Intelligent Inspection Method and System Based on Dynamic Confidence and Duplicate Case Filtering

Abstract

This invention provides an intelligent inspection method and system based on dynamic confidence and intelligent filtering of duplicate cases, comprising the following specific steps: Step 1: Obtain the identification result and original identification confidence of new cases; construct a feature vector for new cases based on the new cases; input the feature vector of new cases into a dynamic threshold prediction model to obtain the optimal confidence threshold; if the original identification confidence is higher than the optimal confidence threshold, the case is recorded in the waiting queue; Step 2: Based on a duplicate judgment rule base including spatial distance threshold, time interval threshold, and image similarity threshold, determine whether the new case is a duplicate case with the cases in the waiting queue; if it is a duplicate case, filter the new case. This invention significantly reduces the workload of manual review of duplicate cases and low-confidence suspected cases through automated filtering and parameter self-evolution mechanism, and significantly reduces the cost of manual review.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology, and in particular to an intelligent inspection method and system based on dynamic confidence and intelligent filtering of duplicate cases. Background Technology

[0002] Currently, intelligent urban inspection systems are widely used in urban management, traffic monitoring, and security patrols. Existing intelligent inspection systems primarily rely on front-end data collection devices (such as vehicles, drones, and robots) to capture images via cameras, utilize deep learning models to identify various urban management issues (such as street vending, illegal parking, and garbage dumping), and report the identification results to the back-end system for review and processing. However, existing technical solutions have the following problems and shortcomings:

[0003] 1. Poor filtering effect of fixed confidence threshold: Traditional systems use fixed confidence thresholds for case filtering, which cannot adapt to the complex changes in different scenarios, time periods, and device types. For example, in low-light environments at night, the recognition confidence is generally low, and a fixed threshold will cause a large number of valid cases to be falsely filtered out; while in bright daylight environments, a fixed threshold cannot effectively filter false alarms.

[0004] 2. Low accuracy in identifying duplicate cases: Existing systems rely mainly on simple time interval and spatial distance rules to judge duplicate cases, lacking comprehensive consideration of multi-dimensional factors such as device type, case type, and image similarity. This results in a large number of duplicate cases when the same problem is reported multiple times by different devices or when the same device passes by multiple times, increasing the cost of manual review.

[0005] 3. Filtering rules cannot evolve automatically: Traditional filtering rules are configured manually and cannot be automatically optimized and adjusted based on historical audit data and user feedback; as the urban environment changes and equipment is updated, fixed rules gradually become ineffective, thus requiring frequent manual intervention and maintenance, resulting in high labor costs. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and to provide an intelligent inspection method and system based on dynamic confidence and intelligent filtering of duplicate cases.

[0007] The objective of this invention is achieved through the following technical solution: an intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases, comprising the following specific steps: Step 1: Obtain the identification results and original identification confidence of the new case. Construct the feature vector of the new case based on the new case. Input the feature vector of the new case into the dynamic threshold prediction model to obtain the optimal confidence threshold. When the original identification confidence is higher than the optimal confidence threshold, the case is recorded in the queue to be reviewed. The dynamic threshold prediction model is trained using historical false alarm negative samples as the training dataset. Step 2: Based on a duplicate judgment rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold, determine whether the new case is a duplicate case with the cases in the pending review queue. If it is a duplicate case, filter the new case. The spatial distance threshold, time interval threshold, and image similarity threshold in the duplicate judgment rule base are updated periodically based on duplicate cases.

[0008] As a preferred approach, for duplicate cases, the earliest case of the same type as the duplicate case within the historical time and space range is set as the earliest primary case, and duplicate case association pairs are constructed by linking the earliest primary case with the corresponding duplicate case; the duplicate case association pairs are periodically clustered to generate candidate optimization parameters for the duplicate judgment rule base, and the parameters of the duplicate judgment rule base are optimized based on the candidate optimization parameters.

[0009] As a preferred method, the confidence threshold prediction model is constructed as follows: S1: Set the sliding window length of historical review data, filter historical cases that were falsely reported within the sliding window length of historical review data as historical false report negative samples, construct the feature vector of the false report negative samples, where the historical cases that were falsely reported are those whose identification results were manually reviewed and marked as incorrect, and take the feature vector of the false report negative samples and the corresponding original output confidence to construct the training dataset. S2: Construct a confidence threshold prediction model based on the XGBoost regression model, and train the dynamic threshold prediction model using the training dataset.

[0010] As a preferred approach, for each new case, threshold comparisons are performed regarding spatial distance and time interval, as well as a determination of whether they are the same case type. Candidate main cases are selected from the queue to be reviewed. Image similarity thresholds are compared between the selected candidate main cases and the new case. If the image similarity threshold comparison conditions are met, the new case is determined to be a duplicate case and is automatically filtered out.

[0011] As a preferred embodiment, in step two, a twin network is used to extract deep semantic features of the case images of the candidate main case and the new case, and the cosine similarity between the two is calculated. The cosine similarity is compared with the image similarity threshold. If the cosine similarity is greater than or equal to the image similarity threshold, the image similarity threshold comparison condition is met.

[0012] As a preferred approach, six data points are obtained from the inspection equipment, including equipment type, case category, time period, day / night status, grid area code, and weather status. Based on these six data points, a feature vector for the new case is constructed.

[0013] As a preferred method, the specific approach for generating candidate optimization parameters for the duplicate detection rule base using cluster analysis and optimizing the parameters of the duplicate detection rule base based on these candidate optimization parameters is as follows: Unsupervised clustering is performed on the recurring case association pairs within a set time range according to the first time period to form typical recurring pattern clusters based on different clustering features, including the spherical distance between the main case and the recurring case, the time interval between the main case and the recurring case, and the image similarity between the main case and the recurring case. Extract candidate optimization parameter sets for the judgment rule base from typical repetition pattern clusters; conduct case repetition judgment tests on the current parameter repetition judgment rule base and the repetition judgment rule base replaced by the candidate optimization parameter set using the validation set, and record the judgment accuracy of each respectively; if the improvement in accuracy of the repetition judgment rule base replaced by the candidate optimization parameter set is greater than or equal to the set improvement threshold, then replace the parameters in the current repetition judgment rule base using the candidate optimization parameter set.

[0014] Preferably, the candidate optimization parameter set includes: The candidate spatial distance threshold is the arithmetic mean of the spherical distances between all primary cases and repeating cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the spherical distances within that typical repeating pattern cluster.

[0015] The candidate time interval threshold is the arithmetic mean of the reported time intervals between all primary cases and duplicate cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the time intervals within that cluster.

[0016] The candidate image similarity threshold is the arithmetic mean of the cosine similarity between all main cases and repeating cases within a typical repeating pattern cluster, minus the standard deviation of the cosine similarity of images within that typical repeating pattern cluster.

[0017] As a preferred approach, the periodic update method for the spatial distance threshold, time interval threshold, and image similarity threshold base in the duplicate judgment rule base is as follows: Aggregate all newly added duplicate case association pairs within the current update period according to the second time period, and group the newly added duplicate case association pairs according to device type and case category; calculate the feature distribution of spatial distance, time interval and image similarity for each group of newly added duplicate case association pairs, and obtain the parameter update rules for the above feature distribution; update the parameters of the duplicate judgment rule base based on the parameter update rules.

[0018] An intelligent inspection system based on dynamic confidence level and intelligent filtering of duplicate cases includes: The front-end acquisition module is used to collect images through various inspection equipment, identify and generate cases, and report case metadata and original output confidence scores. The filtering service layer includes a dynamic confidence filtering unit, a duplicate case determination unit, and an image feature extraction unit. The dynamic confidence filtering unit is used to predict the optimal confidence threshold in the current scenario and to perform confidence screening on newly reported cases based on the optimal confidence threshold in the current scenario. The duplicate case determination unit determines whether a new case is a duplicate case based on a duplicate determination rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold; The manual review module is used to review the filtered cases and mark the review results; The self-learning module is used to train a confidence threshold prediction model based on historical case data, mine duplicate case association pairs, perform cluster analysis, and update and optimize the duplicate judgment rule base parameters. The database module is used to store case data, historical review data, model parameters, and duplicate judgment rule base parameters.

[0019] The beneficial effects of this invention are: 1. This invention integrates six-dimensional feature vectors, including equipment type, case category, time period, day / night status, grid area, and weather conditions, to establish a periodic self-learning mechanism. It utilizes a sliding time window to automatically mine false negative samples in historical reviews and dynamically calculates and updates the optimal confidence threshold under different scenario combinations through machine learning algorithms to achieve adaptive filtering.

[0020] 2. This invention adopts a dual closed-loop filtering system of "rule engine + data self-learning". By monitoring user review behavior, it automatically establishes a "main case - duplicate case" association record, periodically calculates the spatiotemporal distribution pattern, and dynamically updates five key filtering parameters, including device type characteristics, case type sensitivity, spatial distance threshold, time interval window, and image similarity ratio.

[0021] 3. This invention utilizes cluster analysis to perform unsupervised cluster analysis on historical duplicate case data, automatically derives the optimal filtering parameters, and achieves the "self-evolution" of the parameters.

[0022] 4. This invention significantly reduces the workload of manual review of duplicate cases and low-confidence suspected cases through automated filtering and parameter self-evolution mechanism, significantly reducing the cost of manual review. The system can automatically filter more than 80% of invalid cases, improving review efficiency by 3-5 times. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0025] Those skilled in the art should understand that, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0026] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0027] like Figure 1 As shown, the intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases includes the following specific steps: Step 1: Obtain the identification results and original identification confidence of the new case. Construct the feature vector of the new case based on the new case. Input the feature vector of the new case into the dynamic threshold prediction model to obtain the optimal confidence threshold. When the original identification confidence is higher than the optimal confidence threshold, the case is recorded in the queue to be reviewed. The dynamic threshold prediction model is trained using historical false alarm negative samples as the training dataset.

[0028] The confidence threshold prediction model is constructed as follows: S1: Set the sliding window length of historical review data, filter historical cases that were falsely reported within the sliding window length of historical review data as historical false report negative samples, construct the feature vector of the false report negative samples, where the historical cases that were falsely reported are those whose identification results were manually reviewed and marked as incorrect, and take the feature vector of the false report negative samples and the corresponding original output confidence to construct the training dataset. Each historical case includes the following six types of metadata, which are encoded into the following six types of feature vectors: Equipment type d: such as two-wheeled vehicles, four-wheeled inspection vehicles, drones, fixed cameras, robots, etc., using one-hot encoding, with dimensions Nd∈[3,10]; Case category c: Urban management problem type (such as street vending, illegal parking, garbage dumping, etc.), using embedding vectors; Time period t h 24-hour clock, t h ∈[0,23]; Day / night state L: calculated from the light sensor or time + latitude and longitude, with values ​​of {0 (night), 1 (day)}; Grid area coding g: Divide the city into M×M grids (M∈[100,500]), and assign a unique ID to each grid; Weather conditions w: Weather condition information comes from the meteorological API. Weather conditions are classified into 5 categories: sunny, cloudy, rainy, snowy, and foggy, and one-hot encoded.

[0029] The final feature vector set F = [d,c,th,L,g,w] ∈ R D , where the total dimension D∈[30,70].

[0030] Among them, the length T of the sliding window for historical review data is set. win The time frame is [3,6] months; cases marked as "error" or "ignored" by manual review are automatically selected as negative false alarm samples; and each negative sample is associated and stored with its corresponding feature vector Fi and original output confidence score pi ∈ [0,1]. The original output confidence score of each negative sample is calculated by the algorithm module deployed on the inspection equipment based on the current environmental characteristics.

[0031] S2: A confidence threshold prediction model is constructed based on the XGBoost regression model, and a dynamic threshold prediction model is trained using the training dataset. The input of this model is the feature vector set of the current scene; the output is the optimal confidence threshold for this scene.

[0032] The training objective of the confidence threshold prediction model is to minimize the sum of squares of the differences between the model-predicted threshold and the true optimal threshold for all training samples.

[0033] The confidence threshold prediction model is automatically retrained once a week, using the latest false positive negative sample data from the false positive negative sample library.

[0034] In this application, the sliding window length for historical review data is 4 months; the number of XGBoost regression model trees is set to 100–300, and the learning rate is set to 0.05–0.2.

[0035] After training the confidence threshold prediction model, new reported cases are filtered in real time based on their confidence levels.

[0036] When a new case is reported, the feature vector F of the new case is extracted. new The new case obtains six data elements, including the equipment type, case category, time period, day / night status, grid area code, and weather conditions of the inspection equipment. Based on these six data elements, a feature vector F is constructed for the new case. new .

[0037] The feature vector F of the new case new Input the data into the trained confidence threshold prediction model, and output the optimal confidence threshold τ for the current scenario. dyn ; If the original identification confidence level p of the new case new <τ dyn If the case is found to be false, it will be automatically filtered into the "low confidence case database"; otherwise, it will be placed in the pending review queue.

[0038] In this intelligent inspection solution, the original recognition confidence of new cases is directly output by the deep learning target detection model deployed on front-end acquisition devices (inspection vehicles, drones, fixed cameras, etc.) or edge computing nodes. The essence of the original recognition confidence is the product of the target existence probability and the category probability after the model extracts and classifies the features of the input image.

[0039] Step 2: Based on a duplicate judgment rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold, determine whether the new case is a duplicate case with the cases in the pending review queue. If it is a duplicate case, filter the new case. The spatial distance threshold, time interval threshold, and image similarity threshold in the duplicate judgment rule base are updated periodically based on duplicate cases.

[0040] In this step, the duplicate detection rule base includes spatial distance threshold, time interval threshold, and image similarity threshold. The initial settings for these thresholds are as follows: The initial value of the spatial distance threshold is r0 = 50 ± 10 meters; The initial value of the time interval threshold is Δt0 = 30 ± 5 minutes; The initial value of the image similarity threshold is s0 = 0.70 ± 0.05.

[0041] For each new case, threshold comparisons are performed on spatial distance and time interval, and it is determined whether the case is of the same type. Candidate main cases are selected from the queue to be reviewed. Image similarity threshold comparisons are performed between the selected candidate main cases and the new case. If the image similarity threshold comparison conditions are met, the new case is determined to be a duplicate case and is automatically filtered.

[0042] Specifically, a twin network is used to extract deep semantic features from case images and calculate cosine similarity. The cosine similarity is then compared with an image similarity threshold. If the cosine similarity is greater than or equal to the image similarity threshold, the image similarity threshold comparison condition is met.

[0043] When using Siamese networks to extract deep semantic features from case images and calculate cosine similarity, a Siamese network feature extractor is first constructed. This Siamese network feature extractor uses a dual-branch ResNet-50 as the backbone network and shares weights. The input image size is 224×224×3 mm. The output is a 512-dimensional L2 normalized feature vector, which contains 512 consecutive floating-point values, each of which represents a deep semantic feature of the image.

[0044] The loss function for training the Siamese network feature extractor is contrastive loss, which is as follows: L=y D 2 +(1 y) max(0,m D) 2 ; Where D represents the two input samples (x1 and x2) processed by the feature extractor f( The distance after (x1) is given by D = ||f(x1) f(x2)||2; m is a boundary value, a hyperparameter that controls the minimum separation distance of negative sample pairs. In this application, the value of m is 1.0.

[0045] y is the label used to indicate whether a sample pair is a similar sample pair, y∈{0,1}; when the sample pair is a similar pair, the value of y is 1.

[0046] The training data consists of historical case image pairs, and the image labels are determined by manual review results (duplicate / non-duplicate).

[0047] For example: For new case N: Search the pending review queue for candidate main cases that meet the following three conditions: Condition 1: Same case category; Condition 2: The spatial distance between the candidate main case and the new case is less than or equal to the current spatial distance threshold; Condition 3: The time interval between the candidate main case and the new case is less than or equal to the current time interval threshold; For each candidate main case that simultaneously meets the above judgment conditions, calculate the image similarity s between the candidate main case and the new case N; If any candidate main case satisfies s ≥ scurrent(dN,c), then N is determined to be a duplicate case, and the new case is automatically filtered. scurrent(dN,c) is the current image similarity threshold.

[0048] For duplicate cases, the earliest case of the same type as the duplicate case within the historical time and space range is set as the earliest primary case. The earliest primary case and the corresponding duplicate case are used to construct duplicate case association pairs. The duplicate case association pairs are periodically clustered to generate candidate optimization parameters for the duplicate judgment rule base. The parameters of the duplicate judgment rule base are optimized based on the candidate optimization parameters.

[0049] When constructing duplicate case association pairs, when an auditor marks a case A as "duplicate," the system traces back to cases of the same type within a defined historical time range (within the past 7 days) and spatial range (radius of 200 meters); selects the earliest case as the earliest primary case M, constructs an association pair (M,A) between the earliest primary case M and the corresponding duplicate case; and records the following information for the duplicate case association pair (M,A): The spherical distance (Δd) between two cases in a duplicate case association pair is calculated using the latitude and longitude of the two cases; The time interval (Δt) between two cases in a duplicate case association pair; Image similarity (s) between two cases in a duplicate case association pair.

[0050] The specific method for generating candidate optimization parameters for the duplicate detection rule base using cluster analysis and optimizing the parameters of the duplicate detection rule base based on the candidate optimization parameters is as follows: Unsupervised clustering is performed on the recurring case association pairs within a set time range according to the first time period to form typical recurring pattern clusters based on different clustering features, including the spherical distance between the main case and the recurring case, the time interval between the main case and the recurring case, and the image similarity between the main case and the recurring case. Extract candidate optimization parameter sets for the judgment rule base from typical repetition pattern clusters; conduct case repetition judgment tests on the current parameter repetition judgment rule base and the repetition judgment rule base replaced by the candidate optimization parameter set using the validation set, and record the judgment accuracy of each respectively; if the improvement in accuracy of the repetition judgment rule base replaced by the candidate optimization parameter set is greater than or equal to the set improvement threshold, then replace the parameters in the current repetition judgment rule base using the candidate optimization parameter set.

[0051] The clustering analysis in this application is based on the DBSCAN unsupervised clustering algorithm. Daily, it extracts the most recent 30-day duplicate case association pairs from the historical duplicate case database. An initial neighborhood radius ∈ [15, 30] (Euclidean distance) is set; and a minimum sample size MinPts ∈ [5, 20]. Typical duplicate pattern clusters are the core output of the DBSCAN unsupervised clustering algorithm on urban intelligent patrol duplicate case data. These clusters refer to the dense set of data points automatically identified from historical "primary case - duplicate case" association pairs, possessing highly similar spatiotemporal and image features. Each cluster corresponds to a real and frequently occurring duplicate reporting scenario in urban patrol operations.

[0052] In this application, the first time period is 1 day.

[0053] The candidate optimization parameter set includes: Candidate spatial distance threshold r cluster , is the arithmetic mean of the spherical distances between all primary cases and repeating cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the spherical distances within that typical repeating pattern cluster. In this embodiment, the set multiple is 1.5 times.

[0054] Candidate time interval threshold Δt cluster This is the arithmetic mean of the reporting time intervals between all primary and duplicate cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the time intervals within that cluster. In this embodiment, this set multiple is 1.5 times.

[0055] Candidate image similarity threshold s cluster , is the arithmetic mean of the image cosine similarity between all main cases and repeating cases within a typical repeating pattern cluster, minus the standard deviation of the image cosine similarity within that typical repeating pattern cluster.

[0056] The accuracy improvement ΔAcc of the duplicate detection rule base after replacing it with the candidate optimized parameter set is: ΔAcc=Acc new Acc old ; In the formula, Acc new The accuracy of the duplicate decision rule base after replacement with the candidate optimized parameter set; Acc old This represents the accuracy of the duplicate detection rule base for the current parameter.

[0057] In this application, the threshold for raising the threshold is set to 2%. If ΔAcc ≥ 2%, the new parameter is adopted and synchronized to the duplicate detection rule base.

[0058] Specifically, the method for statistically analyzing the spatiotemporal and image feature distributions of duplicate case association pairs and updating the parameters of the duplicate case determination rule base based on these spatiotemporal and image feature distributions is as follows: Aggregate all newly added duplicate case association pairs within the current update period according to the second time period, and group the newly added duplicate case association pairs according to device type and case category; calculate the feature distribution of spatial distance, time interval and image similarity for each group of newly added duplicate case association pairs, and obtain the parameter update rules for the above feature distribution; update the parameters of the duplicate judgment rule base based on the parameter update rules.

[0059] When calculating the feature distributions of spatial distance, time interval, and image similarity, histograms or kernel density estimation (KDE) can be used. The feature distributions of real duplicate cases, grouped by (equipment type, case category), are transformed into intuitive distribution graphs and computable probability density functions. This objectively reflects the actual occurrence patterns of spatial distance, time interval, and image similarity of duplicate cases in a specific scenario, providing a reliable data foundation for subsequent accurate threshold updates.

[0060] The parameter update rules are as follows: The updated spatial distance threshold is the 85th percentile of the spherical distance between all primary and duplicate cases in the spatial distance feature distribution data, ranging from [20, 100] meters; The updated time interval threshold is the 90th percentile of the time interval between all primary and duplicate cases in the time distance feature distribution data; the range is [10, 120] minutes. The updated image similarity threshold is the 10th percentile of the image similarity between all primary and duplicate cases in the feature distribution data of image similarity. The range is [0.60, 0.85].

[0061] In this application, the second time period is one week.

[0062] An intelligent inspection system based on dynamic confidence level and intelligent filtering of duplicate cases includes: The front-end acquisition module is used to collect images through various inspection equipment, identify and generate cases, and report case metadata and original output confidence scores. The filtering service layer includes a dynamic confidence filtering unit, a duplicate case determination unit, and an image feature extraction unit. The dynamic confidence filtering unit is used to predict the optimal confidence threshold in the current scenario and to perform confidence screening on newly reported cases based on the optimal confidence threshold in the current scenario. The duplicate case determination unit determines whether a new case is a duplicate case based on a duplicate determination rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold; The manual review module is used to review the filtered cases and mark the review results; The self-learning module is used to train a confidence threshold prediction model based on historical case data, mine duplicate case association pairs, perform cluster analysis, and update and optimize the duplicate judgment rule base parameters. The database module is used to store case data, historical review data, model parameters, and duplicate judgment rule base parameters.

[0063] This invention employs a front-end device to report cases, with the filtering service layer sequentially performing: dynamic confidence filtering, duplicate case determination, and feature extraction. Cases that pass the filtering process enter a manual review queue, and the review results are fed back to the database. Simultaneously, the system triggers monthly XGBoost regression model retraining, daily cluster analysis, and weekly updates to the duplicate case determination rule parameters, ultimately forming a closed loop of "collection → filtering → review → learning → optimization."

[0064] Compared with the prior art, the present invention has the following advantages: 1. This invention integrates six-dimensional feature vectors, including equipment type, case category, time period, day / night status, grid area, and weather conditions, to establish a periodic self-learning mechanism. It utilizes a sliding time window to automatically mine false negative samples in historical reviews and dynamically calculates and updates the optimal confidence threshold under different scenario combinations through machine learning algorithms to achieve adaptive filtering.

[0065] 2. This invention adopts a dual closed-loop filtering system of "rule engine + data self-learning". By monitoring user review behavior, it automatically establishes a "main case - duplicate case" association record, periodically calculates the spatiotemporal distribution pattern, and dynamically updates five key filtering parameters, including device type characteristics, case type sensitivity, spatial distance threshold, time interval window, and image similarity ratio.

[0066] 3. This invention utilizes cluster analysis to perform unsupervised cluster analysis on historical duplicate case data, automatically derives the optimal filtering parameters, and achieves the "self-evolution" of the parameters.

[0067] 4. This invention significantly reduces the workload of manual review of duplicate cases and low-confidence suspected cases through automated filtering and parameter self-evolution mechanism, significantly reducing the cost of manual review. The system can automatically filter more than 80% of invalid cases, improving review efficiency by 3-5 times.

[0068] This invention is not limited to the preferred embodiments described above. Anyone can derive other products in various forms under the guidance of this invention. However, regardless of any changes in shape or structure, any technical solution that is the same as or similar to this application falls within the protection scope of this invention.

Claims

1. An intelligent inspection method based on dynamic confidence level and intelligent filtering of duplicate cases, characterized in that, The specific steps include the following: Step 1: Obtain the identification results and original identification confidence of the new case. Construct the feature vector of the new case based on the new case. Input the feature vector of the new case into the dynamic threshold prediction model to obtain the optimal confidence threshold. When the original identification confidence is higher than the optimal confidence threshold, the case is recorded in the queue to be reviewed. The dynamic threshold prediction model is trained using historical false alarm negative samples as the training dataset. Step 2: Based on a duplicate judgment rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold, determine whether the new case is a duplicate case with the cases in the pending review queue. If it is a duplicate case, filter the new case. The spatial distance threshold, time interval threshold, and image similarity threshold in the duplicate judgment rule base are updated periodically based on duplicate cases.

2. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 1, characterized in that, For duplicate cases, the earliest case of the same type as the duplicate case within the historical time and space range is set as the earliest primary case. The earliest primary case and the corresponding duplicate case are used to construct duplicate case association pairs. The duplicate case association pairs are periodically clustered to generate candidate optimization parameters for the duplicate judgment rule base. The parameters of the duplicate judgment rule base are optimized based on the candidate optimization parameters.

3. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 1, characterized in that, The confidence threshold prediction model is constructed as follows: S1: Set the sliding window length of historical review data, filter historical cases that were falsely reported within the sliding window length of historical review data as historical false report negative samples, construct the feature vector of the false report negative samples, where the historical cases that were falsely reported are those whose identification results were manually reviewed and marked as incorrect, and take the feature vector of the false report negative samples and the corresponding original output confidence to construct the training dataset. S2: Construct a confidence threshold prediction model based on the XGBoost regression model, and train the dynamic threshold prediction model using the training dataset.

4. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 1, characterized in that, For each new case, threshold comparisons are performed on spatial distance and time interval, and it is determined whether the case is of the same type. Candidate main cases are selected from the queue to be reviewed. Image similarity threshold comparisons are performed between the selected candidate main cases and the new case. If the image similarity threshold comparison conditions are met, the new case is determined to be a duplicate case and is automatically filtered.

5. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 4, characterized in that, In step two, a Siamese network is used to extract deep semantic features from the case images of the candidate main case and the new case, and the cosine similarity between the two is calculated. The cosine similarity is compared with the image similarity threshold. If the cosine similarity is greater than or equal to the image similarity threshold, the image similarity threshold comparison condition is met.

6. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 1, characterized in that, Based on the six data elements of the new case, including the equipment type, case category, time period, day and night status, grid area code, and weather status of the inspection equipment, a feature vector of the new case is constructed.

7. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 2, characterized in that, The specific method for generating candidate optimization parameters for the duplicate detection rule base using cluster analysis and optimizing the parameters of the duplicate detection rule base based on the candidate optimization parameters is as follows: Unsupervised clustering is performed on the recurring case association pairs within a set time range according to the first time period to form typical recurring pattern clusters based on different clustering features, including the spherical distance between the main case and the recurring case, the time interval between the main case and the recurring case, and the image similarity between the main case and the recurring case. Extract candidate optimization parameter sets for the judgment rule base from typical repetition pattern clusters; conduct case repetition judgment tests on the current parameter repetition judgment rule base and the repetition judgment rule base replaced by the candidate optimization parameter set using the validation set, and record the judgment accuracy of each respectively; if the improvement in accuracy of the repetition judgment rule base replaced by the candidate optimization parameter set is greater than or equal to the set improvement threshold, then replace the parameters in the current repetition judgment rule base using the candidate optimization parameter set.

8. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 7, characterized in that, The candidate optimization parameter set includes: The candidate spatial distance threshold is the arithmetic mean of the spherical distances between all primary cases and repeating cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the spherical distances within that typical repeating pattern cluster. The candidate time interval threshold is the arithmetic mean of the reported time intervals between all primary cases and duplicate cases within a typical repeating pattern cluster, plus a set multiple of the standard deviation of the time intervals within that cluster. The candidate image similarity threshold is the arithmetic mean of the cosine similarity between all main cases and repeating cases within a typical repeating pattern cluster, minus the standard deviation of the cosine similarity of images within that typical repeating pattern cluster.

9. The intelligent inspection method based on dynamic confidence and intelligent filtering of duplicate cases according to claim 2, characterized in that, The periodic update methods for the spatial distance threshold, time interval threshold, and image similarity threshold base in the duplicate detection rule base are as follows: Aggregate all newly added duplicate case association pairs within the current update period according to the second time period, and group the newly added duplicate case association pairs according to device type and case category; calculate the feature distribution of spatial distance, time interval and image similarity for each group of newly added duplicate case association pairs, and obtain the parameter update rules for the above feature distribution; update the parameters of the duplicate judgment rule base based on the parameter update rules.

10. An intelligent inspection system based on dynamic confidence level and intelligent filtering of duplicate cases, characterized in that, include: The front-end acquisition module is used to collect images through various inspection equipment, identify and generate cases, and report case metadata and original output confidence scores. The filtering service layer includes a dynamic confidence filtering unit, a duplicate case determination unit, and an image feature extraction unit. The dynamic confidence filtering unit is used to predict the optimal confidence threshold in the current scenario and to perform confidence screening on newly reported cases based on the optimal confidence threshold in the current scenario. The duplicate case determination unit determines whether a new case is a duplicate case based on a duplicate determination rule base that includes spatial distance threshold, time interval threshold, and image similarity threshold; The manual review module is used to review the filtered cases and mark the review results; The self-learning module is used to train a confidence threshold prediction model based on historical case data, mine duplicate case association pairs, perform cluster analysis, and update and optimize the duplicate judgment rule base parameters. The database module is used to store case data, historical review data, model parameters, and duplicate judgment rule base parameters.

Images