An AI-powered, all-domain situational awareness digital cockpit management system

The AI-powered full-domain situational awareness digital cockpit management system solves the problems of data loss and sensor malfunctions in situation assessment schemes, enabling accurate location of anomalies and true reflection of situational indicators, thereby improving the system's reliability and decision-making accuracy.

CN123115973APending Publication Date: 2026-11-17SHANGHAI YUBANG INTELLIGENT TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611574195.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing situation assessment solutions struggle to distinguish between missing data and changes in the situation. When channels go offline or sensors malfunction, metrics also fluctuate, leading to frequent alarm triggers. Managers are unable to determine the source of the anomalies, thus weakening the practical role of situation awareness.

Method used

The AI-powered full-domain situational awareness digital cockpit management system employs a data acquisition and labeling module to assign timestamps to data and perform clock drift compensation, an access buffer module to construct dynamic time boundaries, a beat alignment module to perform data aggregation and transformation, an estimation and completion module to perform extrapolation prediction and state updates, a quality gating module to calculate quality scores, and a fusion decision module to generate a situational safety index and a traceability link.

Benefits of technology

It effectively distinguishes the sources of anomalies, reduces indicator fluctuations, lowers the false alarm and false negative rates, ensures that situation indicators reflect the true changes in the situation, supports retrospective verification of anomaly location, and improves the reliability of the system and the accuracy of decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115973A_ABST
    Figure CN123115973A_ABST
Patent Text Reader

Abstract

This invention relates to the field of multi-source information fusion processing technology, and more particularly to an AI-powered global situational awareness digital cockpit management system, comprising: a data acquisition and labeling module, which acquires sensing data from multiple sources, assigns timestamps including the time of generation, performs clock drift compensation, and outputs observation records; an access buffer module, which receives the observation records, constructs dynamic time boundaries based on the channel lateness thresholds of each data channel, performs time window aggregation and lateness recalculation on out-of-order observation records, and outputs a time window observation set; and a beat alignment module, which reads the time window observation set based on the system beat and calls corresponding operators for aggregation and transformation according to the data type. By judging the reliability of data in each dimension and stopping output when there is insufficient effective information, the system significantly reduces indicator jitter caused by channel disconnection or sensor failure, lowers the false alarm and false negative rates, and enables managers to distinguish whether anomalies originate from the field or from the link.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of multi-source information fusion processing technology, and in particular to an AI-powered global situational awareness digital cockpit management system. Background Technology

[0002] A comprehensive situational awareness digital cockpit needs to aggregate data from IoT sensors, video inference results, and business systems into unified situational indicators, enabling managers to assess the current situation and allocate resources accordingly. This data comes from heterogeneous devices with sampling periods ranging from milliseconds to minutes, reported via wireless or wired links. Transmission time fluctuates depending on network conditions, and some nodes may miss reporting at certain times due to dormancy or disconnection. The system needs to continuously output situational indicators for decision-making at a fixed pace, even under conditions of incomplete data and uncertain arrival times.

[0003] Existing situation assessment solutions struggle to distinguish between missing data and changes in the situation. When channels go offline or sensors malfunction, metrics also fluctuate, making it difficult for managers to determine whether anomalies originate from the field or from the link. Metrics repeatedly fluctuate around thresholds, leading to frequent alarm triggers. Over time, maintenance personnel tend to ignore system alerts, thus weakening the actual effectiveness of situation awareness. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an AI-powered global situational awareness digital cockpit management system, which aims to improve the existing situational assessment schemes that struggle to distinguish between missing data and changes in the situation, and where indicators fluctuate even when channels are offline or sensors malfunction.

[0005] This invention provides the following technical solution: an AI-powered, all-domain situational awareness digital cockpit management system, comprising: The data acquisition and labeling module obtains sensing data from multiple source devices, assigns a timestamp including the time of generation, performs clock drift compensation, and outputs observation records. The access buffer module receives the observation records, constructs a dynamic time boundary based on the channel lateness threshold of each data channel, performs time window aggregation and lateness recalculation on the disordered observation records, and outputs a time window observation set. The beat alignment module reads the time window observation set based on the system beat, calls the corresponding operator to perform aggregation and transformation according to the data type, and outputs a feature vector; The estimation and completion module inputs the feature vector into the continuous filtering model, performs extrapolation prediction and state update on the missing dimension of the feature vector, and outputs the completed feature vector, posterior covariance and time delay. The quality gating module receives the completed feature vector, calculates the individual quality score and weighted quality score by combining the time delay and the posterior covariance, intercepts the time window observation set whose weighted quality score is lower than the grading threshold according to the grading threshold, and outputs the verification feature vector and the corresponding individual quality score. The fusion decision module performs normalization fusion on the verification feature vector to generate a situational security index and a traceability link, and drives the display terminal to render and issue warnings.

[0006] Preferably, in the acquisition and marking module, the execution of clock drift compensation specifically includes: Obtain sample pairs of network heartbeat arrival times and local timestamps of the multi-source devices; The bias coefficient and drift slope are calculated using the recursive least squares method with a forgetting factor. Outlier pairs are removed using a random sampling consensus algorithm. Based on the bias coefficient and the drift slope, the generation time in the observation record is mapped to the timestamp.

[0007] Preferably, in the access buffer module, the construction of dynamic time boundaries based on the channel lateness threshold of each data channel specifically includes: The distribution of transmission delay from generation to reception of the observation records in each data channel is statistically analyzed. Extract the quantiles of the transmission delay distribution as the channel lateness threshold for the corresponding data channel; The dynamic time boundary is generated by subtracting the channel lateness threshold from the maximum timestamp of the observation record received in the current data channel. The time window aggregation is triggered when the right boundary of the current time window observation set is less than or equal to the dynamic time boundary of all the data channels.

[0008] Preferably, in the beat alignment module, the step of calling the corresponding operator for aggregation and transformation based on the data type specifically includes: When the data type is a continuous quantity, the robust aggregation operator based on the truncated mean is invoked; When the data type is a cumulative quantity, the time step-based differential scaling operator is invoked. When the data type is a discrete state quantity, the zero-order hold operator is invoked to extend the last observed state within the time window observation set to the boundary of the system beat.

[0009] Preferably, in the estimation and completion module, the extrapolation prediction and state update of the missing dimensions of the feature vector specifically includes: For the missing dimension for which the observation record was not received, discretization prediction is performed using a continuous-time state transition matrix, and the process noise covariance is monotonically increased with the time delay. For the feature dimensions of the received observation records, the residual variance of the corresponding data channel within a preset time period is extracted as the observation noise covariance for state correction; The trace of the covariance matrix of the corrected continuous filtering model is calculated and used as the posterior covariance output.

[0010] Preferably, in the quality gating module, the calculation of the individual quality score and the weighted quality score by combining the time delay and the posterior covariance specifically includes: The time delay is mapped to a time freshness factor using a decay function; The posterior covariance is transformed into an estimated certainty factor that is inversely proportional to the numerical value; Multiply the time freshness factor by the estimated certainty factor to obtain the individual quality score of each dimension of the completed feature vector; The weighted quality score of the observation set within the time window is calculated based on preset dimensional weights; When the weighted quality score is lower than the grading threshold, the output of the time window observation set is blocked.

[0011] Preferably, obtaining the individual quality scores for each dimension of the completed feature vector specifically includes: The decay coefficient of the decay function is set according to the characteristic time constant of the physical quantity corresponding to each dimension, and the time freshness factor is generated. Perform link heartbeat verification, value range verification, card value verification, and jump rate verification on each of the aforementioned dimensions, and convert the verification results into source health factors; The individual quality score is obtained by multiplying the time freshness factor, the estimated certainty factor, and the source health factor. The proportion of the completed feature vectors for each dimension within a preset statistical period is statistically analyzed. When the percentage exceeds the upper limit threshold, the source health factor is set to zero.

[0012] Preferably, in the fusion decision module, the step of performing normalization fusion on the verification feature vector to generate the situational security index and the tracing link specifically includes: Convert the verification feature vector into a mapping value; The dynamic weight is obtained by multiplying the individual quality score of each dimension of the verification feature vector with the preset dimension weight; The sum of all the generated dynamic weights is used as the normalized denominator; The sum of the products of each dynamic weight and its corresponding mapping value is calculated and divided by the normalized denominator to generate the situational security index. The data source identifier, timestamp, and dynamic weights that the situational security index depends on are recorded to form a mapping dictionary, which is used to construct the tracing link.

[0013] Preferably, the generation of the mapping value and the early warning output specifically include: Call the monotonic utility function to normalize by replacing the extreme values ​​with the sliding window quantile baseline, and output the mapped value; When the normalized denominator is lower than a preset lower limit, the output of the situational safety index is stopped. An uplink threshold and a downlink threshold are set for the aforementioned situational security index; The warning is triggered when the situation security index is higher than the uplink threshold, and the number of time window observation sets in which the situation security index is continuously higher than the uplink threshold reaches a preset number. The warning is lifted when the situational security index is lower than the downlink threshold. When the delayed recalculation causes the traceability link to generate a new version, a correction warning carrying a version identifier is output.

[0014] The present invention has the following beneficial effects: 1. In this invention, by judging the credibility of data in each dimension and stopping the output when there is insufficient effective information, the fluctuation of indicators caused by channel disconnection or sensor failure is greatly reduced, the false alarm and false alarm ratios are reduced, and the managers can distinguish whether the anomaly comes from the field or from the link.

[0015] 2. In this invention, by unifying multi-source data with significant differences in sampling period and transmission time onto the same time scale for evaluation, the situation indicators reflect real situation changes rather than network fluctuations, and the phase misalignment between various dimensions is eliminated.

[0016] 3. In this invention, by retaining the data source and version information on which each assessment depends, the output situation indicators can be traced back and verified, historical conclusions can still be reproduced after the data is supplemented, and the time required for anomaly location is shortened. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of an AI-powered global situational awareness digital cockpit management system proposed in this invention. Figure 2 A flowchart illustrating the implementation of the estimation and completion module provided by this invention; Figure 3 The flowchart illustrates the implementation of the quality gating module provided by this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In a first embodiment of the present invention, the present invention provides an AI-powered global situational awareness digital cockpit management system, such as... Figures 1-3 As shown, it includes: The data acquisition and labeling module obtains sensing data from multiple source devices, assigns a timestamp including the time of generation, performs clock drift compensation, and outputs observation records. Furthermore, in the data acquisition and marking module, the clock drift compensation process specifically includes: Obtain sample pairs of network heartbeat arrival times and local timestamps from multiple source devices; The bias coefficient and drift slope are calculated using the recursive least squares method with a forgetting factor. Outlier pairs are removed using a random sampling consensus algorithm. Based on the bias coefficient and drift slope, the generation time in the observation record is mapped to a timestamp.

[0020] Specifically, in a global situational awareness scenario, the local crystal oscillators of various IoT sensors accumulate errors over time, causing the local occurrence time carried in the generated data packets to deviate from the actual physical occurrence time. To reduce the impact of this clock asynchrony, the acquisition and labeling module obtains sensing data from multiple sources, assigns timestamps including the occurrence time, performs clock drift compensation, and outputs observation records with a unified time reference.

[0021] During clock drift compensation, the system gateway periodically sends heartbeat probe messages to multi-source devices. Each device receives and responds with its local clock reading. The gateway records the arrival time of the system reference clock when the message returns. To avoid overflow in the calculation of the error covariance matrix due to directly using absolute timestamps, the system records the local initial time and the system initial time when the device first connects. Subsequent acquired times are all subtracted from the corresponding initial time to convert them into relative time offsets. Based on this, the relative local time and the relative system arrival time are combined to form a time sample pair sequence.

[0022] In complex wireless network environments, sudden network congestion can cause a surge in latency for some packets. The system utilizes a random sampling consensus algorithm to remove outlier pairs. The system maintains a capacity of [missing information]. A sliding window of sample pairs is used. At each validation, two sample pairs are randomly selected to calculate a temporary slope and bias model, and the absolute residuals of the remaining sample pairs within the window are calculated and substituted into this temporary model. Feasible implementation data for setting the residual tolerance threshold are as follows: In milliseconds, sample pairs with absolute residuals less than this threshold are marked as inliers. The system repeats this random sampling process, with feasible implementation data for the number of iterations being... Next, retain the consensus set with the largest number of in-place points, output the latest in-place sample pair in the set to the model update queue, and directly remove the remaining outliers.

[0023] The system applies recursive least squares with a forgetting factor to dynamically fit the selected intra-sample pairs to calculate the bias coefficient and drift slope. The state vector to be estimated is then constructed. superscript Denotes the matrix transpose, where The iteration number, Indicates the first The drift slope obtained from the next iteration Indicates the first The bias coefficients are calculated in the next iteration. Initial state is set. To avoid dimensional conflicts caused by directly incorporating time values ​​with physical dimensions into matrix operations, the system sets the value to... normalized time constant in milliseconds The initial error covariance matrix is ​​uniformly calibrated using dimensionless numerical values. ,in This indicates that the elements within the parentheses form a diagonal matrix. The dimensionless initial variance of the drift slope has the following feasible implementation data: , The dimensionless initial variance of the bias coefficients and its feasible implementation data are as follows: Definition of the first The dimensionless input vector of each local sample pair ,in This represents the relative local time for this intra-regional sample pair. The corresponding target observations are simultaneously converted into dimensionless values. ,in Let be the relative system arrival time of this intra-regional sample pair. The formula for calculating the gain vector is: ; The recursive update formula for the state vector is: ; The update formula for the error covariance matrix is: ; in the formula For the first The gain vector of the next iteration. For the first The error covariance matrix after the next iteration As the forgetting factor, its feasible implementation data value in the Industrial Internet of Things (IIoT) scenario is [value missing]. .

[0024] When multiple source devices report packets containing service-aware values, the system extracts the dimensionless drift slope of the device's latest iteration. With dimensionless bias coefficient This maps the local occurrence times in the observation records to a globally unified timestamp. The conversion formula is: ; in the formula The absolute moment of creation recorded by the device hardware. This is the local initial time when the device is first connected. This corresponds to the system initial time. This is the compensated system-level absolute timestamp. The data acquisition and tagging module encapsulates the device identification code, service value, and this timestamp in a structured manner and outputs them to the subsequent stages.

[0025] This step reduces the timestamp deviation introduced by crystal oscillator temperature drift and network jitter, providing a unified time reference for subsequent beat alignment.

[0026] The access buffer module receives observation records, constructs dynamic time boundaries based on the channel lateness threshold of each data channel, performs time window aggregation and lateness recalculation on out-of-order observation records, and outputs the time window observation set. Furthermore, in the access buffer module, the construction of dynamic time boundaries based on the channel lateness threshold of each data channel specifically includes: The distribution of transmission delay from generation to reception of observation records within each data channel is statistically analyzed; Extract the quantiles of the transmission delay distribution as the corresponding data channel lateness threshold; The dynamic time boundary is generated by subtracting the channel lateness threshold from the maximum timestamp of the observation records received in the current data channel. Time window aggregation is triggered when the right boundary of the current time window observation set is less than or equal to the dynamic time boundary of all data channels.

[0027] Specifically, the access buffer module receives observation records with a unified time reference output by the acquisition and labeling module. In a heterogeneous network, the transmission path times of different data channels, such as security video analysis streams and ambient temperature sensor streams, differ significantly, leading to out-of-order data arriving at the system. Regarding the first... Each data channel uses a globally unified system reference clock to obtain the actual arrival time of observation records at the gateway. Combined with the globally unified timestamp generated by the mapping in the previous module Calculate the link transmission time for this single observation record. .

[0028] To measure channel congestion characteristics, the system maintains a fixed-capacity first-in-first-out circular queue in memory for each data channel to store the link transmission time of recently arriving records. Feasible implementation data for the queue capacity are as follows: The system uses insertion sort to ensure that newly enqueued time values ​​fall directly into ordered positions, thus maintaining an ascending order within the queue. The index position is then extracted as the queue capacity and a globally set quantile. The element of the product is used as the channel lateness threshold for the corresponding data channel. This threshold satisfies the cumulative distribution function relationship. , where the symbol Quantiles represent the probability of an event occurring. In complex situational awareness scenarios, quantiles... The feasible implementation data value is Subsequently, the system records the current number in real time. The maximum timestamp of the received observation records within each data channel. and subtract the late threshold of that channel. , generate the first Dynamic time boundaries of each data channel The physical meaning of this boundary is that the system is confident that observations generated earlier than this time boundary have likely reached the gateway.

[0029] The system divides the timeline into consecutive evaluation time windows. Let the right boundary time scale of the currently pending processing time window be denoted as . The system continuously monitors the total volume across the entire domain. The dynamic time boundaries of each data channel participating in the fusion. The determination condition is met if and only if the following criteria are satisfied. At that time, among them This indicates that the minimum value among the parameters is taken. The system determines that the data in the entire domain is basically aligned on the time slice, and then triggers time window aggregation, which combines and packages all observation records whose time falls within the time window, and outputs them as the time window observation set.

[0030] To prevent system memory overflow, the system configures a maximum tolerance duration for time window cache maintenance, and its feasible implementation data is as follows: Seconds. If, due to extreme network disconnection, some observation records arrive late after the time window aggregation is triggered, the system calculates the difference between the timestamp contained in the delayed record and the smallest dynamic time boundary value in the current global data channel. If the difference does not exceed the set maximum tolerance time, the system inserts it into the corresponding historical cache window and triggers the late arrival recalculation logic, recombining and generating a new time window observation set with an incrementing version identifier to push to the next level; if the difference exceeds the maximum tolerance time, the system directly discards the record without triggering recalculation.

[0031] This step incorporates the out-of-order and delay phenomena caused by heterogeneous data during transmission, and uses tolerance time limits to avoid the overall decision pipeline being blocked due to waiting for a single slow channel node, thus achieving a balance between data timeliness and computational stability.

[0032] The beat alignment module reads the observation set of the time window based on the system beat, calls the corresponding operator to perform aggregation and transformation according to the data type, and outputs the feature vector. Furthermore, in the beat alignment module, the specific steps of calling the corresponding operator for aggregation transformation based on data type include: When the data type is continuous, invoke the robust aggregation operator based on the truncated mean; When the data type is a cumulative quantity, the time step-based differential scaling operator is invoked; When the data type is a discrete state quantity, the zero-order hold operator is called to extend the state of the last observation in the time window observation set to the boundary of the system tick.

[0033] Specifically, the beat alignment module receives the time window observation set pushed by the access buffer module. Although the data in this set has undergone out-of-order sorting and aggregation at the network transmission layer, the inherent sampling frequency differences among various service devices across the entire domain cause the observation records in the set to still exhibit an asynchronous distribution on the time axis. To meet the matrix alignment input requirements of the subsequent filtering model, the system sets a globally fixed system beat boundary, assuming the first... The time scale of the system's beat boundary is: ,in This serves as the step sequence index for the system's clock cycle. The system reads the data within this set according to the clock cycle and, based on the underlying device's physical data type, calls the corresponding mathematical operators for aggregation and transformation.

[0034] When the data type corresponding to the sensed data is a continuous quantity such as ambient temperature or equipment amplitude, the system invokes a robust aggregation operator based on truncated means. The system extracts all data from this continuous quantity channel within the current time window observation set. The observed values ​​are arranged in ascending order to form a sequence. To suppress spike noise caused by sensor short circuits, the system sets a truncated ratio parameter. Its feasible implementation data is The system calculates the number of single-end rejections. Its value is and Round the product down to the nearest integer. The system removes values ​​from both ends of the sequence. After finding the extreme values, calculate the arithmetic mean of the remaining data using the following formula: . In the formula, This represents the summation operation. The index variable is used for summation, and its value is determined by... Increment to , The first one after ascending order One observed value, Align the feature values ​​for this continuous output. When the total amount of data... Smaller When the value is zero, the system performs a full arithmetic mean calculation.

[0035] When the data corresponding to the sensed data is a cumulative quantity such as cumulative power consumption of equipment or cumulative output of production line, since this data has a monotonically increasing physical property, forcibly performing an averaging operation would destroy its cumulative semantics. Therefore, the system calls a differential scaling operator based on the time step. The system extracts the first observation value with the smallest timestamp within the current time window for this channel. and the time of its generation Simultaneously extract the last observation with the largest timestamp. and the time of its generation If the time window contains only a single record with the same first and last times, the system directly uses that value as the feature value; if it contains multiple records, the system calculates the differential rate of change within the time window and linearly extrapolates it to the current system beat boundary. The calculation formula is: . in the formula This is the alignment feature value of the output cumulative quantity channel at the system beat boundary.

[0036] When the data corresponding to the sensed data is a discrete state quantity such as valve on / off indication or warning level, intermediate interpolation of the values ​​has no physical meaning. The system calls the zero-order hold operator to directly search within the current time window observation set for the closest and earliest generation time to the current system beat boundary. The last observed state The system directly carries over this state value to the current cycle, calculated using the following formula: Output discrete state characteristic values This preserves the step physical properties of state transitions.

[0037] The beat alignment module concatenates the alignment feature values ​​generated from the various channels according to a pre-defined feature dimension order across the entire domain, outputting a feature vector with a synchronized time label. This step maps observation data with varying sampling rates onto a synchronized time grid based on their respective physical logic, reducing phase misalignment errors during multi-source fusion computation.

[0038] The estimation and completion module inputs the feature vector into the continuous filtering model, performs extrapolation prediction and state update on the missing dimensions of the feature vector, and outputs the completed feature vector, posterior covariance and time delay. Furthermore, the estimation and completion module specifically includes extrapolating and predicting the missing dimensions of the feature vector and updating the state, which includes: For the missing dimensions where no observation records were received, discretization prediction is performed using a continuous-time state transition matrix, and the process noise covariance monotonically increases with the time delay. For the characteristic dimensions of the received observation records, the residual variance of the corresponding data channel within a preset time period is extracted as the observation noise covariance for state correction; Calculate the trace of the covariance matrix of the corrected continuous filtering model and output it as the posterior covariance.

[0039] Specifically, the estimation completion module receives the feature vector output by the beat alignment module. In real-world IoT environments, due to some sensors being in low-frequency dormancy periods or experiencing temporary disconnections, this feature vector contains missing dimensions indicating unreceived observations. The system introduces a continuous filtering model to repair data gaps and maintain the spatial integrity of situational features. The system employs a full-dimensional joint filtering structure, meaning the state vector covers all preset dimensions across the entire domain. Each feature dimension performs joint prediction across all dimensions in each beat, and only performs state updates for dimensions with valid observations in the current beat after dimensionality reduction mapping. The system uses the system beat step sequence index from the beat alignment module. And set the interval between the boundaries of two adjacent system beats as the beat step size. Its feasible implementation data is Second.

[0040] The system constructs the evolutionary relationships of each dimension according to a continuous-time state-space description. Let the continuous-time system matrix be... ,in For the first Each feature dimension corresponds to the inertial time constant of a physical quantity, used to characterize the rate at which the physical quantity itself reverts to a steady state. Based on this, the continuous-time state transition relationship is discretized to obtain the state transition matrix. ,in This represents the exponential operation with the natural constant as the base. The matrix is... The diagonal elements are When the first When the physical quantities corresponding to each dimension remain approximately constant within adjacent observation intervals... When the value is much larger than the clock step size, the diagonal elements approach one, and the state transition matrix degenerates into an identity matrix. The state prediction formula is as follows: , in the formula For the first The posterior state vector of each beat For the first The prior prediction vector for each beat.

[0041] The system calculates the noise covariance matrix during the process. . in the formula Let be the diagonal matrix of the noise power spectral density of a continuous process, and its th... The physical dimension of the diagonal element is the th The square of each physical quantity in each dimension is divided by a second; therefore, each dimension must be calibrated separately, and a uniform value cannot be used. For the environmental temperature dimension, measured in degrees Celsius, the feasible implementation data for this diagonal element is: For the pipeline flow rate dimension measured in cubic meters per hour, the feasible implementation data for this diagonal element is as follows: The system utilizes formulas Calculate the first The prior covariance matrix of each beat , in the formula For the first The posterior covariance matrix of each beat, with superscript... This represents the matrix transpose. For a missing dimension where no observation records have been received for several consecutive cycles, the system only performs the above prediction in each cycle without performing a state update. The process noise covariance therefore accumulates cycle by cycle, causing the variance corresponding to that dimension to monotonically increase with the accumulated time without update. The system records the cumulative time difference between the last received valid observation record and the current system cycle for each dimension, which serves as the time delay for that dimension. Output.

[0042] Based on the feature dimensions of the received observation records, the system performs state correction. The system constructs a mapping observation matrix. This matrix is ​​a dimensionality-reduced matrix generated by removing rows corresponding to missing dimensions from the identity matrix. The system calibrates the observation noise covariance matrix by dimension. Its diagonal elements are taken from the residual variance of the corresponding data channel within a preset time period, where the feasible implementation data for this preset time period is the past. Hours. Since the residual variance statistically includes both prediction uncertainty and sensor observation noise, the system subtracts the variance of the difference between the actual observed values ​​and the prior predicted values ​​within this time period. The difference between the corresponding diagonal elements is used as the observation noise variance for that dimension, and a positive lower bound is set for the result. The feasible implementation data for this lower bound is the square of the range of that dimension. The system calculates the Kalman gain matrix: ; superscript in formula This represents the matrix inversion operation. The system extracts the effective observation vector using the current beat. The state correction is performed using the following formula: , in the formula The feature vector output after integrating observation data and prediction models is used to fill in the missing values.

[0043] After completing the state correction, the system calculates the covariance matrix of the corrected continuous filtering model, and the update formula is as follows: , in the formula The identity matrix is ​​of the same order as the prior covariance matrix. This is the corrected multidimensional joint posterior covariance matrix. Since the physical dimensions of the elements on different main diagonals of this matrix are different, directly taking the trace of the entire matrix would lead to the computational fallacy of adding heterogeneous dimensions. Therefore, the system extracts the traces of each element on the main diagonal of the matrix, i.e., it takes the trace of the first-order submatrix corresponding to each feature dimension, denoted as […]. No. Line number The elements of the column are The variance vector is formed by combining the numerical values ​​of each element and used as the posterior covariance output.

[0044] This step fills in the random gaps in the multi-source data on the beat grid and quantifies the inference uncertainty caused by missing data or the accumulation of extrapolation time dimension by dimension.

[0045] The quality gating module receives the supplemented feature vector, calculates the individual quality score and weighted quality score by combining the time delay and posterior covariance, intercepts the observation set with the weighted quality score below the grading threshold according to the grading threshold, and outputs the verification feature vector and the corresponding individual quality score. Furthermore, in the quality gating module, the calculation of individual quality scores and weighted quality scores by combining time delay and posterior covariance specifically includes: The time delay is mapped to a time freshness factor using a decay function; Transform the posterior covariance into an inversely proportional numerical estimate of the certainty factor; Multiply the time freshness factor by the estimated certainty factor to obtain the individual quality scores for each dimension of the feature vector. Calculate the weighted quality score of the observation set within the time window based on preset dimensional weights; When the weighted quality score is below the grading threshold, the output of the observation set within the time window is blocked.

[0046] Furthermore, the individual quality scores for each dimension of the completed feature vector are specifically obtained as follows: The decay coefficient of the decay function is set according to the characteristic time constant of the physical quantity corresponding to each dimension, and the time freshness factor is generated. Perform link heartbeat verification, value range verification, card value verification, and jump rate verification on each dimension, and convert the verification results into source health factors; The time freshness factor, the estimation certainty factor, and the source health factor are multiplied to obtain the individual quality score; The percentage of completed feature vectors for each dimension within a preset statistical period is statistically analyzed. When the percentage exceeds the upper limit threshold, the source health factor will be set to zero.

[0047] Specifically, the quality gating module receives the completed feature vector output by the estimation and completion module, the posterior covariance vector containing the posterior variance values ​​of each dimension, and the time delay of each dimension. To quantify the reliability decay of the completed data due to the lack of real-time observation, the system maps the time delay of each dimension to a time freshness factor with a value between zero and one. Since the rate of change of different physical quantities varies greatly across the entire scene, the system independently sets a decay coefficient according to the characteristic time constant of the physical quantity corresponding to each feature dimension. The calculation formula is as follows: . In the formula, This represents an exponential function with the natural constant as its base. For the received first Time delay in each dimension For the pre-set number Each dimension of the feature time constant, For the generated corresponding number A time freshness factor for each dimension. This characteristic time constant is used to characterize the rate at which the reliability of data in this dimension decays with increasing time delay, and is independently calibrated from the inertial time constant, which characterizes the evolution rate of the physical quantity itself in the estimation and completion module. Since the time delay and the characteristic time constant have the same dimensions, their dimensions cancel each other out after division, making this factor a dimensionless parameter.

[0048] The system extracts the corresponding posterior covariance vector from the received data. Posterior variance in each dimension This is transformed into an estimation certainty factor that is inversely proportional to the numerical value. The calculation formula is as follows: . in the formula For the first The normalized scaling factor for each dimension. To avoid the physical fallacy of directly adding dimensionless numbers with dimensional variance, this scaling factor is set to the reciprocal of the maximum allowable variance threshold for that dimension. Its dimension is the reciprocal of the corresponding variance dimension, so that the multiplication of the two terms results in the cancellation of dimensions, transforming them into a dimensionless ratio. For the temperature sensing dimension in degrees Celsius, the maximum allowable variance is set to... The corresponding scaling factor is the feasible implementation data. The formula yields... The dimensionless estimate of the certainty factor is generated.

[0049] The system performs multi-rule checks on each dimension of the sensing link. The link heartbeat check determines whether the interval between the data channel corresponding to that dimension and the most recent heartbeat response exceeds the heartbeat timeout duration. Feasible implementation data includes... Seconds. The value range check determines whether the current value of this dimension falls within the pre-configured upper and lower bounds of the physical quantity's range. The card value check determines whether the values ​​of this dimension are completely identical across multiple consecutive beats; the feasible implementation data for determining the number of beats is... The jump rate check determines whether the change in value of this dimension between adjacent beats exceeds 20% of the dimension's range. When all four checks pass, the system converts it into a source health factor with a value of one; if any check fails, it is converted to zero. The system also separately calculates the proportion of the supplementary feature vectors generated by model extrapolation for each dimension within a preset statistical period, with feasible data for this period being the past hour. When this extrapolation proportion exceeds the upper limit threshold of 30%, the system forcibly sets the source health factor of that dimension to zero. The rule generates the first... The source health factors in each dimension are .

[0050] The system multiplies the calculated time freshness factor, the estimated certainty factor, and the source health factor one by one to obtain the first complete feature vector. The individual quality scores for each dimension are calculated using the following formula: Since all three terms involved in the multiplication are dimensionless, the generated individual quality scores do not have any physical dimension conflicts. The system retrieves a pre-defined sequence of weights for each physical feature dimension, where the sum of all weights in the sequence is one. Based on this sequence, the system calculates the weighted quality score for the observation set within the current time window using the following formula: . In the formula, This represents the summation operation. The summation index variable for the feature dimensions. This represents the total number of dimensions of the feature vectors. For the first Preset dimension weights for each dimension. For the corresponding number Individual quality scores for each dimension, This is the weighted quality score for the output.

[0051] The system compares the weighted quality score with a globally set grading threshold, the feasible implementation data of which is as follows: When the weighted quality score falls below the grading threshold, the system determines that the overall credibility of the current global data has deteriorated, and immediately blocks the downstream output of the observation set and its corresponding feature vector for that time window. When the weighted quality score is greater than or equal to the grading threshold, the system confirms the current completed feature vector as the verification feature vector, and outputs it along with the individual quality scores of each dimension to the next level.

[0052] This step establishes a multi-factor reliability evaluation and interception mechanism with self-consistent physical dimensions, reducing the proportion of low-reliability extrapolated data entering the fusion process.

[0053] The fusion decision module performs normalization fusion on the verification feature vectors to generate a situational security index and a traceability link, and drives the display terminal to render and issue warnings.

[0054] Furthermore, in the fusion decision module, normalization fusion is performed on the verification feature vectors to generate the situational security index and the traceability link, specifically including: Transform the verification feature vector into a mapping value; The dynamic weight is obtained by multiplying the individual quality scores of each dimension of the verification feature vector with the preset dimension weights. Use the sum of all generated dynamic weights as the normalized denominator; The cumulative sum of the products of each dynamic weight and its corresponding mapping value is calculated and divided by the normalized denominator to generate the situational security index. The data source identifier, timestamp, and dynamic weights that the generated situational security index depend on are recorded to form a mapping dictionary, which is used to construct the traceability link.

[0055] Furthermore, the generation of mapping values ​​and the output of early warnings specifically include: Call the monotonic utility function to normalize by replacing the extreme values ​​with the sliding window quantile baseline, and output the mapped value; When the normalized denominator is lower than the preset lower limit, the output of the situational safety index will stop. Set upward and downward thresholds for the situational security index; An early warning is triggered when the situation security index is higher than the uplink threshold and the number of observation sets within a time window when the situation security index is continuously higher than the uplink threshold reaches a preset number. The warning is lifted when the situational security index falls below the downside threshold. When a delayed recalculation causes a new version to be generated in the traceability chain, a correction warning with a version identifier is output.

[0056] Specifically, the fusion decision module receives the verification feature vector and corresponding individual quality scores for each dimension from the quality gating module, integrating multi-source heterogeneous data across the entire domain into a unified safety indicator to drive the display terminal. For feature values ​​with different physical dimensions, such as temperature, pressure, or network traffic, contained in the verification feature vector, the system first converts them into dimensionless parameters. The system establishes a sliding statistical window for each feature dimension, with the feasible implementation data for the window duration being the past 24 hours. The system uses insertion sort to maintain the ascending sequence of feature values ​​within the window, and locates elements by the product of the sequence length and the target percentile. The 5th percentile is extracted as the lower baseline for that dimension, and the 95th percentile as the upper baseline, replacing absolute extreme values ​​that are easily contaminated by outliers. The system then determines whether the upper and lower baselines for that dimension are equal. If they are equal, it indicates that the dimension's value is constant within the statistical window or its value is discrete, and the system directly sets the mapped value of that dimension to zero. If the two are not equal, the system calls a monotonic utility function with truncation mechanism for linear normalization, calculated using the following formula: . In the formula, and These represent functions that take the maximum and minimum values ​​of each parameter, respectively. To verify the corresponding first feature vector Feature values ​​in each dimension and They correspond to the first Lower and upper baselines for each dimension. For the corresponding output number The mapping value is calculated in one dimension. Since the physical dimensions of the numerator and denominator are the same, their dimensions cancel each other out after division, and the resulting mapping value falls within the dimensionless interval of zero to one.

[0057] After completing the dimensionless mapping, the system multiplies the individual quality scores of each dimension by the preset dimension weights to obtain the dynamic weights that are actually effective within the current time window. The calculation formula is as follows: . In the formula, For the first Dynamic weights generated from each dimension For the received corresponding number Individual quality scores for each dimension, For the corresponding number Each dimension has a preset weight. The system accumulates all generated dynamic weights and calculates the normalized denominator. . In the formula, This represents the summation operation. The summation index variable for the feature dimensions. This represents the total number of dimensions of the feature vectors. This is the normalized denominator obtained from the calculation. When the individual quality fraction of some sensors decreases due to data degradation, this normalized denominator decreases accordingly. The feasible implementation data with a preset lower limit for the system is... When the normalized denominator is below the preset lower limit, the system determines that the effective information currently participating in the fusion is insufficient, and stops outputting the situational security index to avoid generating decision conclusions based on incomplete data. When the normalized denominator is greater than or equal to the preset lower limit, the system calculates the situational security index using the formula: . in the formula The generated dimensionless situational security index, since its denominator is the sum of the dynamic weights of the actual participating fusion items, will not produce systematic shifts when some dimensions are missing. The system synchronously extracts the data source identifiers and globally aligned timestamps of the underlying business devices on which the index depends, and combines them with the dynamic weights of each dimension to form a structured mapping dictionary, which is stored in memory or persistent storage. Based on this, a traceability link is constructed from the top-level index back to the original data.

[0058] The system drives the digital cockpit display terminal to render data based on the generated situational safety index and execute anti-shake warnings. The system sets uplink thresholds and corresponding downlink thresholds for the situational safety index; feasible implementation data for the uplink thresholds are as follows: Feasible implementation data for the downlink threshold are as follows: When the system determines that the generated situation security index is higher than the uplink threshold, a counter is started; when the number of observation sets within a time window where the situation security index continuously exceeds the uplink threshold reaches a preset number, the system confirms a deterioration in the situation and triggers an early warning. This preset number of feasible implementation data is... A time window is defined. Subsequently, when the situational security index falls below the set downlink threshold, the system cancels the warning. Furthermore, if the upstream access buffer module executes delayed recalculation logic due to network jitter and pushes a new verification feature vector with an incrementing version identifier to this module, the system re-executes the above full-process calculation. When the comparison reveals that the content of the new traceability link generated by the delayed recalculation has changed, and the out-of-limit state of the situational security index has reversed, the system drives the display terminal to output a correction warning carrying the version identifier.

[0059] By introducing a dual-threshold hysteresis and continuous counting mechanism, fluctuation interference in the critical state is filtered out, ultimately providing the digital cockpit with a stable and highly reliable decision-making basis.

[0060] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-powered, all-domain situational awareness digital cockpit management system, characterized in that, include: The data acquisition and labeling module obtains sensing data from multiple source devices, assigns a timestamp including the time of generation, performs clock drift compensation, and outputs observation records. The access buffer module receives the observation records, constructs a dynamic time boundary based on the channel lateness threshold of each data channel, performs time window aggregation and lateness recalculation on the disordered observation records, and outputs a time window observation set. The beat alignment module reads the time window observation set based on the system beat, calls the corresponding operator to perform aggregation and transformation according to the data type, and outputs a feature vector; The estimation and completion module inputs the feature vector into the continuous filtering model, performs extrapolation prediction and state update on the missing dimension of the feature vector, and outputs the completed feature vector, posterior covariance and time delay. The quality gating module receives the completed feature vector, calculates the individual quality score and weighted quality score by combining the time delay and the posterior covariance, intercepts the time window observation set whose weighted quality score is lower than the grading threshold according to the grading threshold, and outputs the verification feature vector and the corresponding individual quality score. The fusion decision module performs normalization fusion on the verification feature vector to generate a situational security index and a traceability link, and drives the display terminal to render and issue warnings.

2. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the acquisition and marking module, the clock drift compensation specifically includes: Obtain sample pairs of network heartbeat arrival times and local timestamps of the multi-source devices; The bias coefficient and drift slope are calculated using the recursive least squares method with a forgetting factor. Outlier pairs are removed using a random sampling consensus algorithm. Based on the bias coefficient and the drift slope, the generation time in the observation record is mapped to the timestamp.

3. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the access buffer module, the construction of dynamic time boundaries based on the channel lateness threshold of each data channel specifically includes: The distribution of transmission delay from generation to reception of the observation records in each data channel is statistically analyzed. Extract the quantiles of the transmission delay distribution as the channel lateness threshold for the corresponding data channel; The dynamic time boundary is generated by subtracting the channel lateness threshold from the maximum timestamp of the observation record received in the current data channel. The time window aggregation is triggered when the right boundary of the current time window observation set is less than or equal to the dynamic time boundary of all the data channels.

4. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the beat alignment module, the step of calling the corresponding operator based on the data type for aggregation and transformation specifically includes: When the data type is a continuous quantity, the robust aggregation operator based on the truncated mean is invoked; When the data type is a cumulative quantity, the time step-based differential scaling operator is invoked. When the data type is a discrete state quantity, the zero-order hold operator is invoked to extend the last observed state within the time window observation set to the boundary of the system beat.

5. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the estimation and completion module, the extrapolation prediction and state update of the missing dimensions of the feature vector specifically includes: For the missing dimension for which the observation record was not received, discretization prediction is performed using a continuous-time state transition matrix, and the process noise covariance is monotonically increased with the time delay. For the feature dimensions of the received observation records, the residual variance of the corresponding data channel within a preset time period is extracted as the observation noise covariance for state correction; The trace of the covariance matrix of the corrected continuous filtering model is calculated and used as the posterior covariance output.

6. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the quality gating module, the calculation of individual quality scores and weighted quality scores by combining the time delay and the posterior covariance specifically includes: The time delay is mapped to a time freshness factor using a decay function; The posterior covariance is transformed into an estimated certainty factor that is inversely proportional to the numerical value; Multiply the time freshness factor by the estimated certainty factor to obtain the individual quality score of each dimension of the completed feature vector; The weighted quality score of the observation set within the time window is calculated based on preset dimensional weights; When the weighted quality score is lower than the grading threshold, the output of the time window observation set is blocked.

7. The AI-powered global situational awareness digital cockpit management system according to claim 6, characterized in that, The specific steps of obtaining the individual quality scores for each dimension of the completed feature vector include: The decay coefficient of the decay function is set according to the characteristic time constant of the physical quantity corresponding to each dimension, and the time freshness factor is generated. Perform link heartbeat verification, value range verification, card value verification, and jump rate verification on each of the aforementioned dimensions, and convert the verification results into source health factors; The individual quality score is obtained by multiplying the time freshness factor, the estimated certainty factor, and the source health factor. The proportion of the completed feature vectors for each dimension within a preset statistical period is statistically analyzed. When the percentage exceeds the upper limit threshold, the source health factor is set to zero.

8. The AI-powered global situational awareness digital cockpit management system according to claim 1, characterized in that, In the fusion decision module, the step of performing normalization fusion on the verification feature vector to generate the situational security index and the tracing link specifically includes: Convert the verification feature vector into a mapping value; The dynamic weight is obtained by multiplying the individual quality score of each dimension of the verification feature vector with the preset dimension weight; The sum of all the generated dynamic weights is used as the normalized denominator; The sum of the products of each dynamic weight and its corresponding mapping value is calculated and divided by the normalized denominator to generate the situational security index. The data source identifier, timestamp, and dynamic weights that the situational security index depends on are recorded to form a mapping dictionary, which is used to construct the tracing link.

9. The AI-powered global situational awareness digital cockpit management system according to claim 8, characterized in that, The generation of the mapping value and the early warning output specifically include: Call the monotonic utility function to normalize by replacing the extreme values ​​with the sliding window quantile baseline, and output the mapped value; When the normalized denominator is lower than a preset lower limit, the output of the situational safety index is stopped. An uplink threshold and a downlink threshold are set for the aforementioned situational security index; The warning is triggered when the situation security index is higher than the uplink threshold, and the number of time window observation sets in which the situation security index is continuously higher than the uplink threshold reaches a preset number. The warning is lifted when the situational security index is lower than the downlink threshold. When the delayed recalculation causes the traceability link to generate a new version, a correction warning carrying a version identifier is output.