formal pre-trained repair state prediction model in parallel. S53: The candidate model is used to replace the current pre-trained repair state prediction model only when the prediction accuracy of the candidate model is higher than that of the current official pre-trained repair state prediction model, thus completing incremental learning.

[0019] Preferably, the cyclical execution of steps S3 and S4 also includes a dynamic control step linked to the environmental simulation and prediction steps, specifically: Receive the repair completion score output in real time from the pre-trained repair status prediction model; The real-time repair rate is calculated based on historical data of repair completion scores; When the real-time repair rate is lower than a preset active threshold, a control command is sent to a programmable multi-factor environmental simulation test chamber to reduce the intensity or frequency of the applied environmental stress.

[0020] Preferably, the step between the standardized initial damage preparation step and S1 includes a damage effectiveness verification step, specifically: The self-healing material of the logistics roadside equipment to be tested contains a pre-embedded conductive fiber mesh with a fiber diameter of 50-100μm and a mesh spacing of 1-2 mm. The initial damaged area was magnified and photographed using a 200x optical microscope with a resolution of 0.5μm integrated into the testing system to obtain microscopic morphology images. Based on the analysis of microscopic morphology images using image recognition algorithms, and combined with the detection of conductivity in the damaged area, it is confirmed that the initial damage has penetrated through the pre-embedded conductive fiber mesh; only when the initial damage is confirmed to be an effective penetrating damage will the subsequent self-repair process monitoring in step S1 be initiated.

[0021] The present invention has at least the following beneficial effects: 1. The biomimetic self-healing testing method for logistics roadside equipment of the present invention realizes automated testing and intelligent decision-making of the self-healing process through real-time multi-source monitoring and time-series deep learning model, which significantly improves testing efficiency and result objectivity, and provides a data foundation for process traceability and optimization of pre-trained repair status prediction model.

[0022] 2. The biomimetic self-healing test method for logistics roadside equipment of the present invention constructs a feature vector that can more comprehensively and accurately represent the repair state by adaptively weighting and fusing multi-source features, thereby improving the evaluation accuracy and robustness of the pre-trained repair state prediction model.

[0023] 3. The biomimetic self-healing test method for logistics roadside equipment of the present invention is tested in a programmable environment simulation test chamber, so that the test conditions are highly consistent with the actual working conditions, which effectively improves the reliability and guidance value of the test results.

[0024] 4. The biomimetic self-healing testing method for logistics roadside equipment of the present invention enables the prediction model to continuously optimize itself using new test data through an incremental learning mechanism, thereby realizing the continuous evolution and adaptive improvement of the test system performance.

[0025] 5. The biomimetic self-healing test method for logistics roadside equipment of the present invention ensures the quality of data used for incremental learning by verifying the reliability of data before updating the pre-trained repair state prediction model, thereby guaranteeing the effectiveness of the pre-trained repair state prediction model update and the long-term stability of the system from the source.

[0026] 6. The biomimetic self-healing test method for logistics roadside equipment of the present invention prepares the initial damage through standardized laser etching, which ensures the consistency of the test starting point and greatly improves the comparability and repeatability of the test results.

[0027] 7. The biomimetic self-repair testing method for logistics roadside equipment of the present invention ensures the continuity of key data acquisition and enhances the reliability of the system in long-term unattended testing by using a redundancy mechanism between primary and backup cameras and an automatic fault switching mechanism.

[0028] 8. The biomimetic self-healing test method for logistics roadside equipment of the present invention effectively controls the risk of performance degradation of the pre-trained repair state prediction model while introducing new knowledge through a conservative incremental learning update strategy, thus ensuring the stability of the evolution process of the pre-trained repair state prediction model.

[0029] 9. The biomimetic self-healing testing method for logistics roadside equipment of the present invention achieves energy saving and consumption reduction in the testing process by intelligently regulating the energy consumption of the environmental simulation chamber by utilizing the prediction results, while ensuring the effectiveness of the test.

[0030] 10. The biomimetic self-healing test method for logistics roadside equipment of the present invention combines microscopic morphology observation with conductivity detection to verify the effectiveness of initial damage and ensure that it can trigger the self-healing mechanism, thereby improving the accuracy and success rate of the test.

[0031] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0032] Figure 1 This is a flowchart of the biomimetic self-healing test method for logistics roadside equipment in Embodiment 1 of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0034] It should be understood that terms such as "having," "comprising," and "including" used in the embodiments of this application do not exclude the presence or addition of one or more other elements or combinations thereof. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of components in a specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed to" or "set on" another element, it can be directly on the other element or may have an intervening element present. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element through an intervening element. Descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0035] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0036] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0037] This invention provides a biomimetic self-healing testing method for logistics roadside equipment, which includes the following steps: S1. Real-time acquisition of multi-source time-series monitoring data of the roadside equipment under test during the self-repair process. The multi-source time-series monitoring data includes at least a sequence of visual images characterizing changes in its surface morphology (acquisition frequency 1 frame / second), acoustic emission signals characterizing micro-deformation of the equipment shell and connectors (sampling frequency 1 MHz), and environmental parameter data streams containing roadside high temperature and ultraviolet radiation characteristics (acquisition frequency 1 Hz). Time-series alignment of multi-source data is achieved through a high-precision timestamp module (synchronization accuracy ≤ 1 ms) to ensure consistency of the data time dimension. S2: Extract features from multi-source time-series monitoring data to obtain a time-series feature vector representing the self-healing process; S3: Input the time-series feature vector into the pre-trained repair status prediction model, and output the current self-repair process status assessment and the estimated remaining repair time required to reach the predetermined repair completion standard; wherein, the pre-trained repair status prediction model is a time-series deep learning model trained based on historical self-repair data; the current self-repair process status assessment output is a normalized repair completion score between 0 and 1 that integrates multi-source time-series monitoring data; the estimated remaining repair time is the time span required for the pre-trained repair status prediction model to dynamically extrapolate and predict the repair completion score to reach the preset threshold based on the historical sequence of the time-series feature vector and the current state; S4: Based on the output status assessment and estimated remaining repair time, the system automatically judges the test progress and automatically generates a test termination command and test report when the predetermined repair completion standard is met. The repair completion threshold in the predetermined repair completion standard can be dynamically adjusted according to the type of roadside equipment under test (e.g., monitoring pole housing, sensor protective cover), with an adjustment range of 0.85 to 0.98. The automatic judgment logic is as follows: real-time monitoring of the repair completion score; when the repair completion score reaches or exceeds the preset threshold for the first time, a test termination command is immediately triggered. The test report includes at least: a curve showing the change in repair completion over time (marking key environmental parameter change nodes), the final repair completion score, a comparison of the estimated remaining repair time and the actual repair time, an analysis of the impact of different environmental stresses (e.g., temperature, ultraviolet radiation) on the repair rate (calculating the contribution ratio of each stress using the controlled variable method), and a quality assessment of multi-source monitoring data (e.g., average visual image sharpness, acoustic emission signal-to-noise ratio).

[0038] In the above technical solution, firstly, the sample of the logistics roadside equipment to be tested, such as a metal sign coated with a microcapsule-type self-healing coating, is placed on a test platform with data acquisition capabilities. At the start of the test, the system uses high-definition industrial cameras deployed around the sample to acquire a sequence of images of the pre-set damaged areas on the sample surface at a rate of one frame per second, forming a visual data stream. Simultaneously, an acoustic emission sensor is coupled to the back of the sample to acquire elastic wave signals generated by changes in the internal microstructure of the material during the repair process; the sampling frequency is set to 1 MHz. In addition, temperature and humidity sensors and ultraviolet intensity sensors record the temperature, humidity, and ultraviolet irradiance data of the test environment in real time. The temperature monitoring range is -20℃ to 60℃, and the ultraviolet irradiance monitoring range is 0 to 150 W / m². 2 These visual, acoustic, and environmental data are synchronously timestamped, forming a multi-source time-series monitoring data stream. Specifically, the visual image sequences, acoustic emission signals, and environmental parameter data streams are all marked with their acquisition time using the system-integrated GPS timestamping module (synchronization accuracy ≤ 1 ms). For differences in sampling frequencies (e.g., acoustic emission 1 MHz vs. environmental parameters 1 Hz), linear interpolation is used to upsample the lower-frequency data, ensuring a one-to-one correspondence between sampling points in the time dimension of the multi-source data and avoiding time shifts during fusion.

[0039] Subsequently, the system performs automated feature extraction on the acquired multi-source data. For visual image sequences, a deep learning-based image segmentation algorithm is used to automatically identify and calculate the area and perimeter of the damaged region in each frame. For acoustic emission signals, time-frequency analysis is used to calculate the cumulative number of acoustic emission events per second, the average amplitude of the signal, and the total energy in real time. For environmental parameters, instantaneous readings are recorded and their short-term trends are calculated. Next, these primary visual, acoustic, and environmental features extracted from different sources are fused. Specifically, eight feature values ​​at a given time point—damage area, damage perimeter, acoustic event count, signal amplitude, signal energy, ambient temperature, ambient humidity, and ultraviolet intensity—are combined into a temporal feature vector. This feature vector forms a sequence over time, used to characterize the dynamic changes in the self-healing process.

[0040] Then, the aforementioned temporal feature vector sequence is input into a pre-trained repair status prediction model. This model is a temporal deep learning model based on a long short-term memory network structure, and its training data comes from more than a thousand historical self-repair test records. After receiving the real-time input feature vector sequence, the model outputs two key prediction results: one is a normalized repair completion score between 0 and 1, where 0 indicates no repair and 1 indicates complete repair; for example, an output of 0.85 indicates that 85% of the repair has been completed. The other is the estimated remaining repair time; for example, it is predicted that it will take another 3 hours and 25 minutes to reach the set repair completion threshold of 0.95. This process is dynamic and continuous, and the model continuously updates the prediction results based on the latest data.

[0041] The pre-trained repair state prediction model is a temporal deep learning model based on a long short-term memory (LSTM) network structure. Its specific construction and training methods are as follows: Model structure: The model consists of an input layer, a feature extraction layer, a temporal processing layer, and an output layer.

[0042] Input layer: Receives a temporal feature vector with dimension 11, which is a weighted fusion of visual primary features (2-dimensional), acoustic primary features (3-dimensional), and environmental primary features (6-dimensional, including instantaneous values ​​and trends of temperature, humidity, and ultraviolet intensity).

[0043] Feature extraction layer: A one-dimensional convolutional layer (Conv1D) is connected after the input layer. It has 64 kernels, a kernel size of 3, a stride of 1, and uses the ReLU activation function to initially extract local feature patterns.

[0044] Temporal processing layer: The output of the feature extraction layer is fed into a three-layer stacked LSTM network for temporal modeling. The first LSTM layer has 128 units and returns the complete sequence; the second LSTM layer has 64 units and also returns the complete sequence; the third LSTM layer has 32 units and only returns the output of the final time step.

[0045] Output layer: The output of the temporal processing layer is passed through a fully connected layer with 16 neurons (using ReLU activation) and finally connected to a linear output layer with 2 neurons. The two neurons in this output layer correspond to the normalized repair completion score (constrained to the range of 0-1 using the Sigmoid function) and the estimated remaining repair time (directly outputting the time value in minutes).

[0046] Model training: The model was pre-trained based on a large-scale historical self-healing test dataset. The specific training steps and parameters are as follows: Training data preparation: 1500 complete self-healing test data sequences were selected from the historical database to form a training set. Each set of data contains a full-cycle time-series feature vector sequence, and its label (i.e., true value) is determined in the following way: the repair completion score is obtained by normalizing the mechanical property recovery rate test results (such as microhardness recovery rate) of the samples after the test; the true value of the expected remaining repair time is the actual time span from the current time point to the first time the repair completion score reaches the threshold (e.g., 0.95) in the same test.

[0047] Loss function: The loss function used for training is a composite loss function for dual outputs, specifically defined as: Total loss = MSE (Mean Score for Repair Completion) + 0.5 × MAPE (Estimated Remaining Repair Time) Where MSE is the mean squared error and MAPE is the mean absolute percentage error. This function aims to simultaneously guarantee the accuracy of score prediction and the relative accuracy of time prediction.

[0048] Training hyperparameters and strategies: The Adam optimizer was used, with an initial learning rate of 0.005. The batch size was set to 32. The total number of training epochs was 100. A dynamic learning rate scheduling strategy was employed: if the validation set loss did not decrease for five consecutive epochs, the learning rate was halved. Gradient pruning was also used with a pruning threshold of 5.0 to prevent gradient explosion during training. During training, 85% of the data was used for training, and 15% was used as a validation set to monitor model performance and prevent overfitting.

[0049] Through the above construction and training, the obtained repair state prediction model can effectively capture the temporal dynamic features of the self-repair process and achieve high-precision state assessment and time prediction.

[0050] Finally, the system makes automated decisions based on the model's output. It monitors the repair completion score in real time, and when the score first reaches or exceeds the preset threshold of 0.95, the system immediately and automatically issues a test termination command, stopping all data collection activities. Simultaneously, it automatically generates a structured test report. This report includes a graph showing the repair completion score over time throughout the entire test cycle, the final repair completion score of 0.96, and a comparative analysis between the predicted remaining repair time of 3 hours and 25 minutes and the actual time taken from start to finish of the test of 3 hours and 30 minutes. This completes one automated testing process.

[0051] Existing technologies typically rely on a combination of manual, timed observation and offline analysis. For example, in existing technologies, technicians need to take photos of the damaged area every few hours, such as every four hours, using a portable microscope. They then compare the photos from different time points and subjectively judge, based on their personal experience, whether the crack has closed or to what extent it has been repaired. The entire testing cycle can take 24 to 48 hours or even longer, requiring multiple interventions, resulting in low efficiency and difficulty in accurately determining the testing endpoint.

[0052] This invention achieves full-process automation and intelligence. In data acquisition, it employs multi-sensor synchronous continuous monitoring, replacing the discrete, manual sampling methods of existing technologies. For state assessment, it outputs a quantitative completion score through a deep learning model, completely replacing subjective, qualitative visual judgment. In test process management, it automatically determines and terminates tests based on model predictions, eliminating the need for manual intervention and decision-making, significantly reducing the test cycle from tens of hours to several hours. Regarding output results, the structured report provided by this invention, which includes a comparison of predicted and actual times, far surpasses the information content and objectivity of simple before-and-after photo comparisons or final mechanical performance test reports in existing technologies. Therefore, this invention achieves substantial improvements and enhancements over existing technologies in terms of testing efficiency, result objectivity, automation level, and information output value. This invention differs from existing single-sensor monitoring or manual judgment schemes by using multi-source time-series data fusion of visual, acoustic emission, and environmental parameters, along with a pre-trained time-series deep learning model, to achieve dynamic quantitative assessment of the repair state.

[0053] The method of this invention is applicable to mid-to-high-end logistics roadside equipment with biomimetic self-healing function (cost of a single unit ≥ 5000 yuan), such as intelligent monitoring poles and high-precision sensor housings, but not to low-cost simple signs or other equipment that do not require high-precision self-healing testing.

[0054] In another technical solution, step S2 involves feature extraction from the multi-source time-series monitoring data, specifically including: S21: Extract primary features from visual image sequences, acoustic emission signals, and environmental parameter data streams respectively. From the visual image sequence, extract the area and perimeter of the damaged area as visual primary features using an image segmentation algorithm. From the acoustic emission signal, extract event counts, signal amplitude, and energy as acoustic primary features. From the environmental parameter data stream, extract the instantaneous values ​​and changing trends of temperature and humidity as environmental primary features. S22: The visual primary features, acoustic primary features, and environmental primary features are weighted and fused to generate a temporal feature vector; wherein, the weight coefficients of each primary feature are learned through backpropagation during the training process of the repair state prediction model, and can be dynamically adjusted according to different environmental primary features (such as extreme temperature, strong vibration). The adjustment process adopts a gradient pruning algorithm (pruning threshold 0.5) to avoid weight oscillation and ensure stable model output. The weight coefficients of each primary feature are obtained through backpropagation during the training process of the repair state prediction model, and can be dynamically adjusted according to different environmental primary features.

[0055] In the above technical solution, feature extraction and fusion of multi-source time-series monitoring data is a key step in implementing the biomimetic self-healing testing method for logistics roadside equipment. The specific implementation of this process is as follows: The system first processes the real-time acquired visual image sequence. It employs an image segmentation algorithm based on the U-Net architecture, which has been pre-trained on thousands of labeled damaged images. For each input frame, the algorithm automatically outputs the precise pixel-level contour of the damaged region. Based on this contour, the system calculates two key primary visual features: the area of ​​the damaged region, measured in square pixels (e.g., the current damaged area is calculated to be 1250 square pixels); and the perimeter of the damaged region, measured in pixels (e.g., the current perimeter is calculated to be 520 pixels).

[0056] Simultaneously, the system processes the acquired acoustic emission signals. The raw acoustic emission signal first passes through a bandpass filter to remove frequency components below 50 kHz and above 400 kHz, eliminating low-frequency environmental noise and high-frequency interference. Subsequently, feature extraction is performed on the preprocessed signal: the number of acoustic emission events exceeding a preset threshold of 40 dB per second is counted, for example, 15 events are detected; the root mean square value of the signal amplitude during this time period is calculated, for example, 0.85 V; and the total energy of the signal is calculated by integrating the square of the signal, for example, 120 mJ. These three values ​​together constitute the primary acoustic feature set.

[0057] For environmental parameter data streams, the system records instantaneous values ​​of temperature, humidity, and ultraviolet (UV) intensity, such as temperature 25 degrees Celsius, humidity 60% RH, and UV intensity 80 W / m². 2 Meanwhile, the trend of these parameters is obtained by calculating the slope of their changes over the past minute; for example, the temperature is slowly rising at a rate of 0.1°C per minute.

[0058] After obtaining all the aforementioned primary features, the system performs the core weighted fusion step. It combines 11 feature values—two visual features, three acoustic features, and three instantaneous and three trend values ​​from the environment—into an initial vector. The weight of each feature before fusion is not fixed but is automatically learned during model training via backpropagation and can be dynamically adjusted according to environmental conditions. For example, in environments with high UV intensity, the weight of visual features characterizing the material surface state may be automatically increased; while in environments with large temperature fluctuations, the weight of acoustic emission features may play a more significant role. This dynamic weight allocation mechanism allows the system to adaptively focus on the key indicators that best reflect the repair status in the current environment. Finally, after weighting and normalization, a temporal feature vector that comprehensively characterizes the self-repair process is generated for use by subsequent prediction models.

[0059] Existing techniques typically employ simple data stacking or fixed empirical formulas for feature processing. For example, some existing techniques may simply extract the damage area from an image while simultaneously recording acoustic emission event counts, then directly concatenate these two values, or combine them according to a pre-set, fixed ratio (e.g., 70% visual features and 30% acoustic features) to form the input features. This approach completely ignores the fact that the importance of various features changes under different environmental conditions, and it also fails to optimize the contribution of features through data-driven methods.

[0060] This invention achieves an intelligent adaptive feature fusion. Firstly, regarding the comprehensiveness of feature extraction, this invention not only extracts basic instantaneous values ​​but also captures the changing trends of environmental parameters, providing richer state information. More importantly, in terms of feature fusion strategy, this invention automatically learns the initial weights of each feature through model training and introduces a mechanism for dynamically adjusting weights based on environmental parameters. This enables the feature fusion process to have self-optimization capabilities. For example, the system can learn to rely more on acoustic emission signals to assess internal repair activity in high-temperature environments, while relying more on high-precision visual measurement results in stable laboratory environments. This dynamically weighted fusion method, compared to the rigid fixed weights or simple splicing methods in existing technologies, can generate higher-quality, more representative temporal feature vectors, thus laying a more accurate and reliable data foundation for subsequent state prediction, directly improving the accuracy of the entire testing system's evaluation results and its adaptability under different operating conditions.

[0061] In another technical solution, in step S1, the multi-source time-series monitoring data is processed in a programmable multi-factor environmental simulation test chamber; wherein, the test chamber is configured to dynamically and synchronously apply at least two of the following environmental stresses according to a preset environmental stress spectrum during the test cycle of the self-healing process: Cyclic temperature stress: It cycles between high temperature limit and low temperature limit according to a preset time program; Mechanical vibration stress: Simulates the vibration frequency and amplitude of the roadside environment near logistics transportation vehicles or equipment; Ultraviolet radiation stress: Simulates the ultraviolet components in outdoor sunlight and irradiates at a preset irradiance. The environmental parameter data stream originates from real-time monitoring data of the aforementioned environmental stresses applied within the test chamber.

[0062] In the above technical solution, the test chamber employs a multi-environment stress synchronous application strategy: the inflection point of temperature cycling (such as switching from cooling to heating) and the switching of mechanical vibration frequency (such as switching from 30 Hz to 50 Hz) are triggered synchronously; ultraviolet irradiation is maintained at full power when the temperature is between 30 and 50℃ (simulating the summer noon environment), and the irradiance is reduced to 50% when the temperature is <0℃ or >60℃, ensuring that the synergistic effect of different stresses conforms to the actual roadside environment.

[0063] In the implementation of the biomimetic self-healing testing method for logistics roadside equipment, the construction and operation of the environmental simulation test chamber is a key step. This test chamber is a sealed space measuring 2 m × 2 m × 2 m, with its inner walls covered by a highly reflective material to ensure uniform distribution of environmental stress.

[0064] The core control unit of the test chamber performs the following operations based on a preset environmental stress spectrum. In terms of cyclic temperature stress simulation, the control unit drives the compressor cooling and resistance wire heating system to complete four full cycles of temperature inside the chamber within 24 hours. In each cycle, the temperature decreases uniformly from the high temperature limit of 60°C to the low temperature limit of -10°C within 3 hours (cooling rate ≈ 23.3°C / h), and then rises uniformly from -10°C to 60°C within the next 3 hours (heating rate ≈ 23.3°C / h), ensuring that four cycles are completed within 24 hours.

[0065] Mechanical vibration stress was achieved using an electromagnetic vibration table mounted beneath the sample fixture. The vibration table generated random vibrations in the frequency range of 30 Hz to 50 Hz, a frequency spectrum specifically designed to simulate road surface vibrations caused by heavy logistics transport vehicles. Vibration acceleration was maintained between 0.5 g and 1.2 g and continuously applied to the test sample.

[0066] Ultraviolet radiation stress is provided by ultraviolet lamp arrays arranged on the cabin ceiling. The lamp arrays emit ultraviolet light with wavelengths mainly distributed in the range of 280 nm to 400 nm, and the irradiance is stable at 120 W / m². 2 This intensity is equivalent to the intensity level of the ultraviolet component in midday sunlight during summer.

[0067] Importantly, these environmental stresses were applied simultaneously. For example, at a temperature of 40°C, the sample was simultaneously subjected to vibrations at a frequency of 45 Hz and 120 W / m². 2 The ultraviolet radiation. This multi-stress coupling effect realistically reproduces the complex environmental conditions that roadside equipment in logistics is subjected to outdoors.

[0068] The environmental parameter data stream originates directly from monitoring sensors installed within the test chamber. These include a PT100 platinum resistance temperature sensor with a measurement accuracy of ±0.5℃; a capacitive humidity sensor with a measurement accuracy of ±3%RH; and a UV-A band ultraviolet sensor with a measurement accuracy of ±5%. These sensors continuously collect environmental data within the chamber at a frequency of 1 Hz, forming the environmental parameter data stream.

[0069] Existing technologies typically employ single-factor environmental chambers for testing, such as constant temperature and humidity chambers or independent ultraviolet aging chambers. In these existing technologies, the test conditions are often static and singular; for example, placing the sample in a constant 50°C environment for repair testing, or first completing mechanical testing on a vibration table and then transferring it to an ultraviolet chamber for light exposure testing. This isolated, single-factor testing method completely fails to reproduce the complex situation of multiple environmental stresses acting simultaneously and dynamically changing in a real roadside environment.

[0070] Another significant limitation of existing technologies is the disconnect between environmental parameters and the testing process. In traditional methods, environmental parameters are typically recorded merely as background conditions for testing, rather than as key input data for analyzing remediation actions. The lack of deep integration between environmental simulation devices and monitoring systems leads to substantial discrepancies between test results and actual performance in actual use.

[0071] This invention achieves a breakthrough improvement through a programmable multi-factor environmental simulation test chamber. First, it can simultaneously apply multiple environmental stresses, such as temperature, vibration, and ultraviolet radiation, and these stresses dynamically change according to a preset spectrum, which highly matches the actual working environment of roadside equipment. Second, environmental parameter monitoring is deeply integrated into the entire testing process, with environmental data streams becoming indispensable input information for evaluating self-healing performance. This highly integrated multi-factor environmental simulation method enables test results to more accurately predict the durability and repairability of materials in actual use environments, significantly improving the reliability and practical value of the test.

[0072] In another technical solution, step S4 is followed by: S5. After generating the test report, the multi-source time-series monitoring data with complete self-healing process obtained in this test and the final measured repair time are used as a new set of labeled samples. The pre-trained repair state prediction model is incrementally learned using new labeled samples to update the model parameters for prediction in subsequent test tasks.

[0073] In the above technical solution, after the biomimetic self-healing test process of the logistics roadside equipment is completed, the system executes an incremental learning process for the model. Once a complete test is completed and a test report is generated, the system automatically packages the entire test's data into a new set of labeled samples. This set of samples includes all multi-source time-series monitoring data from the start to the end of the test, as well as the final determined measured repair time. For example, in a test lasting 8.5 hours, the complete visual image sequence, acoustic emission signals, environmental parameter data stream, and the final measured repair time of 8.5 hours together constitute a new training sample with a clear label.

[0074] Subsequently, the system initiates the incremental learning process. The pre-trained repair state prediction model employs a Long Short-Term Memory (LSTM) network architecture. During incremental learning, a small learning rate, such as 0.001, is set to control the magnitude of model parameter updates and avoid overfitting to new samples. The training process uses the Adam optimizer, adding this new labeled sample to the base training set containing 500 historical samples to form the incremental training set. The model is trained on this dataset for 50 epochs. The training focuses on enabling the model to learn the repair patterns and time-related relationships inherent in this new sample, such as the repair rate characteristics of materials under specific combinations of temperature fluctuations and UV intensity.

[0075] After training, the model's parameters are updated. This updated model will be used for predictions in all subsequent new test tasks. For example, in the next round of testing, when the system detects similar environmental conditions and damage characteristics, the updated model can draw on the experience learned from the previous test, potentially predicting the repair completion time earlier or making a more accurate judgment on a specific repair stall phenomenon. In this way, with each test, the system's predictive ability is fine-tuned and enhanced, enabling it to gradually adapt to new self-healing materials or environmental conditions that were not previously adequately covered.

[0076] One common approach in existing technologies is to employ a static model strategy. Under this strategy, once a repair condition prediction model is trained on an initial historical dataset, its parameters are fixed and remain unchanged in all subsequent tests. For example, a model trained on 200 sets of data collected a year ago will be used indefinitely, even though dozens of new test data sets have been accumulated during that time; the model cannot learn new knowledge from them. When encountering entirely new material systems or significantly different environmental stresses, the predictive accuracy of this static model will decrease significantly.

[0077] Another existing technique is to periodically retrain the model completely. This approach involves collecting all new data generated during a fixed long period, such as every six months, merging it with the original historical data to form a larger dataset, and then training a new model from scratch. This process is computationally expensive, takes several days, and carries the risk of "catastrophic forgetting," where the new model may lose important patterns learned in the original data that are less common in the newly merged dataset.

[0078] The incremental learning mechanism implemented in this invention provides an elegant and efficient solution. It abandons the rigidity of static models and the redundancy of complete retraining, employing a gradual learning path. Each test is not the end point, but rather an opportunity for model evolution. The system can absorb new knowledge in near real-time, enabling continuous fine-tuning and performance improvement of the model, while avoiding the computational burden of large-scale retraining and the catastrophic forgetting problem. This allows the testing system to truly adapt to technological developments and environmental changes, consistently maintaining cutting-edge predictive capabilities and accuracy.

[0079] In another technical solution, in step S5, a data credibility verification step is added before incremental learning, specifically as follows: Based on a preset set of data quality rules, the multi-source time-series monitoring data that constitute the new labeled samples are verified. The data quality rule set includes at least: whether the sharpness of the visual image sequence is higher than a set sharpness threshold, whether the signal-to-noise ratio of the acoustic emission signal is higher than a set signal-to-noise ratio threshold, and whether the values ​​of the environmental parameter data stream are within a preset effective range. New labeled samples are only used for incremental learning of the repair status prediction model if they pass the data credibility verification.

[0080] In the aforementioned technical solution, data reliability verification is a crucial preliminary step in the model update stage of the biomimetic self-healing testing method for logistics roadside equipment, ensuring the quality of incremental learning. When a test is completed and the data is ready to be used for model updates, the system will first initiate the data quality verification process.

[0081] The system pre-sets a set of data quality rules to automatically verify the multi-source time-series monitoring data that constitutes the newly labeled samples. First, it evaluates the sharpness of the visual image sequences. The system calculates the edge sharpness of the damaged areas in each frame of the image, requiring that its average value must be higher than a set sharpness threshold of 150. If the average sharpness value of the image sequence is 138, the sample fails this verification.

[0082] Secondly, a signal-to-noise ratio (SNR) analysis is performed on the acoustic emission signal. The system separates the effective signal component and the ambient noise component from the acquired raw acoustic emission signal and calculates the power ratio between the two. This SNR must be higher than a set SNR threshold of 20 dB. For example, if the SNR of a certain segment of acoustic emission signal is detected to be 18.5 dB, the sample fails this verification.

[0083] Simultaneously, the system verifies the validity of environmental parameter data streams. Temperature readings must be within the valid range, such as between -20°C and 80°C. Humidity readings must be within the valid range of 0%RH to 100%RH. Ultraviolet intensity readings must be within 0 W / m². 2 Up to 200 W / m 2 Within the valid range. If the reading of an environmental parameter exceeds its corresponding valid range, such as a detected temperature of 85℃, which exceeds the upper limit, then the environmental parameter data of that sample is invalid.

[0084] The system executes a full-element verification logic. Only when a new labeled sample meets the standards in all verification items can it pass the data reliability verification. Failure in any verification item will result in the sample being excluded from the incremental learning process. For example, even if a sample has a visual sharpness of 180 and an acoustic emission signal-to-noise ratio of 25 dB, it still cannot be used for model updates if the ambient temperature reading is outside the effective range. Samples that pass verification are tagged as quality qualified and enter the subsequent incremental learning process; samples that fail verification are archived in the problem database for later analysis.

[0085] Compared with existing technologies, the data credibility verification method of this invention achieves a significant breakthrough in data quality control. Existing technologies typically employ simple data integrity checks or completely skip the data quality verification process during model updates. A common practice in these existing technologies is to directly use all acquired data for model training once the data acquisition process is technically completed, with at most some basic data format verification.

[0086] More advanced methods in the prior art may perform simple validations on a single type of data, such as checking only whether image files are corrupted or verifying only the continuity of data acquisition timestamps. However, these methods lack in-depth assessment of data content quality and cannot identify substantial data quality problems caused by sensor performance degradation, environmental interference, or equipment failure. A typical example is that existing technologies may receive a set of images that are clear but misfocused, or acoustic emission signals with a severely insufficient signal-to-noise ratio, and use this low-quality data for model updates, leading to a gradual degradation of model performance.

[0087] This invention establishes a systematic, multi-dimensional data credibility verification mechanism based on specific quantitative indicators. By simultaneously and rigorously controlling visual clarity, acoustic signal-to-noise ratio, and the validity of environmental parameters, this invention can effectively identify and filter various data quality issues. This proactive data quality management model fundamentally blocks low-quality data from entering the learning process, ensuring the reliability and validity of the data used for incremental learning and providing a solid guarantee for the stable improvement of model performance. Compared with existing technologies, this invention significantly reduces the risk of model performance degradation due to data contamination and improves the long-term reliability of the entire testing system.

[0088] In another technical solution, a standardized initial damage preparation step is included before step S1, specifically: Using a programmable laser etching device integrated into the testing system, initial damage with consistent geometry and physical dimensions is prepared on the surface of the self-healing material of the logistics roadside equipment under test, based on pre-stored standardized damage patterns and depth parameters.

[0089] The initial damage is the starting target monitored by the visual image sequence and acoustic emission signal in step S1.

[0090] In the above technical solution, the standardized initial damage needs to meet the requirement of "penetrating the pre-embedded conductive fiber mesh" in the subsequent damage effectiveness verification. The etching depth, shape and other parameters need to match the fiber mesh embedding depth and mesh spacing to ensure that the initial damage has the basis for triggering the self-healing mechanism.

[0091] When preparing the initial damage using a programmable laser etching device, the etching depth is set to 50 μm to ensure that this depth is greater than the embedding depth of the pre-embedded conductive fiber mesh (30-40 μm). This avoids damage that cannot penetrate the fiber mesh due to insufficient etching depth, and ensures that subsequent effectiveness verification can proceed normally.

[0092] In the initial stage of the biomimetic self-healing test process for roadside equipment, a standardized initial damage preparation procedure is performed. This step is completed using a programmable laser etching device integrated into the test system. This device uses a fiber laser with a wavelength of 1064 nm, a maximum output power of 50 W, and a focused spot diameter that can be controlled within 20 μm.

[0093] During operation, the sample of the logistics roadside equipment to be tested, such as a 100 mm × 100 mm metal specimen coated with a microcapsule self-healing coating, is first fixed on the worktable of the laser etching equipment. The equipment control software calls up a pre-stored standardized damage pattern, which is typically designed as a straight line with a length of 10 mm. The laser etching depth parameter is set to 50 μm. This depth is carefully selected to ensure effective penetration of the surface coating and triggering of the self-healing mechanism without causing structural damage to the substrate.

[0094] The laser etching process is executed automatically under computer program control. The laser head moves along a preset path at a scanning speed of 300 mm / s, with a pulse frequency of 20 kHz, ultimately creating initial damage on the sample surface with a geometry and physical dimensions highly consistent. The damage is approximately 25 μm wide and 50 μm deep, exhibiting good repeatability, and its morphology, observed under a microscope, displays clear and uniform groove features.

[0095] This standardized initial damage, thus prepared, serves as the explicit starting point for monitoring visual image sequences and acoustic emission signals during subsequent testing. All samples are tested from this perfectly consistent damage state, providing a unified benchmark for monitoring the subsequent self-healing process and comparing data.

[0096] The standardized initial damage preparation method of this invention represents a significant improvement in the standardization of testing and the reliability of results compared to existing technologies. Existing technologies typically employ manual mechanical scratching to prepare initial damage. In this method, an operator uses a handheld scratching tool, such as a diamond indenter of a hardness tester or a specific sized surgical blade, to create cracks or scratches by applying hand force across the material surface.

[0097] This traditional manual method has significant limitations. First, due to the inability to precisely control the applied force, scratch speed, and angle, the depth, width, and even shape of the damage prepared each time are subject to large random human error. For example, the depth of a scratch may fluctuate drastically between 30μm and 70μm, and the scratch edge may appear irregularly jagged rather than an ideal straight line. Second, the consistency of damage cannot be guaranteed between different operators, or even between different operations by the same operator. This significant difference in initial conditions directly leads to a lack of comparability in subsequent self-healing test results. It is difficult to effectively summarize, analyze, and scientifically evaluate test data from different batches, seriously affecting the repeatability and credibility of the test.

[0098] The program-controlled laser etching technology employed in this invention fundamentally eliminates the uncertainties and random errors introduced by human factors. By precisely controlling the laser energy, scanning path, and focusing position, it ensures that the initial damage, with almost identical geometric dimensions and physical morphology, is reproduced on every sample. This highly standardized preparation method establishes a unified and impartial starting point for self-healing testing, enabling rigorous comparative analysis of test results obtained at different times, locations, and for different batches of materials, greatly enhancing the scientific value and engineering guidance significance of the test data.

[0099] In another technical solution, in step S1, the acquisition of the visual image sequence is jointly ensured by a primary camera and a backup camera. Step S1 includes a fault diagnosis step while simultaneously acquiring multi-source time-series monitoring data of the roadside equipment under test during its self-repair process: real-time monitoring of the operating status of the main camera and the quality of the visual image sequence. When the primary camera is determined to be faulty or the image quality is consistently unsatisfactory, the data acquisition source is automatically switched to the backup camera to ensure the continuity of multi-source time-series monitoring data.

[0100] In the aforementioned technical solution, the visual image sequence acquisition system employs a primary / backup redundancy design in the data acquisition stage of the biomimetic self-healing test for logistics roadside equipment. The system is equipped with one primary industrial camera and one backup industrial camera of the same model. Both cameras have a 5-megapixel resolution and are fixed side-by-side on a bracket, aiming at the damaged area of ​​the sample under test from essentially the same perspective. The primary camera acquires images at a rate of 1 frame per second and transmits the image data to the data processing system in real time via a gigabit Ethernet interface.

[0101] While the system acquires multi-source time-series monitoring data in real time, a fault diagnosis step runs in parallel. This diagnostic step continuously monitors two key indicators: first, the operating status of the primary camera, including its power supply voltage, temperature, network connection heartbeat signal, etc.; second, the quality of the visual image sequence transmitted back by the primary camera, with the core evaluation indicator being image sharpness.

[0102] Image sharpness is calculated based on the Laplacian variance algorithm. The system calculates the sharpness for each incoming image frame. If the sharpness values ​​of five consecutive images are all below the preset sharpness threshold of 150, the image quality is deemed unsatisfactory. Simultaneously, if the system detects a network connection interruption of the main camera for more than 3 seconds, or its internal temperature exceeding 75°C, it is considered a camera hardware malfunction.

[0103] Once the fault diagnosis process determines that the primary camera is faulty or the image quality consistently fails to meet requirements, the system will immediately execute an automatic switchover. The switchover process is completed within 100 ms, including: disconnecting the data link with the primary camera, sending a start command to the backup camera and establishing a data connection, and simultaneously adjusting the acquisition parameters to match those of the primary camera. After the switchover is complete, all subsequent visual image data streams will come from the backup camera, ensuring that the continuity of the visual dimension in the multi-source time-series monitoring data is not interrupted by the failure of a single device. Throughout the entire process, the acquisition of acoustic emission signals and environmental parameters remains unaffected and continues normally.

[0104] Compared with existing technologies, the redundant acquisition and fault self-diagnosis scheme of this invention achieves a significant improvement in system reliability. Existing technologies typically use a single camera for image acquisition. In this configuration, if the single camera malfunctions in any way, such as lens damage, focus misalignment, hardware failure, or data transmission interruption, the entire visual monitoring channel becomes ineffective.

[0105] In existing technologies, such faults are often only discovered when processing data after testing, or rely on irregular on-site inspections by operators. Once a fault occurs, the tests already performed may be rendered useless due to the loss of critical data and need to be restarted. This not only wastes time and resources but may also affect the normal execution of the test plan. This single point of failure risk is particularly prominent for long-term self-healing tests that need to last for several days or even weeks.

[0106] This invention constructs a highly available visual acquisition system through hardware redundancy configuration of primary and backup cameras and a real-time fault diagnosis and switching mechanism. This system not only can perceive the working status and data quality of the acquisition devices in real time, but more importantly, it can achieve rapid and automatic switching in the event of a fault, maximizing the continuity of data acquisition. This design significantly reduces the risk of test interruption or data loss due to the failure of a single device, making it particularly suitable for unattended testing scenarios requiring long-term stable operation, and greatly improving the reliability and data integrity assurance capabilities of the entire testing system.

[0107] In another technical solution, step S5, which involves incrementally learning the pre-trained repair state prediction model using new labeled samples, specifically includes: S51. Incrementally update a copy of the remediation status prediction model using new labeled samples, updating only the parameters of the fully connected layers. If the new samples contain uncovered environmental features (such as extreme UV radiation > 150 W / m²), the update will be cancelled. 2 If the learning rate is 0.0001, then the last convolutional layer of the feature extraction layer is fine-tuned to ensure that the model can capture new features while protecting the trained general feature extraction capability; thus obtaining the candidate model. S52: Using historical data corresponding to the new labeled samples, evaluate the prediction accuracy of the candidate model and the current official model in parallel; S53: Replace the current formal model with the candidate model only when the prediction accuracy of the candidate model is higher than that of the current formal model, thus completing incremental learning.

[0108] In the above technical solution, a safe incremental learning strategy is adopted in the model update stage of the biomimetic self-healing test method for logistics roadside equipment. When a new, quality-verified labeled sample is ready for model update, the system first creates a complete copy of the currently used repair state prediction model. This model is a composite architecture containing convolutional layers, long short-term memory network layers, and fully connected layers.

[0109] The first step in incremental updates is fine-tuning the parameters of the model replica. The system combines the new sample with the 50 most recently collected historical samples to form a small incremental training set. Using the Adam optimizer, this model replica is trained at a learning rate of 0.001 for 50 epochs. During this process, the parameters of the convolutional and long short-term memory (LSTM) layers in the model are completely frozen; only the weights and biases of the fully connected layers are updated. This partial update strategy allows the model to learn specific patterns in the new samples while preserving the general feature extraction capabilities learned from large amounts of data. The updated model replica is then called the candidate model.

[0110] The second step is parallel performance evaluation. The system prepares an independent validation set containing 200 historical test data points covering various operating conditions. The candidate model and the current official model are simultaneously tested on this validation set, and their respective prediction accuracy metrics are calculated, primarily examining the root mean square error of the repair completion score and the mean absolute percentage error of the remaining repair time. For example, the candidate model achieves an overall accuracy of 92.5% on the validation set, while the current official model achieves 93.1%.

[0111] The third step is the safe replacement decision. The system compares the evaluation results of the two models, and only replaces the current official model with the candidate model if the candidate model's prediction accuracy is higher than that of the current official model. In this example, because the candidate model's accuracy of 92.5% is lower than the current official model's 93.1%, the system will retain the current official model, abandon this update, and record this update attempt. The entire process ensures that model performance does not degrade during the update process.

[0112] Compared with existing technologies, the secure incremental learning method of this invention achieves significant improvements in the stability and reliability of model updates. Existing technologies typically employ two simple strategies when updating models: one is to completely retrain the model by merging new data with all historical data, and the other is to perform incremental updates on all parameters of the entire model.

[0113] A complete retraining strategy requires significant computational resources and time. For datasets containing thousands of samples, a single retraining session can take tens of hours, failing to meet the rapid adaptation requirements of testing systems. While incremental updates to all parameters are less computationally intensive, they are highly susceptible to catastrophic forgetting, meaning the model rapidly loses previously learned crucial knowledge while adapting to new data, resulting in a sharp decline in historical performance predictions.

[0114] This invention achieves safe and reliable model evolution through a triple guarantee mechanism of model replication, partial parameter updates, and performance verification. Partial parameter updates significantly reduce computational overhead, enabling the model to quickly adapt to new knowledge; while performance comparisons based on independent validation sets ensure that each update represents a truly beneficial improvement. This conservative and robust update strategy effectively protects existing model capabilities while introducing new knowledge, preventing performance fluctuations or regressions during the update process, and guaranteeing the long-term predictive accuracy and stability of the test system.

[0115] In another technical solution, the cyclical execution of steps S3 and S4 also includes a dynamic control step linked to the environmental simulation and prediction steps, specifically: Receive the repair completion score output in real time from the repair status prediction model; Based on historical data of repair completion scores, the real-time repair rate is calculated. A fifth-order polynomial regression is used to fit the repair completion curve of the most recent 10 minutes. The derivative of the curve at the current time point is used as the real-time repair rate to avoid the calculation bias of linear regression for non-linear repair processes (such as fast repair in the early stage and slow repair in the later stage). When the real-time repair rate is lower than a preset active threshold (0.005 / min), a control command is sent to a programmable multi-factor environmental simulation test chamber to reduce the applied environmental stress intensity or frequency; if the repair rate is 0.003-0.005 / min, the environmental stress intensity is reduced by 30%; if the repair rate is <0.003 / min, the environmental stress intensity is reduced by 50% and the stress cycle period is extended to twice the original period.

[0116] In the above technical solution, during the biomimetic self-healing test of the logistics roadside equipment, the system executes a dynamic control step that is linked to the environmental simulation and prediction steps. This step runs continuously within the main test loop, and its specific implementation is as follows: The system receives a repair completion score from the repair status prediction model in real time, updated every minute. Based on historical data from this score, the system calculates the real-time repair rate. Specifically, it takes the score data from the most recent ten minutes, calculates the slope of its linear regression, and uses this slope as the current repair rate value, expressed as the change in score per minute. For example, if the score linearly increases from 0.35 to 0.45 within 10 minutes, the repair rate is 0.01 per minute.

[0117] The system presets an active repair rate threshold of 0.005 / min. When the calculated real-time repair rate is lower than this threshold, it indicates that the self-repair process has entered a slow phase, and the system sends a control command to the programmable multi-factor environment simulation test chamber to put it into a low-power operation mode.

[0118] In low-power mode, the test chamber adjusted its environmental stress application strategy as follows: the variation range of cyclic temperature stress was reduced from ±25℃ to ±5℃, and the cycle period was extended from 4 hours to 12 hours; the frequency range of mechanical vibration stress was adjusted from 30–50 Hz to 35–45 Hz, and the acceleration was reduced from 0.5–1.2 g to 0.2–0.5 g; the ultraviolet radiation stress was reduced from a continuous 120 W / m². 2 Switch to intermittent mode, turning off for 50 minutes after every 10 minutes of operation.

[0119] The system continuously monitors the repair rate. If the repair rate subsequently rises again and exceeds the active threshold, the system will automatically restore the test chamber to normal operating mode, ensuring sufficient environmental stress excitation during the active repair period. The entire dynamic control process requires no manual intervention, achieving intelligent management of energy consumption during the testing process.

[0120] Compared with existing technologies, the dynamic control method of this invention achieves a significant improvement in energy efficiency testing. Existing technologies typically employ a fixed-program approach to environmental stress application. That is, throughout the entire testing cycle, regardless of the state of the self-healing process, the environmental simulation test chamber operates according to a preset fixed program, maintaining a constant stress level and variation cycle.

[0121] For example, in existing technologies, a seven-day test environment simulation chamber would run at full power for all seven days, consuming the same amount of energy even after the self-healing process was largely completed on the third day. This fixed operating mode leads to significant energy waste, especially for materials with fast repair speeds, where most of the test time is spent in an over-testing state, resulting in unnecessary energy consumption and increased testing costs.

[0122] This invention achieves adaptive optimization of test energy consumption through real-time repair rate monitoring and intelligent linkage with environmental stress. The system can identify different stages of the self-repair process, providing sufficient environmental stimulus during active repair periods and automatically reducing environmental stress intensity during slow repair periods. This ensures test effectiveness while significantly reducing energy waste. This dynamic control capability based on the actual repair status makes the testing process more intelligent and economical, particularly suitable for long-term, continuous automated testing scenarios, significantly reducing operating costs while ensuring test quality.

[0123] In another technical solution, a damage effectiveness verification step is included between the standardized initial damage preparation step and S1, specifically: The self-healing material of the logistics roadside equipment under test is pre-embedded with a conductive fiber mesh. The fiber diameter is 50-100 μm and the mesh spacing is 1-2 mm. The fiber mesh is fixed to the material surface 30-40 μm below the surface using a water-soluble adhesive. The adhesive can be naturally degraded during the self-healing process and does not affect the material's repair performance. The initial damaged area was magnified and photographed using a 200x optical microscope with a resolution of 0.5 μm integrated into the testing system to obtain microscopic morphology images. Based on the analysis of microscopic morphology images using image recognition algorithms, and combined with conductivity detection of the damaged area (comparing the conductivity changes caused by material aging before and after damage, excluding interference from non-damage factors), it is confirmed that the initial damage has penetrated through the pre-embedded conductive fiber mesh; only when the initial damage is confirmed to be an effective penetrating damage will the subsequent self-repair process monitoring in step S1 be initiated.

[0124] In the above technical solution, during the biomimetic self-healing test of logistics roadside equipment, after the standardized initial damage preparation is completed, the system performs a damage effectiveness verification step. First, during the manufacturing stage of the self-healing material for the logistics roadside equipment under test, a conductive fiber mesh is pre-embedded. This conductive fiber uses a copper-nickel alloy material, with a fiber diameter controlled at 80 μm and a mesh spacing set at 1.5 mm, and is uniformly distributed within the material in an orthogonal mesh structure.

[0125] During verification, the initial damaged area was examined using a high-precision optical microscope integrated into the testing system. This microscope has a 200x magnification and a resolution of 0.5 μm. The microscope automatically located the damaged area and acquired three sets of microscopic morphology images at different focal lengths, each set containing five sampling points from different viewpoints.

[0126] Simultaneously, the system uses a four-probe detector to measure the conductivity of the damaged area. Before damage preparation, the background conductivity of the material is measured as a baseline value, typically 15 kΩ. After damage preparation, the conductivity is measured again at the same location. If the initial damage has successfully penetrated the pre-embedded conductive fiber mesh, the conductivity reading will change significantly, usually showing a sharp increase in conductivity to below 1 kΩ.

[0127] Based on image recognition algorithms, the system analyzes the acquired microscopic morphology images and automatically identifies the damage contour and determines whether it has reached a predetermined depth. Combined with conductivity detection results, the system classifies the initial damage as a valid through-damage when both of the following conditions are met: first, the microscopic image shows a damage depth exceeding 60 μm; second, the conductivity of the damaged area is less than 1 kΩ. Only when the initial damage is confirmed as a valid through-damage will the system initiate the subsequent self-healing process monitoring step. If the verification fails, the system will automatically mark the sample and terminate the testing process.

[0128] The damage validity verification method of the present invention represents a significant advancement in the reliability of test preparation compared to existing technologies. Existing technologies typically employ only a single method for damage verification, such as relying solely on subjective visual judgment by an operator under a common optical microscope, or verification through simple conductivity continuity tests.

[0129] In existing technologies that rely solely on visual judgment, the lack of precise depth measurement standards and objective criteria leads to differing conclusions among operators regarding whether the same damage is "deep enough." While methods relying solely on conductivity testing can detect whether the fiber network has been severed, they cannot distinguish between a sufficient depth and a change in conductivity caused by other factors, nor can they obtain the specific morphological characteristics of the damage.

[0130] This invention employs a dual verification mechanism combining microscopic morphology observation and conductivity testing. This ensures that the damage meets geometrical requirements while simultaneously confirming its functional effectiveness through changes in electrical properties. This multi-dimensional and quantitative verification method completely eliminates the uncertainty caused by subjective judgment, ensuring that the initial damage of each test sample is truly effective enough to trigger the self-healing mechanism. This lays a solid foundation for the accuracy and reliability of subsequent test results.

[0131] Although the technical solutions of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A biomimetic self-healing testing method for logistics roadside equipment, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-source time-series monitoring data of the roadside equipment under test during the self-repair process. The multi-source time-series monitoring data includes at least a sequence of visual images characterizing changes in its surface morphology, acoustic emission signals characterizing micro-deformation of the equipment housing and connectors, and environmental parameter data streams containing roadside high temperature and ultraviolet radiation characteristics. S2: Extract features from multi-source time-series monitoring data to obtain a time-series feature vector representing the self-healing process; S3: Input the time-series feature vector into the pre-trained repair status prediction model, and output the current self-repair process status assessment and the estimated remaining repair time required to reach the predetermined repair completion standard; wherein, the pre-trained repair status prediction model is a time-series deep learning model trained based on historical self-repair data; the current self-repair process status assessment output is a normalized repair completion score between 0 and 1 that integrates multi-source time-series monitoring data; the estimated remaining repair time is the time span required for the pre-trained repair status prediction model to dynamically extrapolate and predict the repair completion score to reach the preset threshold based on the historical sequence of the time-series feature vector and the current state; S4: Based on the output status assessment and estimated remaining repair time, automatically determine the test progress and automatically generate a test termination command and test report when the predetermined repair completion criteria are met. The automatic judgment logic is as follows: monitor the repair completion score in real time, and trigger the test termination command immediately when the repair completion score reaches or exceeds the preset threshold for the first time. The test report should include at least the repair completion rate change curve over time, the final repair completion score, and the comparison information between the estimated remaining repair time and the actual repair time.

2. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, In step S2, feature extraction is performed on the multi-source time-series monitoring data, specifically including: S21: Extract primary features from visual image sequences, acoustic emission signals, and environmental parameter data streams respectively. From the visual image sequence, extract the area and perimeter of the damaged area as visual primary features using an image segmentation algorithm. From the acoustic emission signal, extract event counts, signal amplitude, and energy as acoustic primary features. From the environmental parameter data stream, extract the instantaneous values ​​and changing trends of temperature and humidity as environmental primary features. S22: Weighted fusion of visual primary features, acoustic primary features, and environmental primary features to generate a temporal feature vector; The weight coefficients of each primary feature are obtained through backpropagation during the training process of the repair state prediction model, and can be dynamically adjusted according to different environmental primary features.

3. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, In step S1, the multi-source time-series monitoring data is processed in a programmable multi-factor environmental simulation test chamber; wherein, the test chamber is configured to dynamically and synchronously apply at least two of the following environmental stresses according to a preset environmental stress spectrum during the test cycle of the self-healing process: Cyclic temperature stress: It cycles between high temperature limit and low temperature limit according to a preset time program; Mechanical vibration stress: Simulates the vibration frequency and amplitude of the roadside environment near logistics transportation vehicles or equipment; Ultraviolet radiation stress: Simulates the ultraviolet components in outdoor sunlight and irradiates at a preset irradiance. The environmental parameter data stream originates from real-time monitoring data of the aforementioned environmental stresses applied within the test chamber.

4. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, Step S4 is followed by: S5. After generating the test report, the multi-source time-series monitoring data with complete self-healing process obtained in this test and the final measured repair time are used as a new set of labeled samples. The pre-trained repair state prediction model is incrementally learned using new labeled samples to update its parameters for prediction in subsequent test tasks.

5. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 4, characterized in that, In step S5, before incremental learning, a data credibility verification step is added, specifically as follows: Based on a preset set of data quality rules, the multi-source time-series monitoring data that constitute the new labeled samples are verified. The data quality rule set includes at least: whether the sharpness of the visual image sequence is higher than a set sharpness threshold, whether the signal-to-noise ratio of the acoustic emission signal is higher than a set signal-to-noise ratio threshold, and whether the values ​​of the environmental parameter data stream are within a preset effective range. New labeled samples are only used for incremental learning of the pre-trained repair state prediction model if they pass the data credibility verification.

6. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, Step S1 is preceded by a standardized initial damage preparation step, specifically: Using a programmable laser etching device integrated into the testing system, initial damage with consistent geometry and physical dimensions is prepared on the surface of the self-healing material of the logistics roadside equipment under test, based on pre-stored standardized damage patterns and depth parameters. The initial damage is the starting target monitored by the visual image sequence and acoustic emission signal in step S1.

7. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, In step S1, the acquisition of the visual image sequence is jointly ensured by a primary camera and a backup camera; Step S1 includes a fault diagnosis step while simultaneously acquiring multi-source time-series monitoring data of the roadside equipment under test during its self-repair process: real-time monitoring of the operating status of the main camera and the quality of the visual image sequence. When the primary camera is determined to be faulty or the image quality is consistently unsatisfactory, the data acquisition source is automatically switched to the backup camera to ensure the continuity of multi-source time-series monitoring data.

8. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 4, characterized in that, Step S5, which involves incrementally learning the pre-trained repair state prediction model using new labeled samples, specifically includes: S51. Use new labeled samples to incrementally update a copy of the pre-trained repair state prediction model. Only update the fully connected layer parameters of the pre-trained repair state prediction model, without retraining the feature extraction layer, to obtain a candidate model. S52: Using historical data corresponding to the new labeled samples, evaluate the prediction accuracy of the candidate model and the current officially pre-trained repair state prediction model in parallel. S53: The candidate model is used to replace the current pre-trained repair state prediction model only when the prediction accuracy of the candidate model is higher than that of the current pre-trained repair state prediction model, thus completing incremental learning.

9. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 1, characterized in that, The cyclical execution of steps S3 and S4 also includes dynamic control steps linked to the environmental simulation and prediction steps, specifically: Receive the repair completion score output in real time from the pre-trained repair status prediction model; The real-time repair rate is calculated based on historical data of repair completion scores; When the real-time repair rate is lower than a preset active threshold, a control command is sent to a programmable multi-factor environmental simulation test chamber to reduce the intensity or frequency of the applied environmental stress.

10. The biomimetic self-healing testing method for logistics roadside equipment as described in claim 6, characterized in that, Between the standardized initial damage preparation step and S1, there is also a damage effectiveness verification step, specifically: The self-healing material of the logistics roadside equipment to be tested contains a pre-embedded conductive fiber mesh with a fiber diameter of 50-100μm and a mesh spacing of 1-2 mm. The initial damaged area was magnified and photographed using a 200x optical microscope with a resolution of 0.5μm integrated into the testing system to obtain microscopic morphology images. Based on the analysis of microscopic morphology images using image recognition algorithms, and combined with the detection of conductivity in the damaged area, it is confirmed that the initial damage has penetrated through the pre-embedded conductive fiber mesh; only when the initial damage is confirmed to be an effective penetrating damage will the subsequent self-repair process monitoring in step S1 be initiated.

Images