Time sequence neural network risk inference method and system for overseas research behavior sequence

Abstract

The application discloses an overseas research behavior sequence timing neural network risk inference method and system, and belongs to the technical field of machine learning and timing data processing. A behavior state sequence is constructed according to a uniform time base, and a behavior increment sequence is obtained by taking the difference value of adjacent time base points; an increment embedding vector is obtained by a timing neural network with a reserved time base sequence and a parameter journey pre-freezing; an increment direction vector is sequentially whitened, unitized and discarded with a module length; the increment direction vectors are added according to the time base sequence, and a vector module length is used as a risk statistic, compared with a zero hypothesis distribution quantile and graded. The distribution is determined by the number of additions and the dimension, and is not calibrated by the measured object observation; the significance level is obtained by dividing the preset value by the total number of journey time base points. The training target contains an isotropic constraint term to make the normal sample increment direction approximate to a spherical surface, and the network must be left out of the set distribution to be consistent with the acceptance. The application makes the criterion zero point not roll with the history of the measured individual, and can be used for research risk inference and graded prompt.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and time series data processing technology, specifically to a method and system for online risk inference of behavioral observation sequences of overseas study participants, detection of continuous and gradual behavioral evolution, and output of graded prompts. Background Technology

[0002] In cross-border study tours, participants are separated from their original home-school guardianship system for a period of time. Activities take place in various locations, including group trips, accommodations, and independent activities. The entities responsible for management are distributed among the organizing institutions, tour leaders, overseas hosts, and guardians. Furthermore, due to geographical distance and time differences, the response radius for management is significantly larger than for similar activities within China. Therefore, management needs to rely on information technology to monitor the participants' behavior throughout the trip.

[0003] Currently, the common practice for such activities is to collect participants' status at predetermined times or locations, such as scheduled check-ins, location reporting, and roll call during group activities, and to issue alerts when participants' locations exceed a preset range. This method of data collection is discrete; no data is generated between adjacent collections. The judgment is also sequential, based on whether the status obtained in a single collection exceeds the preset range. Therefore, the management side obtains several independent snapshots of the status, while what is actually needed is to determine the continuous change in the participant's behavior throughout the entire trip—that is, whether the overall behavior is undergoing a continuous change, even if the status obtained in each collection is within the preset range.

[0004] To overcome the aforementioned limitations, related technologies have begun to organize participants' behavioral data into behavioral sequences chronologically and introduce machine learning models to process these sequences. For example, patent application CN121170898A discloses a deep learning-based method and system for detecting abnormal behavior on campus. This scheme collects and preprocesses time-series data from each student's school uniform RFID tag. Based on the preprocessed time-series data, it constructs a trajectory feature background vector for each student's location. Then, based on the current trajectory feature vector and the trajectory feature background vector, it uses a dynamic attention mechanism to assess the degree of abnormality in real time and obtain an abnormality score. Based on the abnormality score, it generates graded early warning information and pushes it to the management terminal. The trajectory feature background vector is generated by weighted summation of the student's trajectory feature vectors at all times according to their respective trajectory attention weights. The warning threshold used for graded early warning is adaptively generated based on a first-level benchmark threshold, a second-level benchmark threshold, a threshold adjustment coefficient, and a background consistency score. Compared to the common practice of relying on timed check-ins and fixed-range judgments, this scheme changes the reference for judgment from a uniform preset range to a reference generated from the individual's own history, which improves the adaptability to individual differences and the real-time performance of the judgment to a certain extent.

[0005] However, the methods used to achieve the aforementioned effects have structural consequences. First, the trajectory feature background vector, which serves as the judgment reference, is obtained by weighted summation of the trajectory feature vectors of the individual at each time step. The result of the weighted summation is independent of the order of the items involved in the summation within the sequence; therefore, rearranging the order of the feature vectors at the same time step does not change the resulting trajectory feature background vector or the anomaly score derived from it. Second, this trajectory feature background vector is generated from the historical data of the tested individual and is updated as the individual's historical data progresses; therefore, the judgment reference and the tested data are in the same direction of movement. Third, the background consistency score used to adaptively generate the warning threshold measures the closeness between the weighted summation background vector and the arithmetic mean of the trajectory feature vectors at each time step, i.e., the consistency between time steps within the sequence.

[0006] As can be seen from the above structure, for continuous unidirectional behavioral changes whose amplitude accumulates slowly between collections and does not exceed the judgment threshold at any time, this type of scheme can easily include the change in the judgment reference itself before the change is identified, so that the deviation between the current time and the reference tends to decrease rather than increase; at the same time, such continuous and stable changes tend to increase the consistency between the times within the sequence, thereby raising the warning threshold generated adaptively.

[0007] Furthermore, patent application CN118397519B discloses an artificial intelligence-based campus student safety monitoring system and method. This scheme inputs behavioral data and environmental parameters into a multimodal fusion deep learning model based on a variational autoencoder to output a unified behavioral embedding vector sequence. It then uses an improved dual-decision Q-learning algorithm to model normal behavior to generate baseline behavioral patterns. Finally, it calculates anomaly scores and obtains their confidence levels based on a Markov decision algorithm. A judgment is made when both the anomaly score and confidence level reach preset values. Patent application CN108388969A discloses an internal threat risk prediction method based on the temporal characteristics of individual behavior. This scheme performs preprocessing such as quantification sampling, pre-emphasis, and windowing on accumulated historical behavioral information and extracts digital features to train a risk prediction model based on a long short-term memory network, thereby outputting risk prediction assessments and hierarchical early warnings.

[0008] Patent application CN121483597A discloses a method for detecting abnormal vital signs in patients based on artificial intelligence technology. This method uses historical monitoring records to learn the multidimensional distribution of a multidimensional feature dataset, establishing an individual baseline model and determining a comprehensive reference interval reflecting a normal state within this model. The multidimensional feature vectors obtained from real-time monitoring are input into the individual baseline model to calculate their deviation from the comprehensive reference interval. Clustering and trend analysis are then performed on the deviation feature data to calculate the consistency of the direction of change and the magnitude of continuous change of the multidimensional feature vectors, thereby obtaining a cumulative change index and a dynamic confidence score within a continuous time window. When the dynamic confidence score continuously exceeds an adaptive threshold, an abnormal event is determined, and the feature weights and comprehensive reference interval of the individual baseline model are updated according to the characteristics of the abnormal event. Regarding the determination step of this method, what is accumulated is a comprehensive index including the magnitude of change, rather than the direction itself after discarding the magnitude; the threshold used for comparison is calibrated from the individual's historical records and confirmed events; and the individual baseline model continues to be updated after the determination occurs.

[0009] The event calibration has been confirmed, and the individual baseline model continues to be updated after the determination occurs.

[0010] The above-mentioned schemes respectively adopt attention weighting, reinforcement learning modeling, fixed-length windowing and multidimensional distribution learning in the generation of the judgment reference, and adopt attention mechanism, Markov decision, long short-term memory network and clustering and trend analysis in the model form. As for the judgment link that this invention focuses on, they all have in common that they use the deviation of the current state from the reference generated by the test individual's own history as a measure of risk, and output the level after comparing the deviation with the threshold.

[0011] The aforementioned common forms are mathematically insensitive to the temporal order of the sequence. Arbitrarily rearranging the same behavioral sequence does not change the reference or criterion output, thus failing to utilize the distinguishable source of unidirectional cumulative change that grows linearly with observation time, while the cumulative amount of random fluctuations only grows with the square root of observation time. Simultaneously, the zero point of the criterion is generated by the history of the tested individual and rolls accordingly, causing continuous unidirectional behavioral evolution to be absorbed as a new reference before detection. The criterion output decreases instead of increasing, and the adaptive threshold using sequence homogeneity as a confidence indicator is conversely raised under this condition. Both are bound by the same quantity: the following strength of the criterion zero point relative to the tested sequence. Increasing this following strength achieves individual fit and low false alarms, but causes the criterion to lose its temporal cumulativeity; decreasing this following strength allows the criterion to gain temporal cumulativeity, but converts all normal differences between individuals into false alarms. Existing technologies can only choose one of these two parameters and cannot achieve both simultaneously.

[0012] It is evident that existing technologies struggle to detect the gradual and continuous behavioral evolution of participants throughout the entire journey without exceeding the judgment threshold at any given moment, without generating false positives between individuals. A new risk inference method and system are urgently needed to address this issue. Summary of the Invention

[0013] In view of the above-mentioned shortcomings of the prior art, the technical problem to be solved by the present invention is: how to make the risk criterion obtain the time sensitivity of the unidirectional cumulative component in the behavioral sequence without taking the baseline generated by the historical rolling of the tested individual as the zero point, so as to detect the continuous gradual change of behavior that has not crossed the judgment threshold at any time while ensuring that normal differences between individuals are not misjudged as abnormal.

[0014] The crux of the aforementioned technical problem lies in the fact that any judgment made by comparing a statistic with a threshold must answer the question of what distribution the statistic follows under normal circumstances. Current practices answer this question by either observing the object itself or estimating the distribution using a sample set; this is the root cause of why the zero point of the criterion depends on and rolls with the measured data.

[0015] The concept of this invention lies in making it unnecessary to estimate the distribution of statistics under normal conditions, rather than improving the estimation of the distribution. The question of "what distribution the statistic follows under normal conditions," which in existing practices must be answered by data, is transformed in this invention into a fact guaranteed by the construction during the training phase. The remaining elements of this invention are either necessary for the construction to function or are direct consequences of it.

[0016] To address the aforementioned technical problems, this invention provides a temporal neural network-based risk inference method for overseas study tour behavior sequences.

[0017] The temporal neural network of this invention incorporates an isotropic constraint term in its training objective. This constraint term makes the distribution of the incremental direction vector of normal samples approximate a uniform distribution on a unit sphere. The uniform distribution on a unit sphere is a known mathematical distribution completely determined by its dimension. The distribution of the magnitude of the sum of several unit vectors on this sphere can be analytically given by the known results of the hypersphericity test in directional statistics. Therefore, after this shaping, the distribution of the risk statistic under normal circumstances becomes a known object determined solely by the cumulative number and vector dimension. The decision threshold is then calculated solely from the cumulative number, vector dimension, and significance level, no longer requiring estimation from any observed data.

[0018] In order for the above-mentioned shaping to be successful, the present invention defines the object to be shaped and its retention during the inference phase as follows.

[0019] First, to ensure that the object being shaped is independent of the individual being tested, this invention collects multi-source behavioral observations of the study tour participants. After aligning with a unified time base, a multi-dimensional behavioral state vector is constructed for each valid time base point, resulting in a behavioral state sequence. The difference between adjacent time base points is then taken from this behavioral state sequence to obtain a behavioral increment sequence. The inherent differences among different study tour participants in terms of sleep schedule, activity radius, and social frequency are reflected in the overall translation of the behavioral state vector, while the difference between adjacent time base points remains strictly unchanged regarding this overall translation.

[0020] Secondly, to ensure that the object to be shaped falls within the domain of the target distribution, this invention inputs the behavioral increment sequence into the temporal neural network to obtain an incremental embedding vector. This incremental embedding vector is then subjected to whitening and normalization processes sequentially to obtain an incremental direction vector, while discarding the magnitude of the incremental embedding vector. Whitening eliminates the directional bias caused by correlations and scale differences between dimensions; the target distribution to be shaped is defined on a unit sphere, and the magnitude does not belong to the domain of this distribution. Discarding the magnitude also eliminates the behavioral fluctuations of the participants being tested.

[0021] Furthermore, in order to maintain the established distribution during the inference phase, the present invention determines and freezes the network parameters of the temporal neural network and the whitening matrix used in the whitening process before the start of the study tour, and does not update them with the observation data of the study tour participants during the tour.

[0022] Furthermore, to ensure that the accumulation of directions conveys the meaning of "whether the change continues in the same direction over time," the encoding is performed in a manner that preserves the time base order and does not employ summation-type pooling operations that are insensitive to the input order. Any aggregation in the form of weighted summation where each time point represents a weighted summation remains unchanged after any permutation of the time index and simultaneous permutation of the weights.

[0023] Under the above limitations, this invention adds the incremental direction vectors hourly in time base order to obtain a cumulative composite vector. The magnitude of the cumulative composite vector is used as the risk statistic, which is updated synchronously with the observation at each valid time base point. The risk statistic is then compared with the quantile of the aforementioned null hypothesis distribution to obtain the risk quantity, which is used to generate a graded risk warning. Components continuously moving in the same direction accumulate during addition, while components with independent directions cancel each other out. Therefore, a continuous and slow unidirectional evolution can gradually manifest in the statistics as observation continues, without requiring it to exceed the judgment threshold individually at any time base point. The number of accumulated vectors is both a component parameter of the risk statistic and a determining parameter of the null hypothesis distribution. The same number synchronously locks the statistic and its comparison benchmark, ensuring that they remain comparable throughout the entire observation process.

[0024] The significance level corresponding to the quantile is obtained by dividing the preset significance level by the total number of time base points within the study tour; this total number is determined by the tour length and time base length before the start of the tour, and is not estimated by the observation data of the participants being tested. After training, the time-series neural network must pass the test on a set of retained samples that does not overlap with the training sample set: the difference between the empirical quantile of the test statistic calculated from the retained sample set and the corresponding theoretical quantile of the null hypothesis distribution is not greater than the preset tolerance.

[0025] Therefore, the computational inputs for the comparison threshold converge to the cumulative number, vector dimension, significance level, and total number of travel time base points, none of which are estimated from the observational data of the participants in the study.

[0026] As a preferred embodiment, the encoding can employ a one-dimensional causal convolutional network, a gated recurrent network, or a causal self-attention network with positional encoding; the whitening matrix can be calculated and frozen once from the covariance matrix of the training sample set before the start of the journey; the isotropic constraint term can employ a synthetic modulus expectation approximation or a hyperspherical uniform energy formula, and a decorrelation constraint term can be added separately; the graded risk warning can be divided into levels according to thresholds corresponding to at least two significance levels; accumulation can be paused and no interpolation completion can be performed on time base points determined to be invalid; the sliding window can be adjusted during the journey. The incremental direction vector within the vector is tested for uniformity and a criterion premise failure flag is output. The magnitude of the incremental embedding vector can be retained as a bypass quantity, and the prompt wording can be annotated with the quantile of the empirical distribution of the magnitude of the training sample set. At least two accumulation windows of different lengths can be set in parallel, and the length of each window is determined by the analytical inverse solution of the drift intensity to be detected. The deterministic quantity that evolves according to a fixed period can be removed from the behavior state vector and input as a conditional quantity into the temporal neural network. The incremental direction vectors participating in the same cumulative synthesis vector can be obtained by non-overlapping encoding windows.

[0027] This invention also provides a temporal neural network risk inference system for overseas study tour behavior sequences that implements the above-mentioned method. The system includes an observation normalization module, an incremental embedding module, a direction shaping module, a direction accumulation module, a risk determination module, an effectiveness handling module, a read-only configuration storage module, and a management terminal. Each module corresponds one-to-one with each stage of the method and supports each other along a data flow consistent with the above-mentioned limitations. The isotropic constraint applied by the direction shaping module is a prerequisite for the null hypothesis distribution upon which the risk determination module is based to hold; the modulus discarded by the direction shaping module is a prerequisite that the comparison threshold of the risk determination module does not contain any parameters estimated by the participants being tested. This invention also provides an electronic device and a computer-readable storage medium as carriers for implementing the above-mentioned method and system.

[0028] The beneficial effects of this invention are as follows: The beneficial effect of this invention is that, even without any parameters estimated by the participants in the study being tested, the risk criterion still exhibits a time-cumulative property that monotonically increases with the continuation of observation. Under existing judgment methods, the above two requirements are simultaneously determined by a single parameter: the following strength of the criterion zero point relative to the tested sequence. Increasing this following strength results in the former being achieved at the expense of the latter, while decreasing it results in the latter being achieved at the expense of the former; only one can be chosen. This invention enables both to be achieved simultaneously.

[0029] This effect is achieved because the present invention does not estimate the distribution of the risk statistics under normal circumstances, but constructs it as known: after training with the isotropic constraint terms and passing the acceptance criteria, the distribution of the magnitude of the cumulative composite vector under normal circumstances becomes a known object determined only by the number of incremental direction vectors that have participated in the addition and the dimension of the vector; the comparison threshold is then calculated only by the number, the dimension, the significance level and the total number of travel time base points, where the number is the count, the dimension is the network output dimension, the significance level is given by the regulatory strategy, and the total number of travel time base points is determined by the travel arrangement and the time base length, none of which are estimated from the observation data of the participants in the study tour being tested.

[0030] Under this premise, the present invention further achieves the following effects. First, the inherent differences in the absolute levels of behavior among different participants in the study tour, that is, those whose behavior state vectors shift as a whole, are strictly eliminated algebraically by taking the difference between adjacent time base points, and do not enter any subsequent stages. Second, for changes that show a continuous bias in the incremental direction, the expected value of the square of the risk statistic increases quadratically with the cumulative number, while the square of the comparison threshold increases linearly with the cumulative number. Therefore, even if the magnitude of the change does not exceed the judgment threshold at each time base point, it will still inevitably cause the risk statistic to exceed the threshold at a certain cumulative number. Third, the cumulative number mentioned above, and the minimum drift intensity that can be detected at a given cumulative number, can be analytically calculated from the cumulative number, the dimension of the vector, and the significance level, allowing the deployer to determine the required observation length before the observation begins.

[0031] The aforementioned effects are not the sum of the effects of each stage, but rather depend on the mutual support between the functions of each stage for their validity. If the difference between adjacent time base points is not taken, the shaped object contains individual fixed offsets, and the directional distribution centers of each tested individual are different; the same known distribution cannot simultaneously apply to all individuals. If normalization is not performed and the modulus is discarded, the shaped object does not fall on the unit sphere, and the uniform distribution on the unit sphere is inapplicable. If the training target does not contain isotropic constraints, the directional distribution under normal circumstances is unknown, and the null hypothesis distribution cannot be analytically given. If the network parameters and whitening matrix continue to be updated during the training phase, the distribution shaped during the training phase drifts accordingly, and the known null hypothesis distribution no longer holds true during the inference phase. If a directional accumulation stage is not included, no risk statistics are generated, nor is any judgment ability produced. The absence of any of the above stages will not weaken the aforementioned effects, but rather render them invalid.

[0032] Compared to existing practices, the technical concept of this invention lies in constructing a known distribution for the statistics instead of estimating it under normal circumstances. This differs from methods that estimate the distribution based on observations of the object being measured or on a sample set. Furthermore, this invention can analytically calculate the required observation length and the minimum detectable drift intensity before observation begins; whereas methods that generate judgment criteria based on the object's own history cannot, in principle, provide such boundaries before deployment because the criteria rely on observation data that has not yet been obtained.

[0033] Furthermore, the preferred technical solution of the present invention also has the following effects: Accumulation is paused and no interpolation is performed on time base points determined to be invalid, so that the missing observations are not counted as no change in behavior; the uniformity of the incremental direction vector within the sliding window is checked and a criterion premise failure flag is output, and this flag is not used to adjust the comparison threshold, so that the premise for the criterion to be valid can be continuously monitored without destroying the property that the threshold does not contain the measured individual parameters; the magnitude of the incremental embedding vector is retained as a bypass quantity and used only to assist in modifying the determined risk level, so that the amplitude information can be utilized without entering the main criterion; at least two accumulation windows of different lengths are set in parallel and the length of each window is analytically deduced from the drift intensity to be detected, so that when the drift intensity is unknown in advance, drifts of different intensities are detected by their matching windows; the deterministic quantity that evolves according to a fixed period is used as a conditional quantity and is not allowed to enter the behavior state vector that is being differentiated, and each incremental direction vector participating in the same accumulation synthesis vector is obtained by a non-overlapping encoding window, all of which ensure the validity of the aforementioned known distribution under real observation conditions. The calculation inputs for all of the above items do not contain any parameters estimated from the object being measured.

[0034] The quantitative performance of this invention on real-world research and study behavior data, including false alarm frequency, detection rate and detection lead, as well as the actual achievable level of residual anisotropy, must be determined through training on real data and holdout set testing. This specification does not make any assertions on its numerical values. Attached Figure Description

[0035] Figure 1 This is a block diagram illustrating the principle of the method of the present invention; Figure 2 This is a schematic diagram of the overall process of the method of the present invention; Figure 3 This is a structural block diagram of the system of the present invention; Figure 4 A schematic diagram illustrating the principle that baseline offsets between individuals are eliminated as a whole in the behavioral increment space through hourly subtraction; Figure 5 This is a schematic diagram illustrating the relationship between the distribution of the incremental direction vector on a unit sphere, the cumulative composite vector, and the decision threshold. Figure 6 This is a schematic diagram illustrating the training, convergence determination, and freezing process of a temporal neural network and a whitening matrix. Figure 7 This is a schematic diagram illustrating the relationship between risk statistics and decision thresholds as the cumulative number of items changes.

[0036] In the diagram: 1—Observation Normalization Module; 2—Incremental Embedding Module; 21—Differential Unit; 22—Causal Temporal Network Unit; 3—Direction Shaping Module; 31—Whitening Unit; 32—Unitization Unit; 33—Module Length Bypass Unit; 4—Direction Accumulation Module; 5—Risk Judgment Module; 51—Threshold Calculation Unit; 52—Grading Unit; 6—Effectiveness Handling Module; 7—Read-Only Configuration Storage Module; 8—Management Terminal; 41—Behavioral State Trajectory of Individual A; 42—Behavioral State Trajectory of Individual B; 43—Behavioral Increment Sequence of Overlapping Individuals A and B; 44—Fixed Absolute Threshold Applied to Behavioral State Quantity; 45—Unit Sphere; 46—Incremental Direction Vector; 47—Accumulated Composite Vector; 48—Judgment Threshold Circle; 49—Average Direction of Incremental Direction Vector; 61—Curve of Judgment Threshold Variation with Accumulation Quantity; 62—Risk Statistics under Normal Conditions with Accumulation Quantity The curves showing the changes; 63—the curve showing the change of risk statistics with the number of accumulated values ​​when there is directional drift; 64—the detection point where the risk statistics first cross the judgment threshold; 101—collecting multi-source behavioral observations and aligning them according to a unified time base; 102—constructing a multi-dimensional behavioral state vector; 103—determining whether the observation at this time base point is valid; 104—pausing accumulation, not performing interpolation fill, and setting an interruption flag; 105—taking the difference between adjacent time base points; 106—obtaining the incremental embedding vector from the frozen temporal neural network; 107—whitening processing; 108—normalizing processing and discarding the magnitude; 109—adding time-by-time and taking the magnitude of the composite vector; 110—calculating the analytical threshold and comparing it with the risk statistics; 111—outputting the risk quantity and graded risk warning; 112—separating the deterministic quantity that evolves according to a fixed period as a conditional quantity; 113—determining the significance level of this judgment based on the preset significance level and the total number of time base points in the journey; 201—Obtain the behavioral observation sample set of historical itineraries; 202—Filter to obtain the normal sample set; 203—Construct the incremental sequence and calculate the incremental embedding; 204—Construct the loss function; 205—Train the temporal neural network and whitening matrix; 206—Determine whether the convergence criterion is met; 207—Increase the sample size or increase the weight of the constraint terms; 208—Freeze the parameters and solidify them into a read-only configuration; 209—Deliver for online inference. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention. Equivalent substitutions and improvements made without departing from the concept of this invention should be covered within the scope of protection of this invention. In the following embodiments, quantities represented by symbols are exemplary parameters set according to actual deployment conditions; values ​​marked as theoretical calculation values ​​are calculated from the relationships given in the text and well-known mathematical quantile tables, and do not represent the actual measured results of the system.

[0038] Example 1: Acquisition, temporal alignment, and construction of behavioral state sequences for multi-source behavioral observations like Figure 1 and Figure 2 As shown, this embodiment provides a specific implementation method for the observation acquisition and state construction stage, corresponding to... Figure 2 Steps 101, 102 and 103 in the process.

[0039] The study tour participants will be monitored for one or more of the following behaviors during the trip: location coordinates and area identification reported by their personal devices; access records to accommodations, school buildings, and meeting points; check-in records for group activities; arrival and departure records at various points in the itinerary; and the timing of the exchange requests between their personal devices and the tour leader's device. All of these observations will be generated as events, with the generation times of each source being asynchronous and the generation density varying depending on the itinerary.

[0040] The time base length is set to divide the travel period into equal-length time base intervals. The value of the time base length is set according to the actual deployment observation density, preferably such that at least one valid observation can be obtained on average within a single interval; in one specific embodiment, the time base length can be a value between fifteen minutes and one hour.

[0041] A validity flag is assigned to each time base interval. An observation at a given time base point is considered valid if at least one observation within the interval passes integrity verification and has a clock offset within tolerance; otherwise, it is considered invalid. This determination is... Figure 2 Step 103 in the example, the handling of invalid base points is described in Example 6.

[0042] For a time base interval deemed valid, a multidimensional behavioral state vector is constructed from the observations within that interval, with each component valued as follows: First, the spatial proportion component: the locations involved in the itinerary are divided into several categories according to function, such as accommodation, group activity venues, public transportation, and autonomous activity areas. For each category, the ratio of the time spent by participants in that category within the interval to the length of the interval is taken as a component. Second, the movement intensity component: the sum of the displacements between two consecutive location observations within the interval is divided by the length of the interval as a component. Third, the group activity participation component: the ratio of the number of group activity entries that should be attended to the number of actual check-in entries within the interval is taken as a component. If there are no entries that should be attended within the interval, this component takes the value of the previous valid interval. Fourth, the interactive response component: the median time interval between the time of sending an interactive request from the team leader terminal to receiving a response within the interval, truncated by an upper bound, is taken as a component. Fifth, the work-rest phase component: the sine and cosine values ​​of the time period sequence number of the interval in the local time are taken as two components.

[0043] It should be noted that the aforementioned work-rest phase components evolve according to a fixed period in time, and the difference between adjacent time base points is a quantity uniquely determined by the time base position, containing no random components. If it is included in the differenced behavioral state vector, the behavioral increment generated by it is a deterministic quantity. After encoding, a periodically changing systematic structure will be injected into the increment embedding vector. This structure will not cancel each other out due to addition in any non-integer period accumulation interval, causing the distribution of the increment direction vector on the unit sphere to systematically deviate from uniformity, and the null hypothesis based on Example 4 will therefore not hold. Therefore, in a preferred embodiment, the work-rest phase components are removed from the behavioral state vector and do not participate in the difference in Example 2, but are instead input into the causal time-series network as conditional quantities as described in Example 2. The increments of components in the behavioral state vector other than the deterministic quantities that evolve according to a fixed period contain random components and are not the object of this treatment.

[0044] Each of the above components is retained according to its own physical dimensions, and no cross-component normalization is performed in this step. The scale differences across components are uniformly eliminated by the whitening process in Example 3. The above five types of components are exemplary examples and can be added or removed according to the actual available observation sources. This invention does not limit the component composition of the state vector; the success of subsequent steps only requires that the state vector maintains the same component composition and the same arrangement order during the journey. Arranging the state vectors of each effective time base interval in chronological order yields the behavioral state sequence.

[0045] Example 2: Extraction of Behavioral Increments and Temporal Neural Network Encoding with Parameter Freeze like Figure 2 As shown, this embodiment provides a specific implementation method for the behavior increment extraction and increment embedding steps, corresponding to... Figure 2 Steps 105 and 106 in the text.

[0046] For two consecutive time base intervals that are both determined to be valid, the difference between the behavior state vector of the latter interval and the behavior state vector of the former interval is the behavior increment of that time base point. The behavior increments are then arranged in chronological order to obtain the behavior increment sequence. When two consecutive intervals are not valid at the same time, no increment is generated in this interval, as shown in Example 6.

[0047] This step serves the following purpose: the inherent differences among participants in study tours in terms of their daily routines, activity radius, and group participation habits are represented by a fixed offset in their state vectors within the state space. For any fixed offset, the difference between the state vectors at the two preceding and following time points, after simultaneously superimposing this offset, is exactly the same as before the superposition; that is, the increment remains strictly unchanged for this offset. Therefore, the inherent differences among participants do not proceed to any subsequent steps after this step. See the illustration of this property. Figure 4 , Figure 4The behavioral state trajectory 41 of individual A and the behavioral state trajectory 42 of individual B differ by a fixed offset. They are located on two separate trajectories in the state space. After hourly subtraction, they coincide in the increment space as the same behavioral increment sequence 43. The fixed absolute threshold 44 applied to the behavioral state quantity will make trajectory 41 judged as abnormal and trajectory 42 not judged as abnormal. This is the origin of the misjudgment of normal differences between individuals.

[0048] A causal temporal network that preserves temporal order information is used to encode the behavioral increments of the most recent time base points into incremental embedding vectors. This causal temporal network can adopt one of the following three structures.

[0049] Structure A is a one-dimensional causal convolutional network. It consists of several layers of dilated causal convolutions stacked together. The kernel width of each layer is three, and the dilation rate increases in powers of two according to the layer order. The input of each layer is padded causally on the left according to the dilation rate to ensure that the current output does not depend on future time. The last layer takes the channel vector of the current time and obtains the incremental embedding vector through a linear mapping.

[0050] Structure B, Gated Recurrent Network. It employs a unidirectional gated recurrent unit, sequentially reading in increments according to the time base, and obtaining the incremental embedding vector by taking the hidden state at the current time step and applying a linear mapping layer.

[0051] Structure C is a causal self-attention network with positional encoding. The incremental sequence within a sliding window of a predetermined length is used to superimpose sinusoidal positional codes at each time step. This sequence is then passed through a multi-head self-attention layer with causal masking and a feedforward layer. The output at the current time step is then processed through a linear mapping layer to obtain the incremental embedding vector.

[0052] In all three structures described above, pooling operations that are insensitive to the input order, such as summation, averaging, or attention-weighted summation, should not be applied to the representations at each time step within the window. The reason for this is that any aggregation, such as a weighted summation of the representations at each time step, retains its value after any permutation of the time indices and simultaneous permutation of the weights. If such aggregation is used, the incremental embedding vectors obtained by rearranging the same set of increments in any order will be identical, and the accumulation in Example 3 will no longer reflect whether the changes continue in the same direction over time.

[0053] For the deterministic quantity that evolves at a fixed period by shifting out of the self-behavioral state vector as described in Embodiment 1, this embodiment inputs it as a conditional quantity into the causal temporal network. That is, at each time point of the network input, the phase quantity of the time base point is concatenated with the behavioral increment of the time base point and then input. Thus, the network can still use phase information to distinguish the same change occurring in different time periods during encoding, while the phase quantity itself does not generate a behavioral increment and does not enter the directional accumulation of Embodiment 3.

[0054] The parameters of the causal temporal network are determined through training before the start of the study tour. The training method is described in Example 5, and the process is as follows: Figure 6 At the start of the process, the network parameters and the whitening parameters used in Example 3 are written together into a read-only configuration file, i.e. Figure 1 and Figure 3 The system uses a read-only configuration storage module 7 to calculate and store a summary value of the configuration file. During the trip, the summary value is verified after each load; if the verification fails, inference is refused. During the trip, the inference process does not calculate gradients or accept any parameter updates; that is, the network parameters are not updated based on the observation data of the study participants.

[0055] Example 3: Temporal accumulation of whitening, normalization, and incremental direction vectors like Figure 2 , Figure 4 and Figure 5 As shown, this embodiment provides a specific implementation method for the orientation shaping and orientation accumulation stages, corresponding to... Figure 2 Steps 107, 108, and 109 in the text.

[0056] The whitening parameters are calculated once before the start of the process. The mean vector and covariance matrix are calculated from all incremental embedding vectors of the training sample set. Eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvector matrix and the eigenvalue diagonal matrix. The whitening matrix is ​​obtained by right-multiplying the eigenvector matrix by the sum of the eigenvalue diagonal matrix and a small positive number multiplied by the identity matrix to the power of -1 / 2, and then right-multiplying by the transpose of the eigenvector matrix. The small positive number is used to prevent ill-conditioned inversion. The whitening matrix and the mean vector are calculated once and frozen before the start of the process and are not updated during the process.

[0057] For each incremental embedding vector during the journey, the mean vector is first subtracted, and then the whitening matrix is ​​left-multiplied to obtain the whitened vector. The whitened vector is then divided by its own magnitude to obtain the incremental direction vector, whose magnitude is always one. Whitening eliminates the direction bias caused by correlation and scale differences between dimensions; normalization ensures that the incremental direction vector retains only direction information. The distribution of the incremental direction vector on the unit sphere is shown in [reference needed]. Figure 5 ,in Figure 5 Column a shows the normal case where the orientation is approximately uniformly distributed. Figure 5 Column b shows the drift situation where the direction continuously deviates to the same side and the average direction is 49 non-zero.

[0058] The magnitude of the whitened vector is not included in subsequent risk statistics and threshold calculations; it can be stored separately as a bypass quantity, as shown in Example 7. When the magnitude of the whitened vector is less than the set lower bound, no incremental direction vector is generated at that time base point, and it is not included in the cumulative count.

[0059] Directional accumulation is performed as follows: The cumulative composite vector and the number of accumulations are both set to zero at the start of the journey; for each time base point that generates an incremental direction vector, the incremental direction vector is added to the cumulative composite vector, and the number of accumulations is incremented by one; the magnitude of the cumulative composite vector is taken as the risk statistic. The risk statistic is updated synchronously with the observations of each valid time base point.

[0060] As an alternative to the handling method described in Embodiment 1, the dyspepsia phase component can also be retained in the behavioral state vector, while the length of the cumulative window is taken as an integer multiple of the number of time base points corresponding to that cycle. The sum of the increments of the deterministic cycle quantity within a complete cycle is a zero vector, and their contributions to the cumulative composite vector cancel each other out over the entire cycle. When using this method, the risk statistic outputs a judgment conclusion only when the cumulative number is an integer multiple of that number; for the remaining time base points, only accumulation is performed without judgment.

[0061] The mathematical properties of this step are as follows. Since the square of the length of a vector in Euclidean space is equal to the dot product of that vector and itself, expanding the square of the magnitude of the cumulative composite vector according to the bilinearity of the dot product yields: the square of the hazard statistic is equal to the sum of the pairwise dot products of all incremental direction vectors; since the magnitude of each incremental direction vector is always one, the sum of terms with the same subscript is exactly the cumulative number, therefore the square of the hazard statistic is equal to the sum of the cumulative number and the sum of the dot products of all pairs with different subscripts. This expansion is a direct consequence of the definition of vector length and does not depend on any additional assumptions.

[0062] Taking the expectation of the above equation, we consider two cases. Case 1: The incremental direction vectors are independent and uniformly distributed on the unit sphere. In this case, the expectation of the dot product of any two vectors with different indices is equal to the dot product of their expected vectors. Since the expected vector of a uniformly distributed vector is zero, the expectation of the dot product of pairs of different indices is zero, and the expectation of the square of the hazard statistic is exactly equal to the cumulative number. Case 2: The incremental direction vectors are independent and share a common non-zero average direction. Let the magnitude of this average direction vector be denoted as the drift intensity. In this case, the expectation of the dot product of any two vectors with different indices is equal to the square of the drift intensity. Therefore, the expectation of the square of the hazard statistic is equal to the cumulative number, plus the cumulative number multiplied by the cumulative number minus one, and then multiplied by the square of the drift intensity.

[0063] From the above two equations, we can deduce that: in Case 1, the risk statistic increases with the number of cumulative occurrences by the order of the square root of the number of cumulative occurrences; in Case 2, when the number of cumulative occurrences is large, the risk statistic increases with the number of cumulative occurrences by the order of the product of the number of cumulative occurrences and the drift intensity. The ratio of the two increases with the number of cumulative occurrences by the order of the square root of the drift intensity multiplied by the number of cumulative occurrences, meaning that the distinguishability of a continuous change in the same direction relative to random fluctuations increases monotonically with the continuation of observation. For the relationship between the two cases and the decision threshold, see [link to relevant documentation]. Figure 7Curve 62 represents the normal situation, curve 63 represents the situation with directional drift, and curve 61 represents the judgment threshold. Therefore, a continuous unidirectional change in amplitude that does not exceed the judgment threshold at any time base point can be gradually revealed in the statistics through accumulation.

[0064] Example 4: Null Hypothesis Distribution, Decision Threshold, Grading, and Detectable Boundary like Figure 2 and Figure 7 As shown, this embodiment provides a specific implementation method for the risk assessment stage, corresponding to... Figure 2 Steps 110 and 111 in the example are used. The null hypothesis is that the incremental direction vectors are uniformly distributed on a unit sphere, and the expected value of the inner product of any two incremental direction vectors with different indices is zero. It should be noted that, according to the expansion in Example 3, the expected value of the square of the cumulative composite vector magnitude is equal to the sum of the expected values ​​of the cumulative number of vectors and the inner product of all pairs of different indices. Therefore, the validity of the above expectation relation only requires that the expected value of the inner product is zero, and does not require that the direction vectors are independent of each other. This null hypothesis does not involve any statistics of the participants in the research study being tested. Its validity is shaped by the isotropic constraint term described in Example 5 during the training phase and confirmed by the acceptance criterion described in Example 5.

[0065] The significance level is allocated as follows. In this embodiment, the judgment is refreshed synchronously with each valid time base point and compared sequentially with the threshold, constituting multiple consecutive judgments on the same object being measured. If each judgment uses the same significance level, the probability of at least one violation during the entire observation process will be significantly higher than that significance level. Assuming the length of the study tour and the time base length are determined before the tour begins, the total number of time base points within the tour is a known quantity, denoted as the total number N. The significance level used for each judgment is taken as the quotient of the preset significance level divided by N; the sum of the significance levels of N judgments is exactly equal to the preset significance level, therefore the probability of at least one violation during the entire tour does not exceed this preset value. The total number N is determined by the itinerary arrangement and time base length. It belongs to the deployment parameters along with the significance level and is not estimated by the observation data of the participants in the study. Therefore, the property that the threshold established in this embodiment does not contain any parameters estimated by the tested object is not changed by this allocation method. The input for calculating the threshold changes from the three factors of cumulative number, vector dimension and significance level to the four factors of the three factors and the total number.

[0066] This allocation method is insensitive to the value of the total number N. Example of theoretical calculation: With a vector dimension of 32 and a preset significance level of 1%, when N increases from 336 to 1008, the threshold coefficient changes from 1.5267 to 1.5624, a change of approximately 2.3%. Therefore, the impact of the estimation bias of the travel length on the criterion is negligible, and precise prediction of the travel length is not required during deployment.

[0067] The threshold is determined analytically as follows. Based on the known results of the hyperspherical homogeneity test in directional statistics, under the above null hypothesis, the test statistic obtained by dividing the vector dimension by the cumulative number and then multiplying by the square of the hazard statistic asymptotically follows a chi-square distribution with degrees of freedom equal to the vector dimension. Accordingly, for a given significance level, the comparison threshold is taken as: the square root of the quotient obtained by multiplying the cumulative number by the chi-square distribution with degrees of freedom equal to the vector dimension (subtracting the significance level quantile), and then dividing by the vector dimension. The quantiles can be obtained from a well-known chi-square quantile table.

[0068] The inputs to the aforementioned thresholds are limited to three factors: the cumulative count, the vector dimension, and the significance level. The cumulative count is the number of counts, the vector dimension is the network output dimension, and the significance level is given by the monitoring strategy. None of these three factors are derived from the observational data of the participants in the research project. Based on this, a verifiable criterion can be given: any threshold whose calculation input consists solely of the above three factors, and none of these factors are estimated from the data of the tested subjects, falls within the scope of this invention.

[0069] Example of theoretical calculation: Assuming a vector dimension of 32, a preset significance level of 1%, a journey duration of 14 days, and a time base length of 30 minutes, the total number N is 672. The significance level for each determination is approximately 1.49 multiplied by 10 to the power of -5. Referring to a known chi-square quantile table, the quantile of the chi-square distribution with 32 degrees of freedom at this significance level is approximately 76.83. Substituting this into the above formula, the threshold is approximately equal to 1.5495 multiplied by the square root of the cumulative number. The above values ​​are theoretical calculations obtained using this formula and a known quantile table.

[0070] The chi-square distribution described above is an asymptotic result, and its approximation decreases as the cumulative number of elements decreases. When the cumulative number of elements is less than a set lower bound, a pre-prepared quantile table is used: before the start of the process, a pseudo-random number generator independently and uniformly generates a given number of direction vectors on a unit sphere of a given dimension, calculates the magnitude of the composite vector, repeats this process a sufficient number of times, and takes the empirical quantiles of each obtained magnitude as the threshold for that number of elements. Each number less than the lower bound is then tabulated. The inputs to this tabulation process are only the vector dimension, the cumulative number of elements, the significance level, and the pseudo-random number generator; no observational data from any tested object or other individuals are used, therefore the obtained thresholds are still not calibrated from any sample set.

[0071] Risk level and classification are output as follows: The risk level is output as the upper probability obtained by the test statistic relative to the chi-square distribution with degrees of freedom equal to the vector dimension. Two significance levels are set, where the first significance level is greater than the second significance level; when the risk statistic does not exceed the threshold corresponding to the first significance level, level one is output; when the risk statistic exceeds the threshold corresponding to the first significance level but does not exceed the threshold corresponding to the second significance level, level two is output; when the risk statistic exceeds the threshold corresponding to the second significance level, level three is output. The risk classification warning, along with the current cumulative number, is pushed to the management terminal 8.

[0072] This scheme has an analytically computable lower bound on the drift intensity it can detect given the cumulative number of data points, vector dimension, and significance level. This lower bound can be used to determine the parameters and assess the sufficiency of the observations. Figure 7 The intersection point 64 of the corresponding hazard statistic curve 63 and threshold curve 61 is given. Comparing the expected value of scenario 2 in Example 3 with the aforementioned threshold, and ensuring that the expected value of the square of the hazard statistic is not less than the square of the threshold, dividing both sides by the cumulative number and rearranging, we obtain: the minimum detectable drift intensity is equal to the square root of the quotient obtained by dividing the chi-square distribution (degrees of freedom equal to the vector dimension) by the significance level quantile, subtracting one from the quotient, and dividing by the cumulative number minus one. The input to this formula is also only the cumulative number, vector dimension, and significance level.

[0073] From this formula, the following relationships can be directly used in implementation. First, the minimum required number of accumulations. The minimum detectable drift intensity is no greater than the condition that one side has a solution; therefore, the number of accumulations must be no less than the quotient obtained by subtracting the significance level quantile from the chi-square distribution (whose degrees of freedom equal the vector dimension) and dividing by the vector dimension. This gives the minimum observation length required for this scheme to begin having detection capability. Second, the relationship between observation length and sensitivity. The minimum detectable drift intensity decreases with the number of accumulations by the reciprocal of the square root of the number of accumulations; that is, for every fourfold increase in observation time, the minimum detectable drift intensity is halved. Third, the relationship between significance level and sensitivity. The smaller the significance level, the larger the chi-square quantile, and the greater the minimum detectable drift intensity; that is, the stricter the false alarm requirement, the greater the detectable drift intensity. The trade-off between the two can be directly quantified by this formula.

[0074] Example of theoretical calculation: Taking a vector dimension of 32, a preset significance level of 1%, and a total number N of 672, the quotient obtained from the aforementioned quantiles is approximately 2.40, therefore the minimum cumulative number is approximately 3; further, the minimum detectable drift intensity when the cumulative number is 10 is approximately 0.395, when the cumulative number is 100 it is approximately 0.119, and when the cumulative number is 1000 it is approximately 0.037. Compared with the corresponding values ​​without allocation (1.67, 2, 0.273, 0.082, 0.026), the sensitivity loss is on the order of a constant factor, without changing the order of magnitude relationship that the minimum detectable drift intensity decreases with the cumulative number according to the reciprocal of its square root. The above values ​​are all theoretical calculations obtained using this formula and the known chi-square quantile table.

[0075] Example 5: Implementation, training, and parameter freezing of isotropic constraint terms like Figure 6 As shown, this embodiment provides two executable implementations that make the distribution of the incremental direction vector of normal samples approximate a uniform distribution on a unit sphere, as well as the complete process of acquiring training samples, determining convergence, and freezing parameters.

[0076] The acquisition and selection of training sample sets correspond to Figure 6 Steps 201 and 202 in the text. Take behavioral observations of several study tour participants from previously completed itineraries, and construct incremental sequences according to Examples 1 and 2 to serve as the training sample set. The training sample set should consist of observations under normal itinerary conditions, and its selection should be carried out in one or a combination of the following methods. Method 1: Post-confirmation method. Take itineraries where, after the trip, the team leader and management jointly confirm that no safety incidents occurred and there are no records of participants leaving the group or losing contact; all observations from these itineraries are considered normal samples. This method is based on post-trip conclusions and does not involve any point-by-point annotation during the trip. Method 2, the majority normal method, takes a sufficient amount of trip data when it is difficult to confirm each trip individually. Based on the fact that the vast majority of completed trips are normal, it first trains with all the data for one round, and then calculates the average directional magnitude of each trip according to the convergence determination method of this embodiment. After removing the few trips that are significantly larger, it retrains. The basis of this method is that, according to the expectation formula of embodiment 3, the average directional magnitude of trips with continuous unidirectional evolution is systematically larger, so the few abnormal trips can be separated from the majority normal samples.

[0077] The training sample set does not require time-by-time anomaly labels. This is because, under conditions where there is only a single data stream of the tested object itself and no point-by-point annotation, the normal evolution of an individual and the gradual changes that need to be detected are inextricably linked in terms of information. Any training objective relying on point-by-point anomaly annotation is unreliable under these conditions. The training objective of this invention does not aim to distinguish between normal and abnormal, but rather to shape the directional distribution of normal changes into a known distribution. The two screening methods mentioned above only require post-hoc conclusions at the process level or majority in a statistical sense, not point-by-point annotation.

[0078] The construction of isotropic constraint terms corresponds to Figure 6 Step 204. This constraint can take one of the following two forms.

[0079] Form 1, Approximation of the Expected Synthetic Magnitude. Based on the expansion of Example 3, when each direction vector is independent and uniform, the expected value of the square of the synthetic magnitude of a given number of direction vectors is exactly equal to that number. A constraint term is constructed accordingly: Within each training batch, the samples within the batch are randomly divided into several groups. For each group, the square of its synthetic vector magnitude is calculated and divided by the number of samples in that group. The average of these groups is then taken. This average minus one, and the square of the result, is the value of this constraint term. This term is zero when the direction distribution is uniform and positive when it deviates from uniformity.

[0080] Form 2, Hyperspherical Homogeneity Energy Formula. Using a known hyperspherical homogeneity metric, for any two different samples within a batch, the exponent of the square of the magnitude of the difference between their incremental direction vectors multiplied by the negative temperature coefficient is calculated. The average of all different sample pairs is then taken as the logarithm, which is the value of this constraint term. This term takes a minimum value when the directional distribution is uniform. This homogeneity energy itself is a known distribution uniformity metric in the art, and its form is not emphasized in this embodiment. The key point of this invention is that the constraint term is applied to the incremental direction vector, and its function is to ensure that the null hypothesis distribution described in Example 4 holds.

[0081] Both of the above forms suffer from degeneration when used alone: ​​if the network maps all inputs to random directions independent of the inputs, the direction distribution is still uniform, but all incremental information is lost. To eliminate this degeneration, an auxiliary term is incorporated into the training objective, requiring that the main components of the behavior increment can be reconstructed from the incremental embedding vector through a linear decoder. This auxiliary term is taken as the square of the magnitude of the difference between the behavior increment and the decoded vector. The overall training objective is taken as the weighted sum of the isotropic constraint term and this auxiliary term according to a tradeoff coefficient.

[0082] Acceptance judgment, corresponding Figure 6Steps 206 and 207 in the example. The null hypothesis upon which Example 4 is based requires that the incremental direction vectors of normal samples are uniformly distributed on a unit sphere, and that the inner product of any two incremental direction vectors with different indices has an expected value of zero. Item-by-item testing of anisotropy and correlation can only cover the listed deviations; therefore, this example uses an end-to-end acceptance criterion: on the normal sample holdout set that does not overlap with the training set, the test statistic is calculated in the same way as in Examples 3 and 4, its empirical quantile is taken, and compared with the corresponding theoretical quantile of the chi-square distribution whose degrees of freedom equal the vector dimension; when the difference between the two is not greater than the preset tolerance, acceptance is deemed passed; otherwise, it is determined as follows. Figure 6 Step 207 involves adjusting the network structure, increasing the training sample size, or increasing the weights of the isotropic and decorrelation constraints before retraining. Any factor that invalidates the null hypothesis, whether it be anisotropy, adjacency correlation, higher-order dependence, or other unlisted deviations, will necessarily manifest as a deviation of the empirical distribution from the theoretical distribution; therefore, this criterion can cover all of them.

[0083] Alongside this, a separate sensitivity criterion is established. Based on the relationship described in Example 6, residual anisotropy and true drift enter the expected value of the risk statistic in exactly the same way. Therefore, to prevent the true drift from being overwhelmed by residual anisotropy, the former must be significantly smaller than the latter. Specifically, the magnitude of the arithmetic mean vector of the incremental direction vectors on the holdout set must not be greater than the quotient obtained by dividing the minimum drift intensity to be detected by a preset margin factor, where the margin factor can be a value between three and ten. This criterion is independent of the size of the holdout set. If this requirement cannot be met under a given network structure and training sample size, the network structure should be adjusted, the training sample size increased, or the sensitivity requirement lowered, rather than relaxing the requirement itself.

[0084] The determination principles for the parameters involved in this embodiment are summarized in the table below, and the specific values ​​of each parameter are set according to the actual deployment conditions.

[0085]

[0086] Parameter freezing and version fixing correspond to Figure 6 Steps 208 and 209 in the process are followed. After training, the network parameters, whitening matrix, mean vector, vector dimension, lower bound of the magnitude, threshold table, and each saliency level are written to a read-only configuration file. The summary value of this file is calculated and saved separately. During the process, the inference process verifies the summary value after each configuration load. If the verification fails, the process refuses to start. The above configuration is not updated during the process.

[0087] Example 6: Two Boundary Conditions: Observation Interruption and Criterion Premise Failure This embodiment provides two implementation methods for boundary conditions, where the handling of observation interruption corresponds to... Figure 2Step 104. For a time base interval, the observation of that time base point is deemed invalid if any of the following conditions occur: no observations are generated by any source within the interval; the deviation between the timestamp of the observation within the interval and the local time base exceeds the set tolerance; or the observation within the interval fails the integrity check.

[0088] When an item is deemed invalid, no behavior state vector is constructed, no behavior increment is generated, no increment direction vector is generated, the cumulative number is not increased, the cumulative composite vector and risk statistics remain unchanged, and no interpolation, extrapolation or other forms of completion are performed on the observations in that interval.

[0089] The reason for not making a supplement is that if the state vector of the interval is supplemented by interpolation, the resulting increment is the product of the interpolation algorithm rather than the change in the actual behavior; if the increment is normalized and included in the accumulation, a component determined by the algorithm rather than by the behavior will be injected into the direction. This component may either be consistent with the existing accumulation direction and create a false signal, or be unrelated to the existing accumulation direction and dilute the statistics. Both of these make the determination in Example 4 meaningless.

[0090] Simultaneously, an observation interruption flag and its duration are output. This duration is pushed along with the cumulative number of observations, allowing the management end to distinguish between two situations: sufficient observations have been accumulated but no unidirectional trend has been observed, and insufficient observations themselves; the latter can be directly determined by the minimum cumulative number of observations given in Example 4.

[0091] Online isotropic verification is performed as follows: A sliding window of predetermined length is taken, and the magnitude of the arithmetic mean vector of the incremental direction vectors within the window is calculated; when the magnitude exceeds the quotient obtained by dividing the tolerance coefficient by the square root of the window length, a criterion premise failure flag is output.

[0092] Neither the identifier nor the aforementioned modulus is used to adjust the comparison threshold. At the implementation level, the threshold calculation function's input parameters are only the cumulative count, vector dimension, and significance level; its function signature does not contain the modulus or the identifier. The modulus and the identifier are written into a state field independent of the threshold calculation and are read separately by the management terminal. This isolation... Figure 3 This manifests as follows: the output of the effectiveness processing module 6 is not connected to the threshold calculation unit 51. The reason for not feeding it back to the threshold is that if the threshold is adjusted by this amount, the threshold will again contain parameters derived from the observation data of the measured object, and the threshold established in Example 4, which does not contain any parameters estimated from the measured object, will no longer be valid.

[0093] The relationship between residual anisotropy and detection capability is as follows. Assume that after training, a residual average direction bias still exists in the normal incremental direction; simultaneously, assume that the behavioral drift to be detected contributes an average direction in the direction. At this point, the average direction of each incremental direction vector is the vector sum of the two. From the same expansion in Example 3, we get: the expected value of the square of the risk statistic is equal to the cumulative number, plus the cumulative number multiplied by the cumulative number minus one, and then multiplied by the square of the magnitude of the vector sum of the two average directions. That is, the expected value of residual anisotropy and true drift entering the risk statistic in exactly the same form. Based on this, the applicable boundary of this scheme and the acceptance criterion for training are obtained: to prevent true drift from being overwhelmed by residual anisotropy, the residual anisotropy must be much smaller than the minimum drift intensity to be detected. The sensitivity criterion in Example 5 is determined accordingly.

[0094] The above relationship is a theoretical conclusion directly derived from the expansion of Example 3. The extent to which residual anisotropy can be suppressed in real-world research and study behavior data must be determined through training on real data and hold-out set testing; this specification does not assert its numerical value.

[0095] Example 7: Implementation of Modulus Bypass, Multi-Scale Accumulation Window and System The average of the whitened vector magnitudes saved in Example 3 within the sliding window is used as an auxiliary estimate of the rate of behavioral change. The criteria for this estimation are given as follows: Before the start of the trip, the empirical distribution of the whitened vector magnitudes of the training sample set is statistically analyzed, and two quantiles, such as the 50th and 90th quantiles, are fixed together with the frozen configuration; during the trip, when the average magnitude within the sliding window does not exceed the 50th quantile, the prompt wording for that level is marked as slow accumulation; when it exceeds the 90th quantile, it is marked as a large change; when it is in between, no marking is made. This marking only applies to the prompt wording and does not change the risk level itself determined by the risk statistics and thresholds.

[0096] The bypass variable does not participate in the calculation of risk statistics, nor in the determination of thresholds; at the implementation level, the bypass variable and the main criterion are stored separately in the data structure, and the calculation link of the main criterion does not read the field where the bypass variable is located. This separation is... Figure 3 This is represented by the modulus bypass unit 33. The reason is that the modulus contains the fluctuation amplitude information of the tested object itself. If it enters the main criterion or threshold, it will destroy the property established in Example 4. It should be noted that the above two quantiles are taken from the training sample set rather than the observations of the participants in the research and study, and their effect is limited to wording annotation and does not enter the level determination and threshold calculation. Therefore, the property that the threshold established in Example 4 does not contain any parameters estimated from the tested object is not changed by this annotation.

[0097] For the same incremental direction vector sequence, several cumulative windows of different lengths can be maintained in parallel. Each window independently maintains its own composite vector and risk statistics, and compares them with the thresholds for the corresponding number of windows. The highest level obtained from each window is taken as the output level, and the window length that triggers that level is also marked.

[0098] The length of each window is not arbitrarily chosen, but determined by the inverse equation of the drift intensity to be covered. Taking the inverse of the minimum detectable drift intensity relationship given in Example 4 with respect to the cumulative number, we get: the cumulative number equals one, plus the chi-square distribution with degrees of freedom equal to the vector dimension minus the significance level quantile divided by the vector dimension, the quotient minus one, and then divided by the square of the drift intensity to be detected. Therefore, to cover a certain drift intensity interval, we only need to take several drift intensity values ​​within that interval, substitute them into the above equation to calculate their respective cumulative numbers, which is the length of each parallel window. The input to this inverse equation remains only the drift intensity to be detected, the vector dimension, and the significance level; no new parameters to be estimated are introduced.

[0099] Example of theoretical calculation: Taking a vector dimension of 32, a preset significance level of 1%, and a total number N of 672, the quotient obtained from Example 4 is approximately 2.40. Therefore, subtracting one from the quotient results in approximately 1.40. To detect a drift intensity of 0.5, the window length is approximately 7; to detect 0.2, it is approximately 36; and to detect 0.1, it is approximately 141. All the above values ​​are theoretical calculations obtained using this formula and the known chi-square quantile table.

[0100] like Figure 3 As shown, when the above method is implemented in the system, the correspondence between each module and each link is as follows.

[0101] The observation normalization module 1, which implements Example 1, takes the behavioral observations from each source as input and outputs a sequence of behavioral states with validity labels, providing a state vector aligned to a unified time base to the incremental embedding module 2.

[0102] The incremental embedding module 2 implements the second embodiment. It contains an adjacent time base point difference unit 21 and a causal time series network unit 22 with frozen parameters, and outputs the incremental embedding vector to the direction shaping module 3. This module verifies the configuration summary value at startup.

[0103] The orientation shaping module 3 implements the orientation shaping part of Embodiment 3. It contains a whitening unit 31 and a unitization unit 32, outputs an incremental orientation vector to the orientation accumulation module 4, and sends the module length to the module length bypass unit 33.

[0104] The direction accumulation module 4 implements the direction accumulation part of Embodiment 3 and the multi-scale accumulation window of this embodiment, maintains the composite vector and accumulation number of one or more accumulation windows, and outputs the statistics of each window and the corresponding accumulation number to the risk judgment module 5.

[0105] Risk assessment module 5, which implements embodiment four, includes a threshold calculation unit 51 and a grading unit 52. The inputs of the threshold calculation unit 51 are only the cumulative number, vector dimension and significance level, and the outputs risk quantity and graded risk warnings and pushes them to the management terminal 8.

[0106] The validity handling module 6, implementing Embodiment 6, freezes the accumulation and outputs an interruption flag when the observation is invalid, and at the same time independently performs online isotropic verification and outputs a criterion premise failure flag; the output of this module is not connected to the threshold calculation unit 51.

[0107] The read-only configuration storage module 7 stores the frozen network parameters, whitening matrix, and threshold calculation parameters. It provides configuration unidirectionally to the incremental embedding module 2, the orientation shaping module 3, and the risk judgment module 5, and does not accept writes from the above modules.

[0108] The data flow between modules is unidirectional, passing sequentially from the observation normalization module 1 to the risk assessment module 5. The isotropic constraint applied by the orientation shaping module 3 is a prerequisite for the null hypothesis distribution upon which the risk assessment module 5 is based; the modulus discarded by the orientation shaping module 3 is a prerequisite for the threshold of the risk assessment module 5 to not contain the parameters of the measured object.

[0109] The above-described method and system can be implemented by an electronic device, which includes a memory and a processor. The memory stores a computer program and a read-only configuration file, and the processor executes the computer program to implement the above-described method. Alternatively, the computer program can be stored in a computer-readable storage medium, and the processor reads and executes it to implement the above-described method. The electronic device and the storage medium serve as the carriers for implementing the above-described method and system.

[0110] Example 8: Overall Flow of Risk Assessment Method like Figure 2 As shown, the temporal neural network risk inference method for overseas study tour behavior sequences provided by the present invention includes the following steps.

[0111] Step S1, Observation Acquisition and State Construction. Collect multi-source behavioral observations of the study participants and align them according to a unified time base. Construct a multi-dimensional behavioral state vector for each valid time base point to obtain a behavioral state sequence. Determine whether the observation is valid for each time base point. If it is determined to be invalid, pause accumulation, do not perform interpolation filling, and set an interruption flag as in Example 6.

[0112] Step S2, behavior increment extraction. The difference between adjacent time base points is taken from the behavior state sequence to obtain the behavior increment sequence.

[0113] Step S3, Incremental Embedding. The incremental sequence of behavior is input into a causal temporal network whose parameters have been determined and frozen before the start of the trip, to obtain an incremental embedding vector; the network parameters are not updated with the observation data of the participants during the trip.

[0114] Step S4, Orientation Shaping. The incremental embedding vector is sequentially whitened and normalized to obtain the incremental orientation vector, and its magnitude is discarded.

[0115] Step S5, direction accumulation. The incremental direction vectors are added hourly according to the time base order to obtain the cumulative composite vector, and its magnitude is used as the risk statistic. This risk statistic is updated synchronously with the observation of each effective time base point.

[0116] Step S6, Risk Assessment and Alert. The null hypothesis distribution is analytically determined using the number of vectors involved in the addition and the vector dimension. The comparison threshold corresponding to its quantiles is calculated. The risk statistic is compared with this threshold to obtain the risk level, and a graded risk alert is generated and pushed to the management terminal. Then, the process moves to the next time base point and repeats steps S1 to S6.

[0117] Example 9: Theoretical Analysis of the Mechanism of Action To facilitate understanding of the mechanism of action of this invention and its differences from existing methods, a theoretical analysis of this invention is conducted below under idealized assumptions. The following analysis is based on idealized assumptions made for the sake of theoretical explanation and aims to elucidate the concept and mechanism of action of this invention. It does not constitute a limitation on the scope of protection of this invention, nor does it represent the precise quantitative performance of actual systems.

[0118] The idealized assumptions are as follows: Assumption 1: The behavioral state of the observed individual can be represented as a vector in the state space, which under normal circumstances is the superposition of a fixed individual offset and a zero-mean random fluctuation. Assumption 2: The behavioral evolution to be detected is manifested as a continuous unidirectional displacement along a fixed direction in the state space, with the single-step displacement being less than the typical amplitude of the individual's random fluctuation. Assumption 3: There is no point-by-point annotation during the observation process, meaning that no external annotation regarding whether the current state is abnormal can be obtained during the inference phase. Of the above three assumptions, Assumptions 1 and 2 are neutral descriptions of the research object and do not contain any presuppositions about which approach is better; Assumption 3 is an objective constraint of the cross-border unmanned close monitoring scenario.

[0119] Corollary 1: The practice of generating a baseline based on the subject's own history will inevitably fail under Hypothesis 2. This approach is widely used in the field. The various existing technologies cited in the background section, through different forms such as attention-weighted historical trajectory features, baseline behavior patterns fitted by reinforcement learning, model outputs after windowing historical behavior features, and reference intervals established based on individual historical monitoring records, constitute the same paradigm. That is, they all use the quantity generated by the subject's own history as the judgment benchmark, and then use the deviation of the current state from that benchmark as a risk measure. Therefore, the following corollary does not address the implementation defects of a specific solution, but rather the common property of this paradigm under Hypothesis 2.

[0120] Assume this paradigm uses the weighted average of the measured object's own history as the judgment baseline, with the sum of weights at each time point being one and independent of the evolution to be detected. Under Hypothesis Two, the current state is expressed as the sum of the initial state, the product of the single-step displacement and the time index, and the zero-mean random fluctuation. Substituting these into the baseline definition, we get: the baseline equals the initial state, plus the sum of the single-step displacement multiplied by the product of the weights at each time point and their time indices, plus the weighted random term. Therefore, the deviation on which the criterion is based equals the single-step displacement multiplied by the difference between the current time index and the aforementioned weighted sum, plus the random term. The difference between the current time index and the weighted sum is a constant determined by the weight distribution and does not increase with the continuation of observation. Therefore, the expected value of the deviation does not increase with the continuation of observation; the criterion lacks time-cumulative property, and its detection capability does not improve with the continuation of observation. This is inevitable, not possible.

[0121] Further, two of the most obvious circumvention methods are blocked: First, shortening the baseline window would make the weighted sum closer to the current time sequence, resulting in a smaller constant and thus a smaller deviation; second, extending the baseline window would increase the constant, but the baseline's adaptation to individual differences would fail, and the inherent bias between individuals would re-enter the deviation range, leading to an increase in false alarms. Neither approach solves the problem. The difference between this failure scenario and the present invention can be compared... Figure 4 and Figure 7 understand.

[0122] Corollary 2: Under the constraints of Assumption 3, all known substitutions fail one by one, except for this scheme. The key constraints examined in this corollary are: no point-by-point annotation, and the decision threshold must not contain any parameters estimated from the measured object.

[0123] Alternative option 1: Freeze the initial baseline. The baseline is still taken from the observations of the object under test in the initial stage, and the thresholds still include the parameters of the object under test; moreover, the cross-border journey itself is a complete change in the environment, and the baseline established before the journey is not consistent with the normal state during the journey.

[0124] Alternative two uses the group of people traveling together as the benchmark. Travelers share the same itinerary, and the changes caused by the itinerary itself constitute the group common mode. In the calculation of the difference between the common mode and the group mean, the common mode components are completely canceled out, so the situation of the entire group evolving in the same direction cannot be detected.

[0125] Alternative three is the cumulative sum control chart. This type of method requires pre-defined target values ​​and standard deviations; without point-by-point annotation, both can only be taken from historical observations of the measured object, and the thresholds again contain parameters of the measured object.

[0126] Alternative four: the difference between fast and slow baselines. The two baselines are estimated from the same data stream and are both affected by evolution; under the condition of constant single-step displacement, the difference between the two tends to be constant and does not increase with the continuation of observation.

[0127] Alternative five is the concept drift detector. This type of method uses distribution changes as a signal to trigger model updates, which is semantically opposite to that of this invention; and it also cannot distinguish between normal individual evolution and the evolution to be detected without point-by-point annotation.

[0128] The sixth approach retains the full-sequence attention model. Its criterion remains the distance between the current state and the historical summary, and the judgment benchmark is still generated from the history of the object being tested.

[0129] Alternative seven: Manually set absolute thresholds. Assuming that the fixed individual offset from assumption one is directly incorporated into the criterion, and inter-individual differences are considered risk, see [link to relevant documentation]. Figure 4 Column A.

[0130] This solution differs: In Example 2, the incremental shift of individuals is algebraically eliminated; in Example 3, normalization and discarding the modulus eliminate individual fluctuations algebraically; and in Example 5, the isotropic constraint during the training phase shapes the directional distribution of normal increments into a known uniform distribution. Thus, the null hypothesis distribution in Example 4 is determined solely by the cumulative number of samples and the vector dimension. At this point, the threshold input consists only of the cumulative number of samples, the vector dimension, and the significance level, none of which are derived from the tested object. Therefore, under the aforementioned key constraints, this solution holds true, while all the aforementioned substitutions do not. This invention is not a simple replacement of known methods, but rather essential for achieving the goal under these constraints.

[0131] Second inference: Several means employed in this invention are not unprecedented in the art; however, the effects of these means applied by this invention differ from those known in the art. To facilitate understanding of the essence of this invention, they are described below.

[0132] Firstly, regarding the role of isotropic constraints. There are existing practices in the art of imposing hyperspherical homogeneity or isotropic constraints on the embeddings of neural networks during training. The disclosed effects are to prevent representation collapse, enhance the separability between categories, and improve the performance of downstream tasks. The objects of these constraints are typically the centers of different categories or the relative dispersion between different samples. The purpose of this invention is not this, but rather to ensure that the null hypothesis distribution of a specific statistical test holds, thereby allowing the decision threshold to deviate from the tested data; the object of the constraint is the distribution of the incremental direction vector of the same normal individual on a unit sphere itself. The problems solved and the roles played by these two methods are different.

[0133] Secondly, regarding the stage of hypothesis testing. There are existing practices in the field that incorporate hypothesis testing into the training loss function. This testing occurs within the training loop, penalizing the model output to ensure its distribution conforms to expectations. Once training is complete, the testing ceases to exist; it is not performed during the inference phase, nor is it used to determine thresholds. This invention is the opposite: the training phase applies uniformity constraints, not testing. Testing occurs during the inference phase, used to determine the object being tested. Furthermore, the validity of this testing is a direct result of the shaping during the training phase. The two methods involve testing at opposite stages.

[0134] Thirdly, regarding the source of the judgment threshold. There are existing practices in the field that regularize features to a hypersphere and then use a threshold to determine anomalies during the inference stage. This threshold is either manually set, determined by empirical distributions on the samples, or generated from a portion of the training data. The comparison threshold of this invention is analytically calculated from the cumulative number of data points, vector dimension, significance level, and total number of time-based reference points, and is not derived from any observed data. The sources of the two thresholds differ.

[0135] Fourth, regarding whether or not calibration data is relied upon. Existing approaches in the field pursue distribution-independent efficient decision thresholds, such as using inconsistent scores combined with quantiles to provide a prediction set. While these approaches do not assume the data follows a specific distribution, their quantiles must still be derived from calibration data, thus requiring a sample with the same distribution as the tested object. This invention does not use any calibration data; the effectiveness of its threshold is guaranteed by the construction during the training phase. The two differ in whether or not they rely on a calibration set.

[0136] Inference 3: Parameter Range, Detectability Boundary, and Technical Effectiveness. Firstly, distinguishability monotonically increases with the duration of observation. From the expected value formula in Example 3, the expected value of the squared risk statistic under Hypothesis 2 is greater than under normal circumstances by an additional term: cumulative number multiplied by cumulative number minus one, then multiplied by the squared drift intensity. This term increases with the square of the cumulative number, while under normal circumstances, the statistic itself increases with the first power of the cumulative number. Therefore, distinguishability monotonically increases with the square root of the cumulative number. This indicates that the detection capability of this scheme monotonically increases with the duration of observation, and there is no upper limit to the pre-set time window length. See [link to related information]. Figure 7 The separation process of curve 62 and curve 63.

[0137] Secondly, the detectable boundary is analytically calculable. As shown in Example 4, the minimum detectable drift intensity is analytically determined by the cumulative number of observations, the vector dimension, and the significance level. The minimum cumulative number of observations is also analytically determined by these three factors. This allows the deployer to know, before the deployment begins, how many observations are needed to detect a given evolution at a given observation density and significance level. This property is not present in the paradigm that generates the baseline based on the history of the object being measured. From Corollary 1, the expected deviation of this paradigm does not increase with continued observation, and its detectable boundary cannot be given without parameters of the object being measured.

[0138] Thirdly, the selection of dimension. As shown in Example 4, under the ideal condition where isotropy is strictly true, the minimum detectable drift intensity decreases as the vector dimension increases; that is, the higher the dimension, the more sensitive the signal. However, as shown in Example 6, residual anisotropy and drift enter the statistic in the same form, and residual anisotropy increases with increasing dimension. Therefore, the appropriate dimension is determined by the minimum ratio between the two, and the specific method for determining this is shown in the parameter table listed in Example 5.

[0139] Fourth, it provides a basis for the limitation of the claims. The relationship given in Example 6, that the residual anisotropy must be much smaller than the minimum detectable drift intensity, provides a quantifiable source for the limitation in the claims that the isotropic constraint makes the directional distribution approximate a uniform distribution; the analytical expressions of the threshold and detectable boundary given in Example 4 provide a verifiable basis for the limitation in the claims that the null hypothesis distribution is determined analytically only by the cumulative number and dimension, and the threshold does not contain any parameters estimated from the measured object.

[0140] Fifth, the nature of the technical effects. The effects mentioned above are not simply the sum of the effects of each step: taking the increment alone only eliminates the absolute level of an individual without producing any cumulative effect; taking the unitization alone only eliminates the amplitude but leaves the directional distribution unknown and the threshold undeterminable; taking the isotropic constraint alone, without any decision-making ability, produces no result without the cumulative step. Only when these three are sequentially connected can the threshold input converge to the cumulative number, dimension, and significance level, and only then can the detectable boundary become an analytically computable quantity.

[0141] The above theoretical analysis demonstrates the mechanism of action of this invention. To further verify the above effects, the following real-world control experiments should be conducted during implementation: Training with isotropic constraints should be completed on real-world research and study data; the actual values ​​of residual anisotropy should be measured on the reserved normal samples and compared with the detectable minimum drift intensity calculated in Example 4; simultaneously, using real or controlled-injection gradient samples, the detection results of this method and the method using the historical baseline generated by the tested object should be measured on the same data. The original records of the above experiments should be retained. Before the completion of the above experiments, this specification makes no numerical assertions regarding the detection rate, false alarm rate, and detection lead of this method on a real system.

[0142] Other implementation methods and variations The embodiments described above are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. For example, the types and components of the multi-source behavioral observations can be increased, decreased, or adjusted according to the actual available observation sources, and are not limited to the five types of components listed in this specification; the causal temporal network can employ other network structures that preserve temporal order information, and is not limited to one-dimensional causal convolutional networks, gated recurrent networks, and causal self-attention networks; the isotropic constraint terms can adopt other forms that can measure the uniformity of distribution on a unit sphere, and are not limited to the synthetic modulus expectation approximation type and the hyperspherical uniformity energy type; the number of risk levels in the graded risk warning can be more than three levels to provide finer divisions between different significance levels; the number and length of the parallel accumulation windows can be increased or decreased according to the range of drift intensity of interest; the method and system can be implemented by an industrial control computer and its computational program, or by an embedded controller or dedicated computational circuit. All the aforementioned modifications and improvements should be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A temporal neural network risk inference method for overseas study tour behavior sequences, characterized in that, include: Multi-source behavioral observations of study participants were collected and aligned with a unified time base. A multi-dimensional behavioral state vector was constructed for each effective time base point to obtain a behavioral state sequence. The difference between adjacent time base points is taken from the behavioral state sequence to obtain the behavioral increment sequence; The incremental sequence of behaviors is input into a temporal neural network to obtain an incremental embedding vector; the temporal neural network retains the temporal order information and does not use summation-type pooling operations that are insensitive to the input order in the encoding; the network parameters of the temporal neural network are determined and frozen before the start of the study tour and are not updated with the observation data of the study tour participants during the study tour. The incremental embedding vector is then subjected to whitening and normalization processes in sequence to obtain the incremental direction vector, and the magnitude of the incremental embedding vector is discarded. The incremental direction vectors are added hourly in time base order to obtain the cumulative composite vector. The magnitude of the cumulative composite vector is used as the risk statistic. The risk statistic is updated synchronously with the observation of each effective time base point. The risk statistic is compared with the quantile of the null hypothesis distribution to obtain the risk quantity, and a graded risk warning is generated accordingly. The null hypothesis distribution is determined analytically only by the number of incremental direction vectors that have been added and the dimension of the incremental direction vectors, and is not obtained by calibration from the observation data of the tested object or other natural persons; the significance level corresponding to the quantile is obtained by dividing the preset significance level by the total number of time base points in the study tour, and the total number is determined by the tour length and time base length before the start of the study tour; The training objective of the temporal neural network includes an isotropic constraint term, which makes the distribution of the incremental direction vector of normal samples on a unit sphere approximate a uniform distribution, so that the null hypothesis distribution holds. After training, the temporal neural network meets the following acceptance criteria: on a set of retained samples that does not overlap with the training sample set, the difference between the empirical quantile of the test statistic calculated from the incremental direction vector of the retained sample set in the same way as the risk statistic and the corresponding theoretical quantile of the null hypothesis distribution is not greater than a preset tolerance.

2. The method according to claim 1, characterized in that: The temporal neural network is any one of a one-dimensional causal convolutional network, a gated recurrent network, or a causal self-attention network with positional encoding; the whitening matrix and mean vector used in the whitening process are calculated and frozen once by the incremental embedding vector of the training sample set before the start of the study tour, and are not updated during the study tour.

3. The method according to claim 1, characterized in that, The isotropic constraint term is one of the following: the expectation of the square of the magnitude of the cumulative composite vector of the incremental direction vector of the training sample set with a given number of vectors approximates the theoretical expectation of the square of the magnitude when the directions are independent and uniformly distributed under the same dimension. Alternatively, the training objective minimizes the uniform energy of the incremental direction vectors of the training sample set on a unit sphere; and the training objective further includes a decorrelation constraint term, which makes the arithmetic mean of the inner product between the incremental direction vectors at different time base points in the training sample set approach zero.

4. The method according to claim 1, characterized in that: The null hypothesis distribution is a distribution followed by the magnitude of the sum of independent, uniformly random unit vectors whose dimension is equal to the dimension of the incremental direction vector and whose number is equal to the number of incremental direction vectors already added. The comparison threshold corresponding to the quantile is equal to the number of incremental direction vectors already added, multiplied by the quantile of the chi-square distribution with degrees of freedom equal to the dimension at one minus the significance level, and then divided by the dimension, and the square root of the quotient. When the number of incremental direction vectors already added is less than a set lower bound, the comparison threshold is given by a pre-prepared quantile table. The quantile table is obtained by generating a given number of direction vectors independently and uniformly on a unit sphere of a given dimension using a pseudo-random number generator and repeatedly calculating the empirical quantile of the magnitude of their composite vector. Its preparation does not use any observation data of the tested object or other natural persons. The graded risk warning is divided into risk levels according to the comparison thresholds corresponding to at least two different preset significance levels.

5. The method according to claim 1, characterized in that, Also includes: For each time base point, determine whether the observation is valid; for time base points that are determined to be invalid, do not construct a behavior state vector, do not generate an incremental direction vector, do not count the number of incremental direction vectors that have been added, and do not perform interpolation to complete the observation for that time base point; for time base points that are determined to be invalid, keep the cumulative composite vector and the risk statistic unchanged, and output the observation interruption flag and its duration.

6. The method according to claim 1, characterized in that, Also includes: During the study tour, the uniformity of the incremental direction vector within the sliding window is tested on a unit sphere to obtain the isotropic deviation. When the isotropic deviation exceeds the preset tolerance, a criterion prerequisite failure flag is output. Neither the isotropic deviation nor the criterion premise failure flag is used to adjust the comparison threshold.

7. The method according to claim 1, characterized in that, Also includes: The magnitude of the incremental embedding vector is retained as a bypass quantity. Based on the quantile of the mean of the bypass quantity within the sliding window relative to the empirical distribution of the magnitude of the training sample set, auxiliary annotations are made for the prompts of the risk level determined by the risk statistics. The bypass quantity does not participate in the calculation of the risk statistics, nor in the determination of the comparison threshold, nor does it change the determined risk level.

8. The method according to claim 1, characterized in that, Also includes: At least two cumulative windows of different lengths are set in parallel for the same incremental direction vector sequence to obtain their respective risk statistics, and each is compared with the null hypothesis distribution under the corresponding number of windows. The highest risk level obtained from each window is taken as the output. The length of each cumulative window is determined analytically by the drift intensity to be detected, the dimension of the incremental direction vector, and the significance level.

9. The method according to claim 1, characterized in that: The multidimensional behavior state vector does not contain deterministic quantities that evolve according to a fixed period; the deterministic quantities that evolve according to a fixed period are input into the temporal neural network as conditional quantities along with the behavior increment.

10. The method according to claim 1, characterized in that: Each incremental direction vector participating in the same cumulative synthesis vector is obtained from a non-overlapping encoding window.

11. A temporal neural network risk inference system for overseas study tour behavior sequences, characterized in that, include: The observation normalization module is used to collect multi-source behavioral observations of study participants, align them according to a unified time base, construct a multi-dimensional behavioral state vector for each valid time base point, and output a behavioral state sequence with validity labels. The incremental embedding module is used to obtain the behavioral incremental sequence by taking the difference between adjacent time base points of the behavioral state sequence, and output the incremental embedding vector through the temporal neural network; the temporal neural network retains the time base order information, and its network parameters are determined and frozen before the start of the study tour and are not updated with the observation data of the study tour participants during the study tour. The orientation shaping module is used to perform whitening and normalization processing on the incremental embedding vector in sequence, output the incremental orientation vector and discard its magnitude, and send the discarded magnitude to the magnitude bypass unit. The direction accumulation module is used to add the incremental direction vectors hourly in time base order, output the accumulated composite vector, and use its magnitude as the risk statistic that is synchronously refreshed with each effective time base point. The risk assessment module is used to compare the risk statistics with the quantiles of the null hypothesis distribution and output the risk level and risk classification warning. The null hypothesis distribution is determined analytically solely by the number of incremental direction vectors that have been added and the dimension of the incremental direction vectors, and is not obtained by calibration from the observation data of the measured object or other natural persons. The validity handling module is used to keep the cumulative composite vector and the risk statistic unchanged and output the observation interruption flag when the observation is invalid, and to independently perform online uniformity test and output the criterion premise failure flag. The output of the validity handling module is not connected to the threshold calculation unit of the risk judgment module. The read-only configuration storage module is used to store frozen network parameters, whitening matrix and threshold calculation parameters, and provides configuration unidirectionally to the incremental embedding module, the orientation shaping module and the risk determination module, and does not accept writes from the above modules; The management terminal is used to receive the graded risk warning, the observation interruption identifier, and the criterion premise failure identifier; The training objective of the temporal neural network includes an isotropic constraint term that makes the distribution of the incremental direction vector of normal samples on a unit sphere approximate a uniform distribution, and the temporal neural network satisfies the acceptance condition described in claim 1 after training.

12. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method of any one of claims 1 to 10.

Images