Safety grating false alarm suppression method and system based on multispectral fusion

Abstract

The application provides a security grating false alarm suppression method and system based on multispectral fusion, and relates to the technical field of data processing.The method comprises the following steps: acquiring a multi-band detection data set, projecting different band signals to a unified amplitude scale interval, comparing the amplitude change direction and amplitude of different bands on adjacent time slices, generating a band correlation atlas, combining bands with consistent amplitude change direction and similar change amplitude into an aggregation unit, calculating the dynamic change amplitude of each aggregation unit, identifying the concentration degree of interference signals in each aggregation unit, obtaining an interference aggregation index, eliminating interference signals, generating an occlusion feature matrix, calculating the separation degree of occlusion signals according to the matrix, obtaining an occlusion separation ratio, generating a judgment output result and controlling the running state of a mechanical equipment.The application constructs a discrimination channel which is insensitive to interference and sensitive to occlusion, and suppresses the false alarm of a security grating caused by strong light, dust and high-temperature radiation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for suppressing false alarms of safety gratings based on multispectral fusion. Background Technology

[0002] In the field of industrial safety protection, safety light curtains are widely used for personnel protection in machining equipment. When an obstruction of the beam is detected, the control system immediately stops the equipment to prevent injury. Most existing safety light curtains are based on a single-band infrared detection principle, determining obstruction by comparing the beams emitted and received at the transmitting and receiving ends. This method is relatively mature in machining and material handling scenarios, but it may experience false triggering under conditions of strong light interference, floating dust, or high-temperature radiation.

[0003] For example, in a metal processing workshop, high-temperature metal splashes are generated during the workpiece cutting process. The infrared radiation emitted by these splashes overlaps with the grating detection band in a specific band, which may cause the receiver to receive abnormal signals and misjudge that the beam is blocked. Such false alarms will frequently trigger shutdowns, which not only affect the production cycle, but may also cause operators to ignore the real danger signals, posing a dual threat to personnel safety and production efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for suppressing false alarms of safety gratings based on multispectral fusion, in order to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for suppressing false alarms of safety gratings based on multispectral fusion, the method comprising:

[0007] Obtain a multi-band detection dataset and perform band filtering based on it, removing detection data of bands with signal amplitudes lower than the preset signal amplitude to obtain a valid band dataset;

[0008] Band mapping is performed based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain the band mapping matrix;

[0009] Cross-correlation analysis is performed based on the band mapping matrix to compare the direction and magnitude of amplitude changes in different bands in adjacent time slices, and a band correlation map is generated.

[0010] Based on the band correlation map, the bands with the same amplitude change direction and similar change magnitude are grouped into cluster units, and the dynamic change amplitude of each cluster unit is calculated to obtain the interference response data.

[0011] Multi-unit fusion is performed based on interference response data to identify the degree of concentration of interference signals in each aggregation unit and obtain the interference aggregation index.

[0012] The occlusion signal is separated based on the interference aggregation index, the interference signal is removed, an occlusion feature matrix is ​​generated, and the degree of separation of the occlusion signal is calculated based on the matrix to obtain the occlusion separation ratio.

[0013] The system determines the outcome based on the occlusion separation ratio, generates the result, and controls the operation of the mechanical equipment.

[0014] Furthermore, band mapping is performed based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain the band mapping matrix, including:

[0015] Amplitudes are extracted from the effective band dataset, and the minimum and maximum values ​​of each band within the time window are determined to obtain the amplitude range vector.

[0016] The amplitude ratio is calculated based on the amplitude range vector, and the amplitude of each band signal is proportionally converted to its amplitude range to obtain the proportional conversion matrix.

[0017] Amplitude normalization is performed based on the scaling matrix, mapping the scaling values ​​to a uniform amplitude scale range to obtain the amplitude normalization matrix.

[0018] Based on the amplitude unification matrix, the band index is used as the mapping coordinate to assign coordinate values, thus obtaining the band mapping matrix.

[0019] Furthermore, cross-correlation analysis is performed based on the band mapping matrix to compare the direction and magnitude of amplitude changes in different bands on adjacent time slices, generating a band correlation map, including:

[0020] Based on the band mapping matrix, the amplitude difference between adjacent time slices is calculated to obtain the amplitude difference matrix;

[0021] The direction of change is determined based on the amplitude difference matrix. An increase in amplitude is marked as positive, a decrease in amplitude is marked as negative, and no change in amplitude is marked as zero, thus obtaining a direction identification matrix.

[0022] Based on the band mapping matrix, calculate the amplitude ratio of adjacent time slices, compare the amplitude of the later time slice with the amplitude of the previous time slice, and obtain the amplitude ratio matrix.

[0023] Direction and amplitude matching are performed based on the direction identification matrix and amplitude ratio matrix. The correspondence between the changing direction and the changing amplitude of each band pair in adjacent time slices is recorded to obtain the correlation feature matrix.

[0024] Furthermore, pattern aggregation is performed based on the band correlation map, grouping bands with consistent amplitude variation directions and similar amplitudes into aggregation units, and calculating the dynamic variation amplitude of each aggregation unit to obtain interference response data, including:

[0025] Based on the band correlation map, bands with the same direction of change are selected and grouped into a direction set to obtain the direction set data.

[0026] The amplitude variation is calculated based on the directional set data. The amplitude variation of each band within each directional set is calculated within the time window to obtain the amplitude variation data.

[0027] Based on the magnitude change data, the similarity of magnitude is judged, and the set of directions with a difference in magnitude of change that is less than a preset difference threshold is merged into an aggregation unit to obtain the aggregation unit data.

[0028] The dynamic amplitude of the aggregated unit data is accumulated, and the cumulative amplitude, number of fluctuations, and standard deviation of amplitude of each aggregated unit within the time window are calculated to obtain the interference response data.

[0029] Furthermore, multi-unit fusion is performed based on the interference response data to identify the concentration of interference signals within each aggregation unit, resulting in an interference aggregation index, including:

[0030] Based on the interference response data, the amplitude difference between adjacent times of each aggregation unit within the time window is calculated to obtain the difference sequence;

[0031] The features of each aggregation unit are extracted based on the difference sequence, and the cumulative amplitude, number of direction switches, maximum single increment and average deviation are recorded to obtain the unit feature data.

[0032] Based on the unit characteristic data, the allocation of each aggregation unit is calculated, and the allocation of all aggregation units is normalized to determine the weight of each aggregation unit, thus obtaining the unit weight data.

[0033] Based on the unit feature data and unit weight data, the aggregation degree of each aggregation unit is calculated to obtain the unit aggregation degree data;

[0034] Based on the unit weight data, a centralized identification is performed, and the weights of all aggregated units are squared and summed to obtain the unit concentration coefficient.

[0035] The interference aggregation index is obtained by exponential coupling based on the unit weight number, unit aggregation degree data and unit concentration coefficient.

[0036] Furthermore, occlusion signals are separated based on the interference aggregation index, interference signals are removed, an occlusion feature matrix is ​​generated, and the degree of separation of occlusion signals is calculated based on it to obtain the occlusion separation ratio, including:

[0037] Interference aggregation index is used to classify the intervals to obtain interference classification data. The bands in the interference classification are marked as interference items based on the data and the band mapping matrix is ​​used to obtain the interference marking matrix.

[0038] Based on the interference labeling matrix, the interference items are removed and the unlabeled bands are retained to obtain the occlusion band matrix;

[0039] Based on the occlusion band matrix, time-series reconstruction is performed, the bands are rearranged in chronological order, and the missing parts are filled with interpolated values ​​to obtain the occlusion feature matrix.

[0040] Based on the occlusion feature matrix, the concentration and diffusion of each band are extracted, and the occlusion separation ratio is calculated.

[0041] Furthermore, based on the occlusion feature matrix, the concentration and diffusion of each band are extracted, and the occlusion separation ratio is calculated, including:

[0042] Based on the occlusion feature matrix, the quantile benchmark is determined, and the amplitudes of each band are arranged from smallest to largest to obtain the quantile threshold data.

[0043] Based on the quantile threshold data and the occlusion feature matrix, the top segment index set and the remaining segment index set are determined according to the quantile threshold of each band, and the threshold segment index data is obtained.

[0044] Based on the occlusion feature matrix, the ratio of the maximum length of the non-zero continuous segment to the length of the time window is calculated to obtain the discontinuity factor.

[0045] Based on the occlusion feature matrix and threshold index data, the top segment proportion and coverage ratio are calculated, and then weighted and summed with the discontinuity factor to obtain the band score data.

[0046] Based on the band score data, the ratio of each band score to its sum is used as the band weight;

[0047] Based on the threshold index data and the occlusion feature matrix, the amplitude of each band on the top index set is weighted and summed according to the band weight to obtain the concentrated quantity.

[0048] Based on the threshold index data and the occlusion feature matrix, the diffusion is summarized, and the amplitude and discontinuity factor of each band on the residual index set are weighted and summed to obtain the diffusion amount.

[0049] The occlusion separation ratio is obtained by performing a coupled calculation based on the concentration and diffusion amounts.

[0050] Secondly, a false alarm suppression system for safety gratings based on multispectral fusion, the system comprising:

[0051] The filtering module is used to acquire a multi-band detection dataset and perform band filtering based on it, removing detection data of bands with signal amplitudes lower than the preset signal amplitude to obtain a valid band dataset.

[0052] The mapping module is used to perform band mapping based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain the band mapping matrix;

[0053] The correlation module is used to perform cross-correlation analysis based on the band mapping matrix, compare the direction and magnitude of amplitude changes of different bands in adjacent time slices, and generate a band correlation map.

[0054] The interference module is used to perform pattern aggregation based on the band correlation map, combine bands with the same amplitude change direction and similar change magnitude into aggregation units, and calculate the dynamic change amplitude of each aggregation unit to obtain interference response data.

[0055] The aggregation module is used to perform multi-unit fusion based on interference response data, identify the degree of concentration of interference signals in each aggregation unit, and obtain the interference aggregation index.

[0056] The separation module is used to separate the occlusion signal according to the interference aggregation index, remove the interference signal, generate the occlusion feature matrix, and calculate the degree of separation of the occlusion signal based on it to obtain the occlusion separation ratio.

[0057] The judgment module is used to make judgments based on the occlusion separation ratio, generate judgment output results, and control the operating status of mechanical equipment.

[0058] The above-described solution of the present invention has at least the following beneficial effects:

[0059] This invention eliminates dimensional confusion caused by inconsistencies in the gain, detector response curves, and dynamic range of different transmit / receive links by projecting signals from different bands onto a unified scale and establishing a consistent coordinate system using band indices. This ensures that subsequent comparisons of direction and amplitude are in the same metric space, preventing certain high-gain channels from naturally dominating numerically and suppressing real physical changes. At the same time, the unified scale has a common mode compression effect on slow calibration drift, suppressing global offsets caused by temperature drift and aging. This ensures that the differential, ratio, and aggregation statistics of time-adjacent slices have a stable baseline, reducing the threshold resetting cost caused by changes in equipment status.

[0060] This invention captures the synchronous dynamic patterns of multiple bands by cross-comparing different bands in adjacent time slices with varying directions and amplitudes. True occlusion often manifests as a consistent attenuation direction and comparable amplitude levels across multiple bands, while interference such as thermal splashes and strong light flicker tends to show brief spikes or fluctuations in non-consistent directions in a few bands. By using the difference and ratio between adjacent time slices as the observation unit, the influence of slowly varying baselines, such as those used for temperature rise discrimination, is naturally offset, improving the ability to distinguish interference and reducing the risk of misinterpreting global slow drift or local flicker as occlusion.

[0061] This invention combines bands with consistent direction and similar amplitude into aggregation units, transforming the original high-dimensional, multi-channel noise-susceptible features into a small-group, mesoscale description. This aggregation suppresses the weight of occasional spikes in isolated channels in the global criterion, reducing the risk of a few anomalies skewing the overall picture. At the same time, by accumulating and statistically analyzing the group quantity within the time window, the robustness of the features can be significantly improved, avoiding misjudgments caused by short-term disconnections of individual channels or edge jitter due to occlusion.

[0062] This invention introduces weights and fusion at the aggregation unit level to obtain an interference aggregation index, which measures how evidence is distributed among units. Interference usually exhibits energy concentration and direction switching characteristics in a limited frequency band, while true occlusion is closer to a widely consistent energy decay across units. Fusion with concentration as the core can distinguish between abnormally active units and consistent changes in most units, making the index monotonically respond to the increase of interference and the decrease of occlusion, reducing the dimensional curse of downstream judgment, improving the stability and cross-scene transferability of online thresholds, and reducing fluctuations caused by repeated manual parameter tuning.

[0063] This invention further compresses the purified occlusion features into an occlusion separation ratio, which forms the scalar basis for the final decision. This helps to ensure the monotonicity and thresholdability of the decision. As the degree of occlusion increases, the ratio rises steadily, suppressing repeated triggering caused by jitter. The occlusion separation ratio is a coupling of purified occlusion evidence and residual diffusion. It significantly reduces the sensitivity to background intensity and slowly varying radiation in the scene, enhances the robustness of the threshold to environmental drift, and allows for backtracking at each stage of the pipeline. The decision is highly interpretable and easy to trace in security scenarios. Attached Figure Description

[0064] Figure 1 This is a flowchart of a method for suppressing false alarms of a safety grating based on multispectral fusion, provided in an embodiment of the present invention. Detailed Implementation

[0065] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0066] like Figure 1 As shown, embodiments of the present invention propose a method for suppressing false alarms of safety gratings based on multispectral fusion, the method comprising:

[0067] Obtain a multi-band detection dataset and perform band filtering based on it, removing detection data of bands with signal amplitudes lower than the preset signal amplitude to obtain a valid band dataset;

[0068] Band mapping is performed based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain the band mapping matrix;

[0069] Cross-correlation analysis is performed based on the band mapping matrix to compare the direction and magnitude of amplitude changes in different bands in adjacent time slices, and a band correlation map is generated.

[0070] Based on the band correlation map, the bands with the same amplitude change direction and similar change magnitude are grouped into cluster units, and the dynamic change amplitude of each cluster unit is calculated to obtain the interference response data.

[0071] Multi-unit fusion is performed based on interference response data to identify the degree of concentration of interference signals in each aggregation unit and obtain the interference aggregation index.

[0072] The occlusion signal is separated based on the interference aggregation index, the interference signal is removed, an occlusion feature matrix is ​​generated, and the degree of separation of the occlusion signal is calculated based on the matrix to obtain the occlusion separation ratio.

[0073] The system determines the outcome based on the occlusion separation ratio, generates the result, and controls the operation of the mechanical equipment.

[0074] In this embodiment of the invention, a multi-band detection dataset is acquired, and band filtering is performed based on it. Band detection data with signal amplitudes lower than a preset signal amplitude are removed to obtain a valid band dataset. Low-amplitude bands are removed to reduce the noise false consistency of false alarm trigger sources and reduce the computational burden caused by invalid dimensions, thus establishing a clean baseline for cross-band comparability and time-series discrimination. Band mapping is performed based on the valid band dataset, projecting different band signals onto a unified amplitude scale range and using the band index as the mapping coordinate to obtain a band mapping matrix. Different physical channels are projected onto the same metric space, and the amplitude comparison, ratio, and difference across bands are numerically comparable, providing a stable baseline for subsequent direction determination and amplitude registration. Cross-correlation analysis is performed based on the band mapping matrix to compare the amplitude change direction and magnitude of different bands in adjacent time slices, generating a band correlation map. The dual constraints of direction and amplitude make true occlusion and occasional interference separable in the feature space, providing high-purity relation input for subsequent pattern aggregation.

[0075] Pattern aggregation is performed based on band correlation maps, grouping bands with consistent amplitude changes and similar magnitudes into aggregation units. The dynamic amplitude of each aggregation unit is calculated to obtain interference response data. This transforms the fine-grained characteristics of high-dimensional channel-level noise susceptibility into a few mesoscale aggregation units, diluting isolated anomalous peaks through population statistics and ensuring clear contrast of interference / occlusion evidence at the unit level. Multi-unit fusion is then performed based on the interference response data to identify the concentration of interference signals within each aggregation unit, yielding an interference aggregation index. Multidimensional unit-level evidence is compressed into a single, thresholdable index, significantly reducing the complexity of the decision surface and simultaneously improving the impact of occlusion... The blocking shows the opposite trend, improving the discrimination and stability of false alarm suppression and occlusion recognition; the occlusion signal is separated according to the interference aggregation index, the interference signal is removed, the occlusion feature matrix is ​​generated, and the separation degree of the occlusion signal is calculated based on it to obtain the occlusion separation ratio. The mask-reconstruction process with the interference index as the guide decouples the interference mode from the occlusion mode, and the interpolation maintains the temporal continuity of the occlusion process, providing a stable scalar basis for the final judgment; the judgment is made according to the occlusion separation ratio, the judgment output result is generated and the operating state of the mechanical equipment is controlled. The occlusion separation ratio is used as the trigger basis, so that the false alarm triggering is effectively suppressed and high sensitivity to real occlusion is maintained.

[0076] This involves acquiring a multi-band detection dataset, filtering it by removing detection data for bands with signal amplitudes lower than a preset value, and obtaining a valid band dataset, specifically including:

[0077] On the safety grating that is transmitted between the transmitter and receiver, an independent sampling channel is configured for each spectral band. The analog quantities of each band are synchronously acquired with a uniform sampling period. After analog-to-digital conversion, the original amplitude sequence of the multi-band is obtained and arranged according to the time slot index. The sliding data frame with a preset time window length is maintained in the ring buffer of the controller to form a multi-band detection dataset. When any new time slot arrives, the amplitude of each band corresponding to the time slot is written into the buffer and the earliest frame is popped, so as to ensure that the data on which the calculation depends always covers the nearest window interval.

[0078] During the calibration phase before the equipment is installed and put into operation, the safety light curtain is placed in an unobstructed and interference-free state. The amplitude sequence of each band under this state is continuously collected, with a sampling time of not less than 5 seconds to ensure that multiple complete sampling cycles are included. The noise mean is calculated for the sampling data of each band. With noise standard deviation ,in This serves as a band index; simultaneously, under the same calibration conditions, a reference band closest to the working band is selected, and its peak amplitude is obtained. This serves as the benchmark for the maximum effective signal in the given scenario. Then, the preset signal amplitude for this band is determined using the following formula: ,in For the first The preset signal amplitude of the band, For reference band peak amplitude, , The first The noise mean and standard deviation of the band. This is the relative amplitude scaling factor, with a value ranging from 0.05 to 0.15. This is the noise compensation factor, with a value ranging from 0.8 to 1.2. This is the noise tolerance factor, ranging from 2 to 4, used to adjust the degree of immunity to occasional impulse noise. In actual operation, the controller can automatically update this factor every set interval (e.g., 10 minutes) within the time slice where no obstruction is detected. , And recalculate accordingly. It enables adaptive compensation for slowly changing factors such as changes in ambient light and temperature drift, thereby avoiding misjudgment or missed judgment caused by long-term fixed thresholds.

[0079] For each band, a robust amplitude statistic is calculated within the current sliding window to measure the effective energy level of that band within the window period. To avoid the influence of single-point anomalies on the judgment, the statistic can be the quantile of the amplitude within the window or a joint indicator of mean and variance, but the comparison quantity used for final screening is uniformly converted into a scalar with the same dimensions as the actual amplitude. Then, this scalar is compared with the preset signal amplitude one by one: when the effective amplitude of a certain band within the window is consistently lower than the preset threshold, the band is determined to be a low-amplitude channel and marked as to be removed within the current window period.

[0080] To reduce frequent entry and exit caused by channel jitter, a hysteresis and minimum hold time are introduced into the filtering logic: when a band is first marked as to be removed, the removal mark is only removed if it remains above the threshold for several consecutive time slices and exceeds the set recovery count; conversely, if it falls below the threshold, the removal mark is removed. This ensures the stability of the effective band set on the time axis. For bands determined to have low amplitude, their corresponding data are no longer involved in subsequent mapping, association, and aggregation calculations, forming an effective band dataset. This dataset is refreshed once per time slice, allowing the method to maintain an adaptive effective dimension under different operating conditions.

[0081] Specifically, the process of determining the occlusion separation ratio, generating a determination output result, and controlling the operating status of the mechanical equipment includes:

[0082] Configure a real-time determination channel for the occlusion separation ratio within the controller; at each time slice, read the current occlusion separation ratio and compare it with the determination threshold obtained by working condition calibration or online adaptive testing to generate a binary determination quantity: when the ratio meets the determination condition for occlusion to be established, set "occlusion = true"; when the ratio falls back and meets the release condition, reset.

[0083] To ensure the determinism and anti-jitter performance of industrial safety control, a debouncing and hysteresis strategy is introduced in the decision-making process: For setting "Obstruction = True", the ratio must continuously meet the threshold condition for a minimum confirmation time; for releasing "Obstruction = False", the ratio must continuously meet the recovery threshold condition and exceed the hysteresis interval. The controller sends the decision value to the safety loop or the PLC input, driving the emergency stop / permission link interlocked with the mechanical equipment: when "Obstruction = True", a stop command is output, disconnecting the equipment's power circuit and maintaining this until the release condition is met; when "Obstruction = False", a permission command is output, resuming equipment operation.

[0084] In engineering deployment, it is necessary to record the time series of the occlusion separation ratio and the timestamp of each state flip simultaneously, and combine them with the equipment operation log to form auditable traceability evidence; when the threshold is detected to be close to the boundary for a long time or the flip frequency is abnormally increased, a maintenance alarm is output to prompt the field to verify the calibration configuration and optical alignment, thereby ensuring the stability of long-term operation.

[0085] In a preferred embodiment of the present invention, band mapping is performed based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain a band mapping matrix, including:

[0086] Amplitudes are extracted from the effective band dataset, and the minimum and maximum values ​​of each band within the time window are determined to obtain the amplitude range vector.

[0087] The amplitude ratio is calculated based on the amplitude range vector, and the amplitude of each band signal is proportionally converted to its amplitude range to obtain the proportional conversion matrix.

[0088] Amplitude normalization is performed based on the scaling matrix, mapping the scaling values ​​to a uniform amplitude scale range to obtain the amplitude normalization matrix.

[0089] Based on the amplitude unification matrix, the band index is used as the mapping coordinate to assign coordinate values, thus obtaining the band mapping matrix.

[0090] In this embodiment of the invention, amplitude is extracted based on the effective band dataset, and the minimum and maximum values ​​of each band within the time window are determined to obtain an amplitude range vector. This eliminates the cross-band dimension inconsistency problem caused by differences in the gain of different transmit / receive links, laying a unified baseline for subsequent dimensionless processing. Amplitude ratios are calculated based on the amplitude range vector, and the amplitude of each band signal is proportionally converted to its amplitude range to obtain a ratio conversion matrix. This maps the original amplitude of each band to a dimensionless ratio related to its own dynamic range, preserving the discriminative information of relative changes and eliminating absolute amplitude bias across bands. The amplitude is normalized according to the scaling matrix, mapping the scaling values ​​to a unified amplitude scale range to obtain an amplitude unification matrix. This projects the responses of multi-source sensing channels onto the same metric space, ensuring the comparability of values ​​across different bands, equipment batches, and field locations. Based on the amplitude unification matrix, coordinates are assigned using band indices as mapping coordinates to obtain a band mapping matrix. This places the amplitudes at a unified scale into a fixed coordinate system according to the band indices, completing topological alignment in the spatial dimension and sequence alignment in the temporal dimension. This ensures the repeatability and traceability of subsequent feature calculations, providing a data structure foundation for future operations.

[0091] Specifically, the amplitude ratio is calculated based on the amplitude range vector, and the amplitude of each band signal is proportionally converted to its amplitude range to obtain the proportional conversion matrix, which includes:

[0092] Within the set time window, for each band in the valid band dataset Go through each time slice one by one to obtain each time slice. Corresponding original amplitude And combined with the pre-obtained minimum value of this band With the maximum value Through the proportional conversion formula Calculations are performed to obtain the proportion of the band in each time slice. When the denominator is zero or close to zero, outlier handling is performed, such as replacing it with a fixed proportion constant or the value of the previous time slice, to avoid numerical instability.

[0093] Specifically, amplitude normalization is performed based on the scaling matrix, mapping the scaling values ​​to a unified amplitude scale range to obtain an amplitude unification matrix, which includes:

[0094] A normalization mapping function is applied to each scale value of the scaling matrix to map it to a uniform amplitude scale interval, such as the [0,1] interval or an engineering-defined discrete scale interval. This mapping generally chooses a monotonic linear function to preserve the scaling relationship, but a compression mapping can also be applied without changing the monotonicity to suppress extreme peak values. At the same time, during the normalization process, possible out-of-bounds values ​​are truncated to ensure that all results are within the preset interval. The normalized result is the amplitude uniformity matrix.

[0095] Specifically, based on the amplitude unification matrix, coordinates are assigned using band indices as mapping coordinates to obtain the band mapping matrix, which includes:

[0096] A two-dimensional coordinate system with time as the column index and band index as the row index is established. Each element in the amplitude unification matrix is ​​filled into the corresponding position in the matrix according to its band number and corresponding time slice position. At the same time, metadata such as band index and physical wavelength is recorded, so that the matrix not only saves amplitude information, but also retains the physical attributes of the band. A band mapping matrix is ​​generated as the direct input data structure for subsequent cross-band and cross-time slice correlation analysis.

[0097] In a preferred embodiment of the present invention, cross-correlation analysis is performed based on the band mapping matrix to compare the amplitude change direction and magnitude of different bands in adjacent time slices, generating a band correlation map, including:

[0098] Based on the band mapping matrix, the amplitude difference between adjacent time slices is calculated to obtain the amplitude difference matrix;

[0099] The direction of change is determined based on the amplitude difference matrix. An increase in amplitude is marked as positive, a decrease in amplitude is marked as negative, and no change in amplitude is marked as zero, thus obtaining a direction identification matrix.

[0100] Based on the band mapping matrix, calculate the amplitude ratio of adjacent time slices, compare the amplitude of the later time slice with the amplitude of the previous time slice, and obtain the amplitude ratio matrix.

[0101] Direction and amplitude matching are performed based on the direction identification matrix and amplitude ratio matrix. The correspondence between the changing direction and the changing amplitude of each band pair in adjacent time slices is recorded to obtain the correlation feature matrix.

[0102] In this embodiment of the invention, the amplitude difference between adjacent time slices is calculated based on the band mapping matrix to obtain an amplitude difference matrix. The first-order difference cancels out the slowly varying baseline and global drift across time, providing a consistent measurement of dynamics on a unified scale and laying a stable numerical foundation for downstream direction determination and amplitude matching. The direction of change is determined based on the amplitude difference matrix, marking an increase in amplitude as positive, a decrease in amplitude as negative, and no change in amplitude as zero, resulting in a direction identification matrix. The continuously varying difference components are discretized into direction labels, providing conditions for subsequent direction and amplitude matching and avoiding the introduction of false positives when there are no significant changes. Matching; based on the band mapping matrix, the amplitude ratio of adjacent time slices is calculated, and the amplitude of the later time slice is compared with that of the previous time slice to obtain the amplitude ratio matrix, which cancels out the dimensional bias caused by the difference in static gain and dynamic range of different bands; direction and amplitude matching are performed based on the direction identification matrix and the amplitude ratio matrix, and the correspondence between the changing direction and changing amplitude of each band pair in adjacent time slices is recorded to obtain the correlation feature matrix. Through dual-condition filtering of consistent direction and similar amplitude, the coupling mismatch caused by a single index is significantly reduced, ensuring that the aggregation is based on the real cross-band covariance relationship rather than accidental consistency.

[0103] Specifically, direction and amplitude matching is performed based on the direction identification matrix and amplitude ratio matrix to record the correspondence between the changing direction and amplitude of each band pair in adjacent time slices, resulting in an association feature matrix, which includes:

[0104] First, define the input and output. The input is the direction identification matrix calculated according to adjacent time slices. Amplitude Ratio Matrix The output is the correlation feature matrix. This is used to quantify the covariance strength between any two bands over the entire time window. For the number of bands, This refers to the number of time slices. During the data preparation phase, [the following will be used]. and Align with the same time index set {2,...,T} on the time axis, which corresponds to the paired intervals of adjacent time slices; for cases where the amplitude ratio matrix may have zero or very small denominators, continue using the lower limit constant used in the ratio calculation process. The processing ensures that each element satisfies This avoids numerical overflow and amorphous shapes, ensuring the stability of subsequent logarithmic transformations and distance metrics. To enhance the comparability of amplitude levels, [the following is done / implemented]. Apply logarithmic mapping This converts multiplicative changes into additivity, making subsequent thresholds symmetrically distributed on the number axis and easy to set uniformly.

[0105] In the direction consistency screening stage, for any band pair With any time slice Two bands are considered comparable in direction only if their directional labels are exactly the same and both are non-zero in that time slice. Otherwise, it is directly judged as a mismatch and recorded as zero. In the amplitude similarity determination stage, the absolute value of the logarithmic ratio difference is used as the measure of amplitude level difference, and is defined as follows: Set a threshold. This can be obtained from the quantile statistics of the equipment calibration data. At that time, it is assumed that the amplitudes of the two bands are similar in that time slice; in order to suppress the dominance of matching by minimal changes, an amplitude significance threshold can be further introduced. ,Require Only when a valid match is recorded is it included, thus ensuring that the recorded correspondence is contributed by amplitude changes that have engineering significance.

[0106] During the matching scoring and time accumulation phase, for each Constructing instantaneous matching tokens , This is an indicator function. To reflect the relative importance of matching in different time slices, a dependency-only function can be introduced. dimensionless weights This is used to emphasize moments where the amplitude change is more significant. Then, accumulation and normalization are performed along the time dimension to obtain... This structure makes ∈[0,1] has probabilistic interpretability: the closer to 1, the more times the two bands simultaneously satisfy the same direction and similar amplitude within the window, and the more significant the change amplitude at these times, the closer to 0, the opposite is true.

[0107] In the matrix arrangement and output stage, The diagonal elements are set to zero to remove self-matching; symmetry processing is performed using... To eliminate asymmetric micro-errors caused by finite-precision calculations; if a binarized band pair relationship record is required, it can be obtained by... Apply global threshold ∈(0,1) , Let be the characteristic function, where A record matrix representing the correspondence between direction and amplitude within a statistical sense of a time window; it can also be used to maintain... The matrix is ​​a real number and is directly used as the weight input for subsequent pattern aggregation, achieving a natural transition from continuous intensity to discrete clustering. During the parameter setting and calibration phase, The parameters can be determined on representative working condition data by using quantiles or maximizing inter-class separability (e.g., maximizing the correct aggregation probability on labeled samples), thus avoiding reliance on fixed thresholds based on experience. The existence of these parameters does not change the technical essence of this step; they are only used to clarify the engineering implementation path, ensuring the "direction and amplitude matching" algorithm has transferable robustness across different devices and scenarios. Considering that this step only relies on... and Its computational complexity is mainly O(n). In engineering implementation, the computational burden can be reduced by limiting candidate bands to incremental updates of the sliding window, such as comparing only adjacent bands or bands with similar spectral response curves. Simultaneously, a sparse storage structure can be used to hold... Non-zero locations are used to save memory.

[0108] In the anomaly and boundary handling phase, for Short-term burrs and For extreme values, time majority voting and ratio pruning can be performed before proceeding to this step; the strict removal of zero-directed samples and the amplitude constraint in the logarithmic domain within this step jointly suppress spurious matches caused by weak perturbations; when some time slices have missing measurements or saturation values, they can be directly... Setting the center zeros does not account for weights, giving the statistic a natural tolerance for missing data, ultimately leading to... It can stably and quantitatively characterize the cross-band correspondence of adjacent time slices with consistent direction and similar amplitude.

[0109] In a preferred embodiment of the present invention, mode aggregation is performed based on the band correlation map, bands with consistent amplitude change directions and similar amplitudes are grouped into aggregation units, and the dynamic change amplitude of each aggregation unit is calculated to obtain interference response data, including:

[0110] Based on the band correlation map, bands with the same direction of change are selected and grouped into a direction set to obtain the direction set data.

[0111] The amplitude variation is calculated based on the directional set data. The amplitude variation of each band within each directional set is calculated within the time window to obtain the amplitude variation data.

[0112] Based on the magnitude change data, the similarity of magnitude is judged, and the set of directions with a difference in magnitude of change that is less than a preset difference threshold is merged into an aggregation unit to obtain the aggregation unit data.

[0113] The dynamic amplitude of the aggregated data is accumulated, and the cumulative amplitude, number of fluctuations, and standard deviation of amplitude of each aggregated unit within the time window are calculated to obtain the interference response data.

[0114] In this embodiment of the invention, bands with consistent directions are screened based on the band correlation spectrum. Bands with the same direction of change are grouped into direction sets to obtain direction set data. This shields the direction jitter caused by instantaneous noise and avoids misinterpreting the brief reversal of a few bands as differences in physical mechanisms. The amplitude change is calculated based on the direction set data, calculating the change amplitude of each band within each direction set within a time window to obtain amplitude change data. This compresses the discrete difference sequence in the time domain into a robust intensity characterization, ensuring that comparisons between different bands fall on the same numerical semantics, providing a stable and comparable scalar basis for subsequent comparisons. The amplitude variation data is judged for similarity. Directions with amplitude differences below a preset difference threshold are merged into aggregate units to obtain aggregate unit data. This reduces distortion caused by scale inconsistency during aggregation and provides a low-noise carrier for subsequent calculations. Based on the aggregate unit data, the amplitude variation is dynamically accumulated, and the cumulative amplitude, number of fluctuations, and standard deviation of amplitude variation of each aggregate unit within the time window are calculated to obtain interference response data. This presents a separable statistical form for truly consistent cross-band occlusion variations and localized, rapidly changing interference, providing input that is information-rich, appropriately dimensional, and noise-controlled for subsequent calculations.

[0115] Specifically, based on the data from the aggregation units, the dynamic amplitude of changes is accumulated, and the cumulative amplitude, number of fluctuations, and standard deviation of the amplitude of each aggregation unit within the time window are calculated to obtain the interference response data, which includes:

[0116] First, each aggregation unit contains several band indices that satisfy the criteria of "consistent change direction and similar change amplitude". Within the same time window, the normalized amplitude time series of each member band is extracted to form the input dataset of the aggregation unit. A unified configuration of the window start and end times and sampling interval is established to maintain the consistency of subsequent measurement standards. The time window here is consistent with the generation window of the aggregation unit data to ensure the comparability of statistics and avoid the introduction of truncation error. For each aggregation unit, the amplitude difference between adjacent sampling times is calculated on the time series of its member bands, and the absolute value of the difference is taken as the instantaneous change amplitude. To suppress the contamination of statistics by sensor quantization noise and extremely low amplitude disturbances, a small threshold is set for the instantaneous change amplitude, and samples below the threshold are set to zero. When there are isolated anomalous spikes, a median filter with a finite window or a mild truncation strategy is applied to the instantaneous change amplitude to constrain the excessive influence of outliers on accumulation and dispersion estimation, ensuring that the measurement better reflects the true dynamics of the aggregation unit within the window.

[0117] After obtaining the denoised instantaneous change amplitude, the cumulative amplitude is calculated: the instantaneous change amplitudes of all member bands of the aggregation unit at each time within the window are summed to obtain a cumulative measure in both time and member dimensions. When it is necessary to eliminate the scale effect caused by the difference in the number of members, the summation can be normalized once by the number of members or its effective sample number so that aggregation units of different sizes can be compared horizontally under the same threshold system. Then, the number of fluctuations is calculated: first, a unit change sequence is constructed at the aggregation unit level, which can be defined as the average instantaneous change amplitude of the members of the unit at each time or the sum of the members after scale normalization. A Schmitt trigger threshold pair (entry threshold and exit threshold) is introduced into the unit change sequence. When the sequence crosses the entry threshold from below the exit threshold, the fluctuation count is calculated. Each fluctuation event is recorded as a single event. Using a double threshold effectively avoids repeated counting of jitters near the threshold. The number of fluctuations is obtained by counting all rising edge events within the window. Next, the standard deviation of the change amplitude is calculated: taking the unit change sequence as the object, the unbiased sample standard deviation is used to estimate its dispersion within the time window. If necessary, missing samples are interpolated or skipped and the degrees of freedom are corrected accordingly. When the difference in amplitude scale between members of the aggregated unit has been unified through the previous steps, the standard deviation can directly characterize the time volatility of the dynamic strength of the unit. If it is also necessary to take into account the consistency between members, the instantaneous change amplitude of the members can be aligned with the mean of the members before the calculation, and then the standard deviation is calculated in the time dimension to enhance the sensitivity of the measurement to the overall consistent slow change pattern and weaken the occasional fluctuations of individual members.

[0118] After completing the above three metrics, a response vector consisting of "cumulative amplitude - number of fluctuations - standard deviation of amplitude change" is constructed for each aggregation unit, and the structured storage is performed using the aggregation unit index as the key to form the interference response data for that time window.

[0119] In a preferred embodiment of the present invention, multi-unit fusion is performed based on interference response data to identify the concentration of interference signals within each aggregation unit, thereby obtaining an interference aggregation index, including:

[0120] Based on the interference response data, the amplitude difference between adjacent times of each aggregation unit within the time window is calculated to obtain the difference sequence;

[0121] The features of each aggregation unit are extracted based on the difference sequence, and the cumulative amplitude, number of direction switches, maximum single increment and average deviation are recorded to obtain the unit feature data.

[0122] Based on the unit feature data, the allocation of each aggregation unit is calculated, and the allocation of all aggregation units is normalized to determine the weight of each aggregation unit, thus obtaining the unit weight data.

[0123] Based on the unit feature data and unit weight data, the aggregation degree of each aggregation unit is calculated to obtain the unit aggregation degree data;

[0124] Based on the unit weight data, a centralized identification is performed, and the weights of all aggregated units are squared and summed to obtain the unit concentration coefficient.

[0125] The interference aggregation index is obtained by exponential coupling based on the unit weight number, unit aggregation degree data and unit concentration coefficient.

[0126] In this embodiment of the invention, based on the interference response data, the amplitude difference between adjacent times of each aggregation unit within the time window is calculated to obtain a difference sequence, which suppresses occasional spikes to a controllable range and ensures stable convergence of subsequent statistics. Features of each aggregation unit are extracted based on the difference sequence, recording the cumulative amplitude, number of direction switches, maximum single increment, and average deviation to obtain unit feature data, characterizing the disturbance pattern within the unit and laying a data foundation for subsequent concentration identification and exponential coupling. Based on the unit feature data, the allocation of each aggregation unit is calculated, and the allocation of all aggregation units is normalized to determine the weight of each aggregation unit, obtaining unit weight data. The allocation compresses multidimensional features into comparable scalars and forms probabilistic weights through normalization, avoiding the dimensional bias of the naturally dominant high-gain channel. Based on unit feature data and unit weight data, the aggregation degree of each aggregation unit is calculated to obtain unit aggregation degree data. This makes stable and consistent interference / occlusion patterns stand out during subsequent global aggregation, while random noise and occasional spikes are marginalized. Centralized identification is performed based on unit weight data. The weights of all aggregation units are squared and summed to obtain the unit concentration coefficient. This coefficient, along with the unit aggregation degree, enhances the ability to identify concentrated interference and reduces the probability of individual unit anomalies affecting global decisions. An interference aggregation index is obtained by exponentially coupling the unit weights, unit aggregation degree data, and unit concentration coefficient. This index combines the robust aggregation degree within a unit with the global centralized structure, making it highly sensitive to local concentrated interference and relatively insensitive to global consistent changes. This meets the requirements of single-index and monotonicity for subsequent occlusion separation and judgment.

[0127] Specifically, features of each aggregation unit are extracted based on the difference sequence, and the cumulative amplitude, number of direction switches, maximum single increment, and average deviation are recorded to obtain unit feature data, which specifically includes:

[0128] First, within a fixed-length time window, for each aggregation unit obtained from the preceding aggregation step... Read its amplitude time series And calculate the amplitude difference between adjacent time slices. , The time window starts counting from the next moment after the window's starting point; the time window length and sliding step size are configured with a trade-off between system real-time performance and smoothness, and the sampling interval remains constant to avoid distortion of the difference quantity. After the difference is completed, the feature extraction stage can begin, which uses the difference sequence as direct input and reviews it when necessary. To obtain a stable baseline. To suppress the impact of quantization noise and occasional spikes on feature stability, [the following measures are taken]. Perform a light preprocessing step: reduce near-zero small fluctuations to zero by applying a threshold, which can be set based on the median absolute deviation of the difference sequence within the current window; for single-point anomalies, IQR truncation can be used, and the truncation ratio and effective sample number can be recorded to provide a basis for subsequent quality monitoring; without changing the trend, [further details needed]. Applying two-sided three-point smoothing or a single median filter makes subsequent statistics insensitive to peaks but sensitive to continuous changes, thereby ensuring the repeatability and comparability of the four types of features.

[0129] Amplitude accumulation Used to characterize the overall change intensity of a cell within a window, its calculation is based on the absolute summation of differences: To adapt to data sources with different dimensions, adjustments can be made before calculation. Dimensionless scaling is performed using a uniform amplitude scale obtained from the preceding mapping or based on the range within the window, ensuring... It is comparable between different units.

[0130] Direction switching number To capture the stability and perturbation frequency of changing directions, a sign sequence is first obtained by applying a sign function to the preprocessed difference. Then, the number of sign changes at adjacent time points is counted; when there are consecutive zero segments, the most recent non-zero sign is used to maintain a consistent direction determination; the resulting... The value is lower under continuous unidirectional changes (such as continuous attenuation caused by real occlusion), and higher under flickering interference or oscillation.

[0131] Maximum single increment The strength of the most significant mutation within a window is measured as follows: To reduce the impact of occasional noise, it is possible to obtain Simultaneously calculate the first The magnitude of the increment in quantiles is used as a reference (e.g., the 95th percentile). When the ratio to the reference exceeds a given ratio, it is marked as peak-dominant and the mark is appended to the output feature record; in engineering implementation, Locating the corresponding time period Indexes facilitate downstream backtracking and visual auditing.

[0132] Average deviation This is used to measure the steady-state offset of a cell relative to its own baseline within a window. In this embodiment, the original sequence is used. Relative to its window size The mean absolute deviation is achieved. ,in The effective sample size; using the median as the baseline maintains robustness in the presence of local mutations; it can also be used when it is necessary to maintain the same dimensions as the difference domain. First, normalize according to a uniform amplitude scale, and then calculate the deviation. The larger the value, the more likely the cell is to have a continuous overall offset or slow drift within the current window.

[0133] The formula for calculating the interference aggregation index is as follows:

[0134] ;

[0135] in, To interfere with the aggregation index, and For the index of the aggregation unit, The total number of aggregation units. Aggregation unit The weight, , Aggregation unit The cumulative amplitude, Aggregation unit Number of direction switching, Aggregation unit The maximum single increment, Aggregation unit The average deviation Aggregation unit The weight, , These are the weighting coefficients. Aggregation unit The cumulative amplitude, Aggregation unit Number of direction switching, Aggregation unit The maximum single increment, is a coefficient.

[0136] in, These are weighting coefficients, and their sum is 1. In scenarios dominated by transient interference such as strong light / arc flashes or metal spatter, The values ​​are 0.25, 0.6, and 0.15, respectively, because this type of interference has the statistical characteristics of short-term strong peaks, and a single increment... Its contribution to separability is far greater than its contribution to cumulative quantity. And direction switching Sampling jitter can easily lead to artificially high values, and excessive weighting can amplify noise; therefore, let Being dominant helps to quickly highlight the sharp anomalies within a small number of units to the weighting layer, which are then enhanced in subsequent aggregation and concentration coefficients, and has a stronger ability to suppress false alarms caused by strong light and metal splashes as described in the background art.

[0137] In a steady-state disturbance scenario where high-temperature radiation drifts slowly, The values ​​are 0.6, 0.25, and 0.15, respectively, because the background rise caused by thermal radiation exhibits a long-term, slow-changing pattern with an overall increasing trend, and the cumulative amount... Most sensitive to this type of trend; if still based on If the primary evidence is misplaced, occasional spikes may be mistakenly taken as primary evidence, leading to misjudgment; while In slow drift, the weight is usually low, and increasing it does not improve the discriminative power; instead, it introduces sensitivity to sampling noise. Therefore, let leading, Secondly, Minimum size better fits the mechanism.

[0138] In situations where dust particles cause fine obstruction / diffuse reflection, it is recommended to take... The values ​​are 0.4, 0.25, and 0.35, respectively, because this scenario typically involves small, frequent, and inconsistent jitter, with a single increment... With cumulative amount None of them are significant, but the number of direction switches is significant. Significantly increased, moderately improved This allows for the timely highlighting of high-frequency directional flipping units as high-weighted entities, thereby demonstrating localized active interference patterns without cross-unit consistency in subsequent centralized identification stages, helping to distinguish them from the unidirectional consistent attenuation of true occlusion.

[0139] Among them, coefficient Acting on , This compromises the sudden amplitude and directional jitter frequency into a weighted average of a single comparable quantity. These two quantities have different physical meanings and statistical stability—in real-world multispectral scenarios, the number of directional switches is easily inflated by quantization noise and weak flicker disturbances, while the maximum single increment is more separable from transient interferences such as strong light / splashes; therefore, let To improve The relative power of discourse can both enhance sensitivity to spike interference and suppress the increase in false positives caused by high-frequency jitter; Setting it to 0.4 allows Linear weights account for approximately 71.4% of the total. It accounts for approximately 28.6%, which can serve as a robust starting point in offline calibration and can be finely adjusted to 0.6 as needed under conditions where flickering is the primary issue.

[0140] coefficient Acting on The independent variable is located in [0,1], representing the relative stability of the intensity. This process amplifies high quantile regions (values ​​close to 1) and compresses low quantile regions, allowing units with strong accumulation and small average deviation to contribute more, thus favoring a stable and consistent change pattern; if If the value is too large, it will lead to high quantile saturation and excessive suppression of samples with moderate intensity, weakening the separability to mild to moderate interference; therefore, it will... Setting it to 1.2 provides a 20% slope amplification without causing saturation, balancing discriminability and numerical stability, while maintaining a match with the magnitude of the weighted normalization term.

[0141] coefficient Acting on , It is a nonlinear enhancement term for spike / jitter evidence, which occurs when the interference has more localized abrupt increases or frequent switching statistical characteristics compared to occlusion. It should be slightly higher To highlight the passage; Setting it to 1.5 will result in a significantly higher improvement in the mid-to-high quartile range compared to... The intensity stabilization term is applied to create a distance between the concentrated and sudden occlusion patterns and the slow, consistent occlusion patterns, while still avoiding numerical explosions and oversensitivity to occasional spikes.

[0142] coefficient It involves combining the two types of evidence by exponentiation, within the first set of parentheses. This combines unit strength and spike / jitter information, followed by a parenthesis. This is used to measure whether evidence is dominated by a minority of units; The larger the cell, the more the model relies on the strength of evidence within the cell. Smaller, more dependent on a centralized structure across cells; considering that false alarms are often caused by anomalies in a few cells, while true occlusion tends to be consistent across cells, then... Setting it to 0.65 allows strength evidence to dominate while still retaining 35% weight for concentrated structures, achieving a more robust trade-off across multiple operating conditions.

[0143] In a preferred embodiment of the present invention, occlusion signal separation is performed based on the interference aggregation index, interference signals are eliminated, an occlusion feature matrix is ​​generated, and the degree of separation of occlusion signals is calculated based on the matrix to obtain an occlusion separation ratio, including:

[0144] Interference aggregation index is used to classify the intervals to obtain interference classification data. The bands in the interference classification are marked as interference items based on the data and the band mapping matrix is ​​used to obtain the interference marking matrix.

[0145] Based on the interference labeling matrix, the interference items are removed and the unlabeled bands are retained to obtain the occlusion band matrix;

[0146] Based on the occlusion band matrix, time-series reconstruction is performed, the bands are rearranged in chronological order, and the missing parts are filled with interpolated values ​​to obtain the occlusion feature matrix.

[0147] Based on the occlusion feature matrix, the concentration and diffusion of each band are extracted, and the occlusion separation ratio is calculated.

[0148] In this embodiment of the invention, interference aggregation index is used to classify intervals to obtain interference classification data. Based on this data, the band mapping matrix is ​​marked with positions, and the bands in the interference classification are marked as interference items, resulting in an interference marking matrix. The interference aggregation index is discretized and mapped back to band-time coordinates, achieving precise location from a single index to its spatiotemporal position, avoiding direct global thresholding of the original amplitude. Based on the interference marking matrix, interference items are removed, and unmarked bands are retained, resulting in an occlusion band matrix. This suppresses the wake effect of interference in the time-frequency neighborhood, improving the purity and stability of subsequent reconstruction and statistics. Temporal reconstruction is performed based on the occlusion band matrix, rearranging each band in chronological order and filling missing values ​​with interpolation values ​​to obtain an occlusion feature matrix. This restores the continuous trajectory of occlusion events, ensuring that subsequent statistics are built on complete temporal data. Based on the occlusion feature matrix, the concentration and diffusion of each band are extracted, and the occlusion separation ratio is calculated, achieving stability and traceability of the threshold determination. This facilitates maintaining consistent judgment logic under different operating conditions and reduces the maintenance cost of threshold drift.

[0149] Specifically, interference classification data is obtained by dividing the data into intervals based on the interference aggregation index. Then, the bands in the interference classification are marked as interference terms on the band mapping matrix, resulting in an interference labeling matrix, which includes:

[0150] The interference aggregation index sequence is obtained on the time axis. In implementation, the global extreme values ​​and distribution statistics on the time window are used as the benchmark to set hierarchical threshold groups. K-level quantization is then performed to discretize the interference aggregated exponential sequence into level codes. Interference classification data was obtained; subsequently, the band mapping matrix was read. (A two-dimensional time-frequency array with uniform amplitude scale and band index as coordinates), for each time slice If it is determined to be a high level of interference (e.g.) ), then for The corresponding cross-sectional position is marked: first, the robust center of the cross-section in the band dimension is calculated. With dispersion (For example, expressed as median and MAD), then for each band Find the local outlier degree ,when and Time setting Otherwise, set to 0, thus in Obtaining the interference mark matrix from the plane To suppress scattered mislabeling caused by isolated noise, the time dimension can be adjusted. Perform a three-point majority vote to smooth out the sequence and enhance the continuity of the timeline.

[0151] Specifically, temporal reconstruction is performed based on the occlusion band matrix, rearranging each band in chronological order and filling missing values ​​with interpolated values ​​to obtain the occlusion feature matrix, which includes:

[0152] Firstly, according to right Perform element-wise masking, when The element is marked as missing, and the remaining elements are retained, resulting in an occlusion band matrix containing only the unmarked positions. Subsequently, a unified sampling grid was established on the timeline. And for each band The observations were aligned, rearranged in ascending order by timestamp, and deduplicated to ensure that sample columns for all bands were aligned to the same order. Complete the chronological rearrangement; during the interpolation phase, for each band... In its time series Perform two-stage reconstruction on the missing intervals: when the missing span does not exceed the longest missing span of the linear interpolation. Linear interpolation is used to maintain local monotonicity; when the missing span exceeds Piecewise spline interpolation is used and adjacent bands are introduced. In the same The directional consistency regularization term constrains the gradient sign after interpolation to not conflict with the gradient sign of neighboring bands at that time, thus preserving the common occlusion physical characteristics of multiple bands. After interpolation, a median time filter of length 11 is performed on each band to remove residual spikes, and the output is uniformly an occlusion feature matrix. .

[0153] In a preferred embodiment of the present invention, the concentration and diffusion of each band are extracted based on the occlusion feature matrix, and the occlusion separation ratio is calculated, including:

[0154] Based on the occlusion feature matrix, the quantile benchmark is determined, and the amplitudes of each band are arranged from smallest to largest to obtain the quantile threshold data.

[0155] Based on the quantile threshold data and the occlusion feature matrix, the top segment index set and the remaining segment index set are determined according to the quantile threshold of each band, and the threshold segment index data is obtained.

[0156] Based on the occlusion feature matrix, the ratio of the maximum length of the non-zero continuous segment to the length of the time window is calculated to obtain the discontinuity factor.

[0157] Based on the occlusion feature matrix and threshold index data, the top segment proportion and coverage ratio are calculated, and then weighted and summed with the discontinuity factor to obtain the band score data.

[0158] Based on the band score data, the ratio of each band score to its sum is used as the band weight;

[0159] Based on the threshold index data and the occlusion feature matrix, the amplitude of each band on the top index set is weighted and summed according to the band weight to obtain the concentrated quantity.

[0160] Based on the threshold index data and the occlusion feature matrix, the diffusion is summarized, and the amplitude and discontinuity factor of each band on the residual index set are weighted and summed to obtain the diffusion amount.

[0161] The occlusion separation ratio is obtained by performing a coupling operation based on the concentration and diffusion quantities.

[0162] In this embodiment of the invention, a quantile benchmark is established based on the occlusion feature matrix. The amplitudes of each band are arranged from smallest to largest to obtain quantile threshold data, eliminating the incomparability caused by differences in amplitude dimensions and dynamic range between different bands. Based on the quantile threshold data and the occlusion feature matrix, the top segment index set and the remaining segment index set are determined according to the quantile threshold of each band to obtain threshold segment index data. The time axis of each band is divided into a high-confidence evidence region and a background / weak evidence region, which facilitates the subsequent calculation of concentration and diffusion. Based on the occlusion feature matrix, the ratio of the maximum length of the non-zero continuous segment to the time window length of each band is calculated to obtain the discontinuity factor, distinguishing between short, discrete segments easily caused by flicker interference and continuous, coherent segments that conform to the entry and exit patterns of occluded objects. Based on the occlusion feature matrix and the threshold segment index data, the top segment proportion and coverage ratio are calculated and weighted and summed with the discontinuity factor to obtain band score data, making the single band in... The strength of evidence across different dimensions is balanced. Based on band score data, the ratio of each band score to its sum is used as the band weight, transforming the amount of evidence in a single band into a share for global aggregation, forming an environment-adaptive channel weighting strategy. Centralized aggregation is performed based on threshold index data and occlusion feature matrix, with the amplitude of each band on the top-segment index set weighted and summed according to band weight to obtain a centralized quantity. This aggregates segments identified as high-evidence across bands, characterizing the strength of primary evidence during the occlusion entry-continuity-exit stages. Diffusion aggregation is performed based on threshold index data and occlusion feature matrix, with the amplitude of each band on the remaining segment index set weighted and summed with the discontinuity factor to obtain a diffusion quantity, characterizing the energy distribution of non-top segments and discontinuous segments, equivalent to quantifying interfering evidence. Coupled calculations are performed on the centralized quantity and diffusion quantity to obtain the occlusion separation ratio, achieving a lossy but discriminative compression of complex multidimensional statistics.

[0163] Specifically, based on the occlusion feature matrix, a quantile benchmark is established, and the amplitudes of each band are arranged from smallest to largest to obtain quantile threshold data, which includes:

[0164] First, let the input be the occlusion feature matrix. ,in For the number of bands, For the time frame number, the first The time series of each band is denoted as The quantile benchmark is established using a fixed quantile set. Combined with window-based statistics: First, determine the observation window for statistics on the time axis [1, (When implementing online, a sliding window length is used) With step size Fixed configuration, here the entire [1, For example, for each band, first perform linear interpolation or neighborhood median interpolation on the missing values ​​to eliminate holes, then perform saturation clipping on the amplitude (amplitude limiting is performed according to the interval between the 1% and 99th percentiles of that band), and then... Sort by ascending order to obtain an ordered sequence For each quantile proportion ,calculate and take As the first Each band is at the quantile The amplitude threshold at the location; Including the total number of samples in this band The amplitude limiting range and interpolation marker are written together into the quantile threshold data structure to form a band index with the key value as the band index. mapping To ensure consistency between different batches or devices, quantile sets are used. Set as a fixed constant during deployment. Threshold calculation is implemented using sequential statistics or selection algorithms, such as linear time selection algorithm, and a version number and timestamp are attached when writing.

[0165] Specifically, based on the quantile threshold data and the occlusion feature matrix, the top segment index set and the remaining segment index set are determined according to the quantile threshold of each band, resulting in the threshold segment index data, which specifically includes:

[0166] Select =0.90 as the top segment threshold level (and (consistent with the 0.90 quantile in each band) With each time frame Execution judgment Time index that meets the conditions Record in the top segment index set The remaining time indexes are recorded in the remaining segment index set. To improve temporal coherence and suppress single-frame pulses, a morphological opening and closing operation is performed on the time axis using the binary discrimination results: first, the binary discrimination results are processed... The indicator sequence is of length ,like For a single frame of dilation and re-erosion processing, then... Perform dual processing to ensure that the two sets are complementary and cover [1, ]; then and Convert from discrete index table to start-end pair list format This is to facilitate compressed storage and subsequent scanning. After completing the above process for all bands, threshold band index data is constructed. And include a source of thresholds for reproduction. Window range [1, ] and morphological parameters .

[0167] The formula for calculating the occlusion separation ratio is as follows:

[0168] ;

[0169] in, To achieve the occlusion separation ratio, For the index of the band in the occlusion feature matrix, The total number of bands in the occlusion feature matrix. For time indexing, For band In time amplitude, For the top segment index set, For the remaining segment index set, For band The weight, , For band The top segment percentage, , For band coverage ratio , For band discontinuity factor, , For band The maximum length of a non-zero continuous segment. The length of the time window, For the band index, bands The top segment percentage, coverage ratio, and discontinuity factor, is a coefficient.

[0170] Among them, coefficient Used to maintain the numerator term The linear aggregation and dimensional consistency make The response to the top energy level is linearly proportional to the energy itself, which facilitates threshold calibration and cross-device migration; if an introduction is made at this point... Amplification by powers other than 1 will change the relative sensitivity to strong / weak top segments and disrupt the relationship with weights. The linear superposition interpretation. When the value is 1, the statistical significance of the numerator part is consistent with that of the top segment index set and the band weight.

[0171] coefficient Used to ensure the denominator term The measurements of diffusion energy and discontinuity penalty remain linear, thus maintaining consistency with the linear summation in the diffusion summation step and avoiding over-amplification of small diffusion amplitudes. Premature saturation; at the same time The value can be 1 or 2 A value of 1 ensures that when the overall scene amplitude is scaled by a constant, The relative order remains unchanged, which is beneficial for the stability of the threshold during long-term operation.

[0172] coefficient This involves a gentle, non-linear compression of the overall ratio to improve separability under small sample jitter: The numerator / denominator ratio after the action is applied at the output end. The power mapping ∈(0,1) can compress heavy-tailed distributions and suppress a few extremely strong signals. Excessive tension can lead to sensitivity; this setting... The energy of the top-segment polymerization continues to increase monotonically, but the growth curve is smoother, which is beneficial for obtaining a stable trigger bandwidth at a fixed threshold. The specific value is 0.65.

[0173] coefficient The value is 0.35, and Form a normalization constraint that sums to 1. This involves conserving the nonlinear quota of the output layer between enhancing true occlusion evidence and suppressing diffusion / discontinuous evidence: when Applying a sublinear mapping to the growth on the denominator side can avoid misinterpreting the data when a small amount of diffusion energy is present. Excessive suppression; and add Setting it to 1 makes it easier to keep the dynamic range of the output consistent under different operating conditions, reduces the frequency of threshold retuning, and lowers the parameter tuning cost from an engineering and maintenance perspective.

[0174] coefficient The value is 0.25, and it applies to the discontinuous penalty for leaks. The evidence at the top lacks temporal continuity ( When the value is too small, the corresponding top segment energy is proportionally fed back to the denominator as a penalty: when The penalty is 0 when the value is 1 (the top segment is completely continuous); when... When the value approaches 0 (highly discontinuous), up to 25% of the top segment energy is included in the denominator, thus significantly suppressing strong but disjointed interference segments. The rise; A value of 0.25 ensures that the penalty is effective but does not overwhelm the molecular contribution of the actual occlusion.

[0175] Embodiments of the present invention also provide a false alarm suppression system for security gratings based on multispectral fusion, the system comprising:

[0176] The filtering module is used to acquire a multi-band detection dataset and perform band filtering based on it, removing detection data of bands with signal amplitudes lower than the preset signal amplitude to obtain a valid band dataset.

[0177] The mapping module is used to perform band mapping based on the effective band dataset, projecting signals from different bands onto a unified amplitude scale range, and using the band index as the mapping coordinates to obtain the band mapping matrix;

[0178] The correlation module is used to perform cross-correlation analysis based on the band mapping matrix, compare the direction and magnitude of amplitude changes of different bands in adjacent time slices, and generate a band correlation map.

[0179] The interference module is used to perform pattern aggregation based on the band correlation map, combine bands with the same amplitude change direction and similar change magnitude into aggregation units, and calculate the dynamic change amplitude of each aggregation unit to obtain interference response data.

[0180] The aggregation module is used to perform multi-unit fusion based on interference response data, identify the degree of concentration of interference signals in each aggregation unit, and obtain the interference aggregation index.

[0181] The separation module is used to separate the occlusion signal according to the interference aggregation index, remove the interference signal, generate the occlusion feature matrix, and calculate the degree of separation of the occlusion signal based on it to obtain the occlusion separation ratio.

[0182] The judgment module is used to make judgments based on the occlusion separation ratio, generate judgment output results, and control the operating status of mechanical equipment.

[0183] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0184] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0185] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0186] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A false alarm suppression method for security grating based on multispectral fusion, characterized in that, The method comprises: Obtaining a multi-band detection data set, and performing band screening on the multi-band detection data set to remove band detection data with a signal amplitude lower than a preset signal amplitude, and obtaining an effective band data set; Performing band mapping on the effective band data set, projecting different band signals to a unified amplitude scale interval, and taking a band index as a mapping coordinate to obtain a band mapping matrix; Performing cross-correlation analysis on the band mapping matrix, comparing the amplitude change direction and amplitude change of different bands on adjacent time slices, and generating a band correlation graph; According to the band correlation graph, the bands with the same amplitude change direction and similar change amplitude are combined into an aggregated unit, and the dynamic change amplitude of each aggregated unit is calculated to obtain interference response data; According to the interference response data, multi-unit fusion is performed to identify the concentration degree of the interference signal in each aggregated unit to obtain an interference aggregation index; According to the interference aggregation index, the interference signal is removed to generate a shielding feature matrix, and the separation degree of the shielding signal is calculated to obtain a shielding separation ratio; According to the shielding separation ratio, a determination output result is generated and the running state of the mechanical equipment is controlled; According to the interference aggregation index, the interference signal is removed to generate a shielding feature matrix, and the separation degree of the shielding signal is calculated to obtain a shielding separation ratio, comprising: According to the interference aggregation index, interval classification is performed to obtain interference classification data, and position marking is performed on the band mapping matrix according to the interference classification data, the bands of the interference classification are marked as interference items, and an interference marking matrix is obtained; According to the interference marking matrix, the interference items are removed and the bands not marked are retained to obtain a shielding band matrix; According to the shielding band matrix, time sequence reconstruction is performed, each band is rearranged in time sequence, and the missing place is filled with an interpolation value to obtain a shielding feature matrix; According to the shielding feature matrix, the concentration and diffusion of each band are extracted, and the shielding separation ratio is calculated.

2. The multispectral fusion based security raster false alarm suppression method according to claim 1, wherein, According to the effective band data set, band mapping is performed, different band signals are projected to a unified amplitude scale interval, and a band index is taken as a mapping coordinate to obtain a band mapping matrix, comprising: According to the effective band data set, amplitude extraction is performed, and the minimum value and maximum value of each band in the time window are determined to obtain an amplitude range vector; According to the amplitude range vector, amplitude proportion calculation is performed, and the amplitude of each band signal is proportionally converted with the amplitude range to obtain a proportion conversion matrix; According to the proportion conversion matrix, amplitude normalization is performed, and the proportion value is mapped to a unified amplitude scale interval to obtain an amplitude uniform matrix; According to the amplitude uniform matrix, coordinate assignment is performed with the band index as the mapping coordinate to obtain the band mapping matrix.

3. The multispectral fusion based security raster false alarm suppression method according to claim 2, wherein, According to the band mapping matrix, cross-correlation analysis is performed, the amplitude change direction and amplitude change of different bands on adjacent time slices are compared, and a band correlation graph is generated, comprising: According to the band mapping matrix, the amplitude difference of adjacent time slices is calculated to obtain an amplitude difference matrix; According to the amplitude difference matrix, the change direction is determined, the amplitude increase is marked as positive, the amplitude decrease is marked as negative, and the amplitude unchanged is marked as zero, to obtain a direction identification matrix; According to the wave band mapping matrix, the amplitude ratio of adjacent time slices is calculated, the amplitude of the latter time slice is compared with the amplitude of the former time slice, and an amplitude ratio matrix is obtained; According to the direction identification matrix and the amplitude ratio matrix, the direction and amplitude are matched, the corresponding relationship between the change direction and the change amplitude of each wave band pair on adjacent time slices is recorded, and an association feature matrix is obtained.

4. The multispectral fusion based security raster false alarm suppression method according to claim 3, wherein, According to the wave band association graph, the wave bands with consistent change direction and similar change amplitude are combined into an aggregation unit, and the dynamic change amplitude of each aggregation unit is calculated to obtain interference response data, including: According to the wave band association graph, the wave bands with the same change direction are combined into a direction set to obtain direction set data; According to the direction set data, the change amplitude of each wave band in the time window is calculated to obtain change amplitude data; According to the change amplitude data, the direction sets with a difference in change amplitude lower than a preset difference threshold are merged into an aggregation unit to obtain aggregation unit data; According to the aggregation unit data, the amplitude accumulation, fluctuation frequency and change amplitude standard deviation of each aggregation unit in the time window are calculated to obtain interference response data.

5. The multispectral fusion based security raster false alarm suppression method according to claim 4, wherein, According to the interference response data, the concentration degree of the interference signal in each aggregation unit is identified to obtain an interference aggregation index, including: According to the interference response data, the amplitude difference between adjacent times of each aggregation unit in the time window is calculated to obtain a difference sequence; According to the difference sequence, the features of each aggregation unit are extracted, and the amplitude accumulation, direction switching number, maximum single increment and average deviation are recorded to obtain unit feature data; According to the unit feature data, the distribution of each aggregation unit is calculated, and the distribution of all aggregation units is normalized to determine the weight of each aggregation unit to obtain unit weight data; According to the unit feature data and the unit weight data, the aggregation degree of each aggregation unit is calculated to obtain unit aggregation degree data; According to the unit weight data, the concentration is identified, the weight of all aggregation units is squared and summed to obtain a unit concentration coefficient; According to the unit weight number, unit aggregation degree data and unit concentration coefficient, the index coupling is performed to obtain the interference aggregation index.

6. The multispectral fusion based security raster false alarm suppression method according to claim 5, wherein, According to the shielding feature matrix, the concentration and diffusion of each wave band are extracted, and the shielding separation ratio is calculated, including: According to the shielding feature matrix, the amplitude of each wave band is arranged from small to large to obtain quantile threshold data; According to the quantile threshold data and the shielding feature matrix, the top segment index set and the remaining segment index set are determined according to the quantile threshold of each wave band to obtain threshold segment index data; According to the shielding feature matrix, the ratio of the maximum length of the non-zero continuous segment of each wave band to the length of the time window is calculated to obtain the intermittence factor; According to the shielding feature matrix and the threshold segment index data, the top segment proportion and the coverage proportion are calculated, and the weighted sum of the two and the intermittence factor is calculated to obtain wave band score data; According to the wave band score data, the ratio of the score of each wave band to the sum of the scores of all wave bands is taken as the wave band weight; The concentration is obtained by weighting and summing the amplitude of each band on the top index set according to the threshold band index data and the shielding feature matrix, and the band weight is used for weighting; The diffusion is obtained by weighting and summing the amplitude and the discontinuity factor of each band on the residual index set according to the threshold band index data and the shielding feature matrix; The shielding separation ratio is obtained by coupling operation according to the concentration and the diffusion.

7. A false alarm suppression system for security gratings based on multispectral fusion, characterized in that, The system is used for executing the method in any one of claims 1 to 6, and the system comprises: The screening module is used for acquiring the multi-band detection data set, and screening the bands according to the multi-band detection data set, and removing the band detection data with the signal amplitude lower than the preset signal amplitude to obtain the effective band data set; The mapping module is used for performing band mapping according to the effective band data set, projecting different band signals to a unified amplitude scale interval, and taking the band index as the mapping coordinate to obtain the band mapping matrix; The correlation module is used for performing cross-correlation analysis according to the band mapping matrix, comparing the amplitude change direction and the change amplitude of different bands on adjacent time slices, and generating a band correlation graph; The interference module is used for performing mode aggregation according to the band correlation graph, combining the bands with the same amplitude change direction and the similar change amplitude into an aggregation unit, and calculating the dynamic change amplitude of each aggregation unit to obtain interference response data; The aggregation module is used for performing multi-unit fusion according to the interference response data, identifying the concentration degree of the interference signal in each aggregation unit to obtain an interference aggregation index; The separation module is used for separating the shielding signal according to the interference aggregation index, removing the interference signal, generating a shielding feature matrix, and calculating the separation degree of the shielding signal according to the shielding feature matrix to obtain a shielding separation ratio; The determination module is used for determining according to the shielding separation ratio, generating a determination output result, and controlling the running state of the mechanical equipment.

8. A computing device, comprising: Comprise: One or more processors; Storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program, which is executed by the processor to implement the method in any one of claims 1 to 6.

Images