Laser welding OCT image noise suppression method based on multivariate corroboration

Abstract

The application discloses a laser welding OCT image noise suppression method based on multi-element verification, relates to the technical field of laser welding, and comprises the following steps: acquiring a keyhole depth signal in a laser welding process, continuously sampling at a frequency of 200 kHz or above, simultaneously acquiring coordinate information and attitude angle information of a welding end in a three-dimensional space, and completing point-by-point matching according to a uniform time to form an initial information set corresponding to depth information and spatial attitude point by point. Through the point-by-point correspondence between the depth information and the spatial attitude, the application realizes the correlation analysis of the depth change and the motion state, effectively eliminates the interference signals by combining mutation screening and frequency consistency screening. Meanwhile, the spatial statistical processing and the energy modulation mechanism are introduced, so that the depth response and the energy change are kept synchronous, thereby the stability and the reliability of the penetration data are improved, and reliable basis is provided for image noise suppression and anomaly detection.

Description

Technical Field

[0001] This invention relates to the field of laser welding technology, and more specifically to a method for noise suppression in laser welding OCT images based on multivariate verification. Background Technology

[0002] With the widespread application of laser welding technology in high-end manufacturing fields such as automobile manufacturing, power batteries, and aerospace, the demand for online monitoring of welding quality is increasing. Among these, weld penetration depth, as a key indicator of the strength and reliability of welded joints, has become a core parameter in process control. Optical coherence tomography (OCT) technology, due to its advantages such as high resolution, non-contact, and coaxial measurement, has been widely used for real-time detection of keyhole depth during laser welding. This technology uses the low-coherence interference principle of a broadband light source to acquire A-scan depth signals at a sampling frequency of up to hundreds of thousands of times per second, thereby achieving dynamic monitoring of the welding process. However, in actual industrial environments, complex interference factors such as shielding gas flow disturbance, metal spatter, and optical path instability exist during the welding process, inevitably introducing a large amount of random noise into the signals acquired by OCT, affecting the stability and accuracy of the measurement results.

[0003] The existing technology has the following shortcomings: Existing technologies primarily employ single-domain signal processing methods for OCT depth signals, such as median filtering in the time domain or low-pass filtering in the frequency domain, for noise reduction. However, these methods, based on only single-dimensional data, cannot effectively distinguish between actual weld depth variations and noise interference. When fluctuations occur in the weld depth signal, they may originate from process abnormalities during the actual welding process, or be caused by factors such as spatter obstruction or beam deflection, which existing methods struggle to accurately identify. Furthermore, OCT systems only provide one-dimensional depth information and lack the ability to perceive the position and orientation of the welding torch in three-dimensional space. This prevents the correlation analysis between weld depth variations and the motion state of the weld head, resulting in significant deficiencies in anomaly detection and root cause analysis.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a noise suppression method for laser welding OCT images based on multivariate verification, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a noise suppression method for laser welding OCT images based on multivariate verification, comprising the following steps: The keyhole depth signal during laser welding is acquired and continuously sampled at a frequency of 200kHz and above. At the same time, the coordinate information and attitude angle information of the welding end in three-dimensional space are acquired, and point-by-point matching is completed at a unified time to form an initial information set in which the depth information and spatial attitude correspond to each point. A mutation screening operation is performed on the point-to-point corresponding data in the initial information set. When the depth value jumps, the corresponding attitude change amplitude is extracted for comparison. Only the records where the depth change and attitude change are consistent are retained, and the records where the attitude remains stationary but the depth changes abruptly are removed to obtain the associated data. Periodic features are extracted from depth change information and pose change information in the associated data. By comparing the consistency of their frequency distributions, only low-frequency components that coexist are retained, and isolated frequency components that only appear in depth change information are removed to obtain compressed data. Calculate depth statistical features for multiple measurement records corresponding to the same spatial location in the compressed data, and remove records that deviate from the median by more than three times the discrete range and have prominent adjacent changes to obtain a candidate depth information set; Based on the candidate depth information set, the laser welding energy is sinusoidally modulated, and the depth change response over time is extracted simultaneously. Only the measurement records that keep the same frequency response as the energy change are retained, and the effective depth information for noise suppression of laser welding process images is output.

[0007] Preferably, by using a unified time identifier and completing point-by-point matching, a stable correspondence is established between depth information and spatial pose information, resulting in an initial information set with temporal consistency and spatial correlation. The steps are as follows: The keyhole depth signal is acquired and continuously sampled during laser welding. Each depth sampling point is assigned a unique time identifier. At the same time, the three-dimensional spatial coordinate information and attitude angle information of the welding end are read synchronously, and the time identifier source is kept consistent. For data records with consistent time stamps, point-by-point matching processing is performed. The depth values ​​at the corresponding time points are extracted from the depth sampling records, and the three-dimensional coordinate values ​​and attitude angle values ​​with the same time stamps are extracted from the spatial coordinate records. The three are then combined into a data unit. The data unit is organized and processed in a uniform format, arranging the time stamp, depth value, 3D coordinate value and attitude angle value according to fixed field positions, while maintaining the continuity of time sequence; The data units are processed to perform a continuity check, and the time markers are checked one by one and abnormal locations are marked. At the same time, the correspondence between depth values ​​and spatial coordinates and attitude angles is kept unchanged, thus obtaining the initial information set.

[0008] Preferably, by synchronously determining depth and attitude changes, abnormal signals are eliminated and complete depth and spatial attitude information is obtained. The steps are as follows: Read data point by point from the initial information set, calculate the depth difference between adjacent time points, extract the attitude angle change and merge the three directions, and establish a combined record containing time markers, depth change values ​​and attitude change amplitudes; Records showing synchronous changes in depth and attitude are marked as consistent changes, while records showing abrupt changes in depth but static attitude are marked as anomalous changes. Records of consistent changes are extracted and arranged in chronological order, while abnormal records are removed, so that each record meets the condition that depth changes and attitude changes occur simultaneously. By combining the filtered records with the original initial information, a complete set of associated data containing depth values, three-dimensional coordinates, and attitude angles is obtained, reflecting the synchronous relationship between depth changes and spatial attitude changes.

[0009] Preferably, frequency analysis is further performed on the consistent change records. By comparing the periodic consistency of depth changes and attitude changes within the time interval, only records in which depth changes fluctuate synchronously with attitude changes are retained. Abnormal records are identified by the change trends at continuous time points, thereby obtaining a more stable and reliable set of associated data.

[0010] Preferably, periodic feature extraction and frequency consistency screening are performed on the depth and attitude change information in the associated data. Low-frequency components of synchronous changes are retained and isolated interferences are removed to obtain compressed data that can reflect the synchronous characteristics of depth and attitude. This provides a reliable data foundation for subsequent noise suppression of laser welding process images. The steps are as follows: The depth change information in the associated data is read one by one according to the time identifier. The depth change value within the continuous time range is segmented and the complete change process is recorded. The change process is defined as a period and the start time, end time and duration are counted to construct a set of depth change period features. The attitude change information is processed according to the same time division as the depth change, the complete attitude change process is recorded and defined as a period, the start time, end time and duration of each period are registered, and a set of attitude periodic features consistent with the depth change structure is generated. The set of periodic features of depth change and the set of periodic features of attitude change are matched one by one according to the time window. The periods that exist at the same time are identified and marked as common frequency components, and the periods that exist only in depth change are marked as isolated frequency components. The depth change records corresponding to common frequency components are sorted in chronological order, and the records corresponding to isolated frequency components are deleted, so that the remaining data only contains the synchronization period information of depth change and attitude change, thereby obtaining compressed data, which provides reliable input for image noise suppression in the subsequent laser welding process.

[0011] Preferably, a dynamic amplitude comparison is performed on each depth change cycle and attitude change cycle in the compressed data, and only records where the depth response and attitude response amplitudes are in the same trend are retained. These records are then rearranged in chronological order to generate a stable and consistent set of depth information, which is used to enhance the accuracy of noise suppression in laser welding process images.

[0012] Preferably, a stable set of candidate depth information is obtained through depth statistical feature extraction and outlier record removal, providing reliable data for subsequent image noise suppression. The steps are as follows: Extract three-dimensional spatial coordinates from each record in the compressed data, divide the three-directional coordinates into intervals according to preset intervals, and form spatial location identifiers by combining interval numbers. Group records with the same spatial location and arrange them in chronological order. For each spatial location set, depth values ​​are read one by one, the deviation from the median is calculated, and all differences within the set are statistically analyzed to determine the discrete range. The discrete range is then expanded to three times the upper and lower boundaries. Each depth record in the spatial location set is checked for deviations from the median and changes from the preceding and following records. Records that simultaneously exceed three times the discrete range and show prominent adjacent changes are marked as anomalies. Abnormal records are removed from the spatial location set, unmarked records are retained and sorted in chronological order, and the spatial location sets are summarized to obtain a candidate depth information set. Records in the set maintain the continuity of depth values ​​and can reflect the characteristics of spatial location changes.

[0013] Preferably, by analyzing the periodic characteristics of depth changes and spatial attitude changes in the set, the low-frequency components that coexist are extracted, and depth change records that are not synchronized with attitude changes are deleted, and compressed data that maintains dynamic consistency between depth changes and spatial attitude is output to improve the noise suppression accuracy of laser welding images.

[0014] Preferably, effective depth information synchronized with energy changes is obtained through energy modulation and depth response correspondence filtering, which is used for noise suppression in laser welding process images. The steps are as follows: Extract the time stamp and depth value from the candidate depth information set, arrange them in chronological order, and adjust the laser welding energy output in segments so that the energy value gradually increases and decreases according to the cycle, forming a sinusoidal amplitude modulation change. The candidate depth information set aligned with time is segmented and processed, and the depth values ​​within the time interval corresponding to each energy change cycle are recorded one by one to obtain the complete response process of depth change over time. By comparing the deep response period with the energy change period, records with consistent duration and synchronized fluctuation rhythms are marked as synchronous responses, while records that do not meet the corresponding relationship are marked as non-synchronous responses. Extract the synchronous response records and organize them in chronological order, while deleting non-synchronous response records to obtain depth numerical data that only contains data synchronized with energy changes, thus obtaining effective depth information.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a multi-dimensional data relationship between depth information and spatial attitude point-by-point correspondence, and introduces a mutation screening and periodic feature consistency screening mechanism on this basis. This enables the depth change process to be synchronously correlated with the motion state of the welding end, thereby effectively distinguishing the source of change when the depth value fluctuates. This processing method can identify and eliminate abnormal signals caused by spatter obstruction and beam deflection in complex interference environments, making the retained depth data closer to the changes in the actual welding process, thereby improving the stability and reliability of the weld penetration measurement results.

[0016] After completing multidimensional consistency screening, this invention combines spatial location repeatable measurement statistical processing with sinusoidal amplitude modulation of laser welding energy to establish a temporal correspondence between depth changes and energy input, thereby screening out measurement records that maintain the same frequency response as energy changes. This method further enhances the dynamic consistency of depth data, giving the output effective depth information a stronger physical correlation. It provides more stable data support for noise suppression and anomaly alerts in laser welding process images, while also improving the responsiveness to changes in welding quality. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the method for noise suppression of laser welding OCT images based on multivariate verification according to the present invention. Detailed Implementation

[0019] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0020] This invention provides, for example Figure 1 The noise suppression method for laser welding OCT images based on multivariate verification, as shown, includes the following steps: The keyhole depth signal during laser welding is acquired and continuously sampled at a frequency of 200kHz and above. At the same time, the coordinate information and attitude angle information of the welding end in three-dimensional space are acquired, and point-by-point matching is completed at a unified time to form an initial information set in which the depth information and spatial attitude correspond to each point. By uniformly processing the time dimension, a one-to-one correspondence is established between depth information and spatial pose information, thereby constructing an initial information set for subsequent processing. The steps are as follows: The keyhole depth signal is acquired during laser welding, and the sampling frequency of the optical coherence tomography (OCT) device is set to 200kHz or higher for continuous sampling. During the sampling process, each depth sampling point is assigned a unique time identifier, which is generated by the internal clock counter to maintain a fixed time interval between adjacent sampling points. At the same time, within the same sampling period, the position coordinates of the welding end in three-dimensional space are read by the motion control unit, including displacement data along three orthogonal directions, and the attitude angle information of the welding end is read synchronously, including the rotation angle change value around the three coordinate axes. A unified clock source is used for time driving in the depth sampling and spatial attitude acquisition process, so that the depth sampling time identifier is consistent with the spatial coordinate and attitude angle acquisition time identifier, thereby ensuring the synchronization relationship of different types of data in the time dimension. During the data recording process, the depth value, the corresponding three-dimensional coordinate value, and the attitude angle value at each moment are written into the storage unit in chronological order to form a continuous record. Based on the depth data and spatial attitude data that have been unified with time stamps, the depth sampling records and 3D coordinate and attitude angle records are processed point by point. By matching the time stamps one by one, the depth value corresponding to a certain time point is selected from the depth data, and the 3D coordinate value and attitude angle value with the same time stamp are extracted from the spatial coordinate record. The three are combined into a set of data units. During the matching process, the time stamps are executed in ascending order so that each depth sampling point can find a unique corresponding spatial position and attitude angle information.

[0021] When there are records with time difference within the preset time window, the matching is completed by selecting the spatial attitude record with the smallest time difference, thereby avoiding mismatch problems caused by differences in acquisition cycle. After point-by-point matching, a data set is formed consisting of depth values, three-dimensional coordinate values ​​and attitude angle values. The data set after point-to-point correspondence processing is organized in a unified format. Each set of combined data containing depth values, three-dimensional spatial coordinates, and attitude angles is arranged in chronological order, and the data items are expressed in a structured manner so that each record contains a time stamp, depth value, coordinate values ​​in three directions, and attitude angle values ​​in three directions.

[0022] The chronological order is maintained during the arrangement process without disrupting the original record order. At the same time, different types of data are stored in fixed field positions in the data representation, so that depth information is always in the same field position, and spatial coordinate values ​​and attitude angle values ​​occupy independent field positions. Through this unified organization method, the data set can directly extract the corresponding information by field during subsequent processing, thereby ensuring the consistency of data retrieval. After organizing the data structure, all records are continuously processed, and the time markers are checked one by one to ensure that the time markers are arranged in ascending order with a fixed step size. When there are records with abnormal time intervals, the original record order is preserved and the corresponding position is marked to ensure that the data as a whole presents a continuous distribution in the time dimension. At the same time, the correspondence between the depth value, spatial coordinates and attitude angle in each record is kept unchanged. After processing, a complete initial information set is obtained. Each record in this initial information set contains depth information and spatial attitude information under a unique time marker, and can reflect the correspondence between the keyhole depth change and the motion state of the welding end during the laser welding process. This provides a multi-dimensional data foundation with temporal consistency and spatial correlation for subsequent laser welding process image noise suppression.

[0023] A mutation screening operation is performed on the point-to-point corresponding data in the initial information set. When the depth value jumps, the corresponding attitude change amplitude is extracted for comparison. Only the records where the depth change and attitude change are consistent are retained, and the records where the attitude remains stationary but the depth changes abruptly are removed to obtain the associated data. A mutation screening process is performed on the point-to-point corresponding data in the initial information set. This involves judging each record based on the correspondence between depth numerical changes and spatial attitude changes, thereby filtering out records where depth changes and attitude changes are consistent and forming associated data. The steps are as follows: The data corresponding to each point in the initial information set is read one by one in the order of time markers. The difference between the depth values ​​corresponding to two adjacent time points is calculated to obtain the depth change value of the current time point relative to the previous time point. At the same time, the corresponding attitude angle information is extracted under the same time marker, including the rotation angle values ​​around the three coordinate axes. The difference between the attitude angle values ​​of adjacent time points is calculated to obtain the attitude change in the three directions. The attitude change in the three directions is then merged to form a single attitude change amplitude record. At each time point, a combined data record containing the time marker, depth change value, and attitude change amplitude is established and stored one by one in chronological order, so that each record can completely reflect the depth change and spatial attitude change at that time point.

[0024] A mutation screening process is performed on each of the aforementioned combined data records. By continuously scanning the changes in depth values, the time points where depth values ​​jump are identified. During the identification process, a sudden increase or decrease in depth values ​​between adjacent time points is used as the criterion. When there is a significant difference between the depth value at a certain time point and the adjacent time points, that time point is marked as the location where the depth value jumps. At the same time, the already calculated attitude change amplitude is extracted at that time point, and the depth value jumps are compared with the attitude change amplitudes. During the comparison, records where depth and attitude changes simultaneously are marked as consistent change records, and records where the depth value jumps while the attitude change amplitude remains unchanged or changes very little are marked as abnormal change records. The classification process of all data is completed by marking each record individually.

[0025] After the data records are marked, a filtering operation is performed. Data marked as consistent changes are extracted one by one and rearranged while maintaining the original time order, so that the data is continuously distributed in the time dimension. At the same time, data marked as abnormal changes are removed. During the removal process, the depth change value and attitude change amplitude information of the corresponding time point are directly removed, and it is ensured that the remaining data does not contain any records with inconsistent depth and attitude changes.

[0026] This filtering process ensures that each record in the dataset meets the condition that depth changes and attitude changes occur synchronously, thereby eliminating interference from sudden depth changes caused by splash obstruction, airflow disturbance, or optical path offset.

[0027] The data records retained after filtering are linked and integrated. The time stamp in each record is used as an index to extract the original depth value, three-dimensional spatial coordinate information, and attitude angle information at the corresponding time point from the initial information set. These are then combined with the filtered depth change value and attitude change amplitude to form a new data set. In this data set, each record contains complete depth information and spatial attitude information, while maintaining consistency between the time order and the spatial correspondence. The linked data obtained in this way can accurately reflect the synchronous relationship between the depth change process and the spatial attitude change of the welding end, and provide a data foundation with consistent characteristics for subsequent processing.

[0028] Periodic features are extracted from depth change information and pose change information in the associated data. By comparing the consistency of their frequency distributions, only low-frequency components that coexist are retained, and isolated frequency components that only appear in depth change information are removed to obtain compressed data. Periodic feature extraction and frequency distribution consistency screening are performed on depth and pose change information in the associated data. By retaining low-frequency components that coexist in both types of information and removing isolated frequency components that only appear in depth change information, compressed data with consistent dynamic characteristics is formed. The specific steps are as follows: Periodic feature extraction is performed on depth change information in associated data. Depth change records are read one by one according to time identifier order. The depth change values ​​within a continuous time range are divided into fixed time windows. Each time window covers multiple consecutive sampling points. Within each time window, the depth change values ​​are scanned point by point, recording the complete change process of depth change from rising to falling and then rising again. This change process is defined as a period, and the start time, end time, and duration of the period are recorded. At the same time, the total number of periods appearing within the time window is counted, and each period is assigned a unique number. During the recording process, the time position and duration corresponding to each period are written into the periodic feature set, so that the depth change information forms a frequency distribution expression composed of multiple periods in the time dimension. This expression method can reflect the recurrence and distribution pattern of depth change in different time intervals.

[0029] For attitude change information in the associated data, the same processing method as for depth change information is used for periodic feature extraction. The attitude change amplitude is segmented according to the same time window division method as the depth change information. Within each time window, the attitude change amplitude is scanned point by point, recording the complete change process of attitude change from one direction to another and back. This change process is defined as an attitude change cycle. The start time, end time, and duration of each attitude change cycle are recorded. At the same time, the occurrence frequency of the attitude change cycle within the time window is counted, and each cycle is assigned a number. During the recording process, the cycle number, corresponding time position, and duration are stored in a unified manner, so that the attitude change information forms a periodic feature set with the same structure as the depth change information, thereby ensuring the comparability of the two types of information in subsequent processing.

[0030] The frequency distribution consistency of the periodic feature sets of depth change information and attitude change information is compared. The two types of periodic feature sets are matched one by one according to time windows. Within each time window, the depth change cycle and attitude change cycle are matched item by item. By comparing the duration range and occurrence position of the cycle, it is determined whether the depth change cycle and attitude change cycle exist simultaneously in the same time interval. For cycles whose duration ranges overlap within the same time window, they are marked as common frequency components, and the corresponding time position and cycle number are recorded. For records that only appear in the depth change information but are not found in the corresponding time window, they are marked as isolated frequency components. Through the comparison process of each time window, a set of filtering results containing only common frequency components is formed.

[0031] The filtered result set is reconstructed by extracting and rearranging the depth change records corresponding to the time intervals marked as common frequency components in chronological order. At the same time, the depth change records corresponding to the time intervals marked as isolated frequency components are deleted. During the deletion process, the depth change values ​​within the corresponding time range are directly removed, so that the remaining data only contains depth change information that occurs synchronously with the attitude change cycle.

[0032] By continuously organizing the filtered data to maintain a consistent order over time and forming compressed data, the compressed data retains the main dynamic features of depth changes while removing change components that do not correspond to spatial poses. This provides more stable and consistent input data for noise suppression in subsequent laser welding process images.

[0033] Calculate depth statistical features for multiple measurement records corresponding to the same spatial location in the compressed data, and remove records that deviate from the median by more than three times the discrete range and have prominent adjacent changes to obtain a candidate depth information set; For multiple measurement records corresponding to the same spatial location in compressed data, depth statistical feature extraction and outlier record removal are performed. Records that deviate from the median by more than three times the discrete range and have significant adjacent changes are identified and removed, thereby forming a stable set of candidate depth information. The specific process is as follows: For each record in the compressed data, the corresponding three-dimensional spatial coordinate values ​​are extracted. The coordinate values ​​in the three directions are discretized according to a preset interval. For example, the coordinate values ​​in each direction are divided into multiple intervals with a fixed step size, and the coordinate values ​​falling within the same interval are grouped into the same category. By combining the interval numbers in the three directions, a unique spatial location identifier is formed. Based on this spatial location identifier, all records are grouped so that records with the same coordinate position or within the same interval are grouped into the same set. During the grouping process, the time identifier and depth value of each record remain unchanged. Records in the same spatial location set are sorted according to the time identifier from smallest to largest, so that multiple measurement records corresponding to the same spatial location form a data set arranged in chronological order, thereby providing a unified data structure for subsequent statistical processing.

[0034] For each spatial location set, depth statistical feature calculation is performed on multiple measurement records. During the processing, the depth values ​​in the set are read one by one, all depth values ​​are sorted according to their numerical value, and the depth value in the middle position after sorting is selected as the median. At the same time, the difference between each depth value and the median is calculated, all differences are recorded one by one, and these differences are sorted to determine the discrete range.

[0035] When determining the discrete range, all differences are summarized and statistically analyzed. The differences are arranged in ascending order, and the overall distribution range of the differences is used as the basis for determining the discrete range. Based on this, the discrete range is expanded to three times the range to obtain the upper and lower boundary values. This ensures that each depth value corresponds to a deviation distance from the median and a corresponding boundary range, thereby providing clear judgment conditions for the identification of abnormal records.

[0036] Anomaly record filtering is performed on all depth values ​​and their corresponding deviation distances in each spatial location set. During the filtering process, each record's depth value is checked one by one to see if it exceeds the upper and lower boundaries of three times the discrete range. At the same time, the depth values ​​of the previous and next records in the time sequence are extracted. The difference between the current record and its adjacent records is calculated and used as a representation of the adjacent change amplitude. When a record meets both the conditions of its depth value exceeding three times the discrete range and having an abrupt change amplitude with its adjacent records, the record is marked as an anomaly record, and its corresponding time identifier and spatial location identifier are saved. The same processing is performed on all records in all spatial location sets during the filtering process, so that anomaly records are completely marked in all data.

[0037] The marked dataset is processed by removing all depth values ​​marked as abnormal from their corresponding spatial location sets. Unmarked records are rearranged according to their original chronological order, ensuring that the remaining records are continuously distributed within each spatial location set. The retained records from each spatial location set are then aggregated to form a new dataset, which is the candidate depth information set. In this set, each record satisfies the condition that its depth value is within three times the discrete range and that its changes with adjacent records are continuous. This provides stable and consistent data input for noise suppression in the subsequent laser welding process.

[0038] Based on the candidate depth information set, the laser welding energy is sinusoidally modulated, and the depth change response over time is extracted simultaneously. Only the measurement records that keep the same frequency response as the energy change are retained, and the effective depth information for noise suppression of laser welding process images is output. Based on the candidate depth information set, sinusoidal amplitude modulation of laser welding energy and depth response extraction are performed. By selecting measurement records that maintain the same frequency response as the energy change, effective depth information for noise suppression in laser welding process images is formed. The specific steps are as follows: For each record in the candidate depth information set, the corresponding time identifier and depth value are extracted and arranged in ascending order of time identifier. All records are organized into a continuous time data set. Based on this, the energy output in the laser welding process is controlled and processed. The original constant output energy value is segmented according to a fixed time interval. Within each time interval, the energy value is gradually increased or decreased according to a preset change pattern, so that the energy output exhibits periodic fluctuations in the time dimension.

[0039] In each cycle, the energy is gradually increased from the initial value to the peak value, and then gradually decreased from the peak value to the initial value. This process is repeated continuously to form a sinusoidal amplitude modulation process with a fixed period and a fixed amplitude range. At the same time, the start time of the energy change is aligned with the start time in the candidate depth information set, so that the energy output value at each time point can correspond to the specific depth record position in the candidate depth information set.

[0040] For the candidate depth information set that has been aligned with energy changes, depth response extraction processing is performed. During the processing, the depth values ​​are segmented according to a time division method that is completely consistent with the energy changes. The time interval corresponding to each energy change cycle is treated as an independent analysis interval. Within each interval, depth values ​​are read one by one, and the change process of depth values ​​over time is recorded. The change process of depth values ​​from rising to falling and then rising again is completely recorded, and this change process is divided into a depth response cycle. During the recording process, the start time, end time, and duration of each depth response cycle are clearly marked. At the same time, the positions of the maximum and minimum depth values ​​in the time dimension within the cycle are recorded, so that the change response of depth over time can be completely reflected in the corresponding time interval, and the depth response cycle and the energy change cycle are kept consistent in time position.

[0041] A synchronous response screening process is performed based on the correspondence between the depth response period and the energy change period. During the process, the depth response period and the energy change period are compared one by one in each time interval. The duration of the depth response period is matched with the duration of the energy change period. When the two are consistent in time span and the fluctuation rhythm of the depth change is synchronized with the fluctuation rhythm of the energy change, the measurement record corresponding to the time interval is marked as a synchronous response record. When the duration of the depth response period is inconsistent with the energy change period or the depth change does not show a corresponding rhythm with the energy change, the measurement record corresponding to the time interval is marked as a non-synchronous response record. The above judgment operation is performed one by one in all time intervals during the screening process, so that synchronous response records and non-synchronous response records can be completely distinguished in the data.

[0042] The marked dataset is filtered and organized. Depth values ​​marked as synchronous response records are extracted one by one and rearranged in chronological order. At the same time, depth values ​​marked as non-synchronous response records are deleted from the candidate depth information set. This leaves only measurement records that respond synchronously with energy changes. The retained records are then continuously organized to ensure that the time markers are in ascending order and that the depth values ​​correspond one-to-one with the energy change cycle. This forms the final output of effective depth information, which reflects the synchronization relationship between depth changes and laser welding energy changes. This information is used for noise suppression and penetration depth determination in laser welding process images.

[0043] This invention constructs a multi-dimensional data relationship between depth information and spatial attitude point-by-point correspondence, and introduces a mutation screening and periodic feature consistency screening mechanism on this basis. This enables the depth change process to be synchronously correlated with the motion state of the welding end, thereby effectively distinguishing the source of change when the depth value fluctuates. This processing method can identify and eliminate abnormal signals caused by spatter obstruction and beam deflection in complex interference environments, making the retained depth data closer to the changes in the actual welding process, thereby improving the stability and reliability of the weld penetration measurement results.

[0044] After completing multidimensional consistency screening, this invention combines spatial location repeatable measurement statistical processing with sinusoidal amplitude modulation of laser welding energy to establish a temporal correspondence between depth changes and energy input, thereby screening out measurement records that maintain the same frequency response as energy changes. This method further enhances the dynamic consistency of depth data, giving the output effective depth information a stronger physical correlation. It provides more stable data support for noise suppression and anomaly alerts in laser welding process images, while also improving the responsiveness to changes in welding quality.

[0045] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A noise suppression method for laser welding OCT images based on multivariate verification, characterized in that, Includes the following steps: The keyhole depth signal during laser welding is acquired and continuously sampled at a frequency of 200kHz and above. At the same time, the coordinate information and attitude angle information of the welding end in three-dimensional space are acquired, and point-by-point matching is completed at a unified time to form an initial information set in which the depth information and spatial attitude correspond to each point. A mutation screening operation is performed on the point-to-point corresponding data in the initial information set. When the depth value jumps, the corresponding attitude change amplitude is extracted for comparison. Only the records where the depth change and attitude change are consistent are retained, and the records where the attitude remains stationary but the depth changes abruptly are removed to obtain the associated data. Periodic features are extracted from depth change information and pose change information in the associated data. By comparing the consistency of their frequency distributions, only low-frequency components that coexist are retained, and isolated frequency components that only appear in depth change information are removed to obtain compressed data. Calculate depth statistical features for multiple measurement records corresponding to the same spatial location in the compressed data, and remove records that deviate from the median by more than three times the discrete range and have prominent adjacent changes to obtain a candidate depth information set; Based on the candidate depth information set, the laser welding energy is sinusoidally modulated, and the depth change response over time is extracted simultaneously. Only the measurement records that keep the same frequency response as the energy change are retained, and the effective depth information for noise suppression of laser welding process images is output.

2. The method for noise suppression in laser welding OCT images based on multi-factor verification according to claim 1, characterized in that, By unifying the time identifier and completing point-by-point matching, a stable correspondence is established between depth information and spatial pose information to obtain the initial information set. The steps are as follows: The keyhole depth signal is acquired and continuously sampled during laser welding. Each depth sampling point is assigned a unique time identifier. At the same time, the three-dimensional spatial coordinate information and attitude angle information of the welding end are read synchronously, and the time identifier source is kept consistent. For data records with consistent time stamps, point-by-point matching processing is performed. The depth values ​​at the corresponding time points are extracted from the depth sampling records, and the three-dimensional coordinate values ​​and attitude angle values ​​with the same time stamps are extracted from the spatial coordinate records. The three are then combined into a data unit. The data unit is organized and processed in a uniform format, arranging the time stamp, depth value, 3D coordinate value and attitude angle value according to fixed field positions, while maintaining the continuity of time sequence; The data units are processed to perform a continuity check, and the time markers are checked one by one and abnormal locations are marked. At the same time, the correspondence between depth values ​​and spatial coordinates and attitude angles is kept unchanged, thus obtaining the initial information set.

3. The method for noise suppression in laser welding OCT images based on multi-factor verification according to claim 2, characterized in that, By synchronously determining depth and attitude changes, abnormal signals are eliminated and complete depth and spatial attitude information is obtained. The steps are as follows: Read data point by point from the initial information set, calculate the depth difference between adjacent time points, extract the attitude angle change and merge the three directions, and establish a combined record containing time markers, depth change values ​​and attitude change amplitudes; Records showing synchronous changes in depth and attitude are marked as consistent changes, while records showing abrupt changes in depth but static attitude are marked as anomalous changes. Records of consistent changes are extracted and arranged in chronological order, while abnormal records are removed, so that each record meets the condition that depth changes and attitude changes occur simultaneously. By combining the filtered records with the original initial information, a complete set of associated data containing depth values, three-dimensional coordinates, and attitude angles is obtained, reflecting the synchronous relationship between depth changes and spatial attitude changes.

4. The method for noise suppression of laser welding OCT images based on multi-factor verification according to claim 3, characterized in that, Further frequency analysis was performed on the consistent change records. By comparing the periodic consistency of depth changes and attitude changes within the time interval, only records in which depth changes fluctuated synchronously with attitude changes were retained. Abnormal records were identified by the change trends at continuous time points, and a set of related data was obtained.

5. The method for noise suppression in laser welding OCT images based on multi-factor verification according to claim 3, characterized in that, Periodic feature extraction and frequency consistency screening are performed on depth and attitude change information in the associated data. Low-frequency components of synchronous changes are retained and isolated interferences are removed to obtain compressed data that reflects the synchronous characteristics of depth and attitude. The steps are as follows: The depth change information in the associated data is read one by one according to the time identifier. The depth change value within the continuous time range is segmented and the complete change process is recorded. The change process is defined as a period and the start time, end time and duration are counted to construct a set of depth change period features. The attitude change information is processed according to the same time division as the depth change, the complete attitude change process is recorded and defined as a period, the start time, end time and duration of each period are registered, and a set of attitude periodic features consistent with the depth change structure is generated. The set of periodic features of depth change and the set of periodic features of attitude change are matched one by one according to the time window. The periods that exist at the same time are identified and marked as common frequency components, and the periods that exist only in depth change are marked as isolated frequency components. The depth change records corresponding to common frequency components are sorted in chronological order, and the records corresponding to isolated frequency components are deleted, so that the remaining data only contains the synchronization period information of depth change and attitude change, thus obtaining compressed data.

6. The method for noise suppression of laser welding OCT images based on multi-factor verification according to claim 5, characterized in that, For each depth change cycle and attitude change cycle in the compressed data, a dynamic amplitude comparison is performed. Only records where the depth response and attitude response amplitudes are in the same trend are retained and rearranged in chronological order to generate a stable and consistent set of depth information.

7. The method for noise suppression of laser welding OCT images based on multi-factor verification according to claim 5, characterized in that, A stable set of candidate depth information is obtained through deep statistical feature extraction and outlier record removal. The steps are as follows: Extract three-dimensional spatial coordinates from each record in the compressed data, divide the three-directional coordinates into intervals according to preset intervals, and form spatial location identifiers by combining interval numbers. Group records with the same spatial location and arrange them in chronological order. For each spatial location set, depth values ​​are read one by one, the deviation from the median is calculated, and all differences within the set are statistically analyzed to determine the discrete range. The discrete range is then expanded to three times the upper and lower boundaries. Each depth record in the spatial location set is checked for deviations from the median and changes from the preceding and following records. Records that simultaneously exceed three times the discrete range and show prominent adjacent changes are marked as anomalies. Abnormal records are removed from the spatial location set, unmarked records are retained and sorted in chronological order, and the spatial location sets are summarized to obtain a candidate depth information set. Records in the set maintain the continuity of depth values ​​and can reflect the characteristics of spatial location changes.

8. The method for noise suppression of laser welding OCT images based on multi-factor verification according to claim 7, characterized in that, By analyzing the periodic characteristics of depth and spatial attitude changes in the dataset, low-frequency components that coexist are extracted, and depth change records that are not synchronized with attitude changes are deleted, resulting in compressed data that maintains dynamic consistency between depth changes and spatial attitude.

9. The method for noise suppression of laser welding OCT images based on multi-factor verification according to claim 7, characterized in that, By filtering based on the correspondence between energy modulation and depth response, effective depth information synchronized with energy changes is obtained. The steps are as follows: Extract the time stamp and depth value from the candidate depth information set, arrange them in chronological order, and adjust the laser welding energy output in segments so that the energy value gradually increases and decreases according to the cycle, forming a sinusoidal amplitude modulation change. The candidate depth information set aligned with time is segmented and processed, and the depth values ​​within the time interval corresponding to each energy change cycle are recorded one by one to obtain the complete response process of depth change over time. By comparing the deep response period with the energy change period, records with consistent duration and synchronized fluctuation rhythms are marked as synchronous responses, while records that do not meet the corresponding relationship are marked as non-synchronous responses. Extract the synchronous response records and organize them in chronological order, while deleting non-synchronous response records to obtain depth numerical data that only contains data synchronized with energy changes, thus obtaining effective depth information.

Images