Multi-sensor fusion aviation food truck intelligent monitoring and early warning method and device

Abstract

The invention discloses a multi-sensor fusion aviation food truck intelligent monitoring and early warning method and device, and the method comprises the steps: collecting the original monitoring data of multiple types of sensors of an aviation food truck, and recognizing the data inconsistency degree between the sensors through sensor conflict detection; generating a conflict enhancement coefficient based on the resonant frequency characteristics and the periodic conflict mode; cross-sensor time difference features are obtained, a delay utilization matrix is constructed, an inversion reference vector is generated through probability chain reconstruction, and a data feature space is reconstructed to form an enhanced feature set; adopting adversarial feature vector analysis and self-adversarial boundary determination technologies to establish state recognition parameters and a multi-layer state classification matrix; according to the method, the inter-layer response change of a state conversion trajectory diagram is monitored, an alarm intensity coefficient is generated through frequency inversion and correlation analysis, finally, intelligent generation of multi-stage alarm output signals is achieved, and comprehensive and accurate state monitoring and early warning capacity is provided for safe operation of the aviation food truck.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation food cart safety monitoring, in particular to a multi-sensor fusion aviation food cart intelligent monitoring and early warning method and device. BACKGROUND

[0002] The aviation food cart is a mobile food processing and sales device, and real-time monitoring of its running state is crucial to food safety and equipment reliability. Existing aviation food cart monitoring systems usually use multiple sensors to synchronously collect temperature, humidity, pressure, vibration and other parameters. However, due to differences in sensor response characteristics, different installation positions, and environmental interference, data conflicts and time inconsistency often occur between sensors, leading to inaccurate state determination and frequent false alarms.

[0003] Traditional multi-sensor data fusion methods are mainly based on simple weighted average or voting mechanism, lacking deep mining and utilization of conflict characteristics between sensors, and unable to fully utilize the abnormal detection value contained in conflict information. At the same time, existing methods often use simple time synchronization strategy when dealing with sensor time difference, ignoring the system dynamic characteristics reflected by time difference, making it difficult to accurately identify and warn the complex running state of aviation food cart. Therefore, an aviation food cart intelligent monitoring method is needed to solve at least one of the above problems. SUMMARY

[0004] The present application discloses a multi-sensor fusion aviation food cart intelligent monitoring and early warning method and device, aiming to deeply mine conflict characteristics and time difference information between sensors, build an adaptive state recognition mechanism, accurately monitor the running state of aviation food cart, and provide reliable technical support for food safety and equipment maintenance.

[0005] The present application discloses a multi-sensor fusion aviation food cart intelligent monitoring and early warning method and device, aiming to deeply mine conflict characteristics and time difference information between sensors, build an adaptive state recognition mechanism, accurately monitor the running state of aviation food cart, and provide reliable technical support for food safety and equipment maintenance. Collecting original monitoring data of the aviation food cart, detecting sensor conflicts of the original monitoring data to identify the degree of inconsistency between sensors, generating a conflict enhancement coefficient based on the degree of inconsistency; Obtaining cross-sensor time difference characteristics of the original monitoring data, constructing a delay utilization matrix through the cross-sensor time difference characteristics, reconstructing a probability chain of the original monitoring data through the delay utilization matrix to generate a reverse reference vector, generating an enhanced feature set based on the reverse reference vector to reconstruct the data feature space; The conflict enhancement coefficient is used for conflict enhancement processing on the enhanced feature set to generate an adversarial feature vector, and a self-adversarial boundary is determined by analyzing adversarial distribution characteristics of the adversarial feature vector in a feature space; and a dynamic balance discrimination mechanism is established based on the self-adversarial boundary to generate a state recognition parameter; The adversarial feature vector is subjected to state separation processing under the constraint of the state recognition parameter to generate a multi-layer state classification matrix, and a state conversion trajectory graph is formed by monitoring inter-layer response changes of the multi-layer state classification matrix; and a warning trigger parameter is generated by performing trajectory analysis on the state conversion trajectory graph; The state conversion trajectory graph is tracked to obtain a state switching frequency by performing frequency inversion on the state conversion trajectory graph; a warning intensity coefficient is generated by associating the state switching frequency with the warning trigger parameter; and a multi-level warning output signal is generated by performing warning level matching on the warning intensity coefficient and the multi-layer state classification matrix.

[0006] The second aspect of the application provides an intelligent monitoring and early warning device for an aviation food cart based on multi-sensor fusion, comprising: A data acquisition module is configured to acquire original monitoring data of the aviation food cart, detect data inconsistency between sensors by performing sensor conflict detection on the original monitoring data, and generate a conflict enhancement coefficient based on the data inconsistency; A time difference processing module is configured to obtain cross-sensor time difference characteristics of the original monitoring data, construct a delay utilization matrix based on the cross-sensor time difference characteristics, reconstruct a probability chain based on the original monitoring data by using the delay utilization matrix to generate an inverted reference vector, and reconstruct a data feature space based on the inverted reference vector to generate an enhanced feature set; An adversarial analysis module is configured to perform conflict enhancement processing on the enhanced feature set by using the conflict enhancement coefficient to generate an adversarial feature vector, determine a self-adversarial boundary by analyzing adversarial distribution characteristics of the adversarial feature vector in a feature space, and establish a dynamic balance discrimination mechanism based on the self-adversarial boundary to generate a state recognition parameter; A state classification module is configured to perform state separation processing on the adversarial feature vector under the constraint of the state recognition parameter to generate a multi-layer state classification matrix, monitor inter-layer response changes of the multi-layer state classification matrix to form a state conversion trajectory graph, and perform trajectory analysis on the state conversion trajectory graph to generate a warning trigger parameter; A warning output module is configured to track a state switching mode of the state conversion trajectory graph, perform frequency inversion on the state conversion trajectory graph to obtain a state switching frequency, associate the state switching frequency with the warning trigger parameter to generate a warning intensity coefficient, and perform warning level matching on the warning intensity coefficient and the multi-layer state classification matrix to generate a multi-level warning output signal.

[0007] The beneficial effects of the present application are embodied in the following points: first, through sensor conflict detection and resonance frequency analysis technology, the inconsistency of data between sensors is converted into conflict enhancement coefficient, realizing the effective use of conflict data which is traditionally regarded as interference signal. When the refrigeration system of the aircraft food cart appears abnormal, the response conflict of temperature and pressure sensors can be detected and identified in advance, which can gain processing time for abnormal early warning. Second, the cross-sensor time difference feature analysis and delay utilization matrix construction technology are used to reconstruct the data feature space and generate enhanced feature set, combined with the anti-feature vector analysis and self-adversarial boundary determination technology to generate state recognition parameters, which converts the single data stream of multiple independent sensors into high-dimensional feature representation with adversarial characteristics, establishes a complete parameter system from feature enhancement to state discrimination, and makes the system able to distinguish normal operation, equipment start, load change and other operation modes. Finally, a multi-layer state classification matrix and state transition trajectory analysis mechanism are established, and the alarm intensity coefficient is generated through frequency inversion and correlation analysis, realizing multi-level alarm output, which can provide different levels of alarm response according to the degree of abnormality, reducing the false alarm and missing alarm phenomenon, and providing hierarchical abnormal handling strategy for the aircraft food cart operators.

[0008] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0009] The drawings herein show specific examples of the technical solutions described in the present application, and constitute part of the specification together with the specific embodiments, for explaining the technical solutions, principles and effects of the present application.

[0010] Unless specifically stated or defined otherwise, the same reference signs in different drawings represent the same or similar technical features, and different reference signs may also be used to represent the same or similar technical features.

[0011] Figure 1 is a flow diagram of the multi-sensor fusion intelligent monitoring and early warning method of the aircraft food cart of the present application.

[0012] Figure 2 is a structural block diagram of the multi-sensor fusion intelligent monitoring and early warning device of the aircraft food cart of the present application. DETAILED DESCRIPTION

[0013] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, circuits, and

[0014] It is to be understood that the terminology "includes", "comprises", "consists of", or "consisting of" when used in the present specification and in the appended claims, indicates open-ended language such that "comprising" also includes "consisting of".

[0015] Reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. Thus, the appearances of the phrases "in one embodiment" or "in an embodiment" or "in other embodiments" or "in some embodiments" in various places throughout this specification are not necessarily referring to the same embodiment, unless otherwise specified. The terms "including", "comprising", "featuring", and variations thereof are meant to be broad and encompass the terms "consisting of" and "consisting essentially of", unless otherwise noted.

[0016] The technical solutions of the embodiments of the present application are introduced as follows.

[0017] As shown in Figure 1 The method for intelligent monitoring and early warning of an aerial food cart based on multi-sensor fusion provided by the embodiments of the present application includes the following steps S110-S150: In step S110, original monitoring data of the aerial food cart are collected, sensor conflict detection is performed on the original monitoring data to identify the degree of inconsistency between data of different sensors, and a conflict enhancement coefficient is generated based on the degree of inconsistency.

[0018] Specifically, the original monitoring data of the airborne food cart is collected. Through the multi-type sensor network deployed inside the airborne food cart, various monitoring parameters during the operation of the airborne food cart are continuously collected. The temperature sensor array is distributed in the storage area, processing area and display area of the airborne food cart to monitor the temperature changes in each area. The humidity sensor is configured in the air vent and sealed area of the airborne food cart to measure the air humidity and moisture content. The pressure sensor is installed in the refrigeration system and gas pipeline of the airborne food cart to monitor the system pressure and leakage. The vibration sensor is fixed on the chassis and equipment support of the airborne food cart to detect the running vibration and equipment abnormalities. The gas sensor is arranged at the exhaust port and internal space of the airborne food cart to monitor the concentration of carbon monoxide, carbon dioxide and harmful gases. The current sensor is connected to the main circuit and device power line of the airborne food cart to monitor the current change and power consumption. All sensors use a unified data collection frequency of 100 data points per second to ensure time synchronization. The original monitoring data includes sensor identification, timestamp, numerical value, sensor state and environmental conditions. The data collection system uses a distributed architecture, with each sensor node independently collecting and preprocessing data and transmitting it to the central data processing unit through a wireless communication module.

[0019] The original monitoring data is subjected to sensor conflict detection to identify the degree of inconsistency between different sensors. Using the collected original monitoring data of the airborne food cart, the data conflict between different sensors is detected through multi-sensor data comparison and analysis. The measurement results of different sensors at the same time are cross-compared to identify the differences in data values and inconsistencies in trends. The measurement difference of each sensor in the temperature sensor array is calculated, and when the temperature difference between adjacent sensors exceeds 2 degrees Celsius, it is marked as a temperature conflict. The response time difference of the humidity sensor is analyzed, and the time delay between sensors is calculated through cross-correlation analysis. The measurement consistency of the pressure sensor is detected, and the standard deviation and coefficient of variation of each sensor reading are calculated. The frequency response of the vibration sensor is compared to identify the difference characteristics of the frequency spectrum distribution. For example, when the airborne food cart refrigeration system starts, the front temperature sensor shows 18 degrees Celsius and the rear temperature sensor shows 22 degrees Celsius, with a difference of 4 degrees Celsius exceeding the threshold, and the left humidity sensor responds 0.8 seconds slower than the right sensor, which constitutes a data conflict between sensors. The relative difference between sensor pairs is calculated using the data inconsistency measure D = |Vi-Vj| / max(Vi,Vj), where D is the inconsistency degree, Vi and Vj are the measurement values of the two sensors. The inconsistency degrees of all sensor pairs are counted to form a data inconsistency degree matrix.

[0020] In some embodiments, the generating the conflict enhancement coefficient based on the data inconsistency degree comprises: scanning the data inconsistency degree to identify periodic conflict resonance points; extracting resonance frequency features based on the periodic conflict resonance points to generate a system feature signature; performing resonance matching on the system feature signature and the data inconsistency degree to generate a resonance reinforcement factor; and performing coefficient transformation using the resonance reinforcement factor to generate the conflict enhancement coefficient.

[0021] The data inconsistency degree is scanned to identify periodic conflict resonance points. A sliding window method is used to scan the data inconsistency degree between sensors in each time period, with a window length of 60 seconds and a sliding step of 10 seconds. The amplitude variation of the inconsistency degree in each time window is calculated, and the time-domain inconsistency signal is converted to the frequency domain space by fast Fourier transform. The main frequency components of the inconsistency degree are identified in the frequency domain, focusing on the energy distribution in the low and medium frequency bands. A spectral peak detection method is used to identify significant peaks in the frequency domain, and the peaks correspond to the characteristic frequency of data conflict. When the inconsistency degree of multiple sensors appears peaks at the same frequency, the frequency is determined as a periodic conflict resonance point. The quality factor Q = f0 / Δf of the resonance point is calculated, where Q is the quality factor, f0 is the resonance frequency, and Δf is the half-power bandwidth. The quality factor reflects the sharpness and stability of the resonance point. The identified resonance points are tracked in time, and the frequency drift and amplitude variation of the resonance points are recorded. The time delay relationship between the conflicting sensors is determined by phase analysis. A clustering method is used to merge similar resonance frequencies into resonance clusters, and each resonance cluster represents a specific conflict mode. The stability of the resonance points is analyzed, and the recurrence rate and frequency stability of the resonance points in different time periods are calculated.

[0022] Resonance frequency features are extracted based on the periodic conflict resonance points to generate a system feature signature. The frequency position, amplitude intensity, and bandwidth characteristics of the resonance points are analyzed. The frequency interval of each resonance point is calculated, and the frequency interval and harmonic structure are identified. Wavelet transform is used for time-frequency analysis of the resonance points to obtain the time-varying features of the resonance phenomenon. The spectral density distribution of the resonance points is calculated, and the power spectral density function is used to describe the frequency distribution of the resonance energy. Statistical features of the resonance points are extracted, including average frequency, frequency variance, peak amplitude, and frequency bandwidth. Principal component analysis is used to reduce the dimensionality of the extracted frequency features, retaining the main components that explain more than 85% of the variance. The normalized feature vectors after dimensionality reduction are processed to unify the numerical range of each feature component. Feature encoding is used to convert the frequency features into binary encoding sequences to form a compact feature representation. The encoded feature sequence is organized into a system feature signature, which includes position encoding, intensity encoding, and association encoding of the resonance frequency. The integrity of the system feature signature is checked to ensure that the signature can uniquely identify the current conflict mode.

[0023] The resonance matching of the system feature signature and the data inconsistency degree generates a resonance enhancement factor. The system feature signature is used as a template for pattern matching analysis with real-time data inconsistency degree. The template matching method is used to calculate the similarity between the system feature signature and the current inconsistency degree pattern. The normalized cross-correlation function is used to quantify the matching degree, and the correlation coefficient between the two signals is calculated to measure the matching strength. When the correlation coefficient exceeds 0.8, it is determined as strong matching, 0.5-0.8 as moderate matching, and less than 0.5 as weak matching. For strong matching, the phase shift and amplitude ratio of the matching are calculated to determine the parameters of resonance enhancement. The dynamic time warping method is used to process the nonlinear matching on the time axis, allowing local time stretching to improve the matching accuracy. The energy transfer efficiency in the matching process is calculated to quantify the amplification effect of the system feature signature on the data inconsistency degree. The resonance coupling process is simulated by frequency domain convolution operation to calculate the multiplication effect of resonance enhancement. The amplitude and phase of the resonance enhancement effect are decomposed to calculate the amplitude and phase enhancement factors, respectively. The amplitude and phase enhancement factors are combined into a complex resonance enhancement factor F=A·e^(jφ), where F is the resonance enhancement factor, A is the amplitude enhancement factor, φ is the phase enhancement factor, and j is the imaginary unit.

[0024] The resonance enhancement factor is used to perform coefficient transformation to generate conflict enhancement coefficients. The original data inconsistency degree is nonlinearly transformed using the resonance enhancement factor to amplify the performance strength of the conflict feature. The transformation function is calculated as where T(x) is the transformed signal, x is the original inconsistency degree, |F| is the modulus of the resonance enhancement factor, is the phase angle, and a is the nonlinear coefficient. Through the action of the transformation function, the weak conflict signal is amplified to a detectable level, enhancing the recognition ability of the conflict pattern. The envelope of the transformed signal T(x) is extracted, and the Hilbert transform is used to obtain the instantaneous amplitude of the signal. The statistical characteristics of the envelope signal are calculated, including mean, standard deviation, skewness, and kurtosis parameters. An adaptive threshold method is used to segment the envelope signal to identify the effective interval of conflict enhancement. In each effective interval, the maximum value of the envelope amplitude is taken as the conflict enhancement coefficient basis value. The coefficient of variation of the envelope amplitude in the time window is used to modify the basis value, and the coefficient of variation reflects the relative fluctuation degree of the amplitude, and the calculation method is the ratio of the standard deviation to the mean. The final conflict enhancement coefficient K= basis value ×(1+coefficient of variation) is calculated, where K is the conflict enhancement coefficient. The larger the coefficient of variation, the more intense the signal fluctuation, and the more obvious the conflict feature, so a larger enhancement coefficient is needed. The sliding average method is used to smooth the conflict enhancement coefficient to eliminate the influence of sudden noise. The conflict enhancement coefficient is normalized to map the coefficient value to the standard interval of 0-1, ensuring the uniformity of the numerical range of the coefficient.

[0025] In step S120, a cross-sensor time difference feature of the original monitoring data is acquired, a delay utilization matrix is constructed through the cross-sensor time difference feature, a reverse reference vector is generated by performing probability chain reconstruction on the original monitoring data through the delay utilization matrix, and an enhanced feature set is generated by reconstructing a data feature space based on the reverse reference vector.

[0026] Specifically, the cross-sensor time difference feature of the original monitoring data is acquired. The data of the multiple types of sensors is aligned and analyzed according to timestamps. The time delay between any two sensors is calculated, and the time offset corresponding to the maximum correlation is identified by using a cross-correlation analysis method. For example, when the refrigeration system of an aviation food cart is started, the temperature sensor detects a temperature drop within 2 seconds, while the pressure sensor detects a change in refrigerant pressure only after 2.3 seconds, and the time difference between the two is 0.3 seconds. At the same time, the current sensor detects an increase in the current of the refrigeration compressor at 1.8 seconds, forming a response time difference pattern between different sensors. The time difference metric D = |t1-t2| is calculated for each pair of sensors, where D is the time difference, and t1 and t2 are the response times of the two sensors, respectively. The time difference distribution of all sensor pairs is statistically analyzed to form a cross-sensor time difference matrix. The periodicity of the time difference is analyzed to identify the inherent delay pattern caused by differences in sensor physical location, response speed, and signal transmission path. The characteristic parameters of the time difference are extracted, including the average delay, delay variance, maximum delay, and skewness of the delay distribution. The periodic components in the time difference are identified by frequency domain analysis, and the frequency characteristics of the delay are extracted by using power spectrum analysis. The cross-sensor time difference features are organized into a feature vector, and each element corresponds to the time relationship characteristics between a pair of sensors.

[0027] In some embodiments, the delay utilization matrix is constructed through the cross-sensor time difference feature, including: building a time difference analysis grid according to the cross-sensor time difference feature; identifying a reverse feature node from the time difference analysis grid to generate a reverse mark; performing pattern matching transformation on the time difference analysis grid using the reverse mark to generate a time difference layout; and analyzing the time difference layout to form a delay utilization matrix.

[0028] A time difference analysis grid is built based on cross-sensor time difference features. Cross-sensor time difference features are arranged in two dimensions according to sensor types and spatial positions, forming a regular grid structure. The rows of the grid correspond to source sensors, and the columns correspond to target sensors. The grid intersection stores the time difference value of the corresponding sensor pair. The time difference value is quantized and classified, and the continuous time difference value is discretized into a limited time delay level. The time difference analysis grid is visualized using color coding, with different color depths representing the size of the time delay. Time synchronization points are identified in the grid, which are the positions of sensor pairs with time difference close to zero. The spatial distribution pattern of time difference in the grid is analyzed to identify sensor clusters with similar delay characteristics. The gradient change between adjacent nodes in the grid is calculated to identify areas of time difference mutation and smoothness. Spatial filtering is performed on the grid to eliminate time difference jitter caused by measurement noise. Interpolation is used to fill in missing data points in the grid to ensure the integrity and continuity of the time difference analysis grid.

[0029] Reversal markers are identified from the time difference analysis grid. Special node positions where the time difference sign reverses are identified in the time difference analysis grid. Reversal feature nodes exhibit a change in the direction of time delay for adjacent sensor pairs, from positive delay to negative delay or vice versa. For example, under normal operation, the temperature sensor responds 0.5 seconds earlier than the humidity sensor (positive delay), but when the aircraft food cart door is opened, the humidity sensor responds 0.2 seconds earlier than the temperature sensor (negative delay) due to direct contact with external air. This change in delay direction constitutes a reversal feature node. The time difference gradient of each grid node is calculated, and the reversal occurrence position is identified by the gradient sign change. A second derivative analysis method is used to detect the inflection point of the time difference curve, and the inflection point position corresponds to the potential reversal feature node. Cluster analysis is performed on the identified reversal feature nodes, and spatially adjacent reversal nodes are merged into reversal regions. Each reversal feature node is assigned a unique reversal marker, which contains node coordinates, reversal intensity, and reversal direction information. The influence range of the reversal feature node is calculated to determine the influence radius of each reversal node on the surrounding time difference distribution. The stability of the reversal feature node is evaluated through statistical analysis to identify stable reversal points that persist and temporary reversal points that occur occasionally. The reversal markers are sorted according to importance, and the reversal feature nodes with large influence range and high stability are prioritized for processing.

[0030] The time difference layout is generated by performing pattern matching transformation on the time difference analysis grid using the inverted marker. The inverted marker is used as a pattern template to search for a matching time difference pattern in the time difference analysis grid. The similarity between the inverted marker and the local region of the grid is calculated by a template matching algorithm to identify the grid region with similar inverted features. The coordinate transformation is performed on the grid region with a successful match to adjust the local time difference distribution to the standardized inverted pattern. The grid is geometrically adjusted using an affine transformation method, including translation, rotation, and scaling operations, to make the grid layout conform to the spatial characteristics of the inverted marker. The energy function of the transformed grid is calculated to optimize the overall consistency of the time difference layout by minimizing the energy function. An iterative optimization algorithm is used to gradually adjust the grid node positions until a stable time difference layout configuration is achieved. The grid topology changes during the transformation are constrained to maintain the basic connection relationship between sensors unchanged. The quality indicators of the transformed time difference layout are evaluated, including the regularity, symmetry, and stability of the layout. The parameter changes during the transformation are recorded to form a complete transformation path from the original grid to the final time difference layout.

[0031] The delay utilization matrix is formed by analyzing the time difference layout. The optimized time difference layout is converted into a numerical delay utilization matrix representation. The rows and columns of the matrix correspond to the sensor numbers involved in time synchronization, and the matrix element numerical value represents the delay utilization weight of the corresponding sensor pair. The delay utilization weight is calculated according to the time difference distribution in the time difference layout, and the weight value reflects the contribution of the delay relationship to the overall data synchronization. Normalization processing is used to ensure the numerical range of the delay utilization matrix uniform, and the matrix element value is controlled between 0 and 1. The consistency of the delay relationship is verified by the symmetry analysis of the matrix, and the diagonal elements are set to 1 to represent the complete synchronization of the sensor with itself. The condition number and rank of the delay utilization matrix are calculated to evaluate the numerical stability and information integrity of the matrix. The delay utilization matrix is sparsified by setting the matrix elements with small numerical values to zero, reducing the storage space and computational complexity. The delay utilization matrix is saved in a compressed storage format, storing only the non-zero elements and their corresponding row and column indices.

[0032] In some embodiments, the probability chain reconstruction of the original monitoring data using the delay utilization matrix to generate the inverted reference vector includes: performing a probability node extraction operation on the delay utilization matrix to form a node probability graph; detecting a link path from the node probability graph to generate a path weight factor; and performing chain reconstruction on the original monitoring data according to the path weight factor to generate the inverted reference vector.

[0033] The probability node extraction operation is performed on the delay utilization matrix to form a node probability graph. Each non-zero element of the delay utilization matrix is regarded as a probability connection node, and the element value represents the connection probability strength. The probability value P of each node is calculated as P = w / ∑w, where P is the normalized probability, w is the matrix element weight, and ∑w is the total weight of the same row elements. The probability value of each row is normalized to ensure that the sum of the probability values of each row is equal to 1, meeting the basic requirement of probability distribution. A probability transition matrix is constructed, and the matrix elements represent the probability of transition from one sensor node to another node. The statistical characteristics of the probability distribution are analyzed, and the mean, variance, and entropy value of the probability distribution are calculated. High-probability nodes and low-probability nodes in the probability distribution are identified, and the high-probability nodes correspond to important sensor connection relationships. A probability graph model is used to represent the dependence relationship between sensors, with nodes representing sensors and edges representing probability connection strength. The degree distribution of the nodes is calculated, the in-degree and out-degree of each node are counted, and the connection characteristics of the network are analyzed. The probability analysis results are organized into a node probability graph, which contains node probability values, connection relationships, and network topology information.

[0034] The link path is detected from the node probability graph to generate a path weight factor. In the node probability graph, the data propagation path connecting different sensors is searched, and the high-probability connection path is identified. The shortest path algorithm is used to find the data transmission path with the maximum probability weight, and the path length is defined as the reciprocal of the probability weight. The Dijkstra algorithm is used to calculate the optimal probability path from the source node to the target node, optimizing the calculation efficiency of path selection. The probability distribution of multiple parallel paths is analyzed, and the main path and alternative path are identified. The reliability index of the path is calculated, and the transmission reliability of the overall path is evaluated by multiplying the probabilities of each segment on the path. The identified link paths are classified according to path length, probability strength, and transmission delay, and the paths are divided into different types. The weight factor of each path is calculated where Pi is the probability of the i-th segment of the path, and n is the number of path segments. The weight factor reflects the importance and priority of the path in data reconstruction. The normalization of the path weight ensures that the sum of all path weight factors is equal to 1. The mapping relationship between the path weight factor and the corresponding path is recorded to form a complete path weight configuration table.

[0035] For example, the chain reconstruction of the original monitoring data according to the path weight factor to generate an inverted reference vector includes: analyzing the weight jump feature in the path weight factor; locking the available weight imbalance interval in the weight jump feature; evaluating the reconstruction strength of the weight imbalance interval to generate a reconstruction coefficient; and performing chain reconstruction on the original monitoring data according to the weight imbalance interval and the reconstruction coefficient to generate an inverted reference vector.

[0036] The weight jump feature in the path weight factor is analyzed. The calculated path weight factor is analyzed by difference, and the sudden increase or decrease of the weight value is identified. The difference ΔW between the adjacent path weights is calculated, where W_i is the weight factor of the i-th path, and W_i+1 is the weight factor of the i+1-th path. The significance of weight change is judged by setting a jump threshold, and the weight jump is considered to occur when the weight difference exceeds the threshold. The spatial distribution characteristics of the weight jump are analyzed, and the path position and sensor node where the jump occurs are identified. The amplitude and duration of the weight jump are calculated to evaluate the strength and influence range of the jump phenomenon. The sliding window method is used for local analysis of the weight sequence, and the weight jump feature is identified under different window scales. The periodic characteristics of the weight jump are extracted by spectral analysis, and the regularity pattern of the weight change is identified. The weight jump features are clustered and similar jump patterns are merged into jump types. The parameter information of each weight jump feature is recorded, including the jump position, jump amplitude, jump direction, and duration.

[0037] The available weight imbalance interval is locked in the weight jump feature. The imbalance interval with reconstruction value is selected from the identified weight jump feature. The weight imbalance interval is defined as the path segment where the weight distribution deviates significantly from the uniform distribution. These intervals contain rich data reconstruction information. For example, in the 5-path sensor network of the aviation food cart, the ideal uniform distribution should be that each path weight is 0.2, but the actual measurement finds that the weights of paths 1-3 are 0.05, 0.1, and 0.15 (lower), while the weights of paths 4-5 are 0.3 and 0.4 (higher), forming a clear imbalance distribution, among which the high weight interval composed of paths 4-5 is the available weight imbalance interval. The imbalance degree U=σ² / μ² of each interval is calculated, where U is the imbalance degree, σ is the standard deviation of the weights in the interval, and μ is the weight mean. The intervals with imbalance degree exceeding the set threshold are selected as available intervals, and the weight distribution of these intervals has obvious non-uniformity. The spatial continuity of the available interval is analyzed, and adjacent imbalance intervals are merged into extended intervals. The effective length and coverage range of each available interval are calculated to evaluate the potential contribution of the interval to data reconstruction. The boundary of the locked weight imbalance interval is accurately positioned to determine the start and end positions of the interval.

[0038] The reconstruction strength of the weight imbalance interval is evaluated to generate a reconstruction coefficient. The strength indicator of data reconstruction is calculated for each locked weight imbalance interval. The reconstruction strength reflects the ability and influence degree of the interval on the transformation of the original monitoring data. The dynamic range R of the weights in the interval is calculated, where R is the dynamic range, and W_max and W_min are the maximum and minimum weights in the interval, respectively. The change intensity of the interval is calculated by weight gradient analysis, and the greater the gradient, the higher the reconstruction strength. The information complexity of the interval is quantified using the information entropy method, and the interval with higher complexity has stronger data reconstruction ability. The reconstruction coefficient C is calculated as C = R x H, where R is the dynamic range and H is the information entropy. The reconstruction coefficient considers the amplitude and complexity of the weight change, and the larger the value, the higher the reconstruction strength. The reconstruction coefficient is normalized to map the coefficient value to the standard interval of 0 to 1. The intervals with larger reconstruction coefficients are retained by threshold screening, and these intervals play an important role in the generation of the inversion reference vector.

[0039] The chain reconstruction is performed on the original monitoring data according to the weight imbalance interval and the reconstruction coefficient to generate the inversion reference vector. The processing priority of each imbalance interval is determined according to the size of the reconstruction coefficient, and the interval with a larger coefficient is given priority to participate in the reconstruction process. In each imbalance interval, the original monitoring data is weighted and combined according to the weight distribution, and the data component with a larger weight obtains a higher reconstruction weight. A segmented reconstruction strategy is used to decompose the entire data reconstruction process into local reconstruction of multiple intervals, and then the local results are globally integrated. For example, when the system identifies three imbalance intervals of temperature sensor path, humidity sensor path and pressure sensor path, the temperature monitoring data is first locally reconstructed in the temperature path interval to generate a local inversion vector representing the temperature feature; then the humidity data is processed similarly in the humidity path interval to generate a local vector of humidity feature; finally, the data in the pressure path interval is processed; then the three local inversion vectors are weighted and fused according to their respective reconstruction coefficients to form a complete inversion reference vector containing all sensor features. The intermediate results of segmented reconstruction are calculated, and each interval generates a local inversion vector component. The local inversion vectors of each interval are combined into a complete inversion reference vector V = Σ(C_i x V_i) by weighted fusion, where C_i is the reconstruction coefficient of the i-th interval, and V_i is the corresponding local inversion vector. The generated inversion reference vector is normalized in length to ensure that the Euclidean norm of the vector is equal to 1.

[0040] The enhanced feature set is generated based on reconstruction of the data feature space with inverted reference vectors. The inverted reference vectors are used as new coordinate bases to represent the original monitoring data in the reconstructed feature space. The projection components of the original data in the direction of the inverted reference vectors are calculated, and the projection results represent the coordinates of the data in the new feature space. The Schmidt orthogonalization method is used to extend the inverted reference vectors and construct a complete orthogonal basis set. The main variation directions in the reconstructed feature space are identified by principal component analysis, and the most discriminative feature components are extracted. The statistical properties of the reconstructed features, including mean, variance, correlation and distribution characteristics, are calculated. Feature selection algorithms are used to select the most valuable feature components from the reconstructed feature space to form a candidate set of enhanced features. The candidate features are ranked by feature importance scores, and the highest scoring features are selected to form the final enhanced feature set. The enhanced feature set is normalized and standardized to ensure that the feature values have consistent dimensions and reasonable distributions.

[0041] In step S130, the enhanced feature set is processed by a conflict enhancement coefficient to generate an adversarial feature vector, the adversarial distribution characteristics of the adversarial feature vector in the feature space are analyzed to determine a self-adversarial boundary, and a dynamic balance discrimination mechanism is established based on the self-adversarial boundary to generate a state recognition parameter.

[0042] Specifically, the enhanced feature set is processed by a conflict enhancement coefficient to generate an adversarial feature vector. The conflict enhancement coefficients are sorted by value to identify the highest conflict intensity coefficient components as the focus of reinforcement. Each feature component in the enhanced feature set is applied with the corresponding conflict enhancement coefficient, and the reinforced feature value F' = F x K is calculated, where F is the original feature value and K is the conflict enhancement coefficient. The differences between features are amplified by nonlinear transformation, and the reinforced features are activated by a sigmoid function. The difference vector before and after feature reinforcement is calculated, and the difference reflects the effect of conflict enhancement. Adversarial samples are constructed using the idea of adversarial training, and adversarial feature variants are generated by adding perturbations in the feature direction. For example, when the temperature feature and the humidity feature conflict, the conflict enhancement coefficient will amplify the difference between them, making the temperature feature deviate 0.3 standard deviations in the high temperature direction and the humidity feature deviate 0.2 standard deviations in the low humidity direction, forming a clear adversarial feature pair. The gradient direction of the adversarial perturbation is calculated to ensure that the perturbation can maximize the conflict between features. All reinforced feature components are combined into an adversarial feature vector, which contains amplified conflict information and adversarial perturbations.

[0043] In some embodiments, the analyzing the adversarial distribution characteristics of the adversarial feature vectors in the feature space to determine the adversarial boundary comprises: identifying imbalance regions by spatially layering the adversarial feature vectors according to adversarial intensity; extracting imbalance degree from the imbalance regions to generate a warning level factor; performing boundary search transformation on the imbalance regions based on the warning level factor to generate a dynamic safety line; and fusing the dynamic safety line with the warning level factor to determine the adversarial boundary.

[0044] The imbalance regions are identified by spatially layering the adversarial feature vectors according to adversarial intensity. The adversarial intensity is calculated according to the length and direction information of the adversarial feature vectors, and reflects the intensity of feature conflict. The adversarial intensity is divided into multiple levels using a quantile method, with each level corresponding to a different intensity of adversarial conflict. The feature space is layered according to the adversarial intensity levels, forming a hierarchical spatial structure. The spatial clustering patterns of the adversarial feature vectors are analyzed within each intensity layer to identify regions of uneven feature distribution. The variance of feature density within each layer is calculated, with regions of high variance indicating uneven feature distribution. The boundaries between sparse and dense regions are identified through density gradient analysis. The K-means clustering method is used to identify the center positions of the imbalance regions within each intensity layer. The geometric parameters of the imbalance regions, including area, shape coefficient, and boundary curvature, are calculated. The spatial correlation of imbalance regions between different intensity layers is detected through overlap analysis. The identified imbalance regions are labeled and numbered, and their spatial coordinates and intensity level information are recorded.

[0045] The imbalance degree of each identified imbalance region is calculated to generate a warning level factor. The imbalance degree reflects the severity of the region's deviation from an equilibrium state. The dispersion D = √(Σ(xi-x_mean)² / n) of the adversarial feature vectors within the imbalance region is calculated, where xi is the ith feature vector within the region, x_mean is the mean of the feature vectors within the region, and n is the number of vectors. The skewness and kurtosis of the imbalance region are calculated through statistical analysis, with skewness reflecting the asymmetry of the distribution and kurtosis reflecting the sharpness of the distribution. The complexity of the imbalance region is quantified using the information entropy method, with higher entropy values indicating greater complexity. The KL divergence between the imbalance region and the normal region is calculated, with the divergence size reflecting the degree of deviation from the normal state. The warning level is divided into three levels: mild warning, moderate warning, and severe warning, based on the size of the imbalance degree. Each warning level is assigned a corresponding factor value, with the mild warning factor being 0.3, the moderate warning factor being 0.6, and the severe warning factor being 0.9. For example, when the refrigeration system of an aviation food cart is abnormal, the imbalance degree of the temperature and pressure sensor data reaches the severe warning level, and the warning level factor is set to 0.9, indicating that immediate attention should be paid to the imbalance region. The comprehensive warning level factor of the composite imbalance region is calculated by weighted averaging.

[0046] A dynamic safety line is generated by performing a boundary search transformation on the imbalance region based on the early warning level factor. The identified imbalance region is processed by a boundary search using the calculated early warning level factor. The early warning level factor is used as a search intensity control parameter, with higher early warning level factors resulting in a more refined boundary search step size for the imbalance region. The search boundary position is expanded outward from the center point of each imbalance region, with the search radius determined by the early warning level factor. The feature gradient change at the edge of the imbalance region is calculated, and the initial boundary points are determined by the gradient discontinuity positions. The boundary points are connected to form a closed boundary contour, with the contour forming a safety protection line around the imbalance region. For example, when the temperature sensor and humidity sensor data of an aircraft food cart show a severe imbalance, the early warning level factor is 0.9, and the system searches for the boundary around the imbalance region, finding a combination of temperature 35°C and humidity 80% that forms a boundary turning point, and connecting these turning points forms a dynamic safety line. The searched boundary is smoothed using a moving average method to eliminate the jagged shape of the boundary. The tightness of the boundary is adjusted according to the value of the early warning level factor, with the boundary of a high early warning region closer to the imbalance region. The processed boundary is named the dynamic safety line, with the position of the safety line adjusted according to the change in the early warning level factor. The coordinate point set of the dynamic safety line and the corresponding early warning level factor value are recorded.

[0047] A self-adaptive boundary is determined by fusion processing of the dynamic safety line and the early warning level factor. The generated dynamic safety line and the corresponding early warning level factor are fused to form the final self-adaptive boundary. Each coordinate point on the dynamic safety line is assigned a corresponding early warning level factor weight, which reflects the importance of that position. The new coordinates of the fusion boundary point B = S + λ · W are calculated, where B is the fused boundary point, S is the dynamic safety line coordinate, λ is the fusion coefficient, and W is the early warning level factor weighted offset. The fusion processing causes the boundary to shrink inward in high early warning regions and expand outward in low early warning regions, forming an adaptive boundary shape. For example, in an abnormal region of an aircraft food cart refrigeration system, the early warning level factor is 0.8, and the fusion processing causes the boundary of this region to shrink inward by 2 units, while in a normal operating region, the early warning level factor is 0.2, and the boundary expands outward by 1 unit, forming a non-uniform adaptive boundary. All fused boundary points are connected to form a complete self-adaptive boundary contour. The geometric parameters of the self-adaptive boundary are calculated, including the boundary length, enclosed area, and boundary curvature.

[0048] The dynamic balance discrimination mechanism is generated based on the self-adversarial boundary to establish state recognition parameters. The self-adversarial boundary is used as a discrimination benchmark for state classification, and the internal region of the boundary is defined as a normal state, and the external region of the boundary is defined as an abnormal state. The distance d = |P-B| of the adversarial feature vector to the self-adversarial boundary is calculated, wherein P is the coordinate of the adversarial feature vector, and B is the coordinate of the nearest point on the boundary. According to the size of the distance value, the state level is divided, and the distance less than 0.2 units is determined as a normal state, the distance between 0.2-0.5 units is determined as a pre-warning state, and the distance greater than 0.5 units is determined as an abnormal state. The weight distribution rule of state discrimination is established, and different discrimination weights are allocated to the distances of different feature directions. For example, the temperature feature vector of the aviation food cart is 0.15 units away from the self-adversarial boundary, and the humidity feature vector is 0.35 units away from the boundary. After comprehensive calculation, the system state is at the pre-warning level, and the temperature control parameter is set to 0.7 and the humidity control parameter is set to 0.8. The threshold parameters of the balance discrimination are calculated, including the normal state threshold T1 = 0.2, the pre-warning state threshold T2 = 0.5, and the abnormal state threshold T3 = 1.0. According to different state levels, corresponding recognition parameters are generated, and the normal state parameter a = 0.3, the pre-warning state parameter a = 0.6, and the abnormal state parameter a = 0.9. All discrimination thresholds and state parameters are organized into a state recognition parameter set, which includes a threshold vector, a weight vector, and a control parameter vector.

[0049] In step S140, state separation processing is performed on the adversarial feature vector under the constraint of the state recognition parameters to generate a multi-layer state classification matrix, the interlayer response change of the multi-layer state classification matrix is monitored to form a state transition trajectory diagram, and the state transition trajectory diagram is analyzed to generate an alarm triggering parameter.

[0050] Specifically, the state separation processing is performed on the confrontation feature vector under the constraint of the state recognition parameter to generate a multi-layer state classification matrix. Based on the state recognition parameter, the normal state threshold T1, the pre-warning state threshold T2, and the abnormal state threshold T3 in the threshold vector are taken as the separation standard to classify the state of the confrontation feature vector. The comparison result of each confrontation feature vector with each threshold is calculated, and when the feature value is less than T1, it is classified into the normal state layer, when the feature value is between T1 and T2, it is classified into the pre-warning state layer, and when the feature value is greater than T2, it is classified into the abnormal state layer. The different feature components are weighted by using the weight vector, the temperature feature weight is 0.4, the humidity feature weight is 0.3, and the pressure feature weight is 0.3. The separation sensitivity of each state layer is adjusted by using the control parameter vector, the normal state control parameter a = 0.3, the pre-warning state control parameter a = 0.6, and the abnormal state control parameter a = 0.9. For example, when the temperature confrontation feature value of the aircraft food cart at a certain moment is 0.15, the humidity confrontation feature value is 0.35, and the pressure confrontation feature value is 0.25, the comprehensive feature value after weight calculation is 0.24, and it is classified into the pre-warning state layer according to the threshold separation standard. The separated feature vector is organized into a three-layer structure according to the state category, the first layer stores the normal state feature, the second layer stores the pre-warning state feature, and the third layer stores the abnormal state feature. A multi-layer state classification matrix M is constructed, the matrix row corresponds to the time sequence, the column corresponds to the different feature components, and the third dimension corresponds to the state level.

[0051] The interlayer response change of the multi-layer state classification matrix is monitored to form a state transition trajectory diagram. The generated multi-layer state classification matrix is continuously monitored in time to track the change of the feature quantity and distribution in each state layer. The change rate C of the feature vector quantity of each state layer at adjacent time points is calculated, C = (N_t+1-N_t) / N_t, where N_t and N_t+1 are the feature vector quantities of a certain state layer at t time and t+1 time, respectively. The transfer of the feature vector between different state layers is counted, and the number of features transferred from the normal state layer to the pre-warning state layer and the number of features transferred from the pre-warning state layer to the abnormal state layer are recorded. The time characteristics of the interlayer transfer are analyzed, and the residence time and transfer frequency of the feature vector in each state layer are calculated. The main path of state transition is identified, including the conversion modes of normal → pre-warning, pre-warning → abnormal, abnormal → pre-warning, etc. For example, during the start-up process of the aircraft food cart refrigeration system, the temperature feature vector is transferred from the normal state layer (number 45) to the pre-warning state layer (number increased to 60) within 2 minutes, and then part of it is transferred to the abnormal state layer (number 12) within 30 seconds, forming an obvious state transition path. The time coordinates and transfer quantities of each state transition are recorded, and the transition information is connected in time sequence to form a trajectory line. The state transition trajectories of all feature components are integrated and plotted in the same coordinate system, the horizontal axis represents time, the vertical axis represents state level, and the trajectory line reflects the state evolution process.

[0052] The trajectory analysis of the state transition trajectory graph generates the alarm triggering parameters. Numerical analysis is performed on the formed state transition trajectory graph, and the key features of the trajectory are extracted for the calculation of the alarm parameters. The slope change of the trajectory is calculated, and a positive slope value indicates state deterioration, a negative slope value indicates state improvement, and the absolute value of the slope reflects the change speed. The mutation points in the trajectory are identified, and the mutation points correspond to the time of rapid state conversion, which needs to be focused on. The residence time T_abnormal of the trajectory in the abnormal state layer is calculated, and an alarm is triggered when the residence time exceeds the preset threshold. The periodic characteristics of the trajectory are analyzed, and the repetition mode and cycle length of the state transition are identified. The cumulative degree of state deterioration S = Σ(ΔL x Δt) is calculated, where ΔL is the state layer change amount and Δt is the time interval. The threshold parameters for alarm triggering are set, including the deterioration speed threshold V_threshold = 0.5 layers / min, the abnormal residence time threshold T_threshold = 300 seconds, and the cumulative deterioration threshold S_threshold = 10. For example, when the temperature control system of the aviation food cart fails, the state transition trajectory shows that the temperature feature rapidly deteriorates from the normal state to the abnormal state within 5 minutes, and the deterioration speed reaches 0.6 layers / min, which exceeds the threshold 0.5. The system generates high-priority alarm parameters. Different levels of alarm triggering parameters are generated according to the trajectory analysis results, and the first-level alarm parameters correspond to slight abnormalities, the second-level alarm parameters correspond to medium abnormalities, and the third-level alarm parameters correspond to serious abnormalities.

[0053] In step S150, the state switching mode of the state transition trajectory graph is tracked, the state transition trajectory graph is executed, the state switching frequency is generated, the state switching frequency is associated with the alarm triggering parameters, the alarm intensity coefficient is generated by analyzing the alarm intensity coefficient, and the alarm intensity coefficient and the multi-layer state classification matrix are matched to generate a multi-level alarm output signal.

[0054] Specifically, the system tracks state transition patterns on a state transition trajectory map. It identifies transition events in the trajectory map, recording each change in state level as a transition event. The frequency of various transition patterns is statistically analyzed, including transitions from normal to warning state, from warning to abnormal state, and from abnormal to normal state. The duration of each transition pattern is calculated, representing the time interval from the start of the transition to stabilization in the target state. The temporal distribution characteristics of the transition patterns are analyzed, identifying periods of concentrated and sparse transition events. Feature parameters of the transition patterns are extracted, including transition speed, transition amplitude, and transition direction. For example, in the startup phase of an aviation catering truck's refrigeration system, the frequency of temperature transitions from the normal to the warning level is 3 times per minute, with an average transition duration of 15 seconds, while the frequency of humidity transitions is 1.5 times per minute, forming different transition pattern characteristics. Stability indices for the transition patterns are calculated to assess the reproducibility and predictability of various transition behaviors. The tracked transition pattern information is compiled into a pattern feature library, recording the temporal characteristics, frequency characteristics, and transition characteristics of each pattern.

[0055] In some embodiments, the step of performing frequency inversion on the state transition trajectory map to generate state switching frequencies includes: obtaining the starting state and the ending state of the state transition trajectory map; extracting an ideal frequency pattern from the ending state to generate target frequency data; performing reverse tracing processing on the starting state through the target frequency data to form a reverse frequency path; and performing inversion transformation on the reverse frequency path in reverse time order to generate state switching frequencies.

[0056] Obtain the initial and final states of the state transition trajectory diagram. Identify the first and last two key state points of the time series from the state transition trajectory diagram, which will serve as boundary conditions for subsequent inversion processing. Extract the initial state information at the starting position of the trajectory diagram's time axis and record the feature vector distribution of each state layer at the initial moment. Count the number of feature vectors for each state layer in the initial state: N0 for the normal state layer, W0 for the warning state layer, and A0 for the abnormal state layer. Extract the final state information at the ending position of the trajectory diagram's time axis and record the feature vector distribution of each state layer at the ending moment. Count the number of feature vectors for each state layer in the final state: Nf for the normal state layer, Wf for the warning state layer, and Af for the abnormal state layer. Calculate the total change from the initial state to the final state: ΔN = Nf - N0, ΔW = Wf - W0, ΔA = Af - A0. Analyze the differences between the initial and final states to identify the main direction and trend of state transition. Record the time coordinates of the initial and final states to determine the time range and boundary constraints for inversion processing.

[0057] The ideal frequency mode is extracted reversely from the end state to generate target frequency data. The end state of the trajectory diagram is taken as the starting point for analysis, and the ideal switching frequency mode reaching this state is analyzed in reverse. Assuming that the end state is the desired target state of the system, the optimal switching path and frequency distribution reaching this state are analyzed. The stability index of each state layer characteristic vector of the end state is calculated, and the state layer with high stability corresponds to a lower ideal switching frequency. The frequency characteristic parameters of the end state are extracted, including the switching frequency ratio and switching time interval between each state layer. According to the distribution characteristics of the end state, the ideal frequency mode is set, the normal state layer maintains a frequency of 0.1 times per minute, the pre-warning state layer target frequency is 0.5 times per minute, and the abnormal state layer target frequency is 0.2 times per minute. The ideal frequency mode is converted into numerical target frequency data to form a data vector containing the target switching frequency of each state layer. The target frequency data is normalized to ensure that the frequency values are distributed within a reasonable range.

[0058] For example, the reverse frequency path formed by performing reverse tracing processing on the starting state based on the target frequency data includes: constructing a tracing reference coordinate system using the target frequency data to generate a reference positioning index; using the reference positioning index to implement a path search operation on the starting state to form a candidate tracing path set; performing path optimization transformation on the candidate tracing path set to generate an optimal tracing path; and using the optimal tracing path to perform reverse calibration to generate a reverse frequency path.

[0059] A tracing reference coordinate system is constructed using the target frequency data to generate a reference positioning index. Each component of the target frequency data is used as a coordinate axis to establish a multi-dimensional tracing reference coordinate system. A three-dimensional reference coordinate system is constructed with the normal state layer target frequency as the X-axis, the pre-warning state layer target frequency as the Y-axis, and the abnormal state layer target frequency as the Z-axis. The coordinate position of the target frequency data is calibrated in the reference coordinate system, which serves as the target point for reverse tracing. The projection values of the target point on each coordinate axis are calculated, which constitute the components of the reference positioning index. A reference positioning index vector I = [Ix, Iy, Iz] is established, where Ix, Iy, and Iz are the projection values of the X, Y, and Z axes, respectively. The length and direction angle of the reference positioning index are calculated, with the length reflecting the overall strength of the target frequency and the direction angle reflecting the distribution bias of the frequency. The origin of the reference coordinate system is set as the zero frequency state, and the scale unit of the coordinate axis is the basic unit of frequency change. For example, when the target frequency data of the aviation food cart system is 0.1 times per minute for the normal state, 0.5 times per minute for the pre-warning state, and 0.2 times per minute for the abnormal state, the reference positioning index is I = [0.1, 0.5, 0.2], the index length is 0.55, and the main direction is biased towards the pre-warning state axis.

[0060] A set of candidate backtracking paths is formed by performing a path search operation from the start state to the end state using the reference positioning indicator as a search guide. The search step size is set to one-tenth of the time interval to ensure the accuracy of the search process. At each search step, the search path is adjusted according to the direction of the reference positioning indicator. The deviation of the current search position from the reference positioning indicator is calculated, and the direction with the smallest deviation is preferred as the next search direction. A breadth-first search strategy is used to explore multiple possible backtracking paths simultaneously. Constraints for the path search are set, including a reasonable range of frequency variation and a requirement for time continuity. If the frequency deviation of the search path exceeds 20% of the reference indicator, further search of the path is terminated. All search paths that meet the constraints are recorded, each path containing a time series and corresponding frequency values. The search paths are numbered and identified to form a set of candidate backtracking paths. The statistical characteristics of the candidate path set are calculated, including the number of paths, the average length, and the frequency distribution characteristics.

[0061] The optimal backtracking path is generated by performing a path optimization transformation on the set of candidate backtracking paths. All paths in the set of candidate backtracking paths are comprehensively evaluated to select the path that best meets the requirements of the reference positioning indicator. Evaluation indicators for path optimization are established, including the fitting degree of the path to the reference indicator, the smoothness of the path, and the stability of the path. The fitting degree score S1 of each candidate path is calculated where P is the path frequency vector, I is the reference positioning indicator, and ||·|| is the Euclidean distance. The path smoothness score S2 is calculated by evaluating the smoothness of the path through the second derivative of the path frequency. The path stability score S3 is calculated by evaluating the stability of the path through the variance of the path frequency. A comprehensive evaluation function S = w1 × S1 + w2 × S2 + w3 × S3 is established, where w1, w2, and w3 are weight coefficients. The comprehensive evaluation score of all candidate paths is calculated, and the path with the highest score is selected as the optimal backtracking path. The optimal backtracking path is refined through local optimization methods to further improve the quality of the path. The rationality of the optimal backtracking path is verified to ensure that the path meets the physical and logical constraints.

[0062] The optimal backtracking path is used to generate the reverse frequency path. The selected optimal backtracking path is used as the benchmark for reverse calibration to generate a complete reverse frequency path. The frequency of each time point on the optimal backtracking path is calibrated to ensure that the calibration value is consistent with the benchmark positioning index. The correction coefficient of path calibration is calculated to adjust the path frequency to make it closer to the ideal target. The interpolation method is used to interpolate the missing time points in the path to ensure the integrity of the reverse frequency path. The calibrated path is checked for continuity to ensure that the frequency change between adjacent time points remains reasonable. A mathematical expression of the reverse frequency path is established to describe the frequency characteristics of the path in different time periods using a piecewise function. The control point coordinates of the reverse frequency path are recorded, including key time points and corresponding frequency values. The quality of the reverse frequency path is verified by comparing it with the original trajectory to evaluate the accuracy of the path. The final generated reverse frequency path is stored as structured data, including time series, frequency values, and path parameters.

[0063] The reverse frequency path is rearranged in reverse order along the time axis to implement the inversion transformation to generate the state switching frequency. The obtained reverse frequency path is rearranged in reverse order according to the time sequence to restore the normal time flow. The time axis flipping method is used to reorder the time coordinates of the reverse frequency path from large to small to small. The frequency value of each time point remains unchanged, and only the arrangement order of the time coordinates is adjusted. For example, the original time sequence of the reverse frequency path is 90 seconds→75 seconds→60 seconds→45 seconds→30 seconds, and the corresponding frequency values are [2.1, 1.8, 1.5, 1.2, 0.8]. After inversion transformation, the time sequence is rearranged to 30 seconds→45 seconds→60 seconds→75 seconds→90 seconds, but the order of the frequency values remains unchanged, still [2.1, 1.8, 1.5, 1.2, 0.8]. This converts the reverse tracking result into a state switching frequency with a normal time flow, facilitating subsequent time series analysis and control applications. The frequency change rate of each time period after inversion transformation is calculated to identify the time interval with rapid frequency change. The inversion transformation highlights the key frequency characteristics in the trajectory evolution process and strengthens the switching information in important time periods. The inverted frequency sequence is smoothed to eliminate discontinuities that may occur during the data rearrangement process. The final inverted frequency sequence is defined as the state switching frequency, reflecting the internal frequency law of system state evolution. The corresponding relationship between the state switching frequency and time is recorded to form a complete frequency-time mapping table.

[0064] In some embodiments, the correlating the state switching frequency with the alarm triggering parameter to generate an alarm intensity coefficient comprises: converting the state switching frequency into an intensity evaluation vector field; identifying an optimal correlation path in the intensity evaluation vector field; performing intensity measurement based on the optimal correlation path and the alarm triggering parameter to form intensity interval data; and performing coefficient standardization evaluation on the intensity interval data to generate an alarm intensity coefficient.

[0065] The state switching frequency is converted into an intensity evaluation vector field. The time series of the state switching frequency is taken as the time dimension of the vector field, and the frequency value is taken as the amplitude component of the vector. The gradient of the frequency change is calculated where F is the state switching frequency and t is the time. The frequency gradient is taken as the direction component of the vector field, and a positive gradient value indicates the direction of frequency increase, and a negative gradient value indicates the direction of frequency decrease. A two-dimensional vector field is established, with the horizontal axis representing time and the vertical axis representing frequency amplitude. The vector arrow represents the direction and intensity of the frequency change. The divergence and curl of the vector field are calculated, with the divergence reflecting the divergence of the frequency change and the curl reflecting the rotation characteristics of the frequency change. The source and sink points of the frequency are identified in the vector field, with the source point corresponding to the position of rapid frequency increase and the sink point corresponding to the position of rapid frequency decrease. Color coding is used to represent the intensity distribution of the vector field, with warm colors representing high intensity areas and cold colors representing low intensity areas. The vector field is smoothed to eliminate the influence of data noise on the field distribution.

[0066] The optimal correlation path is identified in the intensity evaluation vector field. The optimal path connecting different intensity regions is searched in the constructed intensity evaluation vector field. The steepest descent path in the vector field is found using the gradient descent method, with the path extending along the negative direction of the intensity gradient. The integral intensity of the path is calculated where F(s) is the frequency function on the path and s is the path parameter. The integral intensities of multiple candidate paths are compared, and the path with the maximum integral intensity is selected as the optimal correlation path. The geometric characteristics of the optimal correlation path are analyzed, including path length, curvature, and turning angle. The tangent direction of each point on the path is calculated to determine the main trend direction of the frequency change. Key nodes on the path are identified, including intensity extreme points, curvature extreme points, and inflection point positions. For example, in an aviation food cart monitoring system, the intensity evaluation vector field shows that the integral intensity of the temperature frequency on the time axis is 12.5, and the integral intensity of the humidity frequency is 8.3. The temperature frequency path is selected as the optimal correlation path. The optimal correlation path is parameterized to describe the geometric shape of the path using a parameter equation.

[0067] The intensity interval data is formed by intensity calculation based on the optimal correlation path and the alarm trigger parameters. At each node position on the optimal correlation path, the corresponding state switching frequency value is extracted. The frequency value on the path is matched with the alarm trigger parameter corresponding to the time, and the product of the two is calculated as the local intensity value. The piecewise integration method is used to calculate the intensity integral of each segment of the path, and the integral result of each segment constitutes a component of the intensity interval data. The weight distribution of intensity calculation is calculated, and the weight coefficient of each segment is determined according to the importance of the alarm trigger parameter. The intensity calculation formula S = Σ(Fi × Pi × wi) is established, where S is the calculated intensity, Fi is the path frequency, Pi is the alarm trigger parameter, and wi is the weight coefficient. The intensity calculation result is stored in segments, and each time interval corresponds to an intensity value. The difference between adjacent intensity intervals is calculated to identify the time position of the intensity mutation. The statistical distribution of the intensity interval data is analyzed, and the mean, variance and distribution range are calculated. The intensity interval data is sorted according to the value, and the high intensity interval and the low intensity interval are identified.

[0068] The alarm intensity coefficient is generated by coefficient standardization evaluation of the intensity interval data. The maximum and minimum values of the intensity interval data are calculated to determine the dynamic range of the data. The maximum-minimum normalization method is used to map the intensity data to the 0-1 interval, and the normalization formula is K = (S-Smin) / (Smax-Smin), where K is the alarm intensity coefficient, S is the original intensity value, Smin and Smax are the minimum and maximum values respectively. The normalized coefficients are processed in stages, 0-0.3 is the low intensity coefficient, 0.3-0.7 is the medium intensity coefficient, and 0.7-1.0 is the high intensity coefficient. The distribution proportion of each intensity level is calculated to count the proportion of low, medium and high intensity coefficients. The alarm intensity coefficient is time-smoothed to eliminate random fluctuations of the coefficient using the moving average method. The confidence interval of the alarm intensity coefficient is established to estimate the uncertainty range of the coefficient by statistical method. The consistency of the alarm intensity coefficient is verified to ensure that the coefficient and the original intensity interval data maintain a monotonic corresponding relationship. The final alarm intensity coefficient is stored in association with the corresponding time label to form a complete intensity coefficient time series.

[0069] The alarm intensity coefficient and the multi-layer state classification matrix are matched to generate a multi-level alarm output signal. The alarm intensity coefficient is matched and analyzed with the number of characteristic vectors of each state layer in the multi-layer state classification matrix. When the number of characteristic vectors of an abnormal state layer exceeds 30% of the total number and the alarm intensity coefficient is greater than 0.7, the match is a first-level alarm level. When the number of characteristic vectors of a pre-alarm state layer exceeds 50% and the alarm intensity coefficient is between 0.3 and 0.7, the match is a second-level alarm level. When the number of characteristic vectors of a normal state layer is less than 70% and the alarm intensity coefficient is less than 0.3, the match is a third-level alarm level. The output signal intensity of each alarm level is calculated, the first-level alarm output signal intensity is 1.0, the second-level alarm output signal intensity is 0.6, and the third-level alarm output signal intensity is 0.3. The encoding format of the multi-level alarm output signal is generated, including alarm level identification, signal intensity value, trigger time and duration. The multi-level alarm output signal is prioritized, the first-level alarm has the highest priority, and the priority decreases in turn, and the intelligent monitoring and early warning of the aviation food car is finally completed.

[0070] In order to perform the multi-sensor fusion intelligent monitoring and early warning method of the aviation food car corresponding to the above-mentioned method embodiment, the corresponding functions and technical effects are realized. Referring to Figure 2 , Figure 2 The structure block diagram of the multi-sensor fusion intelligent monitoring and early warning device 200 of the aviation food car provided by the embodiment of the application is shown. For ease of illustration, only the part related to the embodiment is shown, and the multi-sensor fusion intelligent monitoring and early warning device 200 of the aviation food car provided by the embodiment of the application comprises: The data acquisition module 201 is configured to acquire the original monitoring data of the aviation food car, detect the data inconsistency degree among the sensors based on the sensor conflict, and generate a conflict enhancement coefficient based on the data inconsistency degree; The time difference processing module 202 is configured to acquire the cross-sensor time difference feature of the original monitoring data, construct a delay utilization matrix based on the cross-sensor time difference feature, reconstruct a probability chain based on the original monitoring data by using the delay utilization matrix to generate an inverted reference vector, and generate an enhanced feature set based on the reconstructed data feature space based on the inverted reference vector; The adversarial analysis module 203 is configured to perform conflict enhancement processing on the enhanced feature set based on the conflict enhancement coefficient to generate an adversarial feature vector, analyze the adversarial distribution characteristics of the adversarial feature vector in the feature space to determine a self-adversarial boundary, establish a dynamic balance discrimination mechanism based on the self-adversarial boundary to generate a state recognition parameter; The state classification module 204 is configured to perform state separation processing on the confrontation feature vector under the constraint of the state identification parameter to generate a multi-layer state classification matrix, monitor inter-layer response changes of the multi-layer state classification matrix to form a state transition trajectory diagram, and perform trajectory analysis on the state transition trajectory diagram to generate an alarm triggering parameter. The early warning output module 205 is configured to track a state switching mode of the state transition trajectory diagram, perform frequency inversion on the state transition trajectory diagram to generate a state switching frequency, perform correlation analysis on the state switching frequency and the alarm triggering parameter to generate an alarm intensity coefficient, and perform alarm level matching on the alarm intensity coefficient and the multi-layer state classification matrix to generate a multi-level alarm output signal.

[0071] The multi-sensor fusion aviation food cart intelligent monitoring and early warning device 200 described above can implement the multi-sensor fusion aviation food cart intelligent monitoring and early warning method of the method embodiment described above. The optional items in the method embodiment described above are also applicable to the present embodiment, and will not be described in detail here. The remaining content of the present embodiment can be referred to the content of the method embodiment described above, and will not be described in detail in the present embodiment.

[0072] The purpose of the above embodiments is to exemplarily reproduce and deduce the technical solutions of the present application, and to completely describe the technical solutions, purposes and effects of the present application. The purpose is to make the public understand the disclosure of the present application more thoroughly and comprehensively, and does not limit the protection scope of the present application.

[0073] The above embodiments are not based on an exhaustive enumeration of the present application, and there can be many other unlisted embodiments. Any substitution and improvement made without violating the concept of the present application is within the protection scope of the present application.

Claims

1. A multi-sensor fusion intelligent monitoring and early warning method for an aerial food cart, characterized in that, The method comprises the following steps: Collecting original monitoring data of an aerial food cart, performing sensor conflict detection on the original monitoring data to identify the degree of inconsistency between data of each sensor, generating a conflict enhancement coefficient based on the degree of inconsistency; Obtaining a cross-sensor time difference feature of the original monitoring data, constructing a delay utilization matrix through the cross-sensor time difference feature, performing probability chain reconstruction on the original monitoring data through the delay utilization matrix to generate an inverted reference vector, reconstructing a data feature space based on the inverted reference vector to generate an enhanced feature set; Performing conflict strengthening processing on the enhanced feature set through the conflict enhancement coefficient to generate an adversarial feature vector, analyzing the adversarial distribution characteristics of the adversarial feature vector in the feature space to determine a self-adversarial boundary, establishing a dynamic balance discrimination mechanism based on the self-adversarial boundary to generate a state recognition parameter; Implementing state separation processing on the adversarial feature vector under the constraint of the state recognition parameter to generate a multi-layer state classification matrix, monitoring the inter-layer response change of the multi-layer state classification matrix to form a state transition trajectory graph, performing trajectory analysis on the state transition trajectory graph to generate an alarm triggering parameter; Tracking the state switching mode of the state transition trajectory graph, performing frequency inversion on the state transition trajectory graph to generate a state switching frequency, associating and analyzing the state switching frequency and the alarm triggering parameter to generate an alarm intensity coefficient, performing alarm level matching on the alarm intensity coefficient and the multi-layer state classification matrix to generate a multi-level alarm output signal.

2. The method of claim 1, wherein, The method comprises the following steps: Scanning the degree of inconsistency to identify periodic conflict resonance points; Extracting resonance frequency features based on the periodic conflict resonance points to generate system feature signatures; Performing resonance matching on the system feature signatures and the degree of inconsistency to generate a resonance strengthening factor; Performing coefficient transformation using the resonance strengthening factor to generate a conflict enhancement coefficient.

3. The method of claim 1, wherein, The method comprises the following steps: Building a time difference analysis grid according to the cross-sensor time difference feature; Identifying inverted feature nodes from the time difference analysis grid to generate inverted markers; Performing pattern matching transformation on the time difference analysis grid using the inverted markers to generate a time difference layout; Analyzing the time difference layout to form a delay utilization matrix.

4. The method of claim 1, wherein, The method comprises the following steps: Performing probability node extraction on the delay utilization matrix to form a node probability graph; Detecting link paths from the node probability graph to generate path weight factors; Performing chain reconstruction on the original monitoring data according to the path weight factors to generate an inverted reference vector.

5. The method of claim 1, wherein, The method comprises the following steps: Spatially layering the adversarial feature vector according to adversarial intensity to identify imbalance regions; Extracting imbalance degrees from the imbalance regions to generate early warning level factors; Performing boundary search transformation on the imbalance regions based on the early warning level factors to generate a dynamic safety line; The dynamic security line is fused with the early warning level factor to determine a self-antagonistic boundary.

6. The method of claim 1, wherein, The state transition trajectory graph is subjected to frequency inversion to generate a state switching frequency. A starting state and an end state of the state transition trajectory graph are obtained. An ideal frequency mode is extracted from the end state in reverse to generate target frequency data. Reverse tracing is performed on the starting state based on the target frequency data to form a reverse frequency path. The reverse frequency path is rearranged in reverse order along a time axis to implement inversion transformation to generate a state switching frequency.

7. The method of claim 1, wherein, The state switching frequency is associated with the alarm trigger parameter to generate an alarm intensity coefficient. The state switching frequency is converted into an intensity evaluation vector field. An optimal association path is identified in the intensity evaluation vector field. Intensity measurement is performed on the optimal association path and the alarm trigger parameter to form intensity interval data. The intensity interval data is subjected to coefficient standardization evaluation to generate an alarm intensity coefficient.

8. The method of claim 4, wherein, The path weight factor is analyzed to generate a reverse reference vector, including: Weight jump characteristics in the path weight factor are analyzed. Available weight imbalance intervals are locked in the weight jump characteristics. The reconstruction strength of the weight imbalance intervals is evaluated to generate a reconstruction coefficient. The original monitoring data is subjected to chain reconstruction according to the weight imbalance intervals and the reconstruction coefficient to generate a reverse reference vector.

9. The method of claim 6, wherein, The reverse tracing is performed on the starting state based on the target frequency data to form a reverse frequency path, including: A tracing reference coordinate system is constructed based on the target frequency data to generate a reference positioning index. A path search operation is performed on the starting state based on the reference positioning index to form a candidate tracing path set. An optimal tracing path is generated by performing path optimization transformation on the candidate tracing path set. The reverse frequency path is generated by reverse calibration based on the optimal tracing path.

10. The multi-sensor fusion based intelligent monitoring and early warning device for aerial food trolley, characterized in that, It includes: A data acquisition module is configured to acquire original monitoring data of an aviation food cart, detect data inconsistency between sensors based on sensor conflict detection, and generate a conflict enhancement coefficient based on the data inconsistency. A time difference processing module is configured to obtain a cross-sensor time difference characteristic of the original monitoring data, construct a delay utilization matrix based on the cross-sensor time difference characteristic, and generate a reverse reference vector by performing probability chain reconstruction on the original monitoring data based on the delay utilization matrix. An antagonistic analysis module is configured to perform conflict enhancement processing on the enhanced feature set based on the conflict enhancement coefficient to generate an antagonistic feature vector, determine a self-antagonistic boundary based on the antagonistic distribution characteristics of the antagonistic feature vector in the feature space, and generate a state recognition parameter based on the dynamic balance discrimination mechanism established based on the self-antagonistic boundary. The state classification module is used for implementing state separation processing on the confrontation feature vector under the constraint of the state identification parameter to generate a multi-layer state classification matrix, monitoring interlayer response changes of the multi-layer state classification matrix to form a state conversion trajectory graph, and performing trajectory analysis on the state conversion trajectory graph to generate an alarm triggering parameter. The early warning output module is used for tracking a state switching mode of the state conversion trajectory graph, performing frequency inversion on the state conversion trajectory graph to generate a state switching frequency, correlatively analyzing the state switching frequency and the alarm triggering parameter to generate an alarm intensity coefficient, and performing alarm level matching on the alarm intensity coefficient and the multi-layer state classification matrix to generate a multi-level alarm output signal.

Images