A composite container real-time state early warning system based on multi-modal perception

Abstract

The application relates to the technical field of container monitoring, and discloses a composite material container real-time state early warning system based on multi-modal perception. The system comprises four units, namely a multi-modal perception unit, a real-time acceleration processing unit, an abnormality analysis unit and a warning output unit. The multi-modal perception unit collects multi-dimensional real-time data such as vibration, temperature, humidity, pressure and strain, comprehensively captures the container structure state, environmental influence and load condition, the real-time acceleration processing unit quickly converts the collected data into a state characteristic vector through matrix multiplication and addition operation, realizes real-time data processing, the abnormality analysis unit carries out multi-dimensional abnormality evaluation based on the state characteristic vector, identifies the abnormality in combination with the correlation of multi-modal data, and avoids the limitation of single threshold judgment, and the warning output unit generates a warning signal when detecting the abnormality, and feeds back information in time. The system can comprehensively, timely and accurately monitor and warn the state of the composite material container, and is suitable for the whole life cycle state management of the composite material container.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of container monitoring, in particular to a composite container real-time state early warning system based on multi-modal perception. BACKGROUND

[0002] In the modern logistics transportation system, as the core carrier for cargo storage and transfer, the performance stability of the container is directly related to the safety of cargo transportation and the efficiency of the supply chain. With the development of material science, composite materials gradually replace traditional metal materials in container manufacturing due to their advantages of lightweight, corrosion resistance, fatigue resistance, etc. However, during the long-term use of composite material containers, they are easily damaged, cracked, and have reduced sealing performance due to the influence of complex and variable transportation environments, such as jolting and vibration, alternating temperature and humidity, external impact, and stress concentration caused by cargo loading. These problems are often difficult to detect in time through manual inspection, and once they accumulate and spread, they may cause a decrease in the carrying capacity of the container, damage to the cargo, and even safety accidents during transportation, causing huge economic losses to logistics enterprises.

[0003] Currently, there are obvious limitations in the technical solutions for container state monitoring. The traditional manual inspection method relies on the experience of the staff, which is not only inefficient, but also difficult to detect microscopic damage inside the composite material, and there is a high risk of missed detection. Some existing automated monitoring systems use single modal sensors, such as relying only on vibration sensors or temperature sensors to collect data, which cannot fully reflect the complex state of the container. For example, only through vibration data can the impact strength during transportation be determined, but it is difficult to identify internal aging of the composite material due to humidity changes; only relying on pressure sensors cannot detect the fine cracks on the surface of the container. This single-dimensional monitoring method can easily lead to one-sided monitoring data and missing information, resulting in misjudgment or missed judgment of abnormal states.

[0004] The data processing capability of existing monitoring systems also has deficiencies. A large amount of real-time monitoring data is generated during the transportation of the container, and the traditional data processing unit has a slow operation speed, making it difficult to realize real-time analysis and processing of multi-source data. Often, the data needs to be transmitted to a remote server for subsequent processing, which not only increases the delay of data transmission, but also may cause data loss due to network fluctuations, making it impossible to generate an abnormal warning signal in time. In addition, in the abnormal evaluation link, the existing system mostly uses a simple threshold judgment method, i.e., setting a fixed numerical threshold value, and determining that it is abnormal when the monitoring data exceeds the threshold value. This evaluation method cannot consider the correlation and coupling between multi-modal data, and is difficult to cope with multi-dimensional abnormal states of composite material containers under complex working conditions, resulting in low evaluation accuracy and failing to meet the real-time and accurate monitoring needs of modern logistics for container states. SUMMARY

[0005] The present application aims to provide a composite container real-time state early warning system based on multi-modal perception to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a composite container real-time state early warning system based on multi-modal perception, which comprises:

[0007] A multi-modal perception unit is configured to collect real-time multi-modal sensor data of the composite container.

[0008] A real-time acceleration processing unit is configured to perform matrix multiplication and addition operation on the data collected by the multi-modal perception unit to generate a state feature vector.

[0009] An anomaly analysis unit is configured to perform multi-dimensional anomaly evaluation based on the state feature vector.

[0010] An early warning output unit is configured to generate an early warning signal when the anomaly analysis unit detects an anomaly.

[0011] Preferably, the real-time acceleration processing unit comprises:

[0012] A multi-source signal generation module is configured to generate multiple parallel signal sources.

[0013] A weight loading module is configured to load weights on the signals generated by the multi-source signal generation module according to a preset weight vector.

[0014] A signal modulation module is configured to modulate the data collected by the multi-modal perception unit to the signals loaded with weights by the weight loading module.

[0015] A time delay module is configured to apply time delay to the modulated signals.

[0016] A feature fragment extraction module is configured to extract feature fragments within a specific time window from the delayed signals.

[0017] A signal superposition module is configured to linearly superimpose the extracted feature fragments.

[0018] Preferably, the real-time acceleration processing unit further comprises a dynamic state modeling module.

[0019] The dynamic state modeling module is configured to construct a dynamic state model based on historical state data and generate a theoretical state value by taking the output of the signal superposition module as input.

[0020] Preferably, the anomaly analysis unit comprises a multi-dimensional difference calculation module.

[0021] The multi-dimensional difference calculation module is configured to perform multi-dimensional difference analysis on the theoretical state values generated by the dynamic state modeling module and the measured state values, and generate a difference coefficient matrix.

[0022] The multi-dimensional difference analysis includes time-domain cumulative deviation calculation, frequency-domain energy offset detection, and signal sequence similarity evaluation.

[0023] Preferably, the anomaly analysis unit further comprises a spatial positioning module.

[0024] The spatial positioning module is configured to input the difference coefficient matrix generated by the multi-dimensional difference calculation module into a spatial topology model, generate an anomaly probability distribution map, and locate an abnormal physical region.

[0025] Preferably, the spatial positioning module comprises:

[0026] A topology modeling submodule is configured to construct a spatial topology structure of the composite material container.

[0027] An anomaly propagation simulation submodule is configured to simulate an anomaly propagation path based on the difference coefficient matrix.

[0028] A probability distribution generation submodule is configured to generate an anomaly probability distribution map covering the entire region.

[0029] Preferably, the anomaly analysis unit further comprises an adaptive learning module.

[0030] The adaptive learning module is configured to update model parameters of the dynamic state modeling module based on the anomaly probability distribution map.

[0031] The adaptive learning module comprises an activated neuron identification submodule configured to identify activated neurons of a current data point.

[0032] A fuzzy neuron determination submodule is configured to determine fuzzy neurons based on an intersection relationship between reference data points and the activated neurons.

[0033] A learning rate calculation submodule is configured to calculate a learning rate of the fuzzy neurons.

[0034] Preferably, the fuzzy neuron determination submodule comprises a jump data point analysis unit.

[0035] The jump data point analysis unit is configured to calculate an anomaly likelihood of data points, and determine target jump data points based on the anomaly likelihood.

[0036] The fuzzy neuron determination submodule is configured to identify fuzzy neurons according to the target jump data points.

[0037] Preferably, the learning rate calculation submodule comprises a position variation evaluation unit.

[0038] The position variation evaluation unit is configured to calculate a neuron position variation variance;

[0039] The learning rate calculation submodule is configured to calculate a learning rate of the ambiguous neuron based on the position variation variance and a preset learning rate of the activated neuron.

[0040] Preferably, the early warning output unit comprises an anomaly decision module.

[0041] The anomaly decision module is configured to generate an early warning strategy based on the anomaly probability distribution map and the updated model parameters.

[0042] The early warning strategy comprises high-frequency monitoring mode activation and physical disturbance test execution.

[0043] Compared with the prior art, the present application has the following advantages:

[0044] The real-time multi-modal sensor data of the container is collected by the multi-modal perception unit, which can comprehensively capture various state information of the container during use. Compared with the traditional single modal monitoring method, the multi-modal data covers multiple dimensions such as vibration, temperature, humidity, pressure, and strain, which can fully reflect the structural state, environmental impact and load condition of the container, avoiding state misjudgment or omission caused by one-sided data, and making the container state monitoring more comprehensive and complete. Whether it is the internal microscopic damage of the composite material box body, or the performance fluctuation caused by environmental changes, or the local stress anomaly caused by improper loading of goods, it can be effectively captured through the cooperative collection of multi-modal data, providing rich and comprehensive basic information for subsequent state analysis.

[0045] The real-time acceleration processing unit performs matrix multiplication and addition operation on the data collected by the multi-modal perception unit to generate a state feature vector, which significantly improves the efficiency and speed of data processing. During the transportation of the container, monitoring data will continuously and massively generate, and the traditional data processing unit is difficult to cope with such high-concurrency data processing demand, often causing data backlog and analysis delay. However, the real-time acceleration processing unit has high-efficiency matrix multiplication and addition operation capability, which can perform real-time operation and feature extraction on multi-source heterogeneous monitoring data, quickly convert the original data into a feature vector that can reflect the state of the container, without relying on remote servers for data processing, effectively shortening the time interval from data collection to analysis, ensuring real-time monitoring and analysis of the state of the container, and laying a key foundation for timely detection of abnormal state. Even in the scenario of long-distance transportation and continuous increase of data volume, the processing unit can stably and efficiently complete the data processing task, ensuring the dynamic tracking ability of the system to the state of the container.

[0046] The abnormality analysis unit performs multi-dimensional abnormality evaluation based on the state feature vector, breaks through the limitations of traditional threshold judgment method, and fully considers the correlation and coupling between multi-modal data. The abnormal state of the composite container is often not caused by a single factor, but the result of the joint action of multiple dimensional parameters. The traditional single threshold judgment cannot accurately identify such multi-dimensional related abnormal conditions. The multi-dimensional abnormality evaluation can integrate the multi-dimensional information in the state feature vector, analyze the mutual influence and change law between different parameters, and judge whether the container is in an abnormal state as a whole. For example, when the vibration data shows a small abnormality, combined with the change trend of temperature and humidity data, it can accurately determine whether it is a temporary fluctuation caused by short-term impact or a material performance degradation caused by alternating temperature and humidity, greatly improving the accuracy and reliability of abnormality evaluation, effectively reducing the misjudgment and omission of abnormal state, and ensuring accurate identification of potential risks of the container.

[0047] The early warning output unit generates a warning signal when the abnormality analysis unit detects an abnormality, which can timely feedback the abnormal state of the container to relevant staff or control system. In the logistics transportation process, timely discovery and handling of the abnormal state of the container is crucial. If the warning is not timely, it may lead to further deterioration of the abnormal condition, causing more serious safety accidents or economic losses. The early warning output unit can generate a warning signal at the first time when the abnormality is detected. The form of the warning signal can be flexibly set according to actual needs, such as audible and visual alarm, SMS notification, system pop-up prompt, etc., to ensure that relevant personnel can quickly obtain abnormal information and take appropriate measures in time, such as adjusting the transportation route, suspending transportation for maintenance, reinforcing cargo loading, etc., thereby effectively containing the spread of abnormal state, reducing the probability of cargo damage, container damage and even transportation accidents, ensuring the safety and stability of the logistics transportation process, and maintaining the economic interests of logistics enterprises and the smooth operation of the supply chain.

[0048] The whole system works in a coordinated manner, realizing the whole-process automation operation from data acquisition, real-time processing, multi-dimensional evaluation to early warning output, without the need for manual intervention to complete real-time monitoring and abnormality warning of the container state. This automatic operation mode not only reduces the investment in labor cost and avoids subjective errors and human negligence in manual operation process, but also greatly improves the efficiency and stability of monitoring work. Whether in the container yard of the port terminal or on the long-distance transportation vehicle, the system can continuously and stably operate, providing strong support for the whole life cycle state management of the composite container, and promoting the development of modern logistics monitoring technology in the direction of higher efficiency, higher accuracy and higher intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a timing diagram of the composite container real-time state early warning system based on multi-modal perception according to the present application;

[0050] Figure 2 Flowchart for real-time acceleration processing unit signal processing;

[0051] Figure 3 Flowchart for multi-dimensional difference calculation module analysis;

[0052] Figure 4 Flowchart for adaptive learning module parameter update. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0054] Please refer to Figure 1 The present application provides a composite container real-time state early warning system based on multi-modal perception, which comprises a multi-modal perception unit, a real-time acceleration processing unit, an anomaly analysis unit and a warning output unit.

[0055] The multi-modal perception unit is deployed in the key physical area of the composite container, used for collecting real-time multi-modal sensor data, including strain, temperature, vibration and acoustic emission signals. The sensor network is arranged in a distributed manner, covering the overall structure of the container, and the data is transmitted to the real-time acceleration processing unit through a high-speed bus. After receiving the multi-modal sensor data, the real-time acceleration processing unit performs matrix multiplication and addition operation, and converts the multi-source data stream into a state feature vector in a unified format. The matrix multiplication and addition operation involves the multiplication operation of the weight matrix and the input data vector, and the addition operation of the bias vector, generating a high-dimensional feature representation. The state feature vector is transmitted to the anomaly analysis unit, which analyzes the vector elements based on a multi-dimensional anomaly evaluation algorithm and calculates an anomaly index. The multi-dimensional anomaly evaluation includes feature comparison in time domain, frequency domain and spatial domain, and the abnormal state is determined by a pre-defined threshold rule. When an anomaly is detected, the warning output unit generates a warning signal, which is output to an external monitoring platform through a communication interface, triggers a visual or audible alarm, and records the timestamp and location information of the abnormal event. The entire system is implemented on an embedded hardware platform to ensure real-time and low-power operation.

[0056] Embodiment 1: Please refer to Figure 2, through a series of highly coordinated modular operations, the original multi-modal sensor data is transformed into a state feature vector that can be used for advanced analysis. Its implementation embodies the complete process from signal generation to feature integration, and the design of each module is aimed at realizing the signal processing function and ensuring the parallelism and timeliness of the overall processing. The multi-source signal generation module as the starting point of the entire processing flow, its core function is to produce multiple parallel signal sources. In actual deployment, this module is usually implemented by a programmable waveform generator, which can generate multiple independent reference signals synchronously. The waveform type of these signal sources is configured according to the needs of subsequent processing, usually including sine wave, square wave and pulse signal. The number of signal paths strictly corresponds to the number of sensor channels in the multi-modal perception unit, for example, if the system is deployed with sixteen sensor channels, the module will generate sixteen parallel signals. Each signal has the same initial frequency and amplitude, but its phase can be fine-tuned according to the configuration to match the physical characteristics of different sensors. The signal generation process is based on digital frequency synthesis technology, and high-precision waveform output is achieved through the cooperation of lookup table and digital-to-analog converter. The generated signals are temporarily stored in the cache in the form of digital sequences, waiting for the processing of subsequent modules.

[0057] The weight loading module receives the parallel signal stream from the multi-source signal generation module and applies preset weight coefficients to it. These weight coefficients are stored in the system's non-volatile memory, forming a multi-dimensional weight vector. The dimension of the weight vector is exactly the same as the number of signal channels, and each weight value corresponds to the importance coefficient of a specific signal channel. The weight loading process is implemented through a digital multiplier array, and each multiplier is responsible for the multiplication operation of a signal and the corresponding weight. The weight value can be a fixed value, or it can be dynamically adjusted according to the system running state. For example, during the system initialization phase, the weight vector may adopt the default uniform distribution; while in the running process, the system can adjust the weight distribution adaptively according to the signal-to-noise ratio characteristics of the sensor data, giving higher weight coefficients to sensor channels with higher quality. The weighted signal not only retains the original waveform characteristics, but also integrates the system's evaluation of the importance of various signals.

[0058] The signal modulation module is responsible for fusing the actual collected sensor data with the weighted signals. This module receives two input streams: one is the weighted signals from the weight loading module, and the other is the multi-modal sensor data from the multi-modal perception unit. The modulation process uses amplitude modulation, taking the sensor data as the modulation index and multiplying it with the carrier signal. Specifically, each sensor data sample point is multiplied by the carrier signal amplitude at the corresponding time, thus carrying the data information onto the carrier waveform. This modulation method not only maintains the dynamic range of the original data, but also utilizes the periodic characteristics of the carrier signal. During the modulation process, the system normalizes the sensor data to ensure that the modulation index is within an appropriate range, avoiding signal distortion caused by excessive modulation. The modulated signal retains the frequency characteristics of the carrier signal and incorporates the amplitude information of the sensor data, forming a composite signal structure.

[0059] The time delay module applies a controllable time delay to the modulated signal, and the amount of delay is set based on the propagation characteristics of the multi-modal signal and the system sampling rate characteristics. Different sensors may have natural time differences due to differences in physical location or acquisition links, and the time delay module aligns the signals in time through a programmable delay line. The delay amount can be set to a fixed value or dynamically adjusted according to the signal type. For example, for high-frequency vibration signals, a smaller delay amount may be needed; while for low-frequency temperature signals, a larger delay amount can be set. The delay is usually implemented using digital buffering technology, which shifts the signal sequence in time through a first-in-first-out memory. The delayed signal is synchronized in the time axis, providing a time-aligned data basis for subsequent feature extraction.

[0060] The feature segment extraction module extracts feature segments within a specific time window from the time-delayed signal. The length of this time window is carefully designed to contain enough signal periods to reflect signal characteristics while maintaining sufficient timeliness to meet real-time processing requirements. The sliding step of the window is smaller than the window length, ensuring that there is some overlap between adjacent windows to avoid missing important features. The extraction process is achieved through a sliding window algorithm, and the signal segment in each window is independently taken out and analyzed for features. The extracted feature segments not only contain amplitude information but also retain their time-domain waveform characteristics. These segments are temporarily stored in the feature buffer, waiting for subsequent processing.

[0061] The signal superposition module performs linear superposition on the extracted feature segments. This superposition process is not simply averaging, but a linear combination with weight coefficients. Each feature segment is assigned a different superposition coefficient according to its source sensor type and signal quality. The superposition operation is implemented through a digital adder array, with each channel of signal being added point-to-point under the premise of time alignment. The superposed signal forms a composite feature representation that integrates the information from multiple sensors, highlights common features, and suppresses random noise. This composite signal serves as the final output of the real-time acceleration processing unit and is passed to subsequent units for further analysis.

[0062] The implementation of the entire real-time acceleration processing unit embodies the design idea of high parallelization and pipelining. The various modules are connected through high-speed data buses, and the data stream passes through each processing stage in a pipelined manner. This design maximizes processing efficiency and ensures that the system can handle high-speed multi-modal sensor data. Each module is implemented using a digital signal processor or an application-specific integrated circuit to ensure optimal processing speed and energy efficiency. Parameter transfer and coordinated control between modules are managed by the central scheduling unit to ensure the coordination and stability of the entire processing flow.

[0063] Embodiment 2: Referring to Figure 3 , a quantitative judgment of the system's abnormal state is formed by establishing a theoretical state benchmark and comparing it with real-time data in multiple dimensions. Its implementation process integrates time series prediction and multi-domain feature analysis technology, realizing the transformation from raw features to abnormal indicators. The dynamic state modeling module constructs a prediction model based on historical state data. This module receives real-time state feature vectors from the signal superposition module and continuously accesses historical data sequences stored in the ring buffer. Historical data includes state feature vectors within a certain time period (such as 24 hours) in the past, stored in chronological order. The modeling process uses the autoregressive integrated moving average algorithm framework, with the model order dynamically configured according to data characteristics. Model parameters are updated online using the recursive least squares algorithm, recalculating parameter estimates with each new data input. The module outputs the theoretical state value, which is the state feature vector at the current time based on historical regularity prediction. The theoretical state value is represented as a floating-point number array, with each element corresponding to the expected numerical range of a state feature. Prediction calculations are performed on a dedicated digital signal processor, with a single prediction iteration completed every millisecond to meet real-time requirements.

[0064] The multi-dimensional difference calculation module synchronously receives two input data streams: the theoretical state value generated by the dynamic state modeling module and the measured state value collected by the multi-modal perception unit. The measured state value is subjected to the same time alignment processing as the theoretical value to ensure the time consistency of the comparison. The difference analysis is developed in three independent dimensions in parallel: the time domain cumulative deviation dimension calculates the cumulative difference between the theoretical value and the measured value within a fixed time window. This calculation uses a sliding window integration algorithm to independently calculate the sum of the absolute values of the difference for each state feature channel. The time window length can be configured (e.g., 500 milliseconds), and the window sliding step is set to an integer multiple of the sampling interval. The calculation result reflects the deviation degree of the state feature in the time dimension.

[0065] The frequency domain energy offset detection dimension converts the theoretical value and the measured value into the frequency domain for analysis. This process uses the Fast Fourier Transform algorithm to decompose the time domain signal into frequency spectrum components. For each frequency component, the energy ratio of the theoretical frequency spectrum to the measured frequency spectrum is calculated, and the offset is obtained in decibels after taking the logarithm. The energy change in the characteristic frequency band is focused on, such as analyzing the 0-100Hz frequency band for vibration signals and the 20-80kHz frequency band for acoustic emission signals. Energy offset detection can identify periodic abnormalities or changes in resonance frequency.

[0066] The signal sequence similarity evaluation dimension uses the Dynamic Time Warping algorithm, which overcomes the influence of small time axis shifts and calculates the minimum matching distance between the theoretical sequence and the measured sequence. The calculation process establishes a cost matrix to find the optimal bending path, and finally outputs a normalized similarity coefficient. This coefficient reflects the similarity of the overall waveform shape and has the ability to identify both impact-type abnormalities and gradual-type abnormalities.

[0067] The analysis results of the three dimensions are integrated into a difference coefficient matrix, which is a two-dimensional data structure. The row index corresponds to the time point, and the column index includes three dimensions (time domain cumulative deviation, frequency domain energy offset, and signal sequence similarity) and the combination of state feature channels. Each matrix element stores a normalized difference value, which is mapped to the [0, 1] interval through linear transformation. The matrix generation process uses a pipeline architecture, with each dimension independently calculated and then summarized to the matrix construction unit. The final output of the difference coefficient matrix is passed to the subsequent analysis unit as the basic input for abnormal positioning.

[0068] The whole implementation process is realized by a dedicated hardware accelerator for parallel processing. The dynamic state modeling module runs on a floating-point operation unit, supporting matrix operation acceleration. The multi-dimensional difference calculation module is divided into three processing engines, corresponding to time domain, frequency domain and sequence analysis respectively. Data distribution is realized through cross switch interconnection. The difference coefficient matrix generation unit is equipped with a high-speed memory interface, supporting real-time construction and transmission of large matrices. The system uses a clock synchronization mechanism to ensure that the time alignment accuracy of the theoretical value calculation and the measured value acquisition is in the microsecond level. The data path uses a double buffer design to eliminate processing delay.

[0069] Embodiment 3: The spatial positioning module of the anomaly analysis unit is responsible for mapping the multi-dimensional difference analysis results to the physical space. This module receives the difference coefficient matrix from the multi-dimensional difference calculation module, which contains three-dimensional data of time dimension, feature dimension and difference dimension. The core of spatial positioning is to build a spatial topology model of the container structure, generate a probability distribution map through anomaly propagation simulation, and finally realize accurate positioning of the abnormal physical area.

[0070] The topology modeling submodule first abstracts the physical structure of the composite container into a discretized network model. The container is divided into N regular grid areas, each of which corresponds to a topology node. Node attributes include geometric center coordinates, material thickness, sensor distribution density, and other static parameters. The connection relationship between nodes is established based on physical connectivity: adjacent grid areas automatically generate topology edges, and non-adjacent areas establish cross-area connections if there are structural reinforcing ribs or load transmission paths. Each topology edge is assigned a weight coefficient:

[0071]

[0072] wherein: represents the connection strength between nodes i and j, is the connection cross-sectional area (unit: square meter), is the material elastic modulus (unit: gigapascal), is the Euclidean distance between nodes (unit: meters). This weight model reflects the mechanical stress transmission efficiency, which is pre-calculated by finite element analysis and stored in the topology database.

[0073] The anomaly propagation simulation submodule starts the propagation calculation with the difference coefficient matrix as input. In the initialization stage, the difference values of the latest time slice in the matrix are mapped to the corresponding topology nodes to form the initial abnormal intensity distribution. The propagation process uses an improved diffusion algorithm: in each calculation period, the abnormal intensity increment of node i is determined by the intensity gradient of adjacent nodes:

[0074]

[0075] wherein: represents the abnormal intensity of node i at time t, is the set of neighboring nodes of node i, is the propagation attenuation factor (default value 0.25). The partial differential equation is solved iteratively by the explicit Euler method, with a time step set to 1 / 10 of the system sampling interval. A strength threshold constraint is set during the iteration process, which is automatically truncated when the node intensity exceeds 80% of the material yield limit. Global normalization is performed after each iteration to prevent exponential growth of intensity values.

[0076] The probability distribution generation submodule converts the propagation-stable abnormal intensity field into a probability distribution, using a Bayesian inference framework that treats the topological nodes as independent random variables. The prior probability comes from historical abnormal statistical data, and the likelihood function is constructed from the current abnormal intensity value. Posterior probability calculation is realized by Markov Chain Monte Carlo sampling, generating a two-dimensional probability matrix where represents the abnormal probability at grid coordinates (m, n). After smoothing the probability matrix by bicubic spline interpolation, it is converted into a grayscale probability distribution map covering the entire surface of the container, with dark areas corresponding to high-probability abnormal regions.

[0077] Abnormal physical region positioning is realized by probability threshold segmentation, setting a dynamic probability threshold where is the mean probability of the entire image, is the standard deviation. The connected regions in the probability map that exceed are labeled, and the centroid coordinates of each region are calculated and mapped back to the physical coordinate system. The positioning result is output as a list of abnormal regions, with each entry containing the region number, center coordinates, boundary polygon vertices, and confidence score. At the same time, the visualization engine is triggered to overlay a semi-transparent heat map on the three-dimensional container model, with high-probability regions showing a red gradient effect.

[0078] The entire processing flow is deployed on a graphics processor platform, with topological model data stored in texture memory, abnormal propagation calculation mapped to pixel shader programs, and probability sampling performed in parallel using compute shaders. The time consumption of a single spatial positioning is controlled within 5 milliseconds, and the output data is transmitted to the early warning decision unit through the PCIe bus. The system maintains a topological version management mechanism, supporting online model update when the container structure is modified.

[0079] Example 4: see Figure 4The adaptive learning module is responsible for dynamically optimizing the state prediction model. The module takes the anomaly probability distribution map as input and implements online adjustment of model parameters through neuron state analysis. Its implementation process includes three sub-modules that work together to form an adaptive closed loop from neuron identification to learning rate calculation. The activated neuron identification sub-module continuously monitors the intermediate layer output of the dynamic state modeling module. When the real-time state feature vector is input into the prediction model, this sub-module records the activation values of all neurons in the hidden layer. The activation threshold is set to 0.65, and when the neuron output value exceeds this threshold, it is marked as an activated state. The identification process uses a parallel comparator array to complete the screening of 1024 neurons within 200 microseconds. The activated neuron number is recorded in a ring buffer, and a binary bitmap is generated to mark the activated state. For example, the record at a certain time shows that neurons {15, 28, 73, 209} are in an activated state, indicating that these neurons are sensitive to the current input pattern.

[0080] The fuzzy neuron determination sub-module analyzes the stability of activated neurons based on historical reference data points. The reference data point set is stored in a flash database, containing records of activated neurons in typical states within the past 72 hours. This sub-module performs set operations: taking the intersection of the current activated neuron set and the historical reference set. When the intersection ratio is below a preset threshold (such as 40%), it is determined that the current activation pattern is abnormal. At this time, the jump data point analysis unit is started, and the deviation of the current data point from the historical data distribution is calculated. The deviation is measured by multi-dimensional space distance, considering the 12 main components of the feature vector. Data points with a distance value greater than three times the standard deviation are marked as jump points, for example, a certain detection found a distance value of 7.3 (the historical mean is 2.1, and the standard deviation is 1.5).

[0081] Based on the jump data points, the fuzzy neuron determination sub-module performs weight similarity matching. The feature vector of the jump point is extracted and the cosine similarity is calculated with all neuron weight vectors. Neurons with a similarity less than 0.3 are marked as fuzzy neurons, indicating that their weight configuration does not match the current abnormal pattern. For example, in a certain abnormal event, neurons {38, 112, 305} have similarities of 0.18, 0.22, and 0.27, respectively, and are added to the fuzzy neuron list. The system maintains a dynamic neuron state record, as shown in Table 1.

[0082] Table 1: Neuron state record table

[0083]

[0084] The learning rate calculation submodule determines the parameter update strategy according to the neuron state. The submodule first obtains the preset learning rate of the activated neuron (default 0.01), and then calculates the learning rate adjustment coefficient of the fuzzy neuron through the position variation evaluation unit. The position variation evaluation tracks the change trajectory of each neuron weight vector in the last 20 iterations, and calculates the coefficient of variation of the Euclidean distance. The neuron with a coefficient of variation greater than 0.7 is determined to be in an unstable state, and its learning rate is scaled up. For example, the coefficient of variation of neuron 305 is 0.83, and its learning rate is adjusted to 0.018; the coefficient of variation of neuron 112 is 0.58, and the learning rate remains 0.01.

[0085] The learning rate calculation process adopts a hierarchical weighting mechanism: the learning rate of the activated neuron remains the base value; the learning rate of the fuzzy neuron is graded up according to the abnormality probability; the fuzzy neuron with high coefficient of variation is additionally increased in learning rate weight. Finally, a learning rate configuration matrix is generated, which is written into the parameter register of the dynamic state modeling module through a direct memory access channel. Parameter update is performed in the system idle time slot to avoid interfering with real-time prediction tasks, and the time consumption of a single update is controlled within 50 milliseconds.

[0086] The entire adaptive learning process adopts an event-driven mechanism. When an area with a probability value exceeding 0.75 appears in the abnormal probability distribution map, the learning process is automatically triggered. The system records the model prediction error before and after each learning, but does not use it as a basis for learning rate adjustment, only for monitoring the stability of the learning process. Neuron state data is stored in column format, supporting fast retrieval of historical activation patterns, providing pattern evolution references for subsequent anomaly analysis.

[0087] Embodiment 5: The anomaly decision module of the early warning output unit generates an execution strategy based on the anomaly analysis results. The module receives the abnormal probability distribution map from the spatial positioning module and the model parameters updated by the adaptive learning module, and outputs specific operation instructions through multi-level decision logic. Its implementation process integrates the collaborative judgment mechanism of probability analysis, parameter sensitivity evaluation and physical response test.

[0088] The anomaly decision module first analyzes the topological features of the abnormal probability distribution map. The analysis process extracts three key indicators: the maximum probability value and its coordinate position, the geometric shape features of the high probability area, and the change direction of the probability gradient. These indicators are automatically identified through image processing algorithms, such as using morphological operations to extract connected regions and using edge detection algorithms to determine boundary curvature. At the same time, the neuron state record provided by the adaptive learning module is read, focusing on the distribution density and position variation amplitude of the fuzzy neurons. When the fuzzy neurons are concentrated in a specific network layer and the position variation variance exceeds 0.4, it is marked as a model structure sensitive state.

[0089] Based on the above analysis, the decision engine generates a combination of early warning strategies, and the high-frequency monitoring mode activation instruction is sent to the multi-modal sensing unit through the digital control bus. The instruction contains sensor channel selection mask and sampling rate configuration parameters, such as increasing the sampling rate of strain sensors from 1 kHz to 2 kHz and expanding the acoustic emission channel bandwidth from 50 kHz to 100 kHz. The sampling mode switching uses phase synchronization technology to switch the sampling clock at the signal zero crossing point to avoid signal truncation. The data acquisition unit adjusts the buffer strategy at the same time, expanding the original data buffer capacity by three times to ensure data integrity under high sampling rate.

[0090] The physical disturbance test execution instruction triggers the external actuator to work together. The system sends a sweep signal sequence to the piezoelectric exciter through industrial Ethernet, and the signal frequency range is set according to the material characteristics of the abnormal area. For example, a composite laminate is subjected to a linear sweep vibration of 20-200 Hz in the suspected delamination area, with a duration of 500 milliseconds. The thermal excitation unit is started synchronously, and a gradient temperature field is applied to the target area that does not exceed the material tolerance limit. During the test, the multi-modal sensing unit enters synchronous acquisition mode, and all sensor channels record the response signal at the highest sampling rate.

[0091] The disturbance test signal is compared with the normal response in the reference test database, and the comparison dimensions include time domain waveform correlation coefficient, frequency response function amplitude deviation, and damping characteristic change rate. The analysis result generates an abnormal verification report containing a list of key indicator deviation values. When more than three core indicators deviate by more than 15%, it is confirmed that a physical anomaly exists. At this time, the system automatically upgrades the warning level and generates diagnostic information containing abnormal type speculation, such as "Region B3 suspected fiber breakage, confidence 72%".

[0092] The execution process of the early warning strategy has a state feedback mechanism. In high-frequency monitoring mode, the system analyzes the statistical characteristics of the incremental data in real time, and automatically restores the regular sampling rate if the standard deviation continues to decrease. The physical disturbance test sets a safety interrupt condition, which immediately terminates the excitation when the response signal amplitude exceeds the preset safety threshold. All operation records generate event logs containing structured data such as time stamp, execution parameter, and response feature, which are transmitted to the remote monitoring center through a security protocol.

[0093] The main channel sends machine-readable warning messages through industrial Internet of Things protocols. The message structure contains fields such as abnormal coordinates, severity level, and recommended measures. The auxiliary channel generates a visual alarm by superimposing a flashing marker on the three-dimensional container model. The marker color gradually changes from yellow to red according to the probability value. The local audible and visual alarm is started synchronously, and the alarm frequency is positively related to the abnormal probability value. The system retains a manual confirmation interface, and the operator can view the diagnostic details and confirm the alarm cancellation through the human-machine interface. The entire decision execution cycle is controlled within 800 milliseconds, ensuring the timeliness from anomaly identification to response output.

[0094] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify a subject or action, without necessarily requiring or implying any such actual relationship or order between such subjects or actions. Moreover, the terms "comprising", "containing", or any other similar term are intended to encompass non-exclusive inclusions, such that a process, method, article, or apparatus that comprises a list of elements does not include those elements solely, but can include other elements not expressly listed, or can include elements inherent in such process, method, article, or apparatus.

[0095] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations can be undertaken by those skilled in the art without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A real-time status early warning system for composite material containers based on multimodal perception, characterized in that, include: A multimodal sensing unit is used to collect real-time multimodal sensor data from composite material containers. A real-time acceleration processing unit is used to perform matrix multiplication and addition operations on the data collected by the multimodal sensing unit to generate a state feature vector; An anomaly analysis unit is used to perform multidimensional anomaly assessment based on the state feature vector; An early warning output unit is used to generate an early warning signal when the anomaly analysis unit detects an anomaly. The real-time acceleration processing unit includes: Multi-source signal generation module, used to generate multiple parallel signal sources; A weight loading module is used to load weights onto the signals generated by the multi-source signal generation module according to a preset weight vector. The signal modulation module is used to modulate the data collected by the multimodal sensing unit onto the signal after the weight loading module has loaded the weights; The time delay module is used to apply a time delay to the modulated signal; The feature segment extraction module is used to extract feature segments within a specific time window from the delayed signal; The signal superposition module is used to linearly superimpose the extracted feature segments; The real-time acceleration processing unit also includes a dynamic state modeling module; The dynamic state modeling module is used to construct a dynamic state model based on historical state data, and takes the output of the signal superposition module as input to generate theoretical state values. The anomaly analysis unit includes a multidimensional difference calculation module; The multidimensional difference calculation module is used to perform multidimensional difference analysis on the theoretical state values ​​and measured state values ​​generated by the dynamic state modeling module, and generate a difference coefficient matrix. The multidimensional difference analysis includes time-domain cumulative deviation calculation, frequency-domain energy shift detection, and signal sequence similarity assessment; The anomaly analysis unit also includes a spatial positioning module; The spatial positioning module is used to input the difference coefficient matrix generated by the multidimensional difference calculation module into the spatial topology model, generate an anomaly probability distribution map, and locate the abnormal physical region. The spatial positioning module includes: The topology modeling submodule is used to construct the spatial topology of composite container; An anomaly propagation simulation submodule is used to simulate anomaly propagation paths based on the difference coefficient matrix. The probability distribution generation submodule is used to generate anomaly probability distribution maps covering the entire area.

2. The real-time status early warning system for composite material containers based on multimodal perception according to claim 1, characterized in that, The anomaly analysis unit also includes an adaptive learning module; The adaptive learning module is used to update the model parameters of the dynamic state modeling module based on the anomaly probability distribution map; The adaptive learning module includes: an activation neuron identification submodule, used to identify the activation neurons of the current data point; The fuzzy neuron determination submodule is used to determine fuzzy neurons based on the intersection relationship between reference data points and activated neurons; The learning rate calculation submodule is used to calculate the learning rate of the fuzzy neuron.

3. The real-time status early warning system for composite material containers based on multimodal perception according to claim 2, characterized in that, The fuzzy neuron determination submodule includes a jump data point analysis unit; The jump data point analysis unit is used to calculate the probability of anomalies in data points and determine the target jump data point based on the probability of anomalies. The fuzzy neuron identification submodule is used to identify fuzzy neurons based on target jump data points.

4. The real-time status early warning system for composite material containers based on multimodal perception according to claim 3, characterized in that, The learning rate calculation submodule includes a position change evaluation unit; The position change evaluation unit is used to calculate the variance of neuron position change; The learning rate calculation submodule is used to calculate the learning rate of the fuzzy neuron based on the position variation variance and the preset learning rate of the activated neuron.

5. The real-time status early warning system for composite material containers based on multimodal perception according to claim 4, characterized in that, The early warning output unit includes an anomaly decision-making module; The anomaly decision-making module is used to generate early warning strategies based on the anomaly probability distribution map and the updated model parameters; The early warning strategy includes activating high-frequency monitoring mode and executing physical disturbance tests.

Images