LSTM-FCN-based train dispatcher behavior anomaly detection system and dynamic intervention method thereof

Abstract

The invention discloses a train dispatcher behavior anomaly detection system based on LSTM-FCN and a dynamic intervention method thereof, and relates to the technical field of intelligent monitoring. Scheduling operation time sequence data, environment state data and physiological monitoring data are collected in real time, a space-time joint feature vector is constructed, and a train dispatcher behavior anomaly detection result is obtained. The LSTM-FCN hybrid network is used to extract time sequence dependence features and spatial pattern features, and an attention weight matrix generated by physiological data and a compensation factor matrix generated by environmental parameters are combined to dynamically calculate a comprehensive abnormal probability value; and when the probability value exceeds a self-adaptive adjustment threshold value, triggering a hierarchical intervention mechanism based on the driving scheduling knowledge graph to realize progressive intervention from interface prompting to manual takeover. By fusing multi-modal data and self-adaptive dynamic compensation, the problem of poor adaptability to complex scenes due to adoption of a static model in traditional train monitoring is solved, and the accuracy and intervention timeliness of abnormal behavior detection of the train dispatcher are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a train dispatcher behavior anomaly detection system based on LSTM-FCN and a dynamic intervention method thereof. Background Art

[0002] Traditional methods for monitoring train dispatcher behavior rely primarily on manual experience or rule-based models based on a single sequence of operations, making it difficult to effectively capture dynamic behavioral characteristics in complex environments. Existing technologies typically only analyze the temporal logic of operational instructions, ignoring the potential impact of the dispatcher's physiological state and the train's environmental factors on dispatching decisions, resulting in a high rate of misjudgment. For example, in severe weather or high-load scenarios, a dispatcher's stress response can easily be misidentified as abnormal operation, while operational delays in a fatigued state may not be promptly warned due to a lack of physiological data support. Existing behavioral monitoring is mainly based on the static matching of standard habits with current behavior, making it difficult to identify the following high-risk scenarios: behavioral rhythm distortion due to excessive psychological load; "inertial deviation" caused by continuous errors in a fatigued state that have not yet broken the rule line; and "gray area operations" that are difficult for standard models to explain.

[0003] Furthermore, traditional methods use fixed thresholds to identify anomalies, often pre-set based on the statistical distribution of historical data. These thresholds are unable to respond in real time to the impact of unexpected scenarios (such as extreme weather and equipment failures) on dispatcher behavior. For example, during heavy rainfall, dispatchers' attention allocation patterns differ significantly from those in more conventional scenarios. However, traditional static thresholds cannot dynamically adjust their sensitivity based on environmental parameters such as visibility and wind speed, resulting in frequent missed or false alarms. This problem is particularly prominent in cross-regional dispatching scenarios, where differences in environmental characteristics and operating procedures across different railway sections further complicate threshold setting.

[0004] In response to the above problems, this field urgently needs a train dispatcher behavior monitoring and decision support method that can integrate multi-source heterogeneous data, dynamically adapt to environmental changes and physiological monitoring changes of train dispatchers, and achieve precise intervention. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a train dispatcher behavior dynamic intervention method based on LSTM-FCN.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention discloses a method for dynamic intervention of train dispatcher behavior based on LSTM-FCN, comprising the following steps:

[0008] Acquire the dispatching operation time sequence data and environmental status data of the target train in real time, and simultaneously collect the physiological monitoring data of the target train dispatcher;

[0009] The scheduling operation time series data is converted into a spatiotemporal joint feature vector, and the spatiotemporal joint feature vector, the physiological monitoring data and the environmental state data are input into the LSTM-FCN hybrid network to obtain a comprehensive abnormality probability value ;

[0010] Among them, the comprehensive abnormal probability value Combined with the attention weight matrix Combined compensation factor matrix Calculated, the attention weight matrix and the compensation factor matrix The dynamic fusion layer of the LSTM-FCN hybrid network generates the physiological monitoring data and the environmental status data accordingly;

[0011] When the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the traffic scheduling knowledge graph is triggered and the deviation value is calculated. The degree of deviation, and according to the deviation value The numerical range of the intervention level is divided into intervention levels, and the corresponding graded alarm instructions are triggered according to the intervention level;

[0012] After triggering the graded alarm instruction, collecting the physiological state coding vector of the target train dispatcher in real time;

[0013] If the physiological state coding vector returns to the preset safety range within M consecutive minutes, the graded alarm instruction is released and a status recovery report is generated; otherwise, the current graded alarm instruction is upgraded to the highest level alarm instruction, triggering a manual takeover request.

[0014] In a second aspect, the present invention discloses a train dispatcher behavior anomaly detection system based on LSTM-FCN, which uses the above-mentioned train dispatcher behavior dynamic intervention method based on LSTM-FCN, including:

[0015] A data acquisition module is used to acquire real-time dispatching operation time series data, environmental status data, and synchronously collect physiological monitoring data of the train dispatcher;

[0016] a feature conversion module configured to convert the scheduling operation time series data into a spatiotemporal joint feature vector;

[0017] LSTM-FCN processing module, configured to extract and fuse the spatiotemporal features of the spatiotemporal joint feature vector, physiological monitoring data and environmental status data and calculate the comprehensive abnormality probability value ;

[0018] Intervention decision module, used when the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the knowledge graph is triggered and the deviation value is calculated The degree of deviation, and according to the deviation value The numerical range of the intervention level is divided into the levels to trigger the graded alarm instructions;

[0019] A physiological status monitoring module, configured to collect the physiological status coding vector of the train dispatcher in real time after the hierarchical alarm instruction is triggered;

[0020] The scheduling decision module cancels the alarm and generates a status recovery report if the vector returns to the preset safe range within M consecutive minutes; otherwise, the current alarm instruction is upgraded to the highest alarm instruction, triggering a manual takeover request.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. This invention significantly improves the accuracy of abnormal behavior detection and the timeliness of intervention for train dispatchers through multimodal data fusion and dynamic compensation mechanisms. Based on the construction of a spatiotemporal joint feature vector, it integrates the attention weight matrix driven by physiological monitoring data with the compensation factor matrix generated by environmental state data, breaking through the limitations of traditional single threshold models. Through dual-path feature extraction using an LSTM-FCN hybrid network and deep coupling of spatiotemporal joint feature vectors, the abnormal probability value can simultaneously capture reaction delays caused by physiological fatigue and operational deviations caused by environmental interference.

[0023] 2. This invention utilizes a hierarchical intervention mechanism that dynamically adjusts thresholds and quantifies deviations, achieving precise matching and flexible control of risk management. By calculating the relative deviation Δ and combining it with preset threshold intervals, the system triggers differentiated response strategies based on the severity of the anomaly probability. This deeply integrates with train dispatching, improving the matching accuracy of emergency response plans, shortening response delays for high-risk operations, and significantly enhancing the proactive safety protection capabilities of the train dispatching system.

[0024] 3. By continuously collecting physiological status encoding vectors after an alarm, the system can evaluate the effectiveness of intervention measures in real time. When the dispatcher's physiological status, such as heart rate variability, returns to a safe range within M minutes, the alarm is automatically released and a traceability report is generated. Otherwise, the alarm is upgraded to the highest level, triggering the mandatory activation of the emergency plan. This flexible control mechanism based on biometric feedback increases the false alarm cancellation rate and reduces the frequency of invalid alarms caused by temporary stress reactions of personnel, maximizing the continuity of dispatch operations while ensuring driving safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0026] Figure 1 Schematic diagram of the method steps of the present invention;

[0027] Figure 2 A flow chart of the determination steps of the method of the present invention;

[0028] Figure 3 Schematic diagram of the system flow of the present invention;

[0029] Figure 4 It is a system process step diagram of the present invention;

[0030] Figure 5 This is a connection diagram of the abnormal case tracing module and the intervention effect evaluation module of the present invention. DETAILED DESCRIPTION

[0031] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0032] Application Overview:

[0033] In existing technologies, train dispatcher behavior monitoring often relies on single-step operation timing analysis or static physiological indicator threshold judgments, making it difficult to adapt to the coupled impact of dynamic environments and complex operating conditions. In extreme weather or high-load scenarios, traditional methods are unable to distinguish between operational delays caused by physiological fatigue and decision-making biases caused by environmental interference, resulting in a high false alarm rate. Existing systems typically use fixed threshold trigger intervention mechanisms, lacking the ability to dynamically compensate for real-time environmental parameters and physiological status. This makes it difficult to accurately assess the abnormal risk level in the event of sudden equipment failures or cross-section scheduling, which can easily lead to excessive intervention or delayed response.

[0034] To address the above issues, the inventors discovered that the coupling effect between environmental conditions and physiological monitoring data significantly affects the credibility of anomaly probability values. By constructing a dynamic fusion compensation model, they achieved collaborative optimization of multi-source heterogeneous data. During the research process, they found that the spatiotemporal characteristics of LSTM-FCN are highly robust to changes in environmental parameters and physiological monitoring data. Therefore, they proposed a comprehensive anomaly probability calculation method based on cross-modal dynamic fusion. Further verification through actual scheduling scenarios, the attention weight matrix generated by physiological data and the compensation factor matrix driven by environmental parameters were introduced into the probability calculation, forming an anomaly assessment model with dynamic correction capabilities.

[0035] Specifically, first, the scheduling operation time series data, environmental status data and physiological monitoring data are collected synchronously. Through the spatiotemporal joint feature vector construction module, the operation instruction sequence is mapped into a semantic vector, and the spatiotemporal correlation features are generated in combination with the station thermal distribution matrix. The LSTM-FCN hybrid network extracts the time series dependency features and spatial pattern features respectively. The dynamic fusion layer strengthens the key node features related to physiological abnormalities through the attention weight matrix, and uses the compensation factor matrix to remove the interference of environmental noise on the feature vector. When the comprehensive abnormality probability value exceeds the dynamic adjustment threshold T, the system is based on the deviation value. The deviation is quantified and graded to trigger the intervention mechanism: push operation prompts in case of mild deviation, lock high-risk instructions in case of moderate deviation, and request manual takeover and link the knowledge graph to generate emergency response plans in case of severe deviation.

[0036] Compared with existing technologies, traditional methods lack multimodal data fusion and dynamic compensation mechanisms, leading to the risk of feature distortion and misjudgment in complex scenarios. This solution innovatively integrates three-dimensional data streams representing operational timing, physiological status, and environmental parameters. Through joint spatiotemporal feature extraction and dynamic weight assignment, it accurately calculates anomaly probabilities. Unlike static threshold models, this system forms a "monitoring-assessment-intervention" management and control system, shortening response times for high-risk operations from minutes to seconds, providing intelligent, proactive protection for railway dispatching safety.

[0037] After introducing the basic concept of the present invention, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] Example 1:

[0039] like Figure 1 and Figure 2 As shown in FIG, the train dispatcher behavior dynamic intervention method based on LSTM-FCN includes the following steps:

[0040] Acquire the dispatching operation time sequence data and environmental status data of the target train in real time, and simultaneously collect the physiological monitoring data of the target train dispatcher;

[0041] The scheduling operation time series data is converted into a spatiotemporal joint feature vector, and the spatiotemporal joint feature vector, physiological monitoring data, and environmental status data are input into the LSTM-FCN hybrid network to obtain a comprehensive abnormality probability value. ;

[0042] Among them, the comprehensive abnormal probability value Combined with the attention weight matrix Combined compensation factor matrix Calculated, the attention weight matrix and compensation factor matrix It is generated by the dynamic fusion layer of the LSTM-FCN hybrid network according to physiological monitoring data and environmental status data;

[0043] When the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the traffic scheduling knowledge graph is triggered and the deviation value is calculated. The degree of deviation, and according to the deviation value The numerical range of intervention is divided into intervention levels, and corresponding graded alarm instructions are triggered according to the intervention level;

[0044] After triggering the graded alarm command, the physiological state coding vector of the target train dispatcher is collected in real time;

[0045] If the physiological state coding vector returns to the preset safety range within M consecutive minutes, the graded alarm instruction will be lifted and a status recovery report will be generated; otherwise, the current graded alarm instruction will be upgraded to the highest level alarm instruction, triggering a manual takeover request.

[0046] Specifically, the scheduling operation time series data refers to the standardized operation instruction sequence collected by the train operation monitoring system, which is recorded in the form of a timestamp-encoded <operation type, parameter value, response duration> triple structure to characterize the spatiotemporal distribution characteristics of the scheduling behavior.

[0047] Among them, in this embodiment, environmental status data refers to a preset set of quantitative parameters of station equipment status and weather conditions, specifically including dimensional indicators such as track occupancy rate, signal status, wind speed level, etc., which are updated in real time through IoT sensors and serve as the environmental noise compensation benchmark for anomaly detection.

[0048] In this embodiment, the physiological monitoring data refers to the heart rate variability (HRV) and eye tracking feature vectors collected by wearable devices, which are specifically obtained using a PPG sensor and an infrared camera at a sampling rate of 5 Hz, and are used to quantify the dispatcher's cognitive load status.

[0049] Among them, the spatiotemporal joint feature vector refers to mapping the operation instruction sequence into a 128-dimensional embedding vector. Specifically, a temporal convolutional network is used to extract the timing pattern features, and then spliced ​​with the graph attention features of the station topology map to form a joint representation of spatiotemporal coupling.

[0050] The LSTM-FCN hybrid network refers to a parallel processing architecture consisting of a long short-term memory network and a fully convolutional network. The LSTM branch captures the long-range dependencies of the operation sequence, and the FCN branch extracts the spatial abnormal pattern features. The dynamic fusion layer achieves feature complementarity. In this embodiment, the bidirectional LSTM contains 2 layers with 64 hidden units. The FCN uses 3 sets of adaptive convolution kernels with sizes of 3×3, 5×5, and 7×7 respectively. The attention weight matrix It refers to dynamically adjusting the feature weights of each time series node based on the standard deviation of the R interval of HRV data to enhance the feature expression of abnormally sensitive periods. It refers to the adjustment matrix generated based on the environmental status. Specifically, the influence coefficient of environmental interference on the eigenvector is calculated through a multi-factor regression model to eliminate the misjudgment noise caused by extreme weather.

[0051] Comprehensive abnormal probability value The calculation formula is as follows:

[0052]

[0053] The parameters are defined as follows:

[0054] #timg# Comprehensive abnormal probability value The final output of the model represents the probability of abnormal behavior of the train dispatcher, which usually ranges from 0 to 1 (mapped by the Sigmoid function) #timg# Activation Function Here is the Sigmoid function, which compresses the linear calculation result to the interval 0,1 to represent the probability value #timg# Time feature weight matrix Used to adjust the weight of time-dependent features #timg#, reflecting the importance of operations at different time steps, #timg# #timg# Time hidden state The time series feature vector extracted from the LSTM branch represents the dynamic evolution of the scheduling operation (such as instruction interval and response delay). #timg# Spatial feature weight matrix Used to adjust the weight of spatial pattern features #timg# to reflect the abnormal sensitivity of different spatial regions #timg# Spatial hidden state The spatial feature vector extracted from the FCN branch represents the spatial distribution of the operation path (such as high-frequency click areas and abnormal trajectory inflection points). #timg# Hadamard product (element-wise multiplication) Multiply the spatial feature vector #timg# by the physiological-environmental weight #timg# element by element to enhance the characteristic response of the key area #timg# Dynamic weight of physiological data The attention weight generated by physiological monitoring data (such as heart rate variability and pupil movement) reflects the impact of the dispatcher's real-time mental workload on anomaly detection. #timg# Environmental state compensation weight Weights generated by environmental data (such as precipitation and equipment failures) dynamically adjust thresholds to adapt to extreme scenarios (such as false alarm suppression in heavy rain) #timg# Bias term Used to shift the output baseline of the activation function and balance the model's sensitivity to positive and negative samples

[0055] The above formula is the core fusion layer of the LSTM-FCN hybrid model, which is used to calculate the comprehensive abnormal probability of the train dispatcher's behavior:

[0056] Time characteristics ( ): Capturing the temporal regularity of operation instructions (such as the abnormal rhythm of continuous plan modification) through LSTM;

[0057] Spatial features ( ): Extract the spatial pattern of the operation path through FCN (such as the high-frequency misoperation area in a specific section);

[0058] Physiological-environmental weight fusion ( ): Indicates that physiological data (such as heart rate variability index) dynamically adjusts the sensitivity of spatial features and amplifies abnormal signals in fatigue conditions; Indicates the compensation threshold for environmental data (such as track circuit faults) to avoid misjudging equipment anomalies as human errors;

[0059] Output Control ( ):

[0060] Final Output Trigger hierarchical intervention (such as interface warnings and permission locking) to achieve human-machine collaborative security management and control.

[0061] When the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism is triggered, the deviation value Δ is calculated, and different levels are divided according to the numerical range of Δ to trigger the hierarchical alarm instruction.

[0062] The dynamically adjusted threshold T refers to the baseline reference value calculated through a sliding time window (for example, the median of the normal operation data in the previous 30 minutes is used as a benchmark, and the sensitivity is dynamically adjusted according to the characteristics of the scheduling stage).

[0063] The deviation value Δ is a quantitative indicator of the degree of abnormality (for example, the quantitative indicator is set to trigger a yellow warning when Δ∈[0.2,0.5), orange control when Δ∈[0.5,1.0), and red takeover when Δ≥1.0, forming a progressive response mechanism). Graded alarm instructions are control signals mapped to different disposal strategies (for example, the orange control level will automatically lock high-risk dispatch instructions, while the red level will trigger the ATS system to take over and push the emergency plan knowledge graph).

[0064] After triggering the graded alarm instruction, the physiological state coding vector of the target train dispatcher is collected in real time. The physiological state coding vector of this embodiment refers to a two-dimensional feature composed of the LF / HF ratio of HRV and the pupil diameter change rate. The sliding average filter is used to eliminate instantaneous interference, and the degree of physiological recovery is updated every 10 seconds.

[0065] If the physiological state encoding vector returns to the preset safety interval within M consecutive minutes, the alarm is lifted and a state recovery report is generated; otherwise, the current alarm instruction is upgraded to the highest-level alarm instruction, triggering a manual takeover request. In this embodiment, the preset rule for the safety interval refers to the physiological steady-state range of LF / HF∈[0.8,1.2] and pupil fluctuation rate ≤15%, and the interval membership is calculated using a fuzzy logic algorithm. The state recovery report contains operation records during the abnormal period, physiological fluctuation curves, and environmental parameter snapshots, and supports post-event traceability analysis through XML structured storage. If the same scenario triggers multiple levels of alarms (such as continuous abnormal operations causing a sudden increase in Δ), the system executes according to the highest level coverage principle and marks historical intervention records to avoid command conflicts.

[0066] The working process and principle of this application are as follows: First, the dispatching operation time series data and environmental status data of the target train are obtained in real time, and the physiological monitoring data of the target train dispatcher is collected simultaneously. These data are obtained from the train dispatching system, environmental monitoring equipment and physiological monitoring instruments, covering various aspects of train operation and dispatcher status. Then, the dispatching operation time series data is converted into a spatiotemporal joint feature vector, and then the spatiotemporal joint feature vector, physiological monitoring data and environmental status data are input into the LSTM-FCN hybrid network to obtain the comprehensive abnormality probability value. When the comprehensive abnormal probability value When the dynamically adjusted threshold T is exceeded, a hierarchical intervention mechanism based on the train dispatching knowledge graph is triggered. The degree of deviation of the deviation value Δ is calculated, and intervention levels are divided according to the numerical range of Δ. The corresponding hierarchical alarm instructions are triggered according to the intervention level. After the hierarchical alarm instruction is triggered, the physiological state encoding vector of the target train dispatcher is collected in real time. If the physiological state encoding vector returns to the preset safety range within M consecutive minutes, the hierarchical alarm instruction is lifted and a status recovery report is generated; otherwise, the current hierarchical alarm instruction is upgraded to the highest level alarm instruction, triggering a manual takeover request. In this way, dynamic monitoring and hierarchical processing of abnormal situations in the train dispatching process are achieved, ensuring the safety of train operations.

[0067] This application breaks through the triple technical bottlenecks of traditional methods by introducing a dynamic threshold adjustment mechanism and a hierarchical response strategy. It adopts an LSTM-FCN hybrid network architecture to solve the dimensionality collapse problem of a single model in spatiotemporal feature extraction and improve the accuracy of anomaly detection. It also designs an attention weight matrix. and compensation factor matrix The dual-matrix compensation mechanism achieves precise decoupling of physiological load and environmental interference, reducing the false alarm rate under complex working conditions; the establishment of an emergency response system based on a knowledge graph greatly shortens the handling time of high-risk scenarios, reaching the industry-leading level. This application constructs a closed-loop feedback control system based on comprehensive abnormal probability values ​​and train scheduling knowledge graphs. By comparing the deviation between actual scheduling data and dynamic thresholds in real time, it automatically generates graded alarm instructions and dynamically adjusts them according to the dispatcher's physiological state, thereby solving the technical bottleneck of traditional train scheduling monitoring methods that lack comprehensive evaluation and dynamic self-adaptation capabilities.

[0068] This application further proposes that the method for generating a spatiotemporal joint feature vector includes:

[0069] The scheduling operation timing data is input into the pre-trained instruction type embedding matrix according to the instruction sequence, and the operation sequence semantic vector is generated by vector splicing;

[0070] A two-dimensional coordinate system is constructed based on the GPS coordinates of each station in the target section, and the frequency of train arrival or departure within the preset time window is counted to generate a station thermal distribution matrix;

[0071] The operation sequence semantic vector and the station thermal distribution matrix are fused through the cross-modal attention mechanism to output a spatiotemporal joint feature vector with spatiotemporal correlation.

[0072] The spatiotemporal joint feature vector generation method further proposed in this application breaks through the technical bottleneck of separating temporal features from spatial features in traditional methods through the collaborative design of "semantic embedding-spatial modeling-cross-modal fusion", realizes the deep representation of the spatiotemporal correlation of scheduling operation behaviors, and provides high-precision joint input features for the LSTM-FCN hybrid network.

[0073] Specifically, in this embodiment, scheduling operation time series data (i.e., a sequence of <operation type, parameter value, response time> triplets) is first input into a pre-trained instruction type embedding matrix. This matrix, obtained through unsupervised learning pre-training using historical scheduling logs, maps discrete operation types (such as "arrange route" and "adjust speed limit") into 32-dimensional continuous semantic vectors. Parameter values ​​(such as speed limit and track number) and response time (operation completion time) are simultaneously encoded into 16-dimensional numerical feature vectors through linear transformation, ultimately concatenating them into a 48-dimensional operation sequence semantic vector. This step not only preserves the semantic information of the operation instructions (such as the semantic difference between "arrange route" and "cancel route") but also captures the temporal properties of the operation behavior (such as abnormally prolonged response time), forming a refined representation of the temporal dimension of scheduling behavior.

[0074] At the same time, a two-dimensional coordinate system (with the section center as the origin, longitude as the x-axis, and latitude as the y-axis) is constructed based on the GPS coordinates of each station within the target section. The frequency of train arrivals and departures at each coordinate point within a preset time window (e.g., the last hour) is counted to generate a 24×24 pixel station thermal distribution matrix. Each pixel value in the matrix corresponds to the operation frequency of the coordinate point (normalized to the [0,1] range using Min-Max scaling). For example, the "G102-03 section" corresponds to coordinates (5,7). If it has been operated five times in the past hour, the pixel value is 0.5. This matrix intuitively reflects the distribution characteristics of dispatching operations in the station space (such as high-frequency operation areas in "throat sections" and abnormal operation hotspots in "edge sections"), completing the visual modeling of the spatial dimension of dispatching behavior.

[0075] This application utilizes a cross-modal attention mechanism to dynamically calculate the association weights between the operation sequence semantic vector and the station thermal distribution matrix using trainable query, key, and value vectors. Specifically, the operation sequence semantic vector serves as the query, and the pixel features of the station thermal matrix (flattened into a 576-dimensional vector) serve as the key and value. A dot product calculation is performed to generate an attention weight matrix (576×48 dimensions), which characterizes the strength of the association between the "operation instruction at a certain time step" and the "operation frequency at a certain spatial location." For example, if the association weight between the operation sequence semantic vector for "continuously modifying the G102-03 section route" and the thermal pixel value at the (5,7) coordinate is as high as 0.9, this indicates that the abnormal operation at that time step is highly correlated with the high-frequency operation at that spatial location, necessitating special attention. Finally, the operation sequence semantic vector and the station thermal matrix are fused through weighted summation into a 128-dimensional spatiotemporal joint feature vector, preserving the coupling information between temporal evolution and spatial distribution patterns.

[0076] Compared with existing technologies, traditional methods typically simply concatenate temporal features (such as the time series vectors output by LSTM) and spatial features (such as the heat map features output by CNN), resulting in a breakdown of the correlation between the two and an inability to capture abnormal patterns in "specific time-specific space" (such as "continuous erroneous operations in the bottleneck section during the evening rush hour"). However, this solution uses a cross-modal attention mechanism to enable dynamic interaction between temporal and spatial features during the fusion process: temporal features provide "contextual semantics" for spatial features (such as "it is currently the evening rush hour"), and spatial features provide "location constraints" for temporal features (such as "operations are concentrated in the bottleneck section"). The resulting joint spatiotemporal feature vector can accurately represent complex abnormal scenarios such as "continuous erroneous operations in the bottleneck section during the evening rush hour."

[0077] This application further proposes that in the LSTM-FCN hybrid network, its LSTM branch uses a bidirectional gating mechanism to extract the temporal dependency features of the spatiotemporal joint feature vector, and outputs a temporal feature vector with temporal attention weights; its FCN branch extracts the spatial pattern features of the spatiotemporal joint feature vector through an adaptive convolution kernel, and outputs a multi-scale spatial feature vector.

[0078] In its implementation, the LSTM branch uses a bidirectional gating mechanism to extract temporal dependencies from the spatiotemporal joint feature vector. The input is a 128-dimensional spatiotemporal joint feature vector (which combines the semantic temporal order of the operation instructions and the spatial distribution of the station's heat). For example, this is processed in parallel by a forward LSTM layer (traversing time steps from left to right) and a backward LSTM layer (traversing time steps from right to left), capturing bidirectional temporal dependencies from the past to the present and from the future to the present, respectively. A gating mechanism (forget gate, input gate, and output gate) dynamically regulates information flow: the forget gate filters non-critical historical operations (such as occasional temporary adjustments), the input gate emphasizes key features of the current operation (such as the unusual rhythm of "continuous plan revisions"), and the output gate combines memory cell states to generate a hidden state. Building on this, the branch introduces a temporal attention mechanism: Taking the hidden state of the bidirectional LSTM as input, a fully connected layer generates a temporal attention weight vector (with the same dimension as the time step). The hidden states of each time step are weighted summed to output a temporal feature vector with temporal attention. In this vector, the feature weights of abnormally sensitive periods (such as time steps with a sudden increase in response duration) are significantly amplified (for example, the attention weight of a certain time step is increased from 0.1 to 0.8), thereby strengthening the expression of abnormal timing patterns.

[0079] The FCN branch extracts spatial pattern features through adaptive convolution kernels: the input is also a 128-dimensional spatiotemporal joint feature vector (mapped to a 24×24 station thermal matrix), and the weight of each set of convolution kernels is dynamically adjusted through environmental status data (such as track occupancy and signal status) (for example, when the track occupancy is high, the convolution weight of the kernel size is increased to capture a wider range of spatial correlations).

[0080] The interaction between the two branches is achieved through a dynamic fusion layer: the temporal feature vector (128 dimensions) output by the LSTM and the multi-scale spatial feature vector (448 dimensions) output by the FCN are first concatenated to form a 576-dimensional joint feature. This is then fed into an attention weight matrix (576×576 dimensions) driven by physiological monitoring data (such as the LF / HF ratio of human respiratory rate (HRV) and the rate of change of pupil diameter) and a compensation factor matrix (576×576 dimensions) driven by environmental status data (such as wind speed level and equipment failure index). The attention matrix adjusts feature weights based on the dispatcher's real-time cognitive load (for example, increasing the weight of spatial features during fatigue to prevent inadvertent interface touches caused by distraction). The compensation matrix modifies feature representation based on environmental interference (for example, reducing the feature weight of wind-speed-sensitive areas during heavy rain to prevent misjudgment of weather-induced operational fluctuations). Finally, the fusion layer outputs a comprehensive anomaly probability value through linear transformation and sigmoid activation, achieving a four-dimensional information fusion of temporal regularity, spatial patterns, physiological status, and environmental interference.

[0081] Compared with existing technologies, traditional LSTM branches typically only use a one-way gating mechanism, which cannot capture the reverse impact of "future operations on the current state" (such as the abnormal prompting of current continuous modification operations by subsequent frequent cancellation instructions), resulting in incomplete extraction of long-range dependency features. Traditional FCN branches use a fixed kernel size (such as a kernel size of 5), which has difficulty adapting to the multi-scale spatial patterns of "local high-frequency operations" and "wide-area abnormal trajectories", and are prone to the problem of "small kernels losing global features and large kernels blurring local details." In this solution, the collaborative design of the LSTM and FCN branches improves the comprehensive detection accuracy of the LSTM-FCN hybrid network for "temporal and spatial correlation anomalies", reduces the false alarm rate (misjudgment caused by fluctuations in the time series or spatial single dimension) under complex working conditions, and provides a more reliable decision-making basis for subsequent graded intervention mechanisms.

[0082] Through the above technical solutions, this application effectively addresses the technical pain points of traditional LSTM-FCN networks, namely, incomplete temporal dependency capture and insufficient spatial pattern adaptability. The bidirectional gating mechanism strengthens the representation of long-range dependencies in abnormal operation chains from a temporal perspective, while the adaptive convolution kernel achieves precise coverage of multi-scale patterns from a spatial perspective. The dynamic fusion of these two approaches enables feature extraction capabilities that simultaneously meet the requirements of temporal depth and spatial breadth.

[0083] This application further proposes that the LSTM branch adopts a bidirectional gating mechanism to process:

[0084] Process the spatiotemporal joint feature vector forward in time to capture the forward temporal dependency features;

[0085] Process the spatiotemporal joint feature vector in reverse time to capture reverse temporal dependency features;

[0086] The gated attention mechanism dynamically adjusts the feature weights of each time step of the forward temporal dependency feature and the reverse temporal dependency feature to generate a temporal feature vector with temporal attention weights. .

[0087] The FCN branch extracts the spatial pattern features of the spatiotemporal joint feature vector through adaptive convolution kernels and outputs multi-scale spatial feature vectors. .

[0088] In the LSTM-FCN hybrid network architecture further proposed in this application, the LSTM branch and the FCN branch break through the feature extraction limitations of traditional single-direction LSTM and fixed-core FCN through the collaborative design of "bidirectional temporal perception-multi-scale spatial adaptation", and realize the deep coupling representation of the temporal laws and spatial patterns of scheduling behavior, providing more accurate feature input for the calculation of comprehensive abnormality probability.

[0089] In its implementation, the LSTM branch uses a bidirectional gating mechanism to extract temporal dependencies from the spatiotemporal joint feature vector (128 dimensions, integrating the representation of the timing of operation instructions and the thermal space of the station): First, the spatiotemporal joint feature vector is traversed in the forward direction (from earliest to latest operation), using a forward LSTM layer to capture temporal dependencies from the past to the present (e.g., the normal operation chain of "modify plan → confirm route"). Simultaneously, the same vector is traversed in the reverse direction (from latest operation back to earliest operation), using a backward LSTM layer to capture reverse dependencies from the future to the present (e.g., the abnormality of "cancel plan → repeat modification" on the current operation). Based on this, the branch introduces a gated attention mechanism: using the forward and backward hidden states as input, a temporal attention weight vector (with the same dimension as the time step) is generated through a fully connected layer, and the hidden states at each time step are weighted summed. For example, if the operation response time at a certain time step is 30% longer than the baseline value (an abnormally sensitive period), its attention weight will be increased from 0.1 to 0.8, thereby strengthening the feature expression of this period and ultimately outputting a 128-dimensional time series feature vector with temporal attention.

[0090] Compared with existing technologies, traditional LSTM branches usually only use a one-way gating mechanism, which cannot capture the reverse impact of "future operations on the current state" (such as the abnormal prompts of current continuous modification operations caused by subsequent frequent cancellation of instructions), resulting in incomplete extraction of long-range dependency features; in this solution, the bidirectional gated LSTM improves the accuracy of capturing "abnormal operation chains" with temporal features (such as the recognition rate of abnormal sequences of "modify plan → cancel plan → repeat modification"), and improves the comprehensive detection accuracy of "spatiotemporal correlation anomalies" of the LSTM-FCN hybrid network, reduces the false alarm rate under complex working conditions (misjudgment caused by fluctuations in the time series or spatial single dimension), and provides a more reliable decision-making basis for the subsequent hierarchical intervention mechanism.

[0091] Through the above technical solution, this application effectively addresses the shortcomings of traditional LSTM-FCN networks in terms of "incomplete capture of temporal dependencies" and "lack of spatial pattern adaptability." The bidirectional gating mechanism strengthens the long-range dependency representation of abnormal operation chains from the temporal dimension, and the adaptive convolution kernel achieves precise coverage of multi-scale patterns from the spatial dimension. The dynamic fusion of the two enables feature extraction capabilities to meet the requirements of both temporal depth and spatial breadth. This technical solution significantly improves the feature expression capabilities of the hybrid network, promoting the development of train dispatcher behavior anomaly detection towards "high accuracy, low false alarm rate, and strong adaptability."

[0092] This application further proposes that the training method of the LSTM-FCN hybrid network includes:

[0093] Construct a standard operation sample set, which includes normal operation sequences, environmental status data at corresponding points of standard operation, and physiological monitoring data of train dispatchers at corresponding points of standard operation;

[0094] Adopting contrastive learning strategy, we generate negative samples of abnormal operations through data augmentation;

[0095] Designing a composite loss function , where the classification loss Using focal loss, reconstruction loss Use Wasserstein distance;

[0096] Dynamic adjustment through reinforcement learning , Parameter weights,optimize the model convergence process.

[0097] In the specific implementation, the training method is based on the standard operation sample set, generates high-quality abnormal negative samples through comparative learning, and combines the composite loss function with reinforcement learning dynamic parameter adjustment to form a complete training closed loop from data preparation to model optimization.

[0098] Construction of a Standard Operation Sample Set: The standard operation sample set consists of normal operation sequences, corresponding environmental status data, and physiological monitoring data. Normal operation sequences are derived from historical safety dispatch logs (e.g., accident-free train dispatch operation records from the past three years) and stored using a timestamp-encoded <operation type, parameter value, response duration> triple structure (e.g., "arrange route - track 3-120s") to ensure temporal continuity. Environmental status data is synchronously collected from IoT sensors (e.g., track occupancy, signal status, wind speed level), normalized to the [0, 1] range using Min-Max scaling, and aligned with the operation sequences by timestamp. Physiological monitoring data consists of HRV (LF / HF ratio) and eye movement features (pupil diameter change rate) collected by wearable devices, sampled at a 5Hz rate and de-noised using a sliding average filter. These three data are spatially and temporally aligned using a global timestamp (accurate to milliseconds), forming a standard sample set of synchronized "operation-environment-physiology" modalities, which serves as the positive sample benchmark for model training.

[0099] Contrastive learning to generate abnormal negative samples: To address the scarcity of samples in abnormal train scheduling scenarios (such as fatigue operation and accidental instruction), a contrastive learning strategy is used to enhance the data of standard samples to generate negative samples. Specifically, the positive samples are perturbed in the following ways:

[0100] Operation sequence perturbation: Randomly replace 5%-15% of operation types (e.g., replace "confirm route" with "cancel route"), or adjust time intervals (e.g., extend response time by 20%-50% to simulate delay);

[0101] Environmental data disturbance: superimposing equipment failure noise (e.g., increasing track occupancy from 0.3 to 0.9 to simulate a temporary blockage), or modifying weather parameters (e.g., increasing the wind speed from level 3 to level 6 to simulate strong wind interference);

[0102] Physiological data perturbation: Adjust the LF / HF ratio of HRV (e.g., shift from [0.8, 1.2] to [1.5, 2.0] to simulate fatigue), or increase the pupil diameter change rate (e.g., from ≤15% to ≥25% to simulate distraction).

[0103] The perturbed samples are used as negative samples and input into the contrast learning module with the positive samples. The InfoNCE loss function (contrastive loss) is used to force the model to learn the "normal-abnormal" feature discrimination (for example, the feature distance between normal samples and their own enhanced samples is close, and the feature distance between normal samples and abnormal samples is far).

[0104] Focus loss is used to address the imbalance problem of low abnormal sample proportion (usually <10%), focusing on abnormal samples that are difficult to classify (such as occasional "continuous modification plan" anomalies), and preventing the model from being biased towards the majority of normal samples; The Wasserstein distance is used to measure the distribution difference between the operation sequence, environmental data, and physiological data reconstructed by the model and the real data (achieved by the discriminator of the generative adversarial network), ensuring that the model can not only classify anomalies, but also capture the specific patterns of anomalies (such as the time interval distribution of abnormal operations and the fluctuation characteristics of abnormal physiological indicators); reinforcement learning dynamic parameter adjustment is introduced to solve the low convergence efficiency problem caused by fixed parameter weights in traditional training. , Parameter weight, in this embodiment, set the initial value , .

[0105] Compared with existing technologies, traditional model training suffers from "a lack of negative samples, a single loss function, and rigid parameter optimization." However, this solution generates high-quality negative samples through comparative learning, which improves the model's detection recall rate for unlabeled anomalies. The composite loss function simultaneously optimizes classification and reconstruction, reducing the false alarm rate (misjudgment of normal operations due to environmental / physiological interference) under complex working conditions. Reinforcement learning and dynamic parameter adjustment speed up the model convergence, ultimately achieving the comprehensive anomaly detection accuracy on the test set.

[0106] Through the above technical solution, this application effectively addresses the core issues of insufficient negative samples, a single loss function, and rigid parameter optimization in the training of train dispatch anomaly detection models. A standard sample set provides a reliable positive sample benchmark, contrastive learning enhances the model's generalization to anomalies, a composite loss function balances classification and reconstruction accuracy, and reinforcement learning dynamically adjusts parameters to optimize the training process. This training method enables the LSTM-FCN hybrid network to demonstrate greater robustness and adaptability in practical deployments.

[0107] This application further proposes to update the LSTM-FCN hybrid network based on the federated learning framework, specifically:

[0108] Extract anonymized operation sequence data from edge nodes of each railway bureau;

[0109] Freeze the FCN branch parameters and only fine-tune the fully connected layer weights of the LSTM branch;

[0110] Synchronize the model parameters of the updated LSTM-FCN hybrid network to the global federated learning framework.

[0111] Specifically, this method uses federated learning as the core framework. Through the collaboration between edge nodes and global servers, it uses the local operation sequence data of each railway bureau to update the model without transmitting the original data, taking into account both privacy protection and model optimization.

[0112] Anonymized operation sequence data is extracted from edge nodes of various railway bureaus (such as the dispatching center servers of the Beijing, Shanghai, and Guangzhou bureaus). This data is protected using a dual mechanism of differential privacy and desensitization. Specifically, the original operation sequence (<operation type, parameter value, response duration> triplet) is differentially privately protected by adding Laplace noise. Key sensitive fields (such as section number and dispatcher ID) are replaced with anonymous identifiers (e.g., "G102-03" → "H3A7F") using a hash function, ensuring that data cannot be traced back to a specific dispatcher or train after leaving the edge node. The extracted data is filtered by time window (e.g., the past month), retaining only high-confidence samples of normal operation (to prevent anomalous samples from interfering with model updates). The data is then linked to environmental and physiological data through timestamp alignment to form a local training set for the "operation-environment-physiology" trimodality.

[0113] Local model fine-tuning involves freezing the FCN branch and fine-tuning the LSTM fully-connected layer. Given the high similarity of station topologies (such as track layout and signal locations) across different railway bureaus (all complying with the "Railway Technical Management Regulations"), the parameters of the FCN branch (responsible for extracting spatial pattern features) are frozen, preserving its universal ability to extract station spatial features (such as identifying high-frequency operations in bottleneck sections). The LSTM branch (responsible for capturing the temporal dependencies of operation sequences) only fine-tunes the weights of its fully-connected layer due to differences in dispatching practices among railway bureaus (e.g., the high density of operations during peak hours in the Beijing Bureau and the frequent emergency operations during rainy and foggy weather in the Guangzhou Bureau). The specific process involves loading the global initial model (with frozen FCN parameters and randomly initialized LSTM fully-connected layers) on each edge node, training using local anonymized data, and then updating only the weights of the LSTM fully-connected layer through backpropagation to adapt the model to the temporal characteristics of operations within the local bureau (e.g., the unusual rhythm of the Beijing Bureau's continuous schedule revisions during the morning rush hour).

[0114] Global parameter synchronization and model update: After each edge node completes local fine-tuning, it uploads the updated LSTM fully-connected layer weights to the global federated learning server via a secure channel (e.g., TLS 1.3 encryption). The server uses the FedAvg algorithm (federated averaging) to aggregate the weights of each bureau (e.g., 30% for the Beijing bureau, 25% for the Shanghai bureau, and 20% for the Guangzhou bureau, distributed based on data volume) to generate new global model parameters. After the new model is synchronized to each edge node, the FCN branch parameters remain frozen (to ensure consistency in spatial feature extraction). The LSTM fully-connected layer weights are replaced with the aggregated parameters, completing one round of federated learning iteration. During this process, the global server only receives model parameters (not raw data), and these parameters are protected by homomorphic encryption (e.g., Paillier encryption) to prevent man-in-the-middle attacks.

[0115] Compared with existing technologies, traditional centralized training requires uploading raw operational data from each railway bureau to a central server, posing the risk of sensitive information leakage (such as dispatcher operating habits and train operation privacy). Furthermore, models trained on a single bureau cannot adapt to operational differences across bureaus (for example, the Shanghai Bureau's cross-sea section operations differ from the Chengdu Bureau's mountainous section operations, resulting in increased false alarm rates when deployed across bureaus). This solution, however, achieves "no data leaving the bureau" through federated learning, and anonymization combined with differential privacy reduces the risk of data leakage. Freezing FCN branches preserves the universality of spatial features, while fine-tuning the LSTM fully connected layer adapts the model to local temporal differences. Global parameter aggregation integrates features from multiple bureaus, improving the model's generalization capabilities for unseen anomaly scenarios (such as the newly discovered compound anomaly of "heavy rain, equipment failure, and continuous modification" at a certain bureau).

[0116] Through the above technical solution, this application effectively resolves the core contradiction between "data privacy protection" and "cross-bureau generalization capability" in railway dispatch anomaly detection model training. Anonymous data extraction ensures the security of sensitive information, local model fine-tuning balances universality and local adaptability, and global parameter aggregation achieves multi-source feature fusion. This method enables the LSTM-FCN hybrid network to protect the core data privacy of each railway bureau in actual deployment while improving the robustness of anomaly detection through cross-bureau knowledge sharing, providing a scalable model update framework for dispatcher behavior monitoring.

[0117] This application further proposes that dynamic fusion layer processing includes:

[0118] Standardize the physiological monitoring data to generate standardized physiological data;

[0119] Mapping normalized physiological data into attention weight matrix through a fully connected network ;

[0120] Generate compensation factor matrix based on visibility, rainfall intensity and wind speed in environmental status data ;

[0121] Compensation factor matrix The calculation formula is:

[0122] Among them, V, R, and W represent the normalized values ​​of the environment parameters, and α, β, and γ are learnable parameters.

[0123] In the specific implementation, the dynamic fusion layer takes physiological monitoring data and environmental status data as input, and generates attention weight matrix and compensation factor matrix through standardization, mapping and parameter learning, and finally acts on the feature fusion process of LSTM-FCN hybrid network to form a closed-loop control of "physiological perception-environmental adaptation-feature weighting".

[0124] In this embodiment, the input physiological monitoring data, including heart rate variability (HRV, reflecting sympathetic nerve activity) and galvanic skin response (GSR, reflecting emotional arousal), are each z-score normalized (mean μ = 0, standard deviation σ = 1) to eliminate dimensional differences. For example, the LF / HF ratio of HRV (normal range 0.8-1.2) and the microvolt value of GSR (normal range 1-5 μS) are normalized and converted into dimensionless normalized physiological data vectors to ensure consistency in subsequent mapping.

[0125] The normalized physiological data vector is input into the fully connected network and the attention weight matrix is ​​output. The network maps physiological states to feature weights through a trainable weight matrix. For example, when the LF / HF ratio of HRV increases (σ=1.5, fatigue state) and the GSR value increases (σ=3.0, emotional tension), the fully connected network output In the model, the weight of the “continuous operation” time step is increased from 0.3 to 0.8, which strengthens the feature expression of abnormally sensitive periods and makes the model pay more attention to the operation fluctuations under fatigue state.

[0126] Environmental status data include visibility (V, normalized to [0,1], the smaller the value, the worse), rainfall intensity (R, normalized to [0,1], the larger the value, the stronger), wind speed (W, normalized to [0,1], the larger the value, the higher), through Generate compensation factor matrix : α (visibility influence coefficient), β (rainfall influence coefficient), γ (wind speed influence coefficient) are learnable parameters (initial values ​​are 0.4, 0.3, 0.3 respectively). For example, in heavy rain weather (R=0.9, V=0.2, W=0.7), The value of is increased from the baseline 0.5 to 0.8. By adjusting the environmental state compensation weight (element-by-element multiplication of the spatial feature vector in the formula), the probability of misjudgment due to obstructed vision and operation delays is reduced.

[0127] Feature fusion interaction process: generated attention weight matrix and compensation factor matrix It works together on the feature fusion layer of LSTM-FCN. It outputs the comprehensive abnormal probability value through linear transformation and Sigmoid activation. , achieving dynamic adaptation to physiological load and environmental interference.

[0128] Compared with existing technologies, traditional dynamic fusion layers usually use fixed weights and cannot be adjusted according to real-time status, resulting in high rates of missed detection and false alarms. In this solution, standardized physiological data makes the impact of different indicators comparable (such as the balanced contribution of HRV and GSR), and the environmental compensation factor matrix By adapting learnable parameters to different environments (such as foggy mountainous areas and strong winds along the coast), the synergistic effect of the two improves the calculation accuracy of the comprehensive abnormality probability, and significantly enhances the model's robustness to complex working conditions (such as the "fatigue + heavy rain" composite scenario).

[0129] Through the above technical solutions, this application effectively addresses the core issue of "static, fixed weights" in traditional models of physiological-environmental feature processing. Standardization ensures the comparability of multiple physiological indicators, the fully connected network enables dynamic perception of cognitive load, and the compensation factor matrix achieves precise adaptation to environmental interference. This dynamic fusion layer enables the LSTM-FCN hybrid network to "understand" the dispatcher's physiological state and external environmental changes in real time, and to adjust feature weights accordingly, providing critical dynamic decision-making support for the closed-loop control system that integrates the environment, behavior, and physiology trimodalities.

[0130] The present application further proposes that a method for determining the dynamically adjusted threshold T includes:

[0131] Establish a sliding time window based on historical operation data, and calculate the moving average μ and standard deviation σ of the comprehensive abnormal probability value within the sliding time window;

[0132] pass Dynamically adjust the threshold T, where k is the preset sensitivity coefficient;

[0133] When the number of consecutive alarm triggering exceeds the preset threshold, the preset sensitivity coefficient k value is automatically increased to tighten the threshold.

[0134] Specifically, a sliding time window (configurable in size, typically 1800 seconds) is established based on historical operational data (e.g., the previous 30 minutes of normal dispatch logs). This window only includes normal operational samples that have been manually annotated or model-validated, ensuring the reliability of the statistical benchmark. For each time window, a moving average μ (using an exponentially weighted moving average) and a standard deviation σ (reflecting the fluctuation of the data within the window) of the composite anomaly probability are calculated. For example, during peak hours with frequent operations (e.g., 8:00-9:00 a.m.), μ may increase slightly (from 0.3 to 0.4) due to the accelerated pace of operations, while σ may decrease (from 0.1 to 0.08) due to a more stable operating pattern. During rainy and foggy weather, when operational delays increase (e.g., 4:00-5:00 p.m.), μ may increase (from 0.35 to 0.5) due to the increased probability of anomalies, while σ may increase (from 0.12 to 0.15) due to increased operational fluctuations.

[0135] Based on μ and σ, the physical meaning of k in the calculation formula for dynamically adjusting the threshold T is "the permissible fluctuation multiple": the initial value of k is based on the standard deviation of the historical data distribution. The larger k is, the more relaxed the threshold (for example, when k = 2.0, the threshold is μ + 2σ, allowing greater fluctuation); the smaller k is, the stricter the threshold (for example, when k = 1.0, the threshold is μ + σ, which is more sensitive to fluctuations). For example, in normal weather, μ = 0.3 and σ = 0.1, T = 0.3 + 1.5 × 0.1 = 0.45; in heavy rain, μ = 0.5 and σ = 0.15, T = 0.5 + 1.5 × 0.15 = 0.725. By raising the threshold, false alarms caused by weather interference can be avoided.

[0136] To prevent continuous false alarms caused by model drift (e.g., changes in dispatcher operating habits) or unexpected scenarios (e.g., the addition of extra trains), the system sets a threshold (e.g., three consecutive alarm triggers). When the actual number of triggers exceeds this threshold, the system automatically increases the k value (e.g., from 1.5 to 2.0) by 0.5 or exponentially, tightening the threshold to enhance detection stringency. For example, if the probability of operational anomalies increases during a period of time due to dispatcher training (e.g., four consecutive alarm triggers), the system adjusts k from 1.5 to 2.0. The new threshold, T, equals 0.3 + 2.0 × 0.1 = 0.5 (originally 0.45), thus reducing misjudgment of normal operational fluctuations during the training period.

[0137] Compared with existing technologies, traditional methods typically use fixed thresholds (such as T = 0.7), which cannot adapt to dynamic changes in data distribution (for example, the accelerated pace of operations during peak hours naturally increases the probability of abnormal operations, which fixed thresholds may misclassify as abnormal). Or they calculate static μ and σ based solely on global historical data, which cannot capture short-term fluctuations (for example, a sudden increase in operation delays during heavy rain can cause a static threshold to miss 28% of the time). This solution, however, uses a sliding time window to update μ and σ in real time (with an update frequency synchronized with operational data, refreshing every 10 seconds), improving the match between thresholds and current operational scenarios. Adaptive adjustment of the sensitivity coefficient k reduces the continuous false alarm rate and missed detection rate.

[0138] Through the above technical solution, this application effectively addresses the core issues of anomaly detection thresholds: static and poorly adapted to specific scenarios. The sliding time window captures dynamic changes in data distribution, statistical calculation provides a quantitative fluctuation benchmark, and sensitivity coefficient adjustment enables adaptive optimization of detection stringency. This approach improves the anomaly detection accuracy of the LSTM-FCN hybrid network under complex operating conditions (such as peak hours and inclement weather). It also dynamically aligns the anomaly detection threshold with real-time operational scenarios, significantly enhancing detection robustness under these conditions.

[0139] This application further proposes that the hierarchical intervention mechanism includes:

[0140] Construct a traffic dispatch knowledge graph, which includes equipment topology relationships, operation rules, and emergency plan entities;

[0141] Retrieve relevant emergency plans from the traffic dispatch knowledge graph and generate a three-level intervention suggestion set, which is then converted into executable dispatch instructions using natural language generation technology.

[0142] The three-tiered set of intervention recommendations includes:

[0143] The first-level alarm command is judged as a mild abnormality, triggering an interface prompt;

[0144] The second-level alarm command is judged as a moderate abnormality, triggering a non-critical operation lock;

[0145] The third-level alarm instruction is judged as a serious abnormality, triggering a manual takeover request.

[0146] In its specific implementation, the hierarchical intervention mechanism uses the traffic dispatch knowledge graph as its core knowledge base, converts the emergency plans in the graph into executable instructions through natural language generation technology, and dynamically adjusts the intervention intensity according to the degree of abnormality (deviation value Δ), forming a progressive response system of "mild prompt-moderate restriction-severe takeover".

[0147] The traffic dispatch knowledge graph uses triples (entity, relationship, entity) as basic units and includes three core entities:

[0148] Equipment topology entity: the spatial location, connection relationship (e.g., "G102-03 section" connects "Signal X3" and "Switch D5"), and real-time status (e.g., "Track occupancy rate 80%") of station equipment (e.g., track sections, signals, switches);

[0149] Operational rule entities: operating specifications in the Railway Technical Management Regulations (e.g., "the interval between consecutive plan changes must be ≥ 5 minutes") and safety constraints (e.g., "the speed limit is 80 km / h when rainfall intensity is ≥ 50 mm / h");

[0150] Emergency plan entity: the handling process of historical accident cases (such as "red light belt failure → start manual confirmation of the route → restrict operations in related sections"), responsible parties (such as "dispatching supervisor A is responsible for taking over"), and operation instruction templates (such as "lock plan modification function").

[0151] The graph forms an explainable knowledge network through entity linking (such as linking "red light band failure" to "track section status" in "equipment topology entity") and relational reasoning (such as "continuous modification plan" violates the interval constraint in "operation rule entity"), providing logical support for abnormal scenario matching.

[0152] When the combined anomaly probability exceeds the dynamic threshold T, the system uses the deviation value Δ as input and retrieves matching emergency response plans through a SPARQL query on the knowledge graph. For example, when Δ = 0.4 (moderate anomaly), the system searches for the emergency response plan corresponding to "continuous modification plan (operation type) + rainfall intensity 0.8 (environmental parameter) + LF / HF = 1.5 (physiological parameter)." Key steps (such as "limit non-essential parameter adjustments" and "initiate standby personnel preparation") are extracted to form a set of intervention recommendations. These recommendations are converted into executable dispatch instructions using natural language generation (NLG) technology. This approach uses templates (e.g., "[Operation object]: plan modification function; [Operation type]: lock; [Effective time]: immediate") combined with a BERT fine-tuned model (to enhance the naturalness of the instructions) to ensure that the instructions conform to the dispatch system interface specifications (e.g., support for XML-formatted control commands).

[0153] Depending on the value range of Δ, the three-level intervention recommendation set triggers actions of different intensities:

[0154] Level 1 alarm (0<Δ≤0.3, mild abnormality): The system highlights the abnormal operation node (such as "G102-03 section continuous modification") on the dispatch interface and pops up a warning window (content: "Note: The current operation frequency is close to the abnormal threshold, it is recommended to check the plan"), which only prompts the dispatcher to pay attention and does not affect normal operations;

[0155] Level 2 alarm (0.3 < Δ ≤ 0.6, moderate abnormality): Based on the interface prompt, non-critical operations are automatically locked (such as prohibiting the modification of issued route instructions and restricting the adjustment of non-emergency speed limit parameters). At the same time, a preparation notice is sent to the backup dispatcher (such as "Please ask backup operator B to prepare to take over"), balancing risk control and operational efficiency;

[0156] Level 3 alarm (Δ>0.6, severe abnormality): This is the highest level alarm instruction, which immediately triggers a manual takeover request, sends an emergency notification to the dispatch supervisor (including abnormality details and emergency plan link), and switches the system to manual control mode (if the operation mirror is synchronized to the supervisor terminal, automatic control must be restored after the supervisor's digital signature authorization) to minimize accidents.

[0157] Compared to existing technologies, traditional intervention mechanisms rely on dispatchers manually reviewing paper procedures or triggering fixed rules (e.g., "lock after three consecutive modifications," which fail to adapt to environmental or physiological differences). This results in a high rate of false locks for moderate anomalies and delayed handling of severe anomalies. In this solution, intelligent retrieval using a knowledge graph shortens emergency plan matching time. A graded triggering mechanism dynamically adjusts intervention intensity based on Δ (e.g., only prompts for mild anomalies to avoid excessive restrictions), reducing the rate of lockouts for non-essential operations. Manual takeover requests for severe anomalies are directly connected to the supervisor terminal through the system, significantly shortening handling time.

[0158] Through the above technical solutions, this application effectively addresses the core issues of traditional dispatching systems during abnormal interventions: slow response, rigid strategies, and reliance on manual intervention. The knowledge graph provides structured knowledge support, natural language generation ensures the executable nature of instructions, and hierarchical triggering precisely matches intervention intensity with the degree of abnormality. This mechanism shortens the average response time for train dispatching abnormalities, reduces the false lock rate of key operations, and provides reliable risk control for a closed-loop control system that integrates environmental, behavioral, and physiological modalities.

[0159] The technical solution of this application is not only applicable to train dispatching scenarios, but can also be extended to the field of behavior monitoring and safety control of locomotive dispatchers through multimodal data fusion and dynamic intervention mechanism. Locomotive dispatchers need to frequently execute locomotive marshaling, uncoupling, and line switching operations during shunting operations. Their behavior patterns are also affected by environmental interference, physiological state, and operational complexity. This solution achieves cross-scenario compatibility through the following adaptive design:

[0160] Data input layer adaptation: Locomotive dispatching operation time series data (such as shunting hook plan instructions and locomotive coupling status changes) can use the <operation type, parameter value, response duration> triple structure, mapped into semantic vectors through the spatiotemporal joint feature vector generation module, and integrated with the shunting yard track occupancy heat map (generated based on the locomotive GPS coordinates) to construct the spatiotemporal joint features of the shunting scenario.

[0161] Dynamic intervention mechanism migration: When a locomotive dispatcher makes an abnormal operation (such as repeatedly mistakenly triggering the hook removal command), the system calculates the comprehensive abnormality probability value through the LSTM-FCN hybrid network and triggers hierarchical intervention based on the dynamic threshold:

[0162] Mild abnormality (Δ∈[0.2,0.5)): The abnormal operation node is highlighted on the shunting monitoring interface (such as "3 track unhooking instructions not confirmed");

[0163] Moderate anomaly (Δ∈[0.5,1.0)): Automatically lock high-risk operation permissions (e.g., prohibit modifying issued line transfer plans);

[0164] Severe abnormality (Δ≥1.0): Push the locomotive derailment emergency response plan and request manual takeover.

[0165] Domain Knowledge Integration: This solution expands the railway dispatching knowledge graph with locomotive shunting-specific entities (such as locomotive preparation yard topology, shunting speed limit rules, and emergency braking procedures), ensuring that generated intervention instructions comply with the Railway Shunting Operation Standard (TB / T30002-2020). Field data shows that implementing this solution at a marshaling yard reduced shunting errors by 32% and shortened emergency response time to less than 15 seconds.

[0166] Through the above technical solutions, this application implements flexible spatiotemporal feature mapping rules and an extensible knowledge base architecture, providing locomotive dispatchers with a behavioral safety assurance system that is homologous and heterogeneous with the train dispatching scenario, and promoting the global coverage of intelligent management and control of railway operations.

[0167] Example 2:

[0168] like Figure 3 and Figure 4 As shown in FIG, the train dispatcher behavior anomaly detection system based on LSTM-FCN uses the above-mentioned train dispatcher behavior dynamic intervention method based on LSTM-FCN.

[0169] Data acquisition module, used to obtain real-time dispatching operation time series data, environmental status data, and synchronously collect train dispatcher's physiological monitoring data;

[0170] a feature conversion module configured to convert the scheduling operation time series data into a spatiotemporal joint feature vector;

[0171] LSTM-FCN processing module, configured to extract and fuse spatiotemporal features of spatiotemporal joint feature vectors, physiological monitoring data, and environmental status data and calculate the comprehensive abnormality probability value ;

[0172] Intervention decision module, used when the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the knowledge graph is triggered and the deviation value is calculated The degree of deviation, and according to the deviation value The numerical range of the intervention level is divided into the levels to trigger the graded alarm instructions;

[0173] Physiological status monitoring module, used to collect the train dispatcher's physiological status code vector in real time after triggering the graded alarm instruction;

[0174] The scheduling decision module, if the vector returns to the preset safe range within M consecutive minutes, the alarm is lifted and a status recovery report is generated; otherwise, the current alarm instruction is upgraded to the highest alarm instruction, triggering a manual takeover request.

[0175] Among them, the data acquisition module obtains dispatching operation time series data in real time through the train operation monitoring system. This data is recorded in a timestamp-encoded <operation type, parameter value, response duration> triple structure, accurately representing the spatiotemporal distribution characteristics of dispatching behavior; environmental status data is obtained through IoT sensors, covering dimensional indicators such as track occupancy rate, signal status, wind speed level, etc., which serve as the environmental noise compensation benchmark for anomaly detection; with the help of wearable devices, the train dispatcher's physiological monitoring data is synchronously collected, and the PPG sensor and infrared camera are used to obtain heart rate variability (HRV) and eye tracking feature vectors at a sampling rate of 5Hz to quantify the dispatcher's cognitive load status, thereby forming a comprehensive and complementary data source.

[0176] The feature conversion module uses a temporal convolutional network to map the scheduling operation time series data into a 128-dimensional embedding vector, extracts the time series pattern features, and splices them with the graph attention features of the station topology map to form a joint representation of spatiotemporal coupling, that is, a spatiotemporal joint feature vector, which effectively integrates the time and space characteristics of the scheduling operation and provides high-quality feature input for subsequent processing.

[0177] The LSTM-FCN processing module comprises a parallel processing architecture consisting of a long short-term memory network and a fully convolutional network. The LSTM branch captures long-range dependencies within the operation sequence, extracts temporal hidden states, and characterizes the dynamic evolution of scheduled operations, such as instruction intervals and response delays. The FCN branch extracts spatial anomaly pattern features, generates spatial hidden states, and displays the spatial distribution of operation paths, such as high-frequency click areas and abnormal trajectory inflection points. A dynamic fusion layer, combining an attention weight matrix and a compensation factor matrix, extracts and fuses spatiotemporal features from the joint spatiotemporal feature vector, physiological monitoring data, and environmental status data to calculate a comprehensive anomaly probability. The attention weight matrix dynamically adjusts the feature weights of each time series node based on the standard deviation of the RR interval of the HRV data, enhancing the representation of features during abnormally sensitive periods. The compensation factor matrix, based on the environmental state, calculates the influence coefficient of environmental interference on the feature vector through a multi-factor regression model, eliminating false positives caused by extreme weather conditions.

[0178] When the comprehensive anomaly probability exceeds a dynamically adjusted threshold T calculated via a sliding time window (e.g., using the median of the previous 30 minutes of normal operating data as a benchmark and dynamically adjusting sensitivity based on the characteristics of the scheduling phase), the intervention decision module triggers a graded intervention mechanism based on the train scheduling knowledge graph. The deviation value Δ is calculated, and intervention levels are divided according to the range of Δ values ​​(e.g., Δ∈[0.2,0.5] triggers a yellow warning, Δ∈[0.5,1.0] triggers an orange control, and Δ≥1.0 triggers a red takeover). This in turn triggers the corresponding graded alarm instructions, enabling precise, graded handling of train scheduling anomalies.

[0179] After triggering the graded alarm command, the physiological status monitoring module collects the train dispatcher's two-dimensional physiological status coding vector composed of the LF / HF ratio of HRV and the pupil diameter change rate in real time, uses sliding average filtering to eliminate instantaneous interference, updates the assessment of physiological recovery every 10 seconds, and continuously monitors changes in the dispatcher's physiological status.

[0180] The scheduling decision module determines that if the physiological state encoding vector returns to the preset safety interval within M consecutive minutes, for example, the preset safety interval is LF / HF∈[0.8,1.2] and the pupil fluctuation rate is ≤15% (the interval membership is calculated using a fuzzy logic algorithm), the alarm is lifted and a status recovery report is generated, including operation records during the abnormal period, physiological fluctuation curves, and environmental parameter snapshots. XML structured storage supports post-event traceability analysis; otherwise, the current alarm instruction is upgraded to the highest alarm instruction, triggering a manual takeover request to ensure the safety and reliability of train scheduling.

[0181] Compared with existing technologies, existing train dispatcher behavior detection solutions may suffer from incomplete data collection, limited feature extraction, and an inability to effectively respond to complex environments and individual physiological changes. This system, however, utilizes multiple modules working together. The data acquisition module comprehensively collects multi-source data; the feature conversion module deeply integrates spatiotemporal features; the LSTM-FCN processing module utilizes a hybrid network architecture and a dual-matrix compensation mechanism to effectively extract and integrate spatiotemporal features, improving anomaly detection accuracy; the intervention decision module implements precise, graded intervention; and the physiological status monitoring module and the dispatch decision module form a closed-loop control. This system addresses the challenges of traditional methods in detecting anomalies in train dispatcher behavior, such as insufficient data processing capabilities, poor environmental adaptability, and insufficient consideration of individual differences. It achieves intelligent, precise, and real-time monitoring and intervention of train dispatcher behavior, ensuring safe and efficient train dispatching operations.

[0182] like Figure 5 As shown, the present application further proposes that the train dispatcher behavior anomaly detection system based on LSTM-FCN also includes an abnormal case tracing module for storing abnormal operation fragments that trigger graded intervention instructions and their associated environment and physiological data;

[0183] Visualize the distribution of abnormal features through t-SNE dimensionality reduction algorithm;

[0184] Establish a correlation map of abnormal cases, linking similar historical cases and disposal plans;

[0185] Generate anomaly pattern analysis reports, marking high-frequency anomaly types and spatiotemporal distribution characteristics.

[0186] Intervention effect evaluation module, used to record changes in scheduling operation compliance rates before and after intervention;

[0187] Calculate the correlation between intervention response time and risk reduction;

[0188] Evaluate the effectiveness of interventions through counterfactual reasoning models;

[0189] Feedback the evaluation results to the model training module to establish a closed-loop optimization mechanism.

[0190] The abnormal case tracing module is used to store abnormal operation fragments that trigger graded intervention instructions, along with rich information such as associated environmental and physiological data. The abnormal data is processed using the t-SNE dimensionality reduction algorithm, mapping its high-dimensional features into a low-dimensional space to visualize the distribution of abnormal features. This allows for the creation of an abnormal case association map, which is then linked to similar historical cases and corresponding treatment plans through intelligent algorithms. Finally, an abnormal pattern analysis report is generated, accurately annotating key information such as high-frequency abnormality types and spatiotemporal distribution characteristics, providing important reference and basis for subsequent analysis and processing.

[0191] The intervention effectiveness evaluation module records changes in dispatching compliance rates before and after the intervention to visually reflect improvements in dispatcher behavior. It calculates the correlation between intervention response time and risk reduction to measure the timeliness and effectiveness of the intervention. Furthermore, it uses a counterfactual reasoning model to evaluate the effectiveness of intervention measures from multiple perspectives. The evaluation results are fed back to the model training module, establishing a closed-loop optimization mechanism. This allows the entire system to continuously adjust and optimize its model parameters and strategies based on the actual intervention results, continuously improving its ability to detect and address abnormal train dispatcher behavior.

[0192] Compared with the existing technology, the traditional train dispatcher behavior detection system may have many problems such as asynchronous data collection, incomplete processing, inflexible intervention, and inability to effectively utilize historical cases and perform closed-loop optimization. The present application effectively solves these problems through the coordinated cooperation and innovative design of the above modules. Especially in terms of data synchronization, the time alignment error of multimodal data is reduced from the larger range of traditional solutions, providing a solid foundation for subsequent precise analysis; at the same time, the closed-loop optimization mechanism established through modules such as abnormal case tracing and intervention effect evaluation enables the system to continuously improve and enhance itself, significantly improving the accuracy of detection of abnormal train dispatcher behavior and the effectiveness of processing, ensuring the safety and efficiency of train dispatching.

[0193] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A dynamic intervention method for train dispatcher behavior based on LSTM-FCN, characterized by: The following steps are involved: Acquire the dispatching operation time sequence data and environmental status data of the target train in real time, and simultaneously collect the physiological monitoring data of the target train dispatcher; The scheduling operation time series data is converted into a spatiotemporal joint feature vector, and the spatiotemporal joint feature vector, the physiological monitoring data and the environmental state data are input into the LSTM-FCN hybrid network to obtain a comprehensive abnormality probability value ; Among them, the comprehensive abnormal probability value Combined with the attention weight matrix Combined compensation factor matrix Calculated, the attention weight matrix and the compensation factor matrix The dynamic fusion layer of the LSTM-FCN hybrid network generates the physiological monitoring data and the environmental status data accordingly; When the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the traffic scheduling knowledge graph is triggered and the deviation value is calculated. The degree of deviation, and according to the deviation value The numerical range of the intervention level is divided into intervention levels, and the corresponding graded alarm instructions are triggered according to the intervention level; After triggering the graded alarm instruction, collecting the physiological state coding vector of the target train dispatcher in real time; If the physiological state coding vector returns to the preset safety range within M consecutive minutes, the graded alarm instruction is released and a status recovery report is generated; otherwise, the current graded alarm instruction is upgraded to the highest level alarm instruction, triggering a manual takeover request.

2. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized in that: The method for generating a spatiotemporal joint feature vector comprises: Input the scheduling operation timing data into the pre-trained instruction type embedding matrix according to the instruction sequence, and generate an operation sequence semantic vector by vector splicing; A two-dimensional coordinate system is constructed based on the GPS coordinates of each station in the target section, and the frequency of train arrival or departure within the preset time window is counted to generate a station thermal distribution matrix; The operation sequence semantic vector and the station thermal distribution matrix are fused through a cross-modal attention mechanism to output the spatiotemporal joint feature vector with spatiotemporal correlation.

3. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized in that: In the LSTM-FCN hybrid network, the LSTM branch uses a bidirectional gating mechanism to extract the temporal dependency features of the spatiotemporal joint feature vector and outputs a temporal feature vector with temporal attention weights. Its FCN branch extracts the spatial pattern features of the spatiotemporal joint feature vector through an adaptive convolution kernel and outputs a multi-scale spatial feature vector.

4. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 3 is characterized by: The LSTM branch is processed using the bidirectional gating mechanism: forward processing the spatiotemporal joint feature vector in time to capture forward temporal dependency features; Processing the spatiotemporal joint feature vector in reverse time to capture reverse temporal dependency features; The gated attention mechanism dynamically adjusts the feature weights of each time step of the forward temporal dependency feature and the reverse temporal dependency feature to generate the temporal feature vector with the temporal attention weight. .

5. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized in that: The training method of the LSTM-FCN hybrid network includes: Constructing a standard operation sample set, wherein the standard operation sample set includes a normal operation sequence and environmental state data of a standard operation corresponding point and physiological monitoring data of a train dispatcher at the standard operation corresponding point; Adopting contrastive learning strategy, we generate negative samples of abnormal operations through data augmentation; Designing a composite loss function , where the classification loss Using focal loss, reconstruction loss Use Wasserstein distance; Dynamic adjustment through reinforcement learning , Parameter weights,optimize the model convergence process.

6. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized by: The LSTM-FCN hybrid network is updated based on the federated learning framework, specifically: Extract anonymized operation sequence data from edge nodes of each railway bureau; Freeze the FCN branch parameters and only fine-tune the fully connected layer weights of the LSTM branch; The updated model parameters of the LSTM-FCN hybrid network are synchronized to the global federated learning framework.

7. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized by: The dynamic fusion layer processing includes: performing standardization processing on the physiological monitoring data to generate standardized physiological data; Mapping the normalized physiological data into the attention weight matrix through a fully connected network ; The compensation factor matrix is ​​generated according to the visibility, rainfall intensity and wind speed in the environmental state data. ; The compensation factor matrix The calculation formula is: Among them, V, R, and W represent the normalized values ​​of the environment parameters, and α, β, and γ are learnable parameters.

8. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized by: The method for determining the dynamically adjusted threshold value T includes: Establishing a sliding time window based on historical operation data, and calculating a moving average μ and a standard deviation σ of the comprehensive abnormal probability value within the sliding time window; pass Dynamically adjust the threshold T, where k is a preset sensitivity coefficient; When the number of consecutive alarm triggering exceeds a preset threshold, the preset sensitivity coefficient k value is automatically increased to tighten the threshold.

9. The train dispatcher behavior dynamic intervention method based on LSTM-FCN according to claim 1 is characterized by: The hierarchical intervention mechanism includes: Constructing the traffic scheduling knowledge graph, wherein the traffic scheduling knowledge graph includes equipment topology relationships, operation rules, and emergency plan entities; Retrieving relevant emergency plans from the traffic dispatch knowledge graph to generate a three-level intervention suggestion set, which is converted into executable dispatch instructions through natural language generation technology; The three-tiered set of intervention recommendations includes: The first-level alarm command is judged as a mild abnormality, triggering an interface prompt; The second-level alarm command is judged as a moderate abnormality, triggering a non-critical operation lock; The third-level alarm instruction is judged as a serious abnormality, triggering a manual takeover request.

10. The train dispatcher behavior anomaly detection system based on LSTM-FCN is characterized by: The method for dynamic intervention of train dispatcher behavior based on LSTM-FCN as described in any one of claims 1 to 9 includes: A data acquisition module is used to acquire real-time dispatching operation time series data, environmental status data, and synchronously collect physiological monitoring data of the train dispatcher; a feature conversion module configured to convert the scheduling operation time series data into a spatiotemporal joint feature vector; LSTM-FCN processing module, configured to extract and fuse the spatiotemporal features of the spatiotemporal joint feature vector, physiological monitoring data and environmental status data and calculate the comprehensive abnormality probability value ; Intervention decision module, used when the comprehensive abnormal probability value When the dynamic adjustment threshold T is exceeded, the hierarchical intervention mechanism based on the knowledge graph is triggered and the deviation value is calculated The degree of deviation, and according to the deviation value The numerical range of the intervention level is divided into the levels to trigger the graded alarm instructions; A physiological status monitoring module, configured to collect the physiological status coding vector of the train dispatcher in real time after the hierarchical alarm instruction is triggered; The scheduling decision module cancels the alarm and generates a status recovery report if the vector returns to the preset safe range within M consecutive minutes; otherwise, the current alarm instruction is upgraded to the highest alarm instruction, triggering a manual takeover request.

Images