A Multi-Source Traffic Data Quality Assessment and Alarm Method Based on a Time-Series Large Model

Abstract

This application relates to the field of urban intelligent transportation technology, and in particular to a method for quality assessment and alarm of multi-source traffic data based on a time-series large-scale model. The method includes: acquiring historical multi-source traffic flow data for a target road, associating and filtering these data, using a dynamic time warping algorithm to obtain the confidence level of each pair of associated devices, and selectively retaining data based on the confidence level to obtain a historical reliable traffic flow dataset; training a time-series large-scale model based on the historical reliable traffic flow dataset, and inputting the multi-source traffic flow data from the previous time period into the trained time-series large-scale model to obtain the predicted multi-source traffic flow data and the predicted difference value between the two sources at the current time; acquiring the actual multi-source traffic flow data and the difference value between the two sources at the current time, comparing them with the predicted values ​​at the current time, and generating a multi-dimensional anomaly score; and issuing a data anomaly alarm based on the multi-dimensional anomaly score. This application significantly improves the accuracy and real-time response capability of anomaly detection.

Description

Technical Field

[0001] This application relates to the field of urban intelligent transportation technology, and in particular to a method for quality assessment and alarm of multi-source traffic data based on a time-series large model. Background Technology

[0002] With the development of urbanization and intelligent transportation systems, the number of traffic monitoring devices has surged and the types of data collected have become increasingly diverse, covering various types of data such as traffic flow, GPS floating car data, and meteorological information. These multi-source heterogeneous data provide support for real-time perception of traffic operation status, but due to the expansion in scale and complexity of types, higher requirements are placed on data analysis and processing capabilities.

[0003] In practical applications, raw data often exhibits abnormal fluctuations, noise interference, or even complete distortion due to factors such as sensor malfunctions, communication interruptions, severe weather, or sudden traffic incidents. Failure to identify and process such abnormal data in a timely manner will directly impact the decision-making effectiveness of the traffic control system, leading to problems such as inaccurate signal timing schemes, delayed congestion identification, and incorrect route guidance strategies, thereby reducing overall traffic management efficiency and travel service levels.

[0004] To address the aforementioned issues, existing technologies often employ anomaly detection methods based on statistical analysis, threshold rules, or single time series models. These methods typically rely on statistical distribution assumptions (such as Z-score, IQR, etc.) or pre-defined rules to identify outliers in the data, or introduce models such as moving averages, exponential smoothing, or even ARIMA into univariate time series analysis for bias detection. However, due to the high dimensionality, nonlinearity, and strong correlations among multiple variables in traffic data, these methods still have significant limitations in practical applications. On the one hand, traditional methods struggle to effectively capture the complex spatiotemporal dependencies in traffic data; on the other hand, because they fail to fully model the collaborative change mechanisms among multiple source variables, they often perform poorly when dealing with multidimensional, cross-modal data, easily resulting in false positives and false negatives, thus hindering the overall improvement of anomaly identification performance.

[0005] Therefore, there is an urgent need for a multi-source traffic data quality assessment and alarm method based on a time-series large model. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a multi-source traffic data quality assessment and alarm method based on a time-series large model, which solves the technical problems in the prior art of being unable to effectively capture the complex spatiotemporal correlation characteristics of multi-source traffic data, insufficient multi-variable collaborative change relationships, and high false alarm and false negative rates.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the main technical solutions adopted in this application include:

[0010] This application provides a method for quality assessment and alarm of multi-source traffic data based on a time-series large model, including:

[0011] S100. Obtain historical multi-source traffic flow data of the target road, associate and filter the historical multi-source traffic flow data of the target road, and obtain initial reliable data; the historical multi-source traffic flow data includes historical radar detection traffic flow data and historical checkpoint detection traffic flow data.

[0012] S200. Using a dynamic time warping algorithm, obtain the confidence level of each pair of associated devices in the initial trusted data, and retain the initial trusted data differently according to the confidence level to obtain a historical trusted traffic dataset.

[0013] S300. Train a time series large model based on historical reliable traffic data set, and input the multi-source traffic flow from the previous time period into the trained time series large model to obtain the multi-source traffic flow prediction data and the dual-source difference prediction value at the current time.

[0014] S400: Obtain the current multi-source traffic actual flow and dual-source difference value, compare them with the current multi-source traffic flow prediction data and dual-source difference prediction value, generate a multi-dimensional anomaly score, and issue a data anomaly alarm based on the multi-dimensional anomaly score.

[0015] Optionally, in some embodiments of this application, S100 includes:

[0016] S110. Obtain historical multi-source traffic flow of the target road. Based on the intersection-radar device association and intersection-checkpoint association of the target road, establish radar-checkpoint association. Based on the radar-checkpoint association, associate the traffic flow of different devices at the same location during the same time period to obtain the associated historical multi-source traffic flow.

[0017] S120. According to the pre-set traffic filtering rules, remove traffic from the associated historical multi-source traffic flow that does not conform to the traffic filtering rules, and retain the traffic that conforms to the traffic filtering rules as initial reliable data.

[0018] Optionally, in some embodiments of this application, step S200 includes:

[0019] S210. Normalize the initial trusted data by day to obtain the normalized initial trusted data.

[0020] S220. The dynamic time warping algorithm is used to calculate the trend consistency distance between the radar and the normalized traffic sequence at the same location. If the trend consistency distance is greater than a preset distance threshold, the corresponding data is removed from the normalized initial reliable data. If the trend consistency distance is less than or equal to the preset distance threshold, the data pair is retained, and the corresponding dual-source difference value for each time period is obtained.

[0021] S230. Based on the dual-source difference value, combined with the time-series stability index and the inter-source consistency index, obtain the comprehensive confidence level of each data point, and perform differential retention based on the comprehensive confidence level to obtain a historical reliable traffic dataset.

[0022] Optionally, in some embodiments of this application, in S230,

[0023] The historical reliable traffic data set includes historical radar-detected traffic flow data, historical checkpoint-detected traffic flow data, and corresponding dual-source difference values ​​after differential preservation.

[0024] Optionally, in some embodiments of this application, S230 includes:

[0025] The time-series stability index and the inter-source consistency index are weighted and fused to obtain the comprehensive confidence level of each data point;

[0026] The temporal stability index is the ratio of the number of time periods that meet the stability requirements to the total number of time periods. The stability requirements are: |current traffic - mean of historical data for the same period| < 3 times the standard deviation. The inter-source consistency index is 1 - min(dual-source difference value / maximum acceptable difference, 1).

[0027] Optionally, in some embodiments of this application, S230 further includes:

[0028] When the overall confidence level is high, all the corresponding normalized initial confidence data are saved; when the overall confidence level is medium, the corresponding normalized initial confidence data are saved on a monthly rolling basis; when the overall confidence level is low, the corresponding normalized initial confidence data are saved on a weekly rolling basis.

[0029] Optionally, in some embodiments of this application, S300 includes:

[0030] Extract multi-scenario sample data from the historical reliable traffic dataset to construct a multivariate time-series training dataset containing basic features, derived features, and time features; use the multivariate time-series training dataset to fine-tune the time-series large model to obtain the trained time-series large model.

[0031] The basic features are historical radar-detected traffic flow data and historical checkpoint-detected traffic flow data; the derived features are dual-source difference; the time features include time period features, weekday features, and holiday features.

[0032] Optionally, in some embodiments of this application, the time features are incorporated into a multivariate time series training dataset after encoding processing, and the encoding processing method is one-hot encoding, sine-cosine encoding, or binary encoding.

[0033] Optionally, in some embodiments of this application, S400 includes:

[0034] S410. Obtain the actual multi-source traffic flow at the current moment, and obtain the dual-source difference value based on the multi-source traffic flow at the current moment;

[0035] The dual-source difference value = |current checkpoint flow rate - current radar flow rate| / max(current checkpoint flow rate, current radar flow rate);

[0036] S420. Compare the current multi-source traffic actual flow and dual-source difference value with the current multi-source traffic flow prediction data and dual-source difference prediction value to generate a multi-dimensional anomaly score.

[0037] The multi-dimensional anomaly scoring includes checkpoint anomaly scoring, radar anomaly scoring, and difference anomaly scoring.

[0038] S430. Issue a data anomaly alarm based on the multi-dimensional anomaly score.

[0039] Optionally, in some embodiments of this application, S430 includes:

[0040] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is less than the first difference threshold, it is judged as a traffic incident anomaly, the alarm level is low, and video review and traffic incident push response are triggered.

[0041] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is less than the second radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a checkpoint equipment fault anomaly, the alarm level is medium, and a maintenance work order is generated and a data source switching response is triggered.

[0042] When the checkpoint anomaly score is less than the second checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a radar equipment fault anomaly, the alarm level is medium, and a maintenance work order is generated and a data source switching response is triggered.

[0043] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a dual-device fault anomaly, the alarm level is high, and a maintenance work order is generated and a manual emergency verification response is triggered.

[0044] Other scoring combinations are judged as anomalies to be verified, with a low alarm level, triggering a routine manual verification response.

[0045] (III) Beneficial Effects

[0046] The multi-source traffic data quality assessment and alarm method based on a time-series large model provided in this application introduces a multi-source data credibility collaborative verification mechanism, further improving data quality by utilizing the spatiotemporal consistency relationship between data. Subsequently, the multivariate time-series large model is fine-tuned based on cleaned, high-quality data to better reflect the traffic flow evolution patterns in specific scenarios. Finally, the model predicts future traffic flow, and the prediction results are compared with real-time collected data to achieve anomaly detection and alarm. This method significantly improves the accuracy and real-time response capability of anomaly detection, exhibits higher system stability and application adaptability, and provides reliable technical support for data quality assurance and operational security of intelligent transportation systems. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a multi-source traffic data quality assessment and alarm method based on a time-series large model according to an embodiment of this application;

[0048] Figure 2 The traffic flow curve when the trend consistency distance of the multi-source traffic data quality assessment and alarm method based on a time-series large model according to an embodiment of this application is less than or equal to a preset distance threshold;

[0049] Figure 3 The traffic flow curve is shown when the trend consistency distance of the multi-source traffic data quality assessment and alarm method based on a time-series large model according to an embodiment of this application is greater than a preset distance threshold. Detailed Implementation

[0050] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0051] With the acceleration of urbanization and the in-depth development of intelligent transportation systems, the number of traffic monitoring devices has increased significantly, and the collected multi-source traffic data has become increasingly abundant, providing a fundamental support for real-time perception of urban traffic operation status. However, in practical applications, traffic data is susceptible to factors such as sensor failure, communication interference, and sudden environmental changes, resulting in abnormal fluctuations or distortions. If such abnormal data is not identified and processed in a timely manner, it can easily mislead traffic control system decisions, leading to problems such as unreasonable signal timing, failure of congestion warnings, and incorrect route guidance, seriously affecting traffic efficiency and the quality of travel services.

[0052] Traditional traffic data anomaly detection methods mostly rely on statistical analysis, rule matching, or single time series models. When faced with high-dimensional, nonlinear, and strongly correlated multi-source traffic data, they have significant limitations: on the one hand, it is difficult to effectively capture the complex spatiotemporal correlation characteristics of the data; on the other hand, it is impossible to accurately model the collaborative change relationship between multiple variables, resulting in frequent false alarms and missed alarms during the anomaly detection process, which makes it difficult to meet the high requirements of intelligent transportation systems for data quality and anomaly response.

[0053] To address the aforementioned technical issues, this application proposes a multi-source traffic data quality assessment and alarm method based on a large time-series model. This method is implemented through the following core steps: First, historical multi-source traffic flows are correlated, filtered, and their quality assessed. Data is retained differentially based on confidence levels to construct a historical reliable traffic dataset. Second, a Transformer-based multivariate time-series prediction model is fine-tuned based on this dataset. Basic features, derived features, and encoded time features are incorporated into the model training process to improve its adaptability to complex traffic scenarios. Subsequently, the multi-source traffic flows from the previous time period are input into the trained model to obtain the current traffic prediction data and the predicted dual-source difference value. Finally, by comparing the current actual multi-source traffic flow, dual-source difference, and corresponding predicted values, a multi-dimensional anomaly score is generated based on checkpoint, radar, and difference scores. An alarm is then triggered based on the score combination.

[0054] Through the above-described scheme, this application can effectively eliminate false anomaly data. Combined with the learning ability of a multivariate time-series large model to complex traffic patterns, it significantly reduces the false alarm rate and false negative rate of anomaly detection, improving detection accuracy and robustness. Simultaneously, the time-series large model based on the Transformer architecture can accurately capture the spatiotemporal correlation and collaborative change characteristics of multi-source data, adapting to various traffic scenarios such as weekdays, weekends, and holidays, further ensuring the reliability of prediction and anomaly judgment. Furthermore, this application achieves accurate differentiation and alarm of anomaly types through multi-dimensional scoring, providing maintenance personnel with clear decision-making basis, facilitating priority handling of critical faults, optimizing resource scheduling, and reducing maintenance costs. At the same time, through response actions such as data source switching and automatic generation of maintenance work orders, it improves the efficiency of anomaly handling, providing reliable technical support for data quality assurance and operational security of intelligent transportation systems.

[0055] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0056] Example 1

[0057] Figure 1 This is a flowchart illustrating a multi-source traffic data quality assessment and alarm method based on a time-series large model according to an embodiment of this application, as shown below. Figure 1 As shown, the multi-source traffic data anomaly alarm includes:

[0058] Step S100: Obtain historical multi-source traffic flow data for the target road, correlate and filter the historical multi-source traffic flow data for the target road, and obtain initial reliable data; historical multi-source traffic flow data includes historical radar-detected traffic flow data and historical checkpoint-detected traffic flow data.

[0059] Step S100 includes:

[0060] Step S110: Obtain historical multi-source traffic flow of the target road. Based on the intersection-radar device association and intersection-checkpoint association of the target road, establish radar-checkpoint association. Based on the radar-checkpoint association, associate the traffic flow of different devices at the same location during the same time period to obtain the associated historical multi-source traffic flow.

[0061] In the specific implementation process, multi-source traffic flow data for n consecutive days is obtained. The value of n is adjusted according to the complexity of the traffic scenario. For example, 90 days are used for core urban intersections to cover more traffic fluctuation scenarios, while 30 days are sufficient for suburban intersections. The data source is the traffic monitoring system database, and the obtained multi-source traffic flow data must include basic fields such as device ID, vehicle passage timestamp, and detection location (intersection ID + entrance ID).

[0062] Furthermore, when acquiring traffic flow, fixed time intervals are used as statistical units to count the number of vehicles passing through a single radar device and a single checkpoint device within each time interval, obtaining basic statistical data of "device ID - time period - traffic flow value". For example, the number of vehicles passing through radar device ID "1400012345" during the time period of 08:00:00-08:05:00 on 2025-01-01 is counted as radar traffic flow, and the number of vehicles passing through checkpoint device ID "5419000037" during the same time period is counted as checkpoint traffic flow.

[0063] The following table shows an example of device flow rate in one embodiment:

[0064]

[0065] In addition, when acquiring multi-source traffic flow, there may be data gaps, such as when equipment is temporarily offline and there is no traffic data for a certain period. In this case, a weighted average method is used to fill the gaps: the traffic data of the two adjacent normal periods before and after the missing period are used as the basis, and the weights are set according to the principle that "the closer to the missing period, the greater the weight". The weighted average is calculated as the traffic filling value for the missing period. If the missing period is at the beginning of the data sequence, the traffic value of the adjacent normal period is used to fill the gaps.

[0066] After acquiring multi-source traffic flow data, two types of basic association tables are constructed based on the equipment ledgers provided by the traffic management department, and the radar-checkpoint pairing relationship is established accordingly:

[0067] The intersection-radar equipment association table is used to specify the physical location of each radar device, and the intersection-checkpoint equipment association table is used to specify the physical location of each checkpoint device.

[0068] The radar-checkpoint device association table links the two types of association tables mentioned above. The fields include "intersection ID, entrance ID, radar device ID, and checkpoint device ID", ensuring that each pair of equipped devices corresponds to the same entrance at the same intersection and can collect the same traffic flow data.

[0069] Furthermore, based on the traffic flow data after the radar-checkpoint pairing table, time period matching needs to be performed to ensure that the associated historical multi-source traffic flow data fields include: intersection ID, entrance ID, radar device ID, checkpoint device ID, radar flow, checkpoint flow, time period start time, time period end time, and data collection status, ensuring that each record can be traced back to a specific device, specific time period, and specific location.

[0070] In step S110 above, by relying on the intersection-radar equipment association table and the intersection-checkpoint equipment association table, and using "intersection ID + entrance ID" as the matching key, a radar-checkpoint pairing relationship is established. This can accurately associate radar traffic flow data and checkpoint traffic flow data at the same intersection entrance and during the same time period, effectively solving the problems of scattered multi-source traffic data and mismatch in spatiotemporal dimensions. This lays a solid foundation for subsequent collaborative operations such as using dynamic time warping algorithms to verify the consistency of multi-source data trends. At the same time, by filling missing traffic data through weighted averaging, data loss caused by factors such as temporary equipment offline can be reduced, ensuring the continuity of historical multi-source traffic flow data and avoiding interference from missing data in subsequent screening and quality assessment stages.

[0071] Step S120: According to the pre-set traffic filtering rules, remove traffic from the associated historical multi-source traffic traffic that does not conform to the traffic filtering rules, and retain the traffic that conforms to the traffic filtering rules as initial reliable data.

[0072] Specifically, the pre-defined traffic filtering rules are as follows:

[0073] The flow rate is within a reasonable range, which is defined as [0, maximum inlet flow rate]. The maximum inlet flow rate = number of lanes × lane saturation flow rate × 1.5.

[0074] The deviation between the current traffic value and the historical average for the same period is within a reasonable range. Here, a reasonable range means that |current traffic - historical average for the same period| < 3 times the standard deviation.

[0075] If the traffic filtering rules are not met, a message will be displayed indicating that the data is suspected to be abnormal and requires manual verification. The data will be discarded, and the traffic that meets the traffic filtering rules will be retained as the initial reliable data.

[0076] Optionally, data repair can also be performed on traffic that does not conform to the traffic filtering rules, and the repaired traffic can be added to the initial trusted data; specific repair methods include:

[0077] (1) When the abnormal data involves only 1-2 consecutive time periods, and the traffic data for 3-5 time periods before and after the abnormality all meet the filtering rules:

[0078] If the abnormal period is singular, linear interpolation is used, taking the flow rate F from the preceding normal period. t-1 The flow rate F compared to the subsequent normal period t+1 Repair using the following formula:

[0079] After repair, the flow rate F t =(F t+1 +F t-1 ) / 2;

[0080] If the abnormal period consists of two consecutive periods, or if the traffic flow fluctuates significantly between the abnormal and normal periods, take the data from two normal periods before and after the abnormal period, set the weight according to the principle of "the closer to the abnormal period, the greater the weight", and calculate the weighted average as the repair value.

[0081] (2) When the concurrent flow data of another device paired at the same location (such as radar malfunction or checkpoint normal) meets the screening rules, the spatiotemporal consistency of multi-source device data is used for repair, that is, the proportional coefficient method or deviation correction method is used for repair.

[0082] (3) When the abnormal period is located at the beginning or end of the data sequence, such as the first period of the first day or the last period of the last day, and there is no adjacent normal data, the regularity of historical data of the same period is used for repair, that is, the mean repair method or the median repair method is adopted.

[0083] After the repair is completed, the two filtering rules of step S120 are re-executed on the repaired traffic data. Only if both rules are met can the repaired data be included in the initial trusted data. If the rules are not met after repair, the data is determined to be "unrepairable abnormal data" and is discarded according to the rules.

[0084] In step S120, by using the dual rules of "reasonable flow range" and "deviation from historical average", obvious abnormal data that exceeds the maximum inlet flow or deviates from the historical average by more than 3 times the standard deviation can be quickly removed, thus avoiding such distorted data from interfering with subsequent multi-source collaborative verification and reducing the risk of false alarms and missed alarms from the source.

[0085] Step S200: Using a dynamic time warping algorithm, obtain the confidence level of each pair of associated devices in the initial trusted data, and perform differentiated retention of the initial trusted data according to the confidence level to obtain a historical trusted traffic dataset;

[0086] Step S200 includes:

[0087] Step S210: Normalize the initial trusted data by day to obtain the normalized initial trusted data;

[0088] In the specific implementation process, the initial reliable data undergoes daily normalization. This process aligns radar and checkpoint traffic flow data from the same intersection and on the same date according to a 24-hour time series, forming paired time series samples. To eliminate interference caused by differences in the absolute value of traffic flow between different road segments or different dates, and to highlight the morphological characteristics reflecting the trend of traffic flow changes, the system performs normalization on the daily traffic flow series. The normalization method is adaptively selected based on the data distribution characteristics. For data with large traffic fluctuations but strong trend regularity, Z-score normalization is used to concentrate the data distribution around the mean; for scenarios where the consistency of trend shape is more important, the Min-Max method is used to compress the data into a uniform interval.

[0089] Step S220: Calculate the trend consistency distance between the radar and checkpoint normalized traffic sequences at the same location using a dynamic time warping algorithm; if the trend consistency distance is greater than a preset distance threshold (e.g., ... Figure 3 As shown), the corresponding data is removed from the normalized initial reliable data; if the trend consistency distance is less than or equal to the preset distance threshold (e.g. Figure 2 (As shown), then retain the data pair and obtain the corresponding dual-source difference value for each time period;

[0090] Specifically, the dynamic time warping algorithm can effectively address potential local delays or rhythm differences between two sets of devices on the timeline. It calculates the minimum cumulative deviation between two sequences using a non-linear alignment method, thereby measuring the similarity of their overall trends. In this embodiment, the application calculates a normalized trend consistency distance for the daily sequences of each pair of devices. The smaller the distance, the closer the intraday trends of the two source data are. If the trend consistency distance of a certain day's data exceeds a preset distance threshold, the trend of the dual-source data for that day is considered severely deviated and is deemed unreliable data, thus being removed. If the distance is within the preset distance threshold range, the data for that day is retained, and the dual-source difference values ​​for each time period are further extracted for subsequent confidence assessment.

[0091] The calculation method for the two-source difference value is as follows:

[0092] Dual-source difference = |gate flow - radar flow| / maximum value of both;

[0093] The method for obtaining the preset distance threshold is as follows:

[0094] During the initialization phase, an initial threshold is constructed based on multi-source traffic flow data collected during the historical observation period. Specifically, continuous multi-day data from several typical intersections under normal operating conditions, and after manual verification confirming no equipment malfunctions or traffic incident interference, are selected. The dynamic time-normalized distances of the daily normalized radar and checkpoint traffic flow sequences are calculated, and the overall distribution of these distance values ​​is statistically analyzed. Based on this, a robust statistical method is employed, taking the 90th percentile of this distance sequence as the initial preset distance threshold to ensure that, under normal trend consistency, the vast majority of valid data will not be mistakenly excluded. This initial threshold is used as the baseline parameter for system startup.

[0095] When the method in this application begins execution, a dynamic iterative update mechanism runs continuously online. Data pairs deemed "high confidence level" are periodically selected from newly accessed data. These data pairs must meet conditions such as preliminary trend consistency assessment, normal multi-dimensional anomaly scores, and no triggered device alarms, forming a new high-quality sample set. Subsequently, the trend consistency distance of all data pairs in this dynamic sample set is recalculated, and their distribution characteristics are statistically analyzed using a sliding window approach, focusing on extracting the median and median absolute deviation to avoid interference from extreme values ​​in threshold updates. The new preset distance threshold is composed of the current median plus a certain multiple of the MAD (median absolute deviation), forming a dynamic boundary with statistical control limits. When the difference between the newly calculated threshold and the currently used value exceeds a preset percentage, the threshold update process is automatically triggered, and a change log is recorded for audit tracking.

[0096] Furthermore, to prevent drastic fluctuations in the threshold within a short period, a smooth transition strategy is introduced, gradually replacing the old and new thresholds using a weighted progressive approach rather than abrupt switching. Additionally, this application establishes upper and lower limit protection mechanisms to prevent threshold distortion caused by abnormal batch data.

[0097] Through the aforementioned dynamic iterative update mechanism, the preset distance threshold can be adaptively adjusted according to the evolution of traffic flow patterns, equipment performance degradation, or environmental changes, effectively improving the accuracy and robustness of data quality assessment, ensuring the continuous high-quality construction of historical reliable traffic datasets, and providing a reliable data foundation for the training of subsequent time-series large models and anomaly alarms.

[0098] Step S230: Based on the dual-source difference value, combined with the time-series stability index and the inter-source consistency index, obtain the comprehensive confidence level of each data point, and perform differential retention based on the comprehensive confidence level to obtain a historical reliable traffic dataset.

[0099] The historical reliable traffic data set includes historical radar-detected traffic flow data, historical checkpoint-detected traffic flow data after differential preservation, and the corresponding dual-source differential values.

[0100] Specifically, step S230 includes:

[0101] The time series stability index and the inter-source consistency index are weighted and fused to obtain the comprehensive confidence level of each data point;

[0102] The temporal stability index is the ratio of the number of time periods that meet the stability requirements to the total number of time periods. Meeting the stability requirements means that |current traffic - historical mean data for the same period| < 3 times the standard deviation; the inter-source consistency index = 1 - min(dual-source difference value / maximum acceptable difference, 1).

[0103] When the overall confidence level is high, all the corresponding normalized initial confidence data are saved; when the overall confidence level is medium, the corresponding normalized initial confidence data are saved on a monthly rolling basis; when the overall confidence level is low, the corresponding normalized initial confidence data are saved on a weekly rolling basis.

[0104] In step S230, the reliability of each data point is first comprehensively judged based on the calculated dual-source difference value, combined with the time-series stability index and the inter-source consistency index. The time-series stability index measures the operational stability of a detection device within a specific time period. It is calculated as follows: for each data point's time period (e.g., a specific time interval within a day), the current flow rate value for that period is counted to see if it falls within three standard deviations above and below the historical average for the same period. If this condition is met, it is considered to meet the stability requirements. Finally, the ratio of the number of time periods meeting this condition to the total number of time periods within a single day is calculated to obtain the time-series stability score for that day. A higher score indicates smaller operational fluctuations of the device on that day, and more stable and reliable data.

[0105] Meanwhile, the inter-source consistency index is used to assess the degree of agreement between radar and checkpoint detection sources. The calculated dual-source difference value for each time period is compared with a preset maximum acceptable difference threshold. A smaller difference indicates closer data from both sources and higher consistency. This index is quantified by dividing the dual-source difference value by the maximum acceptable difference, taking the minimum value as upper limit of 1, and then subtracting the ratio from 1, generating a score between 0 and 1. A score close to 1 indicates high consistency between the two sources, while a score close to 0 indicates significant deviation. The system integrates these two indices using a weighted approach, assigning a higher weight to temporal stability to reflect the importance of long-term stable operation of a single source, ultimately generating a comprehensive confidence level for each data point.

[0106] Furthermore, this application compares the current day's overall confidence level with the median and dispersion of the device's confidence levels over the past 30 days. If the current score is higher than the median plus 1.5 times the median absolute deviation, it is classified as high-level; if it is lower than the median minus 1.5 times the median absolute deviation, it is classified as low-level; otherwise, it is classified as medium-level.

[0107] In practical applications, data rated "high" in overall confidence indicates stable trends and high consistency between the two sources, classifying it as a high-quality sample. This data, including normalized radar traffic sequences, checkpoint traffic sequences, and corresponding dual-source difference values, is preserved in its entirety for long-term archiving and can be used for subsequent model training, performance backtesting, and algorithm optimization. Data rated "medium" still possesses some reference value despite fluctuations, and is retained on a monthly rolling basis, keeping only data from typical periods each month to conserve storage resources. Data rated "low" indicates poor stability or significant dual-source deviation, potentially influenced by equipment malfunctions or external interference. Only the most recent week's complete data is retained for short-term verification and problem tracing; the original sequences are automatically cleaned up after this period. This mechanism ensures data quality while achieving efficient use of storage resources, ultimately constructing a well-structured and quality-controlled historical reliable traffic dataset, providing solid data support for subsequent large-scale time-series model training and anomaly alerts.

[0108] Step S300: Train a time series large model based on the historical reliable traffic data set, and input the multi-source traffic flow from the previous time period at the current moment into the trained time series large model to obtain the multi-source traffic flow prediction data and the dual-source difference prediction value at the current moment.

[0109] Step S300 includes:

[0110] Extract multi-scenario sample data from historical reliable traffic datasets to construct a multivariate time series training dataset containing basic features, derived features, and time features; use the multivariate time series training dataset to fine-tune the time series large model to obtain a well-trained time series large model.

[0111] The basic features are historical radar-detected traffic flow data and historical checkpoint-detected traffic flow data, and the derived features are dual-source difference; the time features include time period features, weekday features, and holiday features.

[0112] The time features are encoded and then incorporated into the multivariate time series training dataset. The encoding methods are one-hot encoding, sine-cosine encoding, or binary encoding.

[0113] Specifically, the large-scale time series model includes:

[0114] A multimodal feature embedding layer is used to perform hierarchical embedding processing on multiple types of input features: For basic features (radar traffic, checkpoint traffic) and derived features (dual-source difference), a learnable numerical embedding layer is used to map the normalized continuous values ​​to a high-dimensional feature space through linear transformation, preserving the numerical magnitude relationship; For encoded time features (time period, weekday, holiday), a category embedding layer is used to assign a unique embedding vector to each discrete category, and different time granularity features are fused through residual connections and layer normalization to strengthen the expression of temporal periodicity.

[0115] The spatiotemporal attention encoder consists of six stacked encoder blocks, each composed of the following sub-modules: a multi-head spatiotemporal self-attention mechanism sub-module, a position-enhanced feedforward network sub-module, and a residual linking and dropout sub-module. The position-enhanced feedforward network sub-module employs a two-layer linear transformation plus a GELU activation function, introducing a local feature nonlinear transformation after the attention output, and simultaneously incorporating temporal position encoding based on sine and cosine functions to strengthen the temporal sequence relationship.

[0116] The multi-target prediction decoder uses four stacked decoder blocks to achieve multi-step prediction based on the context features of the encoder output. For three prediction targets—radar traffic, checkpoint traffic, and dual-source difference—independent output branches are specially designed. Each branch generates the prediction result of the corresponding target through linear transformation and Softplus activation function (adapted to non-negative traffic values). The branches share decoder features but have independent parameters, realizing differentiated learning between targets.

[0117] Additionally, during training, a pre-configured time-series model is loaded, the underlying model parameters are frozen, and only the top-level prediction module parameters are unfrozen. The data is divided into training, validation, and test sets in chronological order. The training configuration is fine-tuned, and iterative training begins. The loss curve is monitored in real time during training, and the trends of key metrics are recorded. Training ends when the stopping condition is met, and the model is validated and evaluated on the test set.

[0118] For example, this application divides the training set according to a 7:2:1 ratio, that is, the first 70% of the historical reliable data time series is used as the training set, the middle 20% as the validation set, and the last 10% as the test set. The division maintains the complete natural day granularity to avoid the break in the continuity of time series caused by cross-day cutting.

[0119] During optimization, the AdamW optimizer was used with an initial learning rate of 1e-4. Cosine annealing decay was triggered when the validation set loss did not decrease for 5 consecutive rounds. The batch size was set to 32 based on the GPU memory. If the memory was insufficient, gradient accumulation was used. The loss function was MAE (mean absolute error) + L2 regularization, with MAE weight of 0.7 and L2 regularization weight of 0.3, to suppress model overfitting.

[0120] Training will automatically stop when the validation set loss does not decrease within 100 rounds, or the test set accuracy decreases by more than 2%, and the model weights at the point of lowest loss will be saved as the final training model.

[0121] Additionally, it should be noted that the above process does not involve training a completely new large model from scratch, but rather fine-tuning a pre-trained large temporal model that already possesses basic temporal modeling capabilities and a multimodal feature processing framework.

[0122] The aforementioned time-series large model, by integrating the long-term time-series modeling capabilities of Transformer with multimodal feature processing mechanisms, can capture long-term trends in traffic flow, such as weekday / weekend patterns, and learn collaborative change patterns of dual-source data, such as the dynamic deviation between radar and checkpoint traffic flow, providing an accurate predictive benchmark for subsequent anomaly detection.

[0123] Step S400: Obtain the current multi-source traffic actual flow and dual-source difference value, compare them with the current multi-source traffic flow prediction data and dual-source difference prediction value, generate a multi-dimensional anomaly score, and issue a data anomaly alarm based on the multi-dimensional anomaly score.

[0124] Step S400 includes:

[0125] Step S410: Obtain the actual multi-source traffic flow at the current time, and obtain the dual-source difference value based on the multi-source traffic flow at the current time;

[0126] Dual-source difference value = |current checkpoint flow value - current radar flow value| / max(current checkpoint flow value, current radar flow value);

[0127] Step S420: Compare the current multi-source traffic actual flow and dual-source difference value with the current multi-source traffic flow prediction data and dual-source difference prediction value to generate a multi-dimensional anomaly score.

[0128] Multi-dimensional anomaly scoring includes checkpoint anomaly scoring, radar anomaly scoring, and difference anomaly scoring;

[0129] Step S430: Issue a data anomaly alarm based on the multi-dimensional anomaly score.

[0130] When scoring, it is necessary to calculate the absolute error between the actual and predicted values, and also to calculate the standard deviation of the error. Specific scoring methods include:

[0131] Checkpoint anomaly score = |actual traffic flow at checkpoint - predicted traffic flow at checkpoint| / standard deviation of checkpoint prediction error;

[0132] Radar anomaly score = |Actual radar flow - Predicted radar flow| / Standard deviation of radar prediction error;

[0133] Anomaly score = |Actual two-source difference - Predicted two-source difference| / Predicted two-source difference standard deviation;

[0134] Based on the above scores, abnormal situations are categorized, a four-dimensional decision matrix is ​​constructed, and situations such as traffic incidents, checkpoint malfunctions, radar malfunctions, dual-device malfunctions, and situations requiring manual verification are identified. Data anomaly alarms are then triggered, and corresponding responses are implemented, as shown in the table below:

[0135]

[0136] Specifically, the high, medium, and low ratings in the table above are obtained using the following method:

[0137] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is less than the first difference threshold, it is judged as a traffic incident anomaly, the alarm level is low, and video review and traffic incident push response are triggered.

[0138] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is less than the second radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a checkpoint equipment fault anomaly, the alarm level is medium, and a maintenance work order is generated and a data source switching response is triggered.

[0139] When the checkpoint anomaly score is less than the second checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a radar equipment fault anomaly, the alarm level is medium, and a maintenance work order is generated and a data source switching response is triggered.

[0140] When the checkpoint anomaly score is greater than or equal to the first checkpoint threshold, the radar anomaly score is greater than or equal to the first radar threshold, and the difference anomaly score is greater than or equal to the second difference threshold, it is determined to be a dual-device fault anomaly, the alarm level is high, and a maintenance work order is generated and a manual emergency verification response is triggered.

[0141] Other scoring combinations are judged as anomalies to be verified, with a low alarm level, triggering a routine manual verification response.

[0142] When a checkpoint or radar equipment malfunction is detected and triggers a medium-level alarm, the system automatically upgrades the data from the other functioning device to the primary data source and sends a data source switching command to the traffic signal control system to ensure downstream applications are unaffected. Simultaneously, a maintenance task is automatically generated through the integrated maintenance work order system, assigned to the corresponding maintenance personnel, and notified via SMS or app. In the case of a high-level dual-device malfunction alarm, an emergency plan is activated simultaneously, retrieving video surveillance footage from the intersection for AI event identification (such as congestion or accidents), and sending an alarm pop-up to the command center. All alarm events are recorded in the audit log, supporting subsequent traceability and model feedback optimization.

[0143] This application presents a multi-source traffic data quality assessment and alarm method based on a time-series large-scale model. Through a multi-source data credibility collaborative verification mechanism, it effectively integrates spatiotemporal consistency information from different sources and cross-validates the initial screening results from a single source. This enables more accurate identification and removal of false anomalies and interfering data, significantly reducing the inherent false alarm risk of a single data source, thereby significantly lowering both the false alarm and false negative rates. Simultaneously, by fine-tuning the time-series large-scale model using cleaned, high-quality, multi-source fused data, the model can more accurately learn real traffic flow evolution patterns in complex traffic scenarios. The comparative analysis with real-time data serves as a more accurate anomaly criterion, significantly improving the overall accuracy of anomaly detection. Furthermore, this application classifies anomalies through anomaly scoring to generate alarms, providing operation and maintenance personnel with clearer and more operational decision-making support. This facilitates prioritizing critical issues, optimizing resource scheduling, and reducing operation and maintenance costs.

[0144] It is important to note that the anomaly warning logic in this application is not based on the predicted value output by the time-series large model itself, but rather uses the predicted value as the "dynamic normal baseline" for the current traffic flow. This baseline is dynamically generated by the model based on the real-time multi-source traffic flow in the previous time period at the current moment, which can accurately match the real-time evolution trend of the current traffic scenario, rather than relying on fixed historical statistical thresholds. Subsequently, this dynamic prediction baseline is compared in real time with the multi-source traffic flow and dual-source difference value collected at the current moment. Anomalies are identified by calculating whether the deviation between the two exceeds the normal fluctuation range. This "real-time prediction-real-time comparison" mechanism fundamentally avoids the lag that may exist in "early warning based solely on predicted values".

[0145] In summary, this application establishes a complete multi-source traffic data quality assurance system by constructing a multi-source data credibility assessment mechanism, designing a time-series large-scale model that integrates the differences between two sources, and implementing multi-dimensional anomaly scoring and graded response. This solution is not only applicable to the collaborative verification of radar and checkpoint data, but can also be extended to the fusion anomaly detection of various traffic sensing sources such as loop detectors, video, and floating cars. It possesses good versatility and promotional value, and can be widely applied in scenarios such as urban traffic operation monitoring, intelligent signal control, and traffic big data governance, providing core technical support for the safe, reliable, and efficient operation of intelligent transportation systems.

[0146] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0147] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0148] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0149] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0150] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for multi-source traffic data quality assessment and alarm based on a timing large model, characterized in that, The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device.

2. The method of claim 1, wherein the method further comprises: The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device.

3. The method of claim 1, wherein the method further comprises: The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device.

4. The method of claim 3, wherein the method further comprises: The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device.

5. The method of claim 3, wherein the method further comprises: The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction method and device. The application relates to a traffic flow prediction The time sequence stability index is the ratio of the number of time periods meeting the stability requirement to the total number of time periods, and the meeting the stability requirement is satisfying |current flow-historical simultaneous period data mean|<3 times standard deviation; the inter-source consistency index=1-min(bisource difference value / maximum acceptable difference, 1).

6. The method of claim 5, wherein the method further comprises: The S230 further comprises: When the comprehensive confidence level is high, all corresponding normalized initial trusted data are saved; when the comprehensive confidence level is medium, corresponding normalized initial trusted data are saved monthly; when the comprehensive confidence level is low, corresponding normalized initial trusted data are saved weekly.

7. The method of claim 1, wherein the method further comprises: The S300 comprises: Extracting multi-scene sample data from the historical trusted traffic flow data set, constructing a multivariate time sequence training data set containing basic features, derived features and time features; fine-tuning the time sequence large model using the multivariate time sequence training data set to obtain a trained time sequence large model; The basic features are historical radar detected traffic flow data and historical loop detected traffic flow data, and the derived features are bisource difference; the time features include time period features, week features and holiday features.

8. The multi-source traffic data quality evaluation and alarm method based on a time sequence large model according to claim 7, characterized in that, The time features are included in the multivariate time sequence training data set after encoding processing, and the encoding processing mode is one-hot encoding, sine-cosine encoding or binary encoding.

9. The method of claim 1, wherein the method further comprises: The S400 comprises: S410, obtaining multi-source traffic actual flow at the current time, and obtaining bisource difference value according to the multi-source traffic flow at the current time; The bisource difference value=|current loop flow value-current radar flow value| / max(current loop flow value, current radar flow value); S420, comparing the multi-source traffic actual flow at the current time and the bisource difference value with multi-source traffic flow prediction data at the current time and bisource difference prediction value to generate multi-dimensional abnormal score; The multi-dimensional abnormal score includes loop abnormal score, radar abnormal score and difference abnormal score; S430, data abnormality alarm is performed according to the multi-dimensional abnormal score.

10. The method of claim 9, wherein the method further comprises: The S430 comprises: When the loop abnormal score is greater than or equal to the first loop threshold value, the radar abnormal score is greater than or equal to the first radar threshold value, and the difference abnormal score is less than the first difference threshold value, it is determined as traffic event type abnormality, the alarm level is low, and video review and traffic event pushing response are triggered; When the loop abnormal score is greater than or equal to the first loop threshold value, the radar abnormal score is less than the second radar threshold value, and the difference abnormal score is greater than or equal to the second difference threshold value, it is determined as loop equipment failure type abnormality, the alarm level is medium, and maintenance work order generation and data source switching response are triggered; When the loop abnormal score is less than the second loop threshold value, the radar abnormal score is greater than or equal to the first radar threshold value, and the difference abnormal score is greater than or equal to the second difference threshold value, it is determined as radar equipment failure type abnormality, the alarm level is medium, and maintenance work order generation and data source switching response are triggered; When the camera abnormal score is greater than or equal to the first camera threshold value, the radar abnormal score is greater than or equal to the first radar threshold value, and the difference degree abnormal score is greater than or equal to the second difference degree threshold value, it is determined that the dual-device fault type anomaly occurs, the alarm level is high, a maintenance work order is triggered to be generated, and an artificial emergency verification response is triggered; Other score combinations are determined to be to-be-verified anomalies, the alarm level is low, and an artificial routine verification response is triggered.

Images