Casting liquid leakage detection method based on big data model and cross-modal interaction

Abstract

The application discloses a casting liquid leakage detection method based on a big data model and cross-modal interaction, relates to the technical field of industrial detection, and has the advantages that in the system operation, casting monitoring data and mold furnace batch identification data are acquired, and a structure area suspected liquid leakage identification result is generated, which is beneficial to corresponding abnormality to a specific structure area; the structure area suspected liquid leakage identification result is compared with a liquid leakage alarm condition and a suspected abnormality memory condition, an unalarmed suspected abnormality memory event is generated, which is beneficial to retaining data with hidden danger value that does not trigger an alarm; the unalarmed suspected abnormality memory event is written into a structure area hidden danger memory library and forms an unalarmed suspected abnormality memory sequence, which is beneficial to tracking multiple furnace batch abnormalities in the same area; a structure area hidden danger memory result is generated by analyzing repeatedly appearing features, and a liquid leakage hidden danger early warning result is output, which is beneficial to discovering progressive liquid leakage hidden dangers in advance.

Description

Technical Field

[0001] This invention relates to the field of industrial testing technology, specifically to a method for detecting casting leakage based on big data models and cross-modal interaction. Background Technology

[0002] In the casting production process, especially in the manufacturing of complex structural parts such as automotive transmission housings, gearbox housings, and differential housings, molten metal needs to complete the filling, pressure holding, and solidification processes according to a preset path. Due to the fast production cycle, complex mold structure, and frequent changes in on-site working conditions, production companies usually use multi-source information such as visual data, temperature data, sound data, and equipment operation data to monitor and analyze the pouring status in order to promptly detect abnormalities in the production process and ensure the quality of castings and production safety.

[0003] Existing methods for detecting casting leakage typically involve real-time analysis of multi-source data. When the detection results meet preset alarm conditions, a leakage alarm is triggered. However, data that does not meet the alarm conditions but exhibits minor abnormal characteristics is often treated as normal data. For example, localized short-term temperature rises at the parting surface, intermittent abnormal acoustic changes in the gate area, or occasional abnormal hot spots on the mold edge may be identified as suspected anomalies by the detection model, but because the degree of a single anomaly is low, it cannot form an effective alarm. Therefore, this abnormal information is usually not continuously monitored and utilized in subsequent detection processes.

[0004] In fact, before some leakage accidents occur, the same structural area of ​​the same mold often repeatedly exhibits the above-mentioned suspected abnormal phenomena that do not meet the alarm conditions in multiple consecutive production furnaces. Although these abnormalities do not immediately trigger leakage alarms, their frequency and duration will gradually increase with mold wear, changes in structural gaps, or a decline in local sealing performance. If there is a long-term lack of correlation analysis and continuous tracking for such suspected abnormalities that do not trigger alarms, potential leakage hazards may not be identified in advance, eventually evolving into sudden leakage accidents during the production process, leading to problems such as line stoppage, mold damage, increased casting scrap rate, and increased production costs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a casting leakage detection method based on big data models and cross-modal interaction, which solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a casting leakage detection method based on big data models and cross-modal interaction, comprising the following steps: S1: Obtain the casting monitoring data and mold furnace identification data corresponding to the current casting furnace, and input the casting monitoring data into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification, and generate the suspected leakage identification results of each structure area. S2: Compare the suspected leakage identification result of the structural area with the preset leakage alarm condition and the preset suspected abnormal memory condition respectively. When the suspected leakage identification result of the structural area does not meet the leakage alarm condition but meets the suspected abnormal memory condition, generate a non-alarm suspected abnormal memory event for the corresponding structural area. S3: Based on the mold furnace identification data, write the non-alarm suspected abnormal memory events into the corresponding structural area hidden danger memory library, and form a non-alarm suspected abnormal memory sequence of the same mold structural area in multiple casting furnaces in the structural area hidden danger memory library; S4: Perform correlation analysis on the recurring features in the non-alarm suspected abnormal memory sequence to generate structural area hidden danger memory results for the corresponding mold structure area; S5: Based on the memory result of the structural area hazard, determine whether the corresponding mold structural area meets the preset leakage hazard warning conditions. If the leakage hazard warning conditions are met, output the leakage hazard warning result.

[0007] Preferably, step S1 includes: S11. Collect visible light image data and temperature image data corresponding to the current casting furnace using the existing image acquisition device at the casting site, and define the visible light image data and the temperature image data as casting monitoring data; S12. Obtain the mold number, furnace number, and pouring time information corresponding to the current casting furnace from the casting production management system or casting equipment control system, and define the mold number, the furnace number, and the pouring time information as mold furnace identification data; S13. Transmit the casting monitoring data and the mold furnace identification data to the big data model processing terminal, and perform time alignment processing on the visible light image data and the temperature image data according to the pouring time information; S14. Input the casting monitoring data after time alignment processing into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification and generate the suspected leakage identification results of each structure area. The mold structure region division result is obtained through the following steps: A1. Obtain the mold image or mold structure diagram corresponding to the casting mold; A2. Based on the gate area, riser area, parting surface area and mold edge area in the casting mold, the mold image or mold structure diagram is divided into regions to obtain multiple structural regions; A3. Assign a unique structural region number to each structural region; A4. Define the multiple structural regions and the corresponding structural region numbers as the mold structural region division result.

[0008] Preferably, the suspected leakage identification result of the structural region is obtained through the following steps: B1. Based on the mold structure area division result and the structure area number, map the visible light image data and the temperature image data to the corresponding structure areas respectively; B2. Using the big data model, cross-modal anomaly identification is performed on the visible light image data and the temperature image data mapped to the same structural region number; B3. Based on the cross-modal anomaly identification results, obtain the location, duration, range of influence, and intensity of the anomaly region in the corresponding structural region. B4. Combine the location of the abnormal region corresponding to the same structural region number, the duration of the abnormality, the range of the abnormality, and the intensity of the abnormality to generate the suspected leakage identification result of the structural region.

[0009] Preferably, the location of the abnormal region is determined by the overlap of the corresponding abnormal regions in the visible light image data and the temperature image data, the range of the abnormal influence is determined by the range of the image area covered by the location of the abnormal region within the corresponding structural region, and the intensity of the abnormality is jointly determined by the abnormal temperature difference in the temperature image data and the abnormal brightness change in the visible light image data.

[0010] Preferably, step S2 includes: S21. Compare the suspected leakage identification result of the structural area with the preset leakage alarm conditions to obtain the leakage alarm comparison result; S22. The suspected leakage identification result of the structural area that has not met the leakage alarm condition is further compared with the preset suspected abnormal memory condition to obtain the suspected abnormal memory comparison result. S23. When the leakage alarm comparison result is that the leakage alarm condition has not been met, and the suspected abnormal memory comparison result is that the suspected abnormal memory condition has been met, the corresponding structural area is determined as the non-alarm suspected abnormal structural area. S24. The suspected abnormal memory condition is at least two of the following: the abnormal duration reaches the abnormal duration recording condition, the abnormal influence range reaches the abnormal influence range recording condition, and the abnormal intensity reaches the abnormal intensity recording condition.

[0011] Preferably, the unalarmed suspected abnormal memory event is generated through the following steps: C1. Extract the mold number, furnace number, structural area number, abnormal area location, abnormal duration, abnormal influence range, and abnormal intensity corresponding to the non-alarm suspected abnormal structural area; C2. Generate the corresponding event generation time based on the pouring time information; C3. The mold number, the furnace number, the structural area number, the location of the abnormal area, the duration of the abnormality, the range of the abnormality, the intensity of the abnormality, and the event generation time are combined into the non-alarm suspected abnormal memory event. C4. Define the non-alarm suspected abnormal memory event as a data object that did not trigger a leak alarm but entered the hidden danger memory processing.

[0012] Preferably, step S3 includes: S31. Establish a corresponding structural area hazard memory bank according to the mold number and the structural area number; S32. Write the unalarmed suspected abnormal memory event into the structural area hidden danger memory bank corresponding to the mold number and the structural area number; S33. Sort the multiple non-alarm suspected abnormal memory events in the structural area hidden danger memory bank according to the furnace number; S34. Multiple non-alarm suspected abnormal memory events sorted by furnace number under the same mold number and the same structural area number are defined as non-alarm suspected abnormal memory sequences.

[0013] Preferably, step S4 includes: S41. Read multiple non-alarm suspected abnormal memory events from the non-alarm suspected abnormal memory sequence; S42. Count the number of suspected abnormal memory events that did not trigger an alarm and define them as the number of times they occurred repeatedly; S43. Calculate multiple adjacent furnace number differences based on the furnace number corresponding to two adjacent non-alarm suspected abnormal memory events, and define the maximum value among the multiple adjacent furnace number differences as the furnace interval; S44. Calculate the ratio of the overlapping area to the union area between the multiple abnormal area locations based on the abnormal area locations corresponding to the multiple non-alarm suspected abnormal memory events, and define the calculation result as the abnormal area location overlap situation. S45. The recurrence count, the furnace interval, and the overlap of the abnormal area location are defined as recurrence characteristics.

[0014] Preferably, step S4 further includes: S46. Compare the number of recurrences with a preset recurrence count condition to obtain a recurrence count comparison result; compare the furnace interval with a preset maximum furnace interval condition to obtain a furnace interval comparison result; and compare the overlap of abnormal area locations with a preset area overlap condition to obtain an area overlap comparison result. S47. When the repetition count comparison result is that the repetition count reaches the preset repetition count condition, the furnace interval comparison result is that the furnace interval is not greater than the preset maximum furnace interval condition, and the region overlap comparison result is that the abnormal region location overlap reaches the preset region overlap condition, the corresponding structural region is determined as a candidate leakage hazard structural region. S48. Determine the hazard level corresponding to the candidate leakage hazard structural area according to the preset hazard level mapping relationship, and generate a structural area hazard memory result including the candidate leakage hazard structural area and the hazard level, wherein the preset hazard level mapping relationship is a correspondence between the hazard levels determined according to the number of recurrences, the furnace interval, and the overlap of the abnormal area location.

[0015] Preferably, step S5 includes: S51. When the hazard level meets the preset leakage hazard warning conditions, the candidate leakage hazard structural area is determined as the leakage hazard structural area, and a leakage hazard warning result including the leakage hazard structural area, the mold number, the furnace number and the hazard level is output.

[0016] This invention provides a method for detecting casting leakage based on big data models and cross-modal interaction, which has the following beneficial effects: (1) During system operation, by associating visible light image data, temperature image data, mold number, furnace number, and pouring time information during the casting process, cross-modal identification and continuous tracking of abnormal phenomena during the casting process are achieved. Compared with the method of judging leakage based on only a single detection result, this invention can not only identify leakage anomalies that have reached the alarm conditions, but also retain and continuously analyze suspected anomalies that have not reached the alarm conditions but have potential risks, thereby avoiding the problem of minor anomalies being directly ignored because they are not obvious in a single instance. At the same time, this invention uses the mold structure area as the analysis object, so that the anomaly identification results can correspond to the specific gate area, riser area, parting surface area, or mold edge area, improving the accuracy of anomaly positioning, and further combining multiple furnaces. By correlating and judging the occurrence of anomalies, this invention transforms previously scattered and isolated anomaly information into continuous potential hazard information. Furthermore, by utilizing the frequency of recurrence, furnace intervals, and overlapping locations of anomaly areas to perform memory analysis on suspected anomalies that have not yet triggered alarms, this invention can identify recurring and concentrated anomalies in the same mold and structural area over a long period. This allows for the early detection of potential leaks before actual leakage occurs, shifting from single-anomaly detection to long-term hazard identification. This not only improves the accuracy and stability of leak detection but also helps on-site personnel to conduct mold maintenance, process adjustments, and risk assessments in advance, reducing the risk of casting defects, mold damage, production interruptions, and batch scrapping caused by leaks, thereby improving the reliability of the casting production process and the stability of product quality.

[0017] (2) This method combines image information, temperature information, mold information and furnace information in the casting process to detect leakage. It can not only identify leakage anomalies that have reached the alarm level, but also retain data that have not reached the alarm conditions but have suspected leakage. This allows minor anomalies that are easily overlooked to be continuously analyzed, thereby discovering potential leakage hazards in the same mold and structural area in advance, reducing the risk of sudden leakage, line stoppage, mold damage and casting scrap.

[0018] (3) In terms of data acquisition and suspected anomaly identification, this method uses visible light image data and temperature image data to jointly determine casting monitoring data, and combines mold number, furnace number and pouring time information for time alignment, so that data from different sources can be matched to the same casting furnace. At the same time, based on the mold structure area division results, the anomaly identification results are directly matched to the gate area, riser area, parting surface area and mold edge area. The suspected leakage behavior is described by the location of the abnormal area, the duration of the abnormality, the range of the abnormality, and the intensity of the abnormality, which improves the accuracy of anomaly location and the reliability of the data basis for subsequent judgment.

[0019] (4) In terms of hazard memory and early warning, this method identifies structural areas that do not meet the leakage alarm conditions but meet the suspected abnormal memory conditions as non-alarm suspected abnormal structural areas and generates non-alarm suspected abnormal memory events. According to the mold number and structural area number, a non-alarm suspected abnormal memory sequence is formed. Then, the repetition frequency, furnace interval and the overlap of abnormal area location are used to determine whether the abnormality has repetition and location concentration, thereby forming the structural area hazard memory result and leakage hazard early warning result. This extends leakage detection from single abnormal judgment to multi-furnace hazard identification, improving the ability to detect progressive leakage risks in advance. Attached Figure Description

[0020] Figure 1 This is a block diagram of the casting leakage detection method based on big data model and cross-modal interaction of the present invention; Figure 2 This is a block diagram for identifying potential hazards that have not triggered alarms in this invention. Detailed Implementation

[0021] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 To facilitate understanding of the technical solutions provided in this application, the application scenarios of this application will be explained first.

[0023] In this embodiment, an automotive transmission housing casting production line is used as an application scenario for illustration. Automotive transmission housings are typically cast from metal materials such as aluminum alloys or magnesium-aluminum alloys. During the casting process, molten metal enters the mold cavity through the gating system and undergoes filling, pressure holding, and solidification processes to form the target casting. Due to the complex structure of the transmission housing, the mold usually has multiple gating areas, riser areas, parting surface areas, and mold edge areas. During long-term continuous production, factors such as mold wear, changes in mold fitting clearance, decreased local sealing performance, thermal deformation, and process fluctuations may gradually lead to leakage risks in some structural areas.

[0024] In actual production, potential leakage hazards usually do not manifest as obvious leakage in the early stages. Instead, they first appear as minor anomalies in local structural areas, such as abnormal temperature, abnormal brightness, abnormal liquid traces, or short-term abnormal diffusion. These anomalies are often insufficient to trigger a leakage alarm in a single casting furnace, and are therefore easily ignored by existing detection systems as normal fluctuations. As the number of production furnaces continues to increase, these minor anomalies may reappear and accumulate in the same structural area of ​​the same mold, eventually evolving into an actual leakage accident.

[0025] Therefore, this application identifies abnormal events that do not meet the leakage alarm conditions but meet the suspected abnormality memory conditions by jointly analyzing the image and temperature information generated during the casting process. It also performs continuous memory and correlation analysis on suspected abnormalities of the same structural area in multiple casting furnaces, thereby discovering potential leakage hazards in advance and achieving early warning of leakage hazards.

[0026] It should be noted that this embodiment uses an automotive gearbox housing casting production line as an example for illustration, but this application is not limited to this application scenario. The technical solutions provided in this application are also applicable to differential housings, clutch housings, motor housings, reducer housings, and other cast parts produced by molten metal casting processes.

[0027] This invention provides a casting leakage detection method based on big data models and cross-modal interaction. Please refer to [link to relevant documentation]. Figure 1 It includes the following steps: S1: Obtain the casting monitoring data and mold furnace identification data corresponding to the current casting furnace, and input the casting monitoring data into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification, and generate the suspected leakage identification results of each structure area. S2: Compare the suspected leakage identification result of the structural area with the preset leakage alarm condition and the preset suspected abnormal memory condition respectively. When the suspected leakage identification result of the structural area does not meet the leakage alarm condition but meets the suspected abnormal memory condition, generate a non-alarm suspected abnormal memory event for the corresponding structural area. S3: Based on the mold furnace identification data, write the non-alarm suspected abnormal memory events into the corresponding structural area hidden danger memory library, and form a non-alarm suspected abnormal memory sequence of the same mold structural area in multiple casting furnaces in the structural area hidden danger memory library; S4: Perform correlation analysis on the recurring features in the non-alarm suspected abnormal memory sequence to generate structural area hidden danger memory results for the corresponding mold structure area; S5: Based on the memory result of the structural area hazard, determine whether the corresponding mold structural area meets the preset leakage hazard warning conditions. If the leakage hazard warning conditions are met, output the leakage hazard warning result.

[0028] Example 2 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, step S1 includes: S11. Collect visible light image data and temperature image data corresponding to the current casting furnace using the existing image acquisition device at the casting site, and define the visible light image data and the temperature image data as casting monitoring data; In this embodiment, the application scenario is set as an automotive gearbox housing casting production line. The image acquisition devices are an industrial camera and an infrared thermal imaging camera installed above or to the side of the pouring station. The industrial camera acquires visible light images of the mold surface, the gate area, the riser area, the parting surface area, and the mold edge area. The infrared thermal imaging camera acquires temperature images within the same field of view. The acquisition time covers the preset observation period before pouring begins, during pouring, and after pouring ends. The preset observation period can be set according to the pouring cycle of a single casting, for example, from 5 seconds before pouring begins to 30 seconds after pouring ends, to ensure that short-term hot spots or local abnormal brightness changes that do not meet the alarm conditions but are suspected of being abnormal can be identified later.

[0029] S12. Obtain the mold number, furnace number, and pouring time information corresponding to the current casting furnace from the casting production management system or casting equipment control system, and define the mold number, the furnace number, and the pouring time information as mold furnace identification data; Specifically, the mold number is obtained from the mold file registered in the casting production management system, the furnace number is obtained from the production record corresponding to the current pouring batch or the current pouring cycle, and the pouring time information is determined by the pouring start signal, pouring end signal or operation confirmation signal in the pouring equipment control system. The mold number is used to distinguish different molds, the furnace number is used to determine the production sequence of the same mold, and the pouring time information is used to map visible light image data and temperature image data to the same casting furnace, thereby avoiding data from different furnaces from being mixed into the same hidden danger memory process.

[0030] S13. Transmit the casting monitoring data and the mold furnace identification data to the big data model processing terminal, and perform time alignment processing on the visible light image data and the temperature image data according to the pouring time information; During data processing, visible light image data, temperature image data, and mold furnace identification data are transmitted to the big data model processing end through the on-site industrial network or local data interface. The big data model processing end uses the pouring start time in the pouring time information as the alignment benchmark, and matches the image frames in the visible light image data and the temperature frames in the temperature image data according to the acquisition time. If the acquisition frequencies of the two types of images are different, the image frame and temperature frame closest to the same acquisition time are used as a set of alignment data to ensure that the visible light changes and temperature changes at the same time can be used together for cross-modal anomaly identification.

[0031] S14. Input the casting monitoring data after time alignment processing into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification and generate the suspected leakage identification results of each structure area. Specifically, the big data model processing end first determines the position range of each structural region in each set of aligned data based on the mold structure region division results. Then, it extracts the visible light change features and temperature change features in each structural region and jointly identifies the visible light change features and temperature change features in the same structural region to obtain whether there is suspected leakage in the structural region and the corresponding abnormal region location, abnormal duration, abnormal impact range and abnormal intensity, thus forming the suspected leakage identification result of the structural region.

[0032] The mold structure region division result is obtained through the following steps: A1. Obtain the mold image or mold structure diagram corresponding to the casting mold; In this embodiment, the mold image can be acquired by the image acquisition device in the empty mold state or the non-casting state, and the mold structure diagram can be obtained from the mold design documents, mold maintenance drawings or on-site calibration diagram. The implementer can choose one of them as the basis for area division, or the mold image and the mold structure diagram can be superimposed and corrected as the basis for area division to improve the accuracy of the correspondence between the area location and the on-site image.

[0033] A2. Based on the gate area, riser area, parting surface area and mold edge area in the casting mold, the mold image or mold structure diagram is divided into regions to obtain multiple structural regions; Specifically, the implementers divide the area according to the location where the molten metal in the automotive gearbox housing mold passes through or may leak out. At least the gate area, riser area, parting surface area, and mold edge area are divided into independent structural areas. The gate area is used to correspond to the location where the molten metal enters the mold cavity, the riser area is used to correspond to the location of shrinkage and high temperature residence, the parting surface area is used to correspond to the location where the upper and lower molds are joined, and the mold edge area is used to correspond to the location where leakage may first appear. This division can ensure that subsequent suspected anomalies are not just based on image coordinates, but can be mapped to specific mold structural locations.

[0034] A3. Assign a unique structural region number to each structural region; Specifically, the structural area number can be generated according to the mold number plus the area category plus the area sequence. For example, different structural area numbers can be configured for the gate area, riser area, parting surface area and mold edge area in the same mold. The structural area number is kept unique within the same mold. It is used to continuously remember and sort suspected anomalies of the same structural area in different furnaces, so as to avoid suspected anomalies of different areas being mistakenly merged.

[0035] A4. Define the multiple structural regions and the corresponding structural region numbers as the mold structural region division result.

[0036] It should be noted that the results of the mold structure area division can be reused when the same mold installation state has not changed. When the mold is replaced, the camera position is adjusted, or the mold is repaired, resulting in a change in the image perspective, it is necessary to re-acquire the mold image or reload the mold structure diagram, and reconfirm the correspondence between each structure area and the structure area number to ensure that the subsequent suspected leakage identification results of the structure area can be accurately assigned to the corresponding structure area.

[0037] The suspected leakage identification result of the structural region is obtained through the following steps: B1. Based on the mold structure area division result and the structure area number, map the visible light image data and the temperature image data to the corresponding structure areas respectively; Specifically, the big data model processing end extracts the image regions corresponding to each structural region from the visible light image data and temperature image data according to the region boundaries in the mold structure region division results. It then establishes a correspondence between the extracted visible light image regions and temperature image regions and the same structural region number, thereby forming structural region-level data input. This ensures that subsequent identification can directly determine which type of location the anomaly occurred in, such as the gate region, riser region, parting surface region, or mold edge region.

[0038] B2. Using the big data model, cross-modal anomaly identification is performed on the visible light image data and the temperature image data mapped to the same structural region number; During data processing, the big data model simultaneously analyzes visible light image data and temperature image data under the same structural region number. Visible light image data is used to identify sudden changes in brightness, changes in liquid morphology, changes in smoke obscuration, or changes in reflection. Temperature image data is used to identify local temperature rises, abnormally high temperature areas, and temperature diffusion ranges. When visible light changes and temperature changes occur simultaneously in the same structural region and within the same time period, the big data model treats that region as a suspected abnormal region and includes it in subsequent calculations.

[0039] B3. Based on the cross-modal anomaly identification results, obtain the location, duration, range of influence, and intensity of the anomaly region in the corresponding structural region. Specifically, the abnormal region location is the image region location where anomalies are observed in both visible light image data and temperature image data; the abnormal duration is the length of time the same abnormal region location remains in an abnormal state in consecutive image frames or consecutive temperature frames; the abnormal influence range is the image region range covered by the abnormal region location within the corresponding structural region; and the abnormal intensity is the degree of abnormality determined by the abnormal temperature difference in the temperature image data and the abnormal brightness change in the visible light image data. The abnormal temperature difference can be obtained by the difference between the abnormal region temperature and the normal background temperature of the same structural region, and the abnormal brightness change can be obtained by the difference between the abnormal region brightness and the normal background brightness of the same structural region.

[0040] B4. Combine the location of the abnormal region corresponding to the same structural region number, the duration of the abnormality, the range of the abnormality, and the intensity of the abnormality to generate the suspected leakage identification result of the structural region.

[0041] Specifically, the suspected leakage identification results for a structural region are generated based on the structural region number. A suspected leakage identification result for a structural region includes at least the structural region number, the location of the abnormal region, the duration of the abnormality, the range of the abnormality, and the intensity of the abnormality. This is used to indicate whether the structural region exhibits suspected leakage in the current casting furnace and to provide a data basis for subsequent judgment on whether the suspected leakage meets the leakage alarm conditions or whether it enters the suspected abnormality memory conditions.

[0042] The location of the abnormal region is determined by the overlap of the corresponding abnormal regions in the visible light image data and the temperature image data. The range of the abnormal influence is determined by the range of the image area covered by the location of the abnormal region within the corresponding structural region. The intensity of the abnormality is jointly determined by the abnormal temperature difference in the temperature image data and the abnormal brightness change in the visible light image data.

[0043] It should be noted that when there is a brightness anomaly in the visible light image data but no corresponding temperature anomaly in the temperature image data, the area can be treated as ordinary visible light interference. When there is a temperature anomaly in the temperature image data but no corresponding morphological change in the visible light image data, the area can be treated as a single temperature anomaly. Only when the corresponding anomaly areas in the visible light image data and the temperature image data overlap will the overlapping area be regarded as the anomaly area location, thereby reducing the risk of misidentification caused by reflection, smoke, high background temperature or image noise.

[0044] Example 3 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, step S2 includes: S21. Compare the suspected leakage identification result of the structural area with the preset leakage alarm conditions to obtain the leakage alarm comparison result; In the scenario set in Example 1, the leakage alarm condition is used to determine whether the current structural area has reached the level that requires an immediate alarm. Specifically, it can be composed of an abnormal duration alarm condition, an abnormal impact range alarm condition, and an abnormal intensity alarm condition. For example, when the abnormal duration, abnormal impact range, and abnormal intensity in the same structural area all reach the corresponding alarm conditions, the leakage alarm comparison result is determined to be that the leakage alarm condition has been met. When at least one of them does not reach the corresponding alarm condition, the leakage alarm comparison result is determined to be that the leakage alarm condition has not been met. Each alarm condition can be set according to the lower limit of abnormal data in historical real leakage furnaces and the upper limit of abnormal data in normal furnaces to ensure that the alarm conditions can correspond to the actual leakage risk on site.

[0045] S22. The suspected leakage identification result of the structural area that has not met the leakage alarm condition is further compared with the preset suspected abnormal memory condition to obtain the suspected abnormal memory comparison result. Specifically, for suspected leakage identification results in structural areas that do not meet the leakage alarm conditions, they are not directly discarded as normal data. Instead, they are further judged to see if they meet the suspected anomaly memory conditions. The suspected anomaly memory conditions are used to filter out minor abnormal data that, although insufficient to trigger a leakage alarm, still have repeatable tracking value. Implementers can use the background fluctuation range in historical normal furnace cycles as a lower limit reference and the leakage alarm conditions as an upper limit boundary, so that the suspected anomaly memory conditions are between normal fluctuations and leakage alarms. This ensures that the data entering the hidden danger memory processing is neither ordinary noise nor leakage data that has already been clearly alarmed.

[0046] S23. When the leakage alarm comparison result is that the leakage alarm condition has not been met, and the suspected abnormal memory comparison result is that the suspected abnormal memory condition has been met, the corresponding structural area is determined as the non-alarm suspected abnormal structural area. During data processing, a suspected abnormal structure area without alarm indicates that the structure area has not yet reached the level of immediate alarm, but has shown suspected leakage behavior that exceeds the normal fluctuation range. For example, if the parting surface area shows a short-term high-temperature overlap area and local brightness abnormality in the current furnace, but the duration or range of abnormality has not yet reached the leakage alarm condition, the corresponding parting surface area is identified as a suspected abnormal structure area without alarm and is used for subsequent hidden danger memory processing.

[0047] S24. The suspected abnormal memory condition is at least two of the following: the abnormal duration reaches the abnormal duration recording condition, the abnormal influence range reaches the abnormal influence range recording condition, and the abnormal intensity reaches the abnormal intensity recording condition.

[0048] Specifically, the recording conditions for abnormal duration can be determined based on the duration of short-term noise in normal furnace cycles. The recording conditions for abnormal impact range can be determined based on the coverage of occasional bright spots or local hot spots in normal furnace cycles. The recording conditions for abnormal intensity can be determined based on the background temperature difference and background brightness changes in normal furnace cycles. When at least two of the abnormal duration, abnormal impact range, and abnormal intensity meet the corresponding recording conditions, it indicates that the suspected abnormality is no longer a single occasional noise and has the necessity to be included in the hidden danger memory processing. The reason for adopting at least two of the corresponding recording conditions is to avoid data being mistakenly remembered due to a single image reflection, a single temperature disturbance, or a single frame of noise.

[0049] The unreported suspected abnormal memory events are generated through the following steps: C1. Extract the mold number, furnace number, structural area number, abnormal area location, abnormal duration, abnormal influence range, and abnormal intensity corresponding to the non-alarm suspected abnormal structural area; Specifically, after identifying the suspected abnormal structural area without alarm, the big data model processing end extracts the mold number and furnace number from the data record of the current furnace, extracts the structural area number from the mold structural area division result, and extracts the abnormal area location, abnormal duration, abnormal impact range and abnormal intensity from the suspected leakage identification result of the structural area, so that each suspected abnormal without alarm can be simultaneously matched with a specific mold, specific furnace, specific structural area and specific abnormal manifestation.

[0050] C2. Generate the corresponding event generation time based on the pouring time information; Specifically, the event generation time can be determined by the image frame time of the first appearance of a suspected non-alarm anomaly, or by the time when a suspected non-alarm anomaly reaches the suspected anomaly memory condition. In this embodiment, the time when a suspected non-alarm anomaly reaches the suspected anomaly memory condition is preferred as the event generation time, because this time can indicate the starting point of the suspected anomaly formally entering the hidden danger memory processing, which is convenient for subsequent tracing according to the furnace order and time order.

[0051] C3. The mold number, the furnace number, the structural area number, the location of the abnormal area, the duration of the abnormality, the range of the abnormality, the intensity of the abnormality, and the event generation time are combined into the non-alarm suspected abnormal memory event. Specifically, unalarmed suspected abnormal memory events can be stored in the form of database records, structured data tables, or event objects. Each unalarmed suspected abnormal memory event corresponds to an unalarmed suspected abnormality recorded in the current furnace batch. During implementation, the mold number, furnace batch number, structural area number, and event generation time can be used as index fields, and the abnormal area location, abnormal duration, abnormal impact range, and abnormal intensity can be used as abnormal content fields, thereby supporting subsequent querying, sorting, and statistics by mold, by area, and by furnace batch.

[0052] C4. Define the non-alarm suspected abnormal memory event as a data object that did not trigger a leak alarm but entered the hidden danger memory processing.

[0053] It should be noted that unalarmed suspected abnormal memory events are different from alarmed leakage events. Alarmed leakage events are used for immediate on-site handling, while unalarmed suspected abnormal memory events are used for long-term hazard identification. Their technical role is to retain critical abnormal data that are usually ignored by existing detection methods, so that minor abnormalities that repeatedly occur in the same structural area in multiple furnace cycles can be continuously tracked.

[0054] Step S3 includes: S31. Establish a corresponding structural area hazard memory bank according to the mold number and the structural area number; In this embodiment, the structural area hazard memory library can be set in the big data model processing terminal, local industrial computer or production management server. The structural area hazard memory library establishes storage units according to mold number and structural area number. For example, the parting surface area, gate area, riser area and mold edge area of ​​the same automobile gearbox housing mold have independent storage units, thereby ensuring that unalarmed suspected abnormalities of different molds and different structural areas will not be confused with each other.

[0055] S32. Write the unalarmed suspected abnormal memory event into the structural area hidden danger memory bank corresponding to the mold number and the structural area number; Specifically, during the writing process, the storage location of the corresponding mold is first found based on the mold number in the non-alarm suspected abnormal memory event. Then, the storage unit of the corresponding structural area is found based on the structural area number. The non-alarm suspected abnormal memory event is then written to the corresponding storage unit. If the corresponding storage unit does not exist, it is created first and then written, thereby ensuring that new molds or newly added structural areas can also be included in the hazard memory processing.

[0056] S33. Sort the multiple non-alarm suspected abnormal memory events in the structural area hidden danger memory bank according to the furnace number; During data processing, sorting is based on the production sequence indicated by the furnace number. When there are multiple suspected abnormal memory events that have not been alarmed within the same furnace, they can be sorted again according to the event generation time. By sorting first by furnace number and then by event generation time, the process of changes in suspected abnormalities in the same mold and structural area as the production furnace progresses can be formed, which facilitates subsequent judgment on whether there are repeated occurrences.

[0057] S34. Multiple non-alarm suspected abnormal memory events sorted by furnace number under the same mold number and the same structural area number are defined as non-alarm suspected abnormal memory sequences.

[0058] Specifically, the unalarmed suspected anomaly memory sequence is an event sequence formed with the same mold and the same structural area as the object. The elements in the unalarmed suspected anomaly memory sequence are multiple unalarmed suspected anomaly memory events. Each unalarmed suspected anomaly memory event includes mold number, furnace number, structural area number, location of the abnormal area, duration of the abnormality, range of the abnormality, intensity of the abnormality, and time of event generation. The unalarmed suspected anomaly memory sequence is used to calculate the number of recurrences, furnace intervals, and overlap of abnormal area locations, thereby determining whether there is a potential for gradual leakage in the structural area.

[0059] Example 4 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, step S4 includes: S41. Read multiple non-alarm suspected abnormal memory events from the non-alarm suspected abnormal memory sequence; In the scenario set in Example 1, the big data model processing terminal reads the corresponding non-alarm suspected abnormal memory sequence from the structural area hidden danger memory bank according to the mold number and structural area number that need to be analyzed. The reading object is multiple non-alarm suspected abnormal memory events under the same mold number and structural area number that have been sorted by furnace number. When reading, non-alarm suspected abnormal memory events within the range of the most recent preset number of furnaces can be selected, such as events within the range of the most recent twenty furnaces. This number can be set according to the mold state change rate, single-shift production quantity and historical leakage precursor occurrence cycle in the automotive gearbox housing casting production line, and is not limited to a fixed value.

[0060] S42. Count the number of suspected abnormal memory events that did not trigger an alarm and define them as the number of times they occurred repeatedly; Specifically, after reading the non-alarm suspected anomaly memory sequence, the non-alarm suspected anomaly memory events contained in the non-alarm suspected anomaly memory sequence are counted one by one to obtain the recurrence count. The recurrence count indicates the number of times the same structural area of ​​the same mold has a non-alarm suspected anomaly within the selected furnace batch. For example, if there are five non-alarm suspected anomaly memory events in a certain parting surface area within the last twenty furnace batches, then the recurrence count for that parting surface area is five. The recurrence count is used to indicate whether the suspected anomaly in that structural area has the characteristic of repeated occurrence.

[0061] S43. Calculate multiple adjacent furnace number differences based on the furnace number corresponding to two adjacent non-alarm suspected abnormal memory events, and define the maximum value among the multiple adjacent furnace number differences as the furnace interval; During data processing, multiple non-alarm suspected abnormal memory events are first arranged in ascending order of furnace number. Then, the difference between the furnace numbers corresponding to two adjacent non-alarm suspected abnormal memory events is calculated sequentially to obtain multiple differences between adjacent furnace numbers. The maximum value among the multiple differences between adjacent furnace numbers is taken as the furnace interval. For example, if the non-alarm suspected abnormal memory events corresponding to a certain structural area appear in furnaces 101, 103, 104 and 107 respectively, then the differences between adjacent furnace numbers are 2, 1 and 3 respectively, and the furnace interval is 3. The furnace interval is used to indicate whether the non-alarm suspected abnormalities in this structural area are continuously and closely appearing in the furnace.

[0062] S44. Calculate the ratio of the overlapping area to the union area between the multiple abnormal area locations based on the abnormal area locations corresponding to the multiple non-alarm suspected abnormal memory events, and define the calculation result as the abnormal area location overlap situation. Specifically, the location of the abnormal region in each unalarmed suspected abnormal memory event can be represented as the outline of the image region or the set of pixel regions within the corresponding structural region. When calculating the overlap of abnormal region locations, first obtain the image area where multiple abnormal region locations overlap, then obtain the image area after merging multiple abnormal region locations, and then divide the overlapping image area by the merged image area to obtain the overlap of abnormal region locations. For example, if the overlapping image area of ​​multiple abnormal region locations is the image area corresponding to 20 square millimeters, and the merged image area is the image area corresponding to 50 square millimeters, then the overlap of abnormal region locations is 40%. This overlap of abnormal region locations is used to indicate whether multiple unalarmed suspected abnormalities are concentrated in the same local location.

[0063] S45. The recurrence count, the furnace interval, and the overlap of the abnormal area location are defined as recurrence characteristics.

[0064] It should be noted that the recurrence feature does not simply indicate that an anomaly has occurred in a certain structural area, but also indicates the number of times that a structural area has a suspected anomaly that has not been alarmed within the selected furnace batch, the proximity of adjacent furnace batches, and the degree of spatial overlap of the anomaly locations. Therefore, the recurrence feature can be used to distinguish between sporadic noise and repeated suspected anomalies with continuous hidden danger significance, providing a data basis for the subsequent determination of the hidden danger memory results of structural areas.

[0065] Example 5 For a casting leakage detection method based on big data models and cross-modal interaction, please refer to [reference needed]. Figure 2 Specifically, step S4 further includes: S46. Compare the number of recurrences with a preset recurrence count condition to obtain a recurrence count comparison result; compare the furnace interval with a preset maximum furnace interval condition to obtain a furnace interval comparison result; and compare the overlap of abnormal area locations with a preset area overlap condition to obtain an area overlap comparison result. In the scenario set in Example 1, the preset repetition count condition, the preset maximum furnace interval condition, and the preset area overlap condition can be set according to the historical normal furnace data and historical furnace data before leakage of the same automotive gearbox housing mold. The preset repetition count condition is used to determine whether the occurrence of unalarmed suspected abnormalities in the same structural area is frequent enough. The preset maximum furnace interval condition is used to determine whether multiple unalarmed suspected abnormalities occur consecutively in adjacent or close furnaces. The preset area overlap condition is used to determine whether multiple unalarmed suspected abnormalities occur in the same local location.

[0066] It should be noted that the preset repetition count condition can be taken as the lower limit of the number of times the same structural area of ​​suspected abnormality without alarm in the previous furnace batches. The preset maximum furnace interval condition can be taken as the upper limit of the furnace batch difference between adjacent suspected abnormalities without alarm in the previous furnace batches. The preset area overlap condition can be taken as the lower limit of the overlap of abnormal area positions in the previous furnace batches. If historical leakage data is lacking in the initial stage, suspected abnormality samples after manual review can be used for setting first, and then updated according to the confirmed samples before leakage and normal samples in the subsequent production process.

[0067] S47. When the repetition count comparison result is that the repetition count reaches the preset repetition count condition, the furnace interval comparison result is that the furnace interval is not greater than the preset maximum furnace interval condition, and the region overlap comparison result is that the abnormal region location overlap reaches the preset region overlap condition, the corresponding structural region is determined as a candidate leakage hazard structural region. Specifically, a structural area is only identified as a candidate leakage hazard area when it simultaneously meets the following conditions: the number of recurrences reaches the preset recurrence count, the furnace interval is not greater than the preset maximum furnace interval, and the overlap of abnormal areas meets the preset area overlap condition. This is because the number of recurrences alone only indicates the frequency of abnormal occurrences, the furnace interval alone only indicates the proximity of abnormal occurrences, and the overlap of abnormal areas alone only indicates the concentration of abnormal locations. Only when all three conditions are met can it be said that the structural area has stable and recurring non-alarm suspected abnormal characteristics, thereby reducing misjudgments caused by occasional reflections, local smoke and dust, and short-term thermal disturbances.

[0068] S48. Determine the hazard level corresponding to the candidate leakage hazard structural area according to the preset hazard level mapping relationship, and generate a structural area hazard memory result including the candidate leakage hazard structural area and the hazard level, wherein the preset hazard level mapping relationship is a correspondence between the hazard levels determined according to the number of recurrences, the furnace interval, and the overlap of the abnormal area location.

[0069] During data processing, the preset hazard level mapping relationship can be set as a lookup table relationship or a rule relationship. For example, the more times it occurs, the smaller the furnace interval, and the higher the overlap of abnormal area locations, the higher the corresponding hazard level. Implementers can classify different hazard levels based on the combination distribution of the number of occurrences, furnace intervals, and overlap of abnormal area locations in historical samples before leakage. If the number of occurrences corresponding to a candidate leakage hazard structure area reaches a high frequency range, the furnace interval is in a small interval range, and the overlap of abnormal area locations is in a high overlap range, then the candidate leakage hazard structure area is determined to be a higher hazard level.

[0070] It should be noted that the structural area hazard memory result is not the alarm result for a single current anomaly, but rather the memory judgment result for repeated occurrences of suspected anomalies without alarm in multiple furnace cycles in the same structural area of ​​the same mold. The structural area hazard memory result includes at least the candidate leakage hazard structural area and the hazard level, which is used to indicate that the structural area has the basis for judgment of leakage hazard warning.

[0071] Step S5 includes: S51. When the hazard level meets the preset leakage hazard warning conditions, the candidate leakage hazard structural area is determined as the leakage hazard structural area, and a leakage hazard warning result including the leakage hazard structural area, the mold number, the furnace number and the hazard level is output.

[0072] Specifically, the leakage hazard warning conditions can be set to a level where the hazard level requires manual review or production intervention. When the hazard level of a candidate leakage hazard structural area meets the leakage hazard warning conditions, the big data model processing terminal converts the candidate leakage hazard structural area into a leakage hazard structural area and outputs the leakage hazard warning result. The leakage hazard warning result can be sent to the on-site display terminal, production management system, or equipment control system to remind operators to focus on checking the corresponding mold number, furnace number, and leakage hazard structural area.

[0073] In this embodiment, the leakage hazard warning result is used to alert potential hazards before an actual leakage alarm occurs. For example, when the parting surface area of ​​the same automotive gearbox housing mold repeatedly shows suspected abnormalities without alarm in multiple similar furnace batches, and the abnormal areas are concentrated in the same local area, the system outputs the leakage hazard warning result for the corresponding parting surface area. On-site personnel can use this to check the mold locking status, parting surface sealing status, local wear status, or casting process stability, thereby reducing the risk of sudden leakage in subsequent furnace batches.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A casting leakage detection method based on big data models and cross-modal interaction, characterized in that: Includes the following steps: S1: Obtain the casting monitoring data and mold furnace identification data corresponding to the current casting furnace, and input the casting monitoring data into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification, and generate the suspected leakage identification results of each structure area. S2: Compare the suspected leakage identification result of the structural area with the preset leakage alarm condition and the preset suspected abnormal memory condition respectively. When the suspected leakage identification result of the structural area does not meet the leakage alarm condition but meets the suspected abnormal memory condition, generate a non-alarm suspected abnormal memory event for the corresponding structural area. S3: Based on the mold furnace identification data, write the non-alarm suspected abnormal memory events into the corresponding structural area hidden danger memory library, and form a non-alarm suspected abnormal memory sequence of the same mold structural area in multiple casting furnaces in the structural area hidden danger memory library; S4: Perform correlation analysis on the recurring features in the non-alarm suspected abnormal memory sequence to generate structural area hidden danger memory results for the corresponding mold structure area; S5: Based on the memory result of the structural area hazard, determine whether the corresponding mold structural area meets the preset leakage hazard warning conditions. If the leakage hazard warning conditions are met, output the leakage hazard warning result.

2. The casting leakage detection method based on big data model and cross-modal interaction according to claim 1, characterized in that: Step S1 includes: S11. Collect visible light image data and temperature image data corresponding to the current casting furnace using the existing image acquisition device at the casting site, and define the visible light image data and the temperature image data as casting monitoring data; S12. Obtain the mold number, furnace number, and pouring time information corresponding to the current casting furnace from the casting production management system or casting equipment control system, and define the mold number, the furnace number, and the pouring time information as mold furnace identification data; S13. Transmit the casting monitoring data and the mold furnace identification data to the big data model processing terminal, and perform time alignment processing on the visible light image data and the temperature image data according to the pouring time information; S14. Input the casting monitoring data after time alignment processing into the big data model according to the preset mold structure area division results to perform cross-modal anomaly identification and generate the suspected leakage identification results of each structure area. The mold structure region division result is obtained through the following steps: A1. Obtain the mold image or mold structure diagram corresponding to the casting mold; A2. Based on the gate area, riser area, parting surface area and mold edge area in the casting mold, the mold image or mold structure diagram is divided into regions to obtain multiple structural regions; A3. Assign a unique structural region number to each structural region; A4. Define the multiple structural regions and the corresponding structural region numbers as the mold structural region division result.

3. The casting leakage detection method based on big data model and cross-modal interaction according to claim 2, characterized in that: The suspected leakage identification result of the structural region is obtained through the following steps: B1. Based on the mold structure area division result and the structure area number, map the visible light image data and the temperature image data to the corresponding structure areas respectively; B2. Using the big data model, cross-modal anomaly identification is performed on the visible light image data and the temperature image data mapped to the same structural region number; B3. Based on the cross-modal anomaly identification results, obtain the location, duration, range of influence, and intensity of the anomaly region in the corresponding structural region. B4. Combine the location of the abnormal region corresponding to the same structural region number, the duration of the abnormality, the range of the abnormality, and the intensity of the abnormality to generate the suspected leakage identification result of the structural region.

4. The casting leakage detection method based on big data model and cross-modal interaction according to claim 3, characterized in that: The location of the abnormal region is determined by the overlap of the corresponding abnormal regions in the visible light image data and the temperature image data. The range of the abnormal influence is determined by the range of the image area covered by the location of the abnormal region within the corresponding structural region. The intensity of the abnormality is jointly determined by the abnormal temperature difference in the temperature image data and the abnormal brightness change in the visible light image data.

5. The casting leakage detection method based on big data model and cross-modal interaction according to claim 4, characterized in that: Step S2 includes: S21. Compare the suspected leakage identification result of the structural area with the preset leakage alarm conditions to obtain the leakage alarm comparison result; S22. The suspected leakage identification result of the structural area that has not met the leakage alarm condition is further compared with the preset suspected abnormal memory condition to obtain the suspected abnormal memory comparison result. S23. When the leakage alarm comparison result is that the leakage alarm condition has not been met, and the suspected abnormal memory comparison result is that the suspected abnormal memory condition has been met, the corresponding structural area is determined as the non-alarm suspected abnormal structural area. S24. The suspected abnormal memory condition is at least two of the following: the abnormal duration reaches the abnormal duration recording condition, the abnormal influence range reaches the abnormal influence range recording condition, and the abnormal intensity reaches the abnormal intensity recording condition.

6. The casting leakage detection method based on big data model and cross-modal interaction according to claim 5, characterized in that: The unreported suspected abnormal memory events are generated through the following steps: C1. Extract the mold number, furnace number, structural area number, abnormal area location, abnormal duration, abnormal influence range, and abnormal intensity corresponding to the non-alarm suspected abnormal structural area; C2. Generate the corresponding event generation time based on the pouring time information; C3. The mold number, the furnace number, the structural area number, the location of the abnormal area, the duration of the abnormality, the range of the abnormality, the intensity of the abnormality, and the event generation time are combined into the non-alarm suspected abnormal memory event. C4. Define the non-alarm suspected abnormal memory event as a data object that did not trigger a leak alarm but entered the hidden danger memory processing.

7. The casting leakage detection method based on big data model and cross-modal interaction according to claim 6, characterized in that: Step S3 includes: S31. Establish a corresponding structural area hazard memory bank according to the mold number and the structural area number; S32. Write the unalarmed suspected abnormal memory event into the structural area hidden danger memory bank corresponding to the mold number and the structural area number; S33. Sort the multiple non-alarm suspected abnormal memory events in the structural area hidden danger memory bank according to the furnace number; S34. Multiple non-alarm suspected abnormal memory events sorted by furnace number under the same mold number and the same structural area number are defined as non-alarm suspected abnormal memory sequences.

8. The casting leakage detection method based on big data model and cross-modal interaction according to claim 7, characterized in that: Step S4 includes: S41. Read multiple non-alarm suspected abnormal memory events from the non-alarm suspected abnormal memory sequence; S42. Count the number of suspected abnormal memory events that did not trigger an alarm and define them as the number of times they occurred repeatedly; S43. Calculate multiple adjacent furnace number differences based on the furnace number corresponding to two adjacent non-alarm suspected abnormal memory events, and define the maximum value among the multiple adjacent furnace number differences as the furnace interval; S44. Calculate the ratio of the overlapping area to the union area between the multiple abnormal area locations based on the abnormal area locations corresponding to the multiple non-alarm suspected abnormal memory events, and define the calculation result as the abnormal area location overlap situation. S45. The recurrence count, the furnace interval, and the overlap of the abnormal area location are defined as recurrence features.

9. The casting leakage detection method based on big data model and cross-modal interaction according to claim 8, characterized in that: Step S4 further includes: S46. Compare the number of recurrences with a preset recurrence count condition to obtain a recurrence count comparison result; compare the furnace interval with a preset maximum furnace interval condition to obtain a furnace interval comparison result; and compare the overlap of abnormal area locations with a preset area overlap condition to obtain an area overlap comparison result. S47. When the repetition count comparison result is that the repetition count reaches the preset repetition count condition, the furnace interval comparison result is that the furnace interval is not greater than the preset maximum furnace interval condition, and the region overlap comparison result is that the abnormal region location overlap reaches the preset region overlap condition, the corresponding structural region is determined as a candidate leakage hazard structural region. S48. Determine the hazard level corresponding to the candidate leakage hazard structural area according to the preset hazard level mapping relationship, and generate a structural area hazard memory result including the candidate leakage hazard structural area and the hazard level, wherein the preset hazard level mapping relationship is a correspondence between the hazard levels determined according to the number of recurrences, the furnace interval, and the overlap of the abnormal area location.

10. The casting leakage detection method based on big data model and cross-modal interaction according to claim 9, characterized in that: Step S5 includes: S51. When the hazard level meets the preset leakage hazard warning conditions, the candidate leakage hazard structural area is determined as the leakage hazard structural area, and a leakage hazard warning result including the leakage hazard structural area, the mold number, the furnace number and the hazard level is output.

Images