Anti-theft door abnormal intrusion behavior recognition method and system based on multi-modal perception

Abstract

The application relates to the technical field of multi-modal perception, and particularly provides a burglary-resistant door abnormal intrusion behavior recognition method and system based on multi-modal perception. The method comprises the following steps: collecting tensile length data groups of strain fiber bundles embedded in different areas of a door body, and outputting a deformation gradient matrix; extracting a local abnormal area coordinate set; performing spatial superposition operation on the obtained local abnormal area coordinate set and breakpoint distribution of a conductive paint grid pre-printed on an inner layer of a door panel; outputting a contact pressure distribution atlas of an intrusion action; performing time domain alignment operation on the obtained contact pressure distribution atlas and airflow disturbance sequences collected by micro piezoelectric films arranged at gaps between a door leaf and a door frame; performing feature fusion calculation on the aligned composite feature sequences to obtain a composite feature tensor; and performing classification mapping calculation on the composite feature tensor to output a classification label of abnormal intrusion behavior. The application realizes high-precision recognition of abnormal intrusion behavior of a burglary-resistant door.

Description

Technical Field

[0001] This invention relates to the field of multimodal sensing technology, and in particular to a method and system for identifying abnormal intrusion behavior of security doors based on multimodal sensing. Background Technology

[0002] As the first physical barrier for residential security, the ability of security doors to resist illegal intrusion is crucial. Traditional security doors mainly rely on their own mechanical structural strength and the protection level of their locks for passive defense. However, with the diversification of intrusion methods, simple physical protection is no longer sufficient to meet the ever-increasing security demands. Therefore, the development of intelligent security door technology that can actively sense and identify abnormal intrusion behavior has become a research hotspot in this field. Existing intelligent security door intrusion detection technologies typically rely on a single type of sensor, such as vibration sensors, magnetic induction sensors, or simple pressure switches; the sensors determine whether illegal intrusion has occurred by detecting whether the door body vibrates beyond a threshold, whether the door magnet separates, or whether there is excessive pressure in a localized area.

[0003] Existing technologies mainly rely on single physical quantities, such as vibration, circuit continuity, and airflow, for judgment. They lack the ability to comprehensively perceive and correlate multi-dimensional physical information generated by intrusion behavior, including mechanical deformation fields, electrical fracture modes, and airflow dynamics characteristics. This results in low recognition accuracy, susceptibility to interference, and inability to identify behavior types. Therefore, to solve the above technical problems, this invention provides a method and system for identifying abnormal intrusion behavior in security doors based on multimodal perception. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In one aspect, the present invention provides a method for identifying abnormal intrusion behavior of security doors based on multimodal perception, comprising the following steps:

[0006] The tensile length data of strain fiber bundles embedded in different areas of the door are collected. The difference comparison operation is performed on the tensile length data to output the deformation gradient matrix. Then, based on the position of the elements in the deformation gradient matrix that exceed the preset threshold, the coordinate set of local abnormal areas is extracted.

[0007] The obtained set of coordinates of local abnormal areas is spatially superimposed with the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel; the fracture mode vector of the conductive paint mesh is generated from the superimposition result, and then pressure inversion calculation is performed on the fracture mode vector to output the contact pressure distribution map of the intrusion action.

[0008] The obtained contact pressure distribution map is aligned in the time domain with the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame. The aligned composite feature sequence is then subjected to feature fusion calculation to obtain a composite feature tensor. The composite feature tensor is then subjected to classification mapping calculation to output the classification label of abnormal intrusion behavior.

[0009] Another aspect of the present invention provides a multimodal perception-based system for identifying abnormal intrusion behavior in security doors, comprising:

[0010] The abnormal area confirmation module is used to collect tensile length data sets of strain fiber bundles embedded in different areas of the door body, perform differential comparison operations on the tensile length data sets, and output the deformation gradient matrix; then, based on the position of the elements in the deformation gradient matrix that exceed the preset threshold, the coordinate set of the local abnormal area is extracted.

[0011] The overlay operation module is used to perform spatial overlay operation on the obtained set of coordinates of local abnormal areas and the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel; the overlay operation result generates the fracture mode vector of the conductive paint mesh, and then performs pressure inversion calculation on the fracture mode vector to output the contact pressure distribution map of the intrusion action.

[0012] The temporal alignment module is used to perform temporal alignment operations on the obtained contact pressure distribution map and the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame. The aligned composite feature sequence is then used to perform feature fusion calculation to obtain a composite feature tensor. Finally, the composite feature tensor is used to perform classification mapping calculation to output the classification label of abnormal intrusion behavior.

[0013] This invention achieves high-precision identification of abnormal intrusion behavior in security doors through multimodal sensing collaboration. Utilizing differential comparison operations of strain fiber bundles, the global deformation of the door body is transformed into a gradient matrix. Local abnormal coordinate sets are extracted through threshold filtering, enabling preliminary localization and spatial quantification of the intrusion contact area. By introducing spatial superposition operations of conductive paint mesh breakpoint distribution and abnormal coordinate sets, a fracture mode vector is generated and then subjected to pressure inversion calculations. This maps discrete coordinates into a continuous pressure distribution map, revealing the mechanical characteristics and contact patterns of the intrusion action. Through temporal alignment and feature fusion of the pressure distribution map and airflow disturbance sequence, a composite feature tensor is constructed. Finally, behavior labels are output through classification mapping, achieving spatiotemporal correlation and joint analysis of multi-source signals from mechanics, electricity, and airflow dynamics. The overall technical solution significantly improves the spatial resolution, mechanical inversion accuracy, and dynamic behavior discrimination robustness of intrusion behavior identification through progressive coupling analysis of multi-level sensor data, while also enhancing the system's ability to generalize and identify complex intrusion techniques. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0015] Figure 1 This is a flowchart of the method for identifying abnormal intrusion behavior of a security door based on multimodal perception provided in Embodiment 1 of the present invention;

[0016] Figure 2 This is a schematic diagram of the method for identifying abnormal intrusion behavior of a security door based on multimodal perception provided in Embodiment 1 of the present invention.

[0017] Figure 3 This is a process diagram of extracting the coordinate set of local abnormal regions provided in Embodiment 2 of the present invention;

[0018] Figure 4 This is a process diagram of the contact pressure distribution spectrum of the output intrusion action provided in Embodiment 4 of the present invention;

[0019] Figure 5 This is a diagram illustrating the process of obtaining the composite feature tensor provided in Embodiment 7 of the present invention;

[0020] Figure 6 This is a block diagram of the anti-theft door abnormal intrusion behavior recognition system based on multimodal perception provided in Embodiment 13 of the present invention;

[0021] Figure 7 A block diagram of the electronic device provided by the present invention;

[0022] Figure 8 A block diagram of a computer-readable storage medium provided for this invention.

[0023] Reference numerals: 1. Abnormal region confirmation module; 2. Overlay operation module; 3. Time domain alignment module; 4. Central processing unit / microprocessor / main control chip; 5. Storage medium; 6. Data bus; 7. Input / output bus / external bus / device bus; 8. Display; 9. Input / output device; 10. Computer-readable instructions; 11. Non-transitory computer-readable storage medium. Detailed Implementation

[0024] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0026] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, "connection" can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0027] In this embodiment of the invention, directional terms such as "up," "down," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0028] This invention constructs an identification framework that can jointly analyze intrusion behavior from three dimensions—space, mechanics, and time—by integrating strain deformation field, conductive mesh fracture mode, and airflow disturbance time series data, thereby overcoming the inherent defects of single sensors in terms of positioning accuracy and mechanical information inversion.

[0029] Example 1: As Figure 1 As shown in the figure, this embodiment of the invention provides a method for identifying abnormal intrusion behavior of a security door based on multimodal perception, comprising the following steps:

[0030] Step S100: Collect tensile length data sets of strain fiber bundles embedded in different regions of the door body, perform differential comparison operation on the tensile length data sets, and output the deformation gradient matrix; then, based on the position of the elements in the deformation gradient matrix that exceed the preset threshold, extract the coordinate set of local abnormal areas.

[0031] Assume that the different regions of the door are divided according to a two-dimensional planar grid, with the number of rows being... The number of columns is Each grid point ( This corresponds to the installation position of a strain fiber bundle, where , The tensile length measured at this point is denoted as L( (The unit is meters.) The center-to-center distance between adjacent areas in the horizontal direction is defined as... The vertical distance between the centers of adjacent regions is All units are meters.

[0032] For each internal grid point ( ),Right now , Calculate its horizontal difference Difference with vertical direction :

[0033]

[0034] ;

[0035] Then calculate the deformation gradient magnitude of the grid points. :

[0036] ;

[0037] For boundary grid points ( or or or ), using one-sided difference:

[0038]

[0039]

[0040]

[0041] , ;

[0042] Boundary points are also calculated using the above formula for the sum of squares and the square root. .all Arranging the original mesh forms the deformation gradient matrix, whose elements are dimensionless and represent the spatial rate of change of the tensile length on the door surface. The above formula is a synthesis of central difference and amplitude in the discrete domain. Strain fiber bundles are deployed in different regions of the door, and their tensile length data sets... It reflects the static deformation distribution of the door body after being subjected to external force; through central difference calculation, the rate of change of tensile length in the two orthogonal directions, horizontal and vertical, is obtained. and Then synthesize the gradient magnitude This process transforms the original one-dimensional stretch length value into a two-dimensional spatial variation intensity. It does not rely on the absolute stretch length but highlights the differences between adjacent regions, thus eliminating the overall uniform deformation caused by temperature changes or the door's own gravity, retaining only the abrupt deformation caused by local intrusion actions. In the output deformation gradient matrix, larger element values ​​indicate more drastic stretch length changes near that location, meaning the door exhibits a significant deformation gradient in that region, corresponding to the concentrated action point of intrusion actions such as prying or impact. Based on the positions of elements exceeding a preset threshold, a set of coordinates for local anomaly regions is extracted. This matrix serves a dual purpose: spatial high-pass filtering and anomaly localization. On the one hand, it suppresses global uniform deformation; on the other hand, it discretizes the continuous deformation field into quantifiable anomaly candidate regions, providing precise spatial constraints for spatial superposition with the conductive paint mesh breakpoint distribution.

[0043] Step S200: Perform a spatial overlay operation on the obtained set of coordinates of the local abnormal area and the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel; generate the fracture mode vector of the conductive paint mesh from the overlay operation result, and then perform pressure inversion calculation on the fracture mode vector to output the contact pressure distribution map of the intrusion action.

[0044] Step S300: Perform time-domain alignment operation between the obtained contact pressure distribution map and the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame. Perform feature fusion calculation on the aligned composite feature sequence to obtain the composite feature tensor. Then perform classification mapping calculation on the composite feature tensor to output the classification label of abnormal intrusion behavior.

[0045] The different areas of the door refer to multiple sub-regions of the security door panel divided geometrically, such as the upper area, middle area, lower area, hinge side area, and lock side area. These sub-regions are used to distinguish the installation positions of the strain fiber bundles, allowing deformation data from different areas to be collected independently, thereby pinpointing the extent of intrusion on the door. A strain fiber bundle is a slender sensing unit embedded within the door material; its tensile length changes linearly with the deformation of the door surface. When the security door is subjected to external pressure, prying, or impact, the fiber bundle is stretched or compressed, converting mechanical deformation into a measurable change in length, serving as the basic signal source for sensing intrusion actions. The tensile length data set consists of the current tensile length values ​​output by all strain fiber bundles deployed in different areas of the door; each value corresponds to the deformation of a specific door area, and multiple values ​​together form a data set reflecting the overall deformation state of the door. The row and column indices corresponding to each element in the element position deformation gradient matrix have a pre-defined mapping relationship with the actual area coordinates on the door. This is used to locate specific door positions where the deformation value exceeds a preset threshold, thereby extracting the coordinates of local abnormal areas. The conductive paint mesh is a mesh of conductive material formed by a printing process on the inner layer of the door panel. When the mesh is intact, it is conductive. When the door is subjected to external pressure, impact, or prying, causing local deformation, the corresponding position of the mesh breaks, interrupting the electrical connection. The breakpoint position reflects the contact position of the intrusion action. The breakpoint distribution execution space refers to the operation object that aligns and superimposes the breakpoint coordinates of the conductive paint mesh with the obtained local abnormal area coordinates in the same two-dimensional spatial reference system. The spatial coordinate domain corresponds to the actual geometric dimensions of the door, ensuring that the two types of sensing information can be fused within the same spatial frame. A micro-piezoelectric film is installed in the gap between the door leaf and the door frame as a thin-film sensor. This film can respond to airflow fluctuations and mechanical vibrations within the gap, generating an electrical signal proportional to the airflow velocity and disturbance amplitude. This is used to sense airflow disturbances caused by changes in the door gap due to illegal intrusion. The airflow disturbance sequence is a continuous voltage signal sequence output by a miniature piezoelectric film over time, recording the time-varying pattern of airflow fluctuations at the gap, such as periodic or sudden disturbances caused by prying, pushing, or cutting the door. This serves as an important temporal characteristic for determining the type of intrusion behavior. The classification label for abnormal intrusion behavior is the final output category identifier used to distinguish different types of abnormal contact behavior; common labels include forced prying, static pressure pushing, abnormal airflow disturbance, and localized impact; this label is used for alarm, recording, or coordinated response.

[0046] In the above embodiments, the principle is referenced in the appendix. Figure 2This embodiment achieves high-precision identification of abnormal intrusion behavior of security doors through multimodal sensing collaboration. Utilizing differential comparison operations of strain fiber bundles, the global deformation of the door body is transformed into a gradient matrix. Local abnormal coordinate sets are extracted through threshold filtering, enabling preliminary localization and spatial quantification of the intrusion contact area. Spatial superposition operations of conductive paint mesh breakpoint distribution and abnormal coordinate sets are introduced to generate fracture mode vectors. Pressure inversion calculations then map discrete coordinates into continuous pressure distribution maps, revealing the mechanical characteristics and contact patterns of the intrusion actions. Through temporal alignment and feature fusion of the pressure distribution map and airflow disturbance sequences, a composite feature tensor is constructed. Finally, behavior labels are output through classification mapping, achieving spatiotemporal correlation and joint analysis of multi-source signals from mechanics, electricity, and airflow dynamics. The overall technical solution significantly improves the spatial resolution, mechanical inversion accuracy, and dynamic behavior discrimination robustness of intrusion behavior identification through progressive coupling analysis of multi-level sensor data, while also enhancing the system's ability to generalize and identify complex intrusion techniques.

[0047] Example 2: Figure 3 As shown, based on Example 1, the process of extracting the coordinate set of local anomaly regions in step S100 provided in this embodiment of the invention specifically includes the following steps:

[0048] Step S101: Extract all elements whose values ​​exceed a preset threshold from the deformation gradient matrix, and use the row and column indices corresponding to each element as candidate positions; then perform spatial merging on the candidate positions according to the adjacency determination rule. The adjacency determination rule is defined as two candidate positions belonging to the same block when the difference between their row indices does not exceed 1 and the difference between their column indices does not exceed 1. After merging, output several separate abnormal blocks, each of which contains a set of row and column indices.

[0049] Step S102: For each abnormal block, calculate the minimum and maximum row values, and the minimum and maximum column values ​​for all row and column indices within the abnormal block. Use the minimum row value and minimum column value to form the coordinates of the top-left corner of the abnormal block; use the maximum row value and maximum column value to form the coordinates of the bottom-right corner of the block. The coordinates of the top-left corner and bottom-right corner together generate the location coordinate range of the abnormal block. At the same time, calculate the difference between the maximum and minimum row values ​​as the row span, and the difference between the maximum and minimum column values ​​as the column span. Output the location coordinate range and span parameters for each abnormal block.

[0050] Step S103: Compare the location coordinate range of each abnormal block with the pre-calibrated door area mapping table. The door area mapping table records the row and column index ranges corresponding to each actual area of ​​the door. At the same time, the product of the row span and the column span is used as the coverage area. Abnormal blocks with a coverage area smaller than the pre-calibrated minimum intrusion contact area are removed. The location coordinate ranges of the remaining abnormal blocks are output as a set of local abnormal area coordinates in the original arrangement order.

[0051] In the above embodiments, this embodiment achieves accurate identification, spatial clustering, geometric feature quantification, and target relevance filtering of abnormal regions in the deformation gradient matrix through multi-level screening and geometric parameter calculation, ultimately generating a set of local abnormal coordinates that meet the requirements of door intrusion detection. First, by threshold screening and spatial merging based on adjacency rules, discrete abnormal elements are aggregated into continuous abnormal blocks, avoiding misjudgments caused by noise or scattered points and enhancing the spatial coherence of abnormal regions. Second, by calculating the row and column extreme values ​​of each abnormal block, a location coordinate interval and span parameters are generated, transforming the abnormal region into a rectangular description with clear geometric boundaries, facilitating region comparison and area calculation. Finally, by comparing with the door region mapping table and screening with the minimum intrusion contact area, abnormal blocks unrelated to the door structure or with excessively small areas are excluded, ensuring that the output coordinate set only contains local abnormal regions related to the door and with actual intrusion significance, improving the relevance and reliability of the detection results.

[0052] Example 3: Based on Example 2, the process of calculating the minimum and maximum row values, and the minimum and maximum column values ​​of all row and column indices within the abnormal block in step S102 of this embodiment of the invention specifically includes the following steps:

[0053] Step S1021: Sort all row and column indices in each abnormal block in ascending order by row index value to obtain a row index sequence; then traverse the row index sequence, and whenever the row index value changes, record the minimum and maximum values ​​of all column indices under the row index value to generate a row compression pair group for the block; then perform overlapping merging on the column minimum and column maximum values ​​corresponding to adjacent row index values ​​in the row compression pair group, that is, if the column maximum value of the previous row is greater than or equal to the column minimum value of the current row minus 1, then merge the column intervals of the two rows, and output the merged continuous row segment list. Each continuous row segment contains the starting row index, the ending row index, and the global minimum and global maximum values ​​of all column indices in the continuous row segment;

[0054] Step S1022: Compare the minimum and maximum global column values ​​of each row segment in the output list of consecutive row segments with the global column intervals of adjacent row segments. If the global column intervals of two adjacent row segments overlap or the interval does not exceed the preset intrusion gap threshold, merge the two row segments into one extended row segment. At the same time, update the starting row index of the extended row segment to the smaller starting row index of the original two row segments, the ending row index to the larger ending row index, the minimum global column value to the smaller of the minimum global column values ​​of the two row segments, and the maximum global column value to the larger value. Repeat the merging operation until merging is no longer possible, and output the extended row segment set.

[0055] Step S1023: From the output set of extended row segments, extract the minimum starting row index of all extended row segments as the minimum row value and the maximum ending row index as the maximum row value; at the same time, extract the minimum of the global column minimum values ​​of all extended row segments as the minimum column value and the maximum of the global column maximum values ​​as the maximum column value; output the minimum row value, maximum row value, minimum column value, and maximum column value as the location coordinate interval parameters of the abnormal block, and use the difference between the maximum row value and the minimum row value as the row span and the difference between the maximum column value and the minimum column value as the column span.

[0056] In the above embodiments, this embodiment transforms the discrete set of abnormal locations into a description of one or more continuous rectangular regions by merging continuous or adjacent parts in the row and column indices layer by layer. While preserving the main shape of the abnormal region, it significantly compresses the data size and outputs uniform and regular coordinate parameters, which is beneficial to improving the operability and computational efficiency of abnormal blocks in detection, labeling and subsequent processing.

[0057] Example 4: Figure 4 As shown, based on Example 1, the process of outputting the contact pressure distribution map of the intrusion action in step S200 of this embodiment of the invention specifically includes the following steps:

[0058] Step S201: Map each location coordinate interval in the local anomaly region coordinate set to the two-dimensional spatial coordinate system of the conductive paint mesh to obtain the set of mesh cells covered by the location coordinate interval; then traverse all breakpoint positions of the conductive paint mesh. If the breakpoint position falls inside any conductive paint mesh cell, mark the breakpoint as a spatially matched breakpoint; otherwise, mark it as an unmatched breakpoint; arrange all spatially matched breakpoints according to their original row and column order in the conductive paint mesh, output the spatially matched breakpoint sequence, and record the number of adjacent connected edges of each spatially matched breakpoint.

[0059] Step S202: For the spatial matching breakpoint sequence, according to the rule that the number of adjacent connected edges is greater than or equal to two and the spatial position is continuous, merge consecutive breakpoints with the number of adjacent edges greater than or equal to two into a break segment. The interval between breakpoints within the same break segment does not exceed one grid spacing. Calculate the length and orientation angle of each break segment. The unit of length is grid step size, and the orientation angle is taken as the angle of the line connecting the first and last ends of the segment, with the horizontal direction to the right as the reference. The length and orientation angle of all break segments constitute two sub-vectors of the break pattern vector. Each element of the length sub-vector corresponds to the length value of a break segment, and the corresponding position of the orientation sub-vector records the orientation angle of the segment.

[0060] Step S203: For each fracture segment in the fracture mode vector, extract the corresponding local pressure amplitude from the pre-stored pressure inversion mapping matrix according to its length and orientation angle. The pressure inversion mapping matrix is ​​established through offline calibration and records the pressure values ​​under different combinations of length and orientation angle. Distribute all extracted local pressure amplitudes bilinearly according to the spatial position of each fracture segment in the conductive paint grid, that is, assign weights to the four neighboring grid cells centered on the fracture segment, with the weights decreasing linearly with distance. Generate a continuous contact pressure distribution map covering the entire door area.

[0061] In the above embodiments, this embodiment realizes the visual reconstruction from abnormal signal location to physical pressure distribution; local abnormal coordinates are mapped to the conductive paint grid coordinate system, and a spatial matching breakpoint sequence is formed by matching and filtering breakpoint positions with grid cells and recording the number of connected edges, realizing the accurate mapping of abnormal signals in physical space while preserving the original grid topology; based on the number of adjacent connected edges and spatial continuity rules, discrete breakpoints are aggregated into fracture segments, and the length and direction angle are extracted to form fracture mode vectors, realizing the transformation from discrete abnormal points to continuous geometric features, quantifying the morphological features of intrusion actions into a computable vector expression, and providing a geometric parameter basis for pressure inversion; through a pre-calibrated pressure inversion mapping matrix, the fracture mode vector is converted into local pressure amplitude, and then a continuous pressure distribution map is generated through bilinear scattering, realizing the mapping from geometric features to physical pressure field, while ensuring the spatial continuity of pressure distribution through a weight allocation mechanism, ultimately forming a visual expression of contact pressure covering the entire door area.

[0062] In summary, this embodiment transforms discrete abnormal signals into a continuous pressure field through three-level processing of coordinate mapping, fracture mode extraction, and pressure inversion distribution, achieving a complete reconstruction of the contact pressure distribution of intrusion actions and providing quantitative basis for safety monitoring.

[0063] Example 5: Based on Example 4, the process of bilinearly distributing all extracted local pressure amplitudes according to the spatial position of each fracture segment in the conductive paint grid in step S203 of this embodiment of the invention specifically includes the following steps:

[0064] Step S2031: Extract the coordinate values ​​of all breakpoints on the fractured line segment, calculate the arithmetic mean of the coordinate values, and obtain the centroid coordinates of the fractured line segment; map the centroid coordinates to the unit coordinate system of the conductive paint mesh, and determine the reference mesh unit to which the centroid coordinates belong; with the reference mesh unit as the center, select its horizontal left neighbor, horizontal right neighbor, vertical top neighbor, and vertical bottom neighbor four adjacent units, together with the reference unit, to form a five-unit scatter window;

[0065] Step S2032: Calculate the spatial offset from the centroid coordinates of the five-element scatter window to the center point of each grid cell within the window. The offset includes horizontal and vertical offset components. Based on the absolute values ​​of the horizontal and vertical offset components, generate linear weight values ​​for the grid cells. The weight value of the reference cell is one, and the weight value of the four neighboring cells is one minus the ratio of the corresponding absolute value of the horizontal or vertical offset to the grid spacing. Multiply the local pressure amplitude of the extracted fracture segment by the linear weight value of each cell to obtain the pressure contribution value of the grid cell.

[0066] Step S2033: Establish a null value matrix with the same mesh dimension as the conductive paint. Each element of the null value matrix corresponds to a mesh cell. Accumulate the pressure contribution values ​​of all fracture segments into the null value matrix according to their respective mesh cell indices. If the same mesh cell receives multiple contribution values ​​from different fracture segments, sum them up. After accumulation, find the maximum value of all elements in the null value matrix. Using the maximum value as a reference, linearly scale the value of each element to between the pre-set lower and upper limits of the pressure value range. The scaled matrix is ​​a continuous contact pressure distribution map.

[0067] In the above embodiments, this embodiment not only solves the problem of the superposition of multi-source pressure contributions from different fracture segments to the same grid cell, but also maps the pressure data to a standardized range by linear scaling based on the maximum value. The resulting continuous contact pressure distribution map has clear spatial resolution and comparability, providing a quantitative basis for the contact mechanical state analysis of conductive paint grid structures.

[0068] Example 6: Based on Example 5, the process of establishing a null matrix with the same dimensions as the conductive paint mesh in step S2033 of this embodiment of the invention specifically includes the following steps:

[0069] Step S20331: Extract the row and column parameters of the conductive paint grid, and use the product of the row and column numbers as the total number of elements in the null matrix; allocate a contiguous storage space according to the total number, and write zero values ​​into each storage cell in the contiguous storage space to generate an initialized null matrix; each element of the null matrix corresponds to a grid cell;

[0070] Step S20332: For each element position in the generated null matrix, calculate the linear index value of the element in the null matrix, and then calculate the offsets of the four adjacent positions of the element's horizontal left neighbor, horizontal right neighbor, vertical top neighbor, and vertical bottom neighbor in the linear index; store the four offsets and the element's linear index value together as a quintuple; after traversing all element positions, arrange all quintuples in order of linear index value, and output the neighborhood index lookup table;

[0071] Step S20333: Bind the null value matrix to the neighborhood index lookup table so that the location of the element's own storage unit and its four neighboring storage units can be accessed simultaneously through any element's index value; after binding, output the null value matrix.

[0072] In the above embodiments, this embodiment allocates continuous storage space based on the number of rows and columns of the conductive paint grid and initializes it uniformly to zero values, ensuring that the null value matrix strictly corresponds to the original grid in terms of dimension; it provides a structurally consistent container for the accumulation of pressure contribution values, avoiding data mapping errors caused by dimension mismatch, while zero-value initialization ensures the mathematical rigor of the accumulation operation; it transforms the calculation of neighborhood positions from real-time calculation to pre-calculation, reducing the overhead of repeatedly calculating neighborhood indexes during the pressure contribution value dispersion process and improving the efficiency of data processing; the binding mechanism provides fast index support for the distribution and accumulation of pressure contribution values ​​among adjacent grid cells, ensuring the operability of spatial weight allocation during bilinear dispersion, and laying the data structure foundation for generating continuous contact pressure distribution maps.

[0073] Example 7: Figure 5 As shown, based on Example 1, the process of obtaining the composite feature tensor in step S300 of this embodiment of the invention specifically includes the following steps:

[0074] Step S301: Extract the acquisition time of the contact pressure distribution map as the time reference, align the time reference with the starting timestamp of the airflow disturbance sequence to obtain the time offset; perform a translation operation on the airflow disturbance sequence based on the time offset, so that each sampling point of the airflow disturbance sequence is in the same time coordinate system as the generation time of the contact pressure distribution map; after the translation operation is completed, repeat the contact pressure distribution map according to a fixed time window to form a spatial map sequence with the same length as the airflow disturbance sequence. The spatial map remains unchanged within each time window. Output the time-domain aligned spatial map sequence and the translated airflow disturbance sequence. The two together constitute the aligned composite feature sequence; each time point in the composite feature sequence contains both a spatial map and an airflow disturbance value.

[0075] Step S302: Expand all element values ​​of the spatial map row by row into a one-dimensional spatial feature vector, and then perform a Cartesian product operation on the one-dimensional spatial feature vector and the airflow disturbance value to generate a two-dimensional fusion matrix; the number of rows in the two-dimensional fusion matrix is ​​the dimension of the spatial feature vector, the number of columns is one, and each matrix element is the product of the corresponding component of the spatial feature vector and the airflow disturbance value; stack the two-dimensional fusion matrices generated at all time points in chronological order along the third dimension to obtain a three-dimensional composite feature tensor; the three dimensions of the three-dimensional composite feature tensor are the spatial feature dimension, the time point dimension, and the fusion value dimension, respectively.

[0076] Step S303: Perform classification mapping calculation on the three-dimensional composite feature tensor. According to the pre-generated behavior classification template library, which contains multiple template tensors, each template tensor has the same dimensional structure as the composite feature tensor. Calculate the element-wise product of the composite feature tensor and each template tensor in sequence and then sum them to obtain the matching score of the classification template. Take the classification label corresponding to the classification template with the highest matching score as the output. The classification label is the classification label of abnormal intrusion behavior.

[0077] In the above embodiments, the generation process of the composite feature tensor in this embodiment achieves effective integration and behavior recognition of the contact pressure distribution map and the airflow disturbance sequence through three steps: time alignment, feature fusion, and classification mapping. First, by extracting the acquisition time of the contact pressure distribution map as a time reference, the starting timestamp of the airflow disturbance sequence is aligned with it, the time offset is calculated, and a translation operation is performed to place the sampling points of the two data in the same time coordinate system. Subsequently, the contact pressure distribution map is repeatedly arranged according to a fixed time window to form a spatial map sequence with the same length as the airflow disturbance sequence. This ensures the synchronization of the two data sources in the time dimension and provides a unified time framework for feature fusion. Secondly, the elements of the spatial map are expanded row-wise into one-dimensional spatial feature vectors, and then Cartesian products are performed with airflow disturbance values ​​to generate a two-dimensional fusion matrix. Rows correspond to the spatial feature dimensions, and there is one column. Each element is the product of a spatial feature component and an airflow disturbance value. The two-dimensional fusion matrices of all time points are stacked along the third dimension to form a three-dimensional composite feature tensor, with its three dimensions corresponding to spatial features, time points, and fusion values, respectively. This achieves deep fusion of spatial distribution information and airflow disturbance information, unifying the two heterogeneous data into a structured tensor representation and enhancing the feature representation capability. Finally, using a pre-generated behavior classification template library, the three-dimensional composite feature tensor is summed after element-wise multiplication with each template tensor to calculate the matching score. The classification label corresponding to the template with the highest matching score is selected as the output to complete the classification and identification of abnormal intrusion behavior. Through the template matching mechanism, the composite feature tensor is compared with known behavior patterns, achieving effective classification of multi-source data fusion features.

[0078] In summary, this embodiment achieves time synchronization, feature fusion, and classification of multimodal data, improving the accuracy and robustness of abnormal intrusion behavior detection. By integrating contact pressure distribution and airflow disturbance information into a unified tensor representation and classifying based on template matching, it can more comprehensively capture the spatial and temporal characteristics of intrusion behavior, thereby enhancing the reliability of behavior recognition.

[0079] Example 8: Based on Example 7, the process of generating a two-dimensional fusion matrix in step S302 of this embodiment of the invention specifically includes the following steps:

[0080] Step S3021: Extract the pressure value of each grid cell in the spatial map. Take out all pressure values ​​in the order of increasing row index and increasing column index to form the original row vector. Calculate the sum of all pressure values ​​in the original row vector. Divide each pressure value in the original row vector by the sum to obtain the normalized row vector. Then extract the coordinate set of local anomaly regions. Use the grid cell index covered by each positioning coordinate interval in the coordinate set of local anomaly regions as a mask. Multiply the component at the corresponding index position in the normalized row vector by a weighting coefficient of two, and multiply the components at other positions by a weighting coefficient of one to obtain the weighted feature vector.

[0081] Step S3022: Perform a Cartesian product operation on the weighted feature vector and the airflow disturbance value at the current time point in the translated airflow disturbance sequence; specifically: create an empty matrix with the number of rows equal to the dimension of the weighted feature vector and the number of columns two. For the coordinate value of the i-th grid point of the weighted feature vector, fill the product of the component and the airflow disturbance value into the first column of the i-th row of the empty matrix, and fill the product of the square of the component and the square of the airflow disturbance value into the second column of the i-th row. After all rows are filled, output a two-dimensional fusion matrix.

[0082] Step S3023: For each row of the two-dimensional fusion matrix, extract the four neighboring grid cells of the corresponding grid cell in the original spatial map, obtain the corresponding components of the four neighboring grid cells in the weighted eigenvector, and calculate the arithmetic mean of the four components; then add the element of the first column of the row to the arithmetic mean and divide by two to obtain the updated element value of the first column; add the element of the second column of the row to the square of the arithmetic mean and divide by two to obtain the updated element value of the second column; after all rows are updated, output the final two-dimensional fusion matrix.

[0083] In the above embodiments, this embodiment extracts the pressure values ​​of grid cells in the spatial map and performs normalization processing. Combined with a masking weighting operation on the coordinate set of local anomaly regions, a weighted feature vector is formed. This strengthens the signal contribution of anomaly regions while maintaining the overall pressure distribution ratio, providing a discriminative feature representation for subsequent fusion. Secondly, the weighted feature vector is combined with the airflow disturbance value via a Cartesian product to construct a two-dimensional matrix containing first- and second-order interaction terms. This dynamically correlates spatial pressure features with time-dimensional airflow disturbances, forming a structured data foundation that simultaneously characterizes the spatial distribution intensity and the coupling relationship between disturbances. Finally, by introducing weighted feature components from four neighboring grid cells to locally smooth each row of the matrix, the output value of each grid cell is integrated with its local spatial context information. This suppresses the influence of isolated noise points, enhances spatial continuity, and preserves the physical consistency of first- and second-order features.

[0084] In summary, this embodiment achieves high-dimensional fusion of spatial pressure distribution and time-series airflow disturbance, generating a two-dimensional matrix that reflects the local characteristics of the anomalous region and possesses spatial smoothness and spatiotemporal coupling expression, providing a feature representation with both discriminative power and robustness for analysis.

[0085] Example 9: Based on Example 8, the process of calculating the arithmetic mean of the four components in step S3023 of this embodiment of the invention specifically includes the following steps:

[0086] Step S30231: For the current grid cell, read the row and column offsets of the horizontal left neighbor, horizontal right neighbor, vertical top neighbor, and vertical bottom neighbor from the pre-stored neighborhood offset record table; add the row index and column index of the current grid cell to the four offsets respectively to generate four neighborhood index pairs; if the row index of any neighborhood index pair is less than one row or greater than the maximum number of rows, or the column index is less than one column or greater than the maximum number of columns, then replace the neighborhood index pair with the index pair of the current grid cell itself;

[0087] Step S30232: The four neighborhood index pairs output are used as lookup addresses in sequence to extract the component values ​​of the corresponding positions from the weighted feature vector. During extraction, the storage order of the weighted feature vector is followed. The storage order is a one-dimensional arrangement with the row index increasing and the column index increasing. After converting each neighborhood index pair into a linear index, the component values ​​are extracted to obtain four neighborhood component values.

[0088] Step S30233: The output four neighborhood component values ​​are summed sequentially to obtain the sum; the sum is divided by four to obtain the arithmetic mean; at the same time, the effective count of the four neighborhoods of the current grid cell is recorded. If there is a case where the boundary is replaced by its own index, the arithmetic mean is still calculated as four values ​​without changing the denominator; the arithmetic mean is output as the neighborhood modulation value of the current row.

[0089] In the above embodiments, this embodiment achieves stable extraction and aggregation of neighborhood components of any cell in the grid-weighted feature vector by combining boundary processing, index transformation, and mean calculation. By using a preset neighborhood offset table and index out-of-bounds judgment, invalid neighborhoods outside the boundary are replaced with the current cell's own index, ensuring that four valid component values ​​can be obtained at any position (including the grid edge), avoiding numerical anomalies or calculation interruptions caused by missing boundaries in traditional neighborhood operations. Utilizing the mapping relationship between row and column indices and linear indices, components are directly extracted from the weighted feature vector through a continuous one-dimensional memory address, reducing the overhead of multi-dimensional array access; at the same time, the consistency of storage order and calculation logic is maintained, improving data reading efficiency. A cumulative averaging strategy with a fixed denominator (four) is adopted, so even if boundary replacement causes some neighborhood values ​​to be their own components, the denominator remains unchanged, allowing the output value to transition smoothly at the boundary, avoiding calculation deviations introduced by dynamic changes in the number of valid neighborhoods, and ensuring the spatial consistency of neighborhood modulation values. By recording the valid counts of four neighborhoods, auxiliary information is provided for possible weighting adjustments or anomaly detection, maintaining the traceability and scalability of the calculation process.

[0090] In summary, this embodiment, while ensuring computational efficiency, achieves stable and smooth output of neighborhood feature values ​​across the entire grid range through boundary normalization and fixed aggregation rules, providing spatially coherent modulated data.

[0091] Example 10: Based on Example 9, the process of extracting the component value after converting each neighborhood index pair into a linear index in step S30232 of this embodiment of the invention specifically includes the following steps:

[0092] Step S302321: Obtain the number of rows and columns of the original spatial map corresponding to the weighted feature vector, and use the number of rows as the step size parameter; for each neighborhood index pair, subtract one from its row index, multiply by the step size parameter, and add its column index to obtain the linear index value of the index pair, and store the linear index value in the linear index cache sequence.

[0093] Step S302322: Sort the four linear index values ​​in the output linear index cache sequence in ascending order according to their numerical values ​​to obtain the sorted linear index sequence; then traverse the sorted linear index sequence. If two adjacent index values ​​are the same, delete the duplicate and record the neighborhood index pair corresponding to the duplicate position. Use the sequence after deleting the duplicate as the unique linear index sequence.

[0094] Step S302323: Using each linear index value in the output unique linear index sequence as the access address, retrieve the corresponding component value from the continuous storage space of the weighted feature vector in sequence; if there are duplicate indexes that have been deleted, copy the retrieved corresponding component value to the original duplicate position, and finally arrange the four component values ​​in the order of the original four neighborhood index pairs, and output the four neighborhood component values.

[0095] In the above embodiments, this embodiment combines linear addressing, index deduplication, and data reconstruction to significantly improve memory access efficiency and computing performance while ensuring accurate extraction of neighborhood component values; at the same time, it maintains the structural consistency of data output.

[0096] Example 11: Based on Example 10, the process of obtaining the linear index value of the index pair in step S302321 of this embodiment of the invention specifically includes the following steps:

[0097] Step S3023211: Extract the row index value and column index value of each neighborhood index pair, subtract one from the row index value to obtain the zero-based row number, and subtract one from the column index value to obtain the zero-based column number; at the same time, read the partition coefficient corresponding to the neighborhood index pair from the pre-stored door partition weight table; the partition coefficient is pre-calibrated according to different areas of the door hinge side, lock side, upper and lower parts, and the value range is from zero to one;

[0098] Step S3023212: Multiply the output zero-based row number by the step size parameter to obtain the row offset; then add the row offset to the zero-based column number to obtain the basic linear index; then multiply the basic linear index by the partition coefficient to obtain the weighted linear index; at the same time, calculate the door edge distance value between the row and column of the neighborhood index, take the minimum value of the edge distance values ​​in the row direction and column direction as the edge distance, divide the edge distance by the maximum edge distance of the door and round it to obtain the edge attenuation value;

[0099] Step S3023213: Add the output weighted linear index to the edge decay value to obtain the final linear index value; perform a modulo operation on the linear index value and the total length of the weighted feature vector. If the linear index value is greater than or equal to the total length, subtract the total length to ensure that the index does not go out of bounds; store the final adjusted linear index value into the linear index cache sequence, and at the same time record the row index and column index of the original neighborhood index pair corresponding to the linear index value.

[0100] In the above embodiments, this embodiment generates a linear index that combines physical meaning and storage security by combining structural partition weights and geometric edge attenuation, enabling the feature extraction process to adapt to the regional characteristics and boundary morphology of the gate, thereby improving the physical consistency and computational robustness of the modulation operation.

[0101] Example 12: Based on Example 11, the process of calculating the distance value between the row and column of the gate body in step S3023212 of this embodiment of the invention specifically includes the following steps:

[0102] Step S30232121: Take the row index of the neighborhood index pair, calculate the distance value from the row index to the first row or the row index minus one and the distance value to the maximum number of rows or the maximum number of rows minus the row index, and take the smaller value as the row edge distance; similarly, take the column index, calculate the distance value to the first column and the distance value to the maximum number of columns, and take the smaller value as the column edge distance;

[0103] Step S30232122: Read the row stiffness coefficient and column stiffness coefficient corresponding to the neighborhood index pair from the pre-stored door stiffness distribution table; wherein the row stiffness coefficient increases linearly along the row direction from the door hinge side to the lock side, and the column stiffness coefficient decreases linearly along the column direction from the top to the bottom of the door; multiply the row edge distance by the row stiffness coefficient to obtain the weighted row edge distance; multiply the column edge distance by the column stiffness coefficient to obtain the weighted column edge distance;

[0104] Step S30232123: Add the weighted row edge distance and the weighted column edge distance to obtain the weighted edge distance sum; then extract the partition coefficient corresponding to the neighborhood index pair, multiply the weighted edge distance sum by the reciprocal of the partition coefficient to obtain the final edge distance value; divide the edge distance value by the maximum edge distance of the gate.

[0105] In the above embodiments, this embodiment generates edge distance values ​​that reflect both spatial location and structural characteristics by combining geometric boundaries, stiffness gradients, and partition weights; it provides input parameters with clear physical meaning and adaptability to the actual working conditions of the door for edge attenuation calculation, thereby enhancing the structural authenticity and working condition adaptability of the overall feature modulation process.

[0106] Example 13: As Figure 6 As shown, based on Embodiments 1-12, the anti-theft door abnormal intrusion behavior recognition system based on multimodal perception provided in this embodiment of the invention includes:

[0107] The abnormal area confirmation module 1 is used to collect tensile length data sets of strain fiber bundles embedded in different areas of the door body, perform differential comparison operations on the tensile length data sets, and output the deformation gradient matrix; then, based on the position of the elements in the deformation gradient matrix that exceed the preset threshold, the coordinate set of the local abnormal area is extracted.

[0108] The superposition operation module 2 is used to perform spatial superposition operation on the obtained set of coordinates of local abnormal areas and the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel; the result of the superposition operation generates the fracture mode vector of the conductive paint mesh, and then performs pressure inversion calculation on the fracture mode vector to output the contact pressure distribution map of the intrusion action.

[0109] The temporal alignment module 3 is used to perform temporal alignment operation on the obtained contact pressure distribution map and the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame. The aligned composite feature sequence is used to perform feature fusion calculation to obtain a composite feature tensor. Then, the composite feature tensor is used to perform classification mapping calculation to output the classification label of abnormal intrusion behavior.

[0110] In the above embodiments, this embodiment achieves high-precision, multi-dimensional identification of intrusion behavior by integrating data from three types of sensors: strain fiber bundles, conductive paint mesh, and micro piezoelectric films. First, the tensile length data collected by the strain fiber bundles is used to generate a deformation gradient matrix through differential comparison calculation, which can quickly locate local abnormal deformation areas on the door surface; effectively eliminating minor deformations caused by environmental noise or normal use, improving the targeting of anomaly detection; by spatially superimposing the coordinates of local abnormal areas with the breakpoint distribution of the conductive paint mesh, the system can generate vector data reflecting the physical fracture mode; combined with pressure inversion calculation, the contact pressure distribution map of the intrusion action on the door panel can be further reconstructed, thereby distinguishing force application modes of different strengths and areas, such as prying, impact, or cutting behaviors; using time-domain alignment operation, the contact pressure distribution map is synchronously fused with the airflow disturbance sequence. By fusing and calculating composite feature sequences, the system can correlate the mechanical effects of intrusion actions with the accompanying airflow characteristics, enhancing its ability to capture the temporal correlation of behaviors. The generated composite feature tensor, after classification mapping, can output classification labels for specific abnormal intrusion behaviors, such as forced entry, technical unlocking, or tentative touching.

[0111] In summary, this embodiment achieves a progressive analysis from local deformation detection and physical action inversion to multimodal temporal feature fusion, which improves the robustness, accuracy, and interpretability of abnormal intrusion behavior identification, while reducing the false alarm rate. It is suitable for real-time monitoring and early warning in security scenarios.

[0112] Figure 7 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0113] The electronic device may include a central processing unit / microprocessor / main control chip 4; and a storage medium 5 coupled to the central processing unit / microprocessor / main control chip 4 and storing computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by the processor.

[0114] The central processing unit / microprocessor / main control chip 4 may include, but is not limited to, one or more processors or microprocessors.

[0115] Storage medium 5 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0116] In addition, the electronic device may include (but is not limited to) a data bus 6, an input / output bus / external bus / device bus 7, a display 8, and input / output devices 9 (e.g., keyboard, mouse, speaker, etc.).

[0117] The central processing unit / microprocessor / main control chip 4 can communicate with external devices (8, 9, etc.) via wired or wireless networks (not shown) through the input / output bus / external bus / device bus 7.

[0118] The storage medium 5 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when the central processing unit / microprocessor / main control chip 4 is running.

[0119] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0120] Figure 8 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0121] like Figure 8 As shown, the non-transitory computer-readable storage medium 11 stores instructions, such as computer-readable instructions 10. When the computer-readable instructions 10 are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions 10 stored on the non-transitory computer-readable storage medium 11, the various methods described above can be performed.

[0122] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods of the various embodiments of this invention through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying abnormal intrusion behavior of security doors based on multimodal perception, characterized in that, Includes the following steps: The contact pressure distribution map of the intrusion action of the security door is obtained and the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame is aligned in the time domain. The aligned composite feature sequence is then subjected to feature fusion calculation to obtain a composite feature tensor. The composite feature tensor is then subjected to classification mapping calculation to output the classification label of abnormal intrusion behavior.

2. The method for identifying abnormal intrusion behavior of security doors based on multimodal perception as described in claim 1, characterized in that, The process of obtaining the composite feature tensor includes the following steps: The acquisition time of the contact pressure distribution map is extracted as the time reference. The time reference is aligned with the starting timestamp of the airflow disturbance sequence to obtain the time offset. The airflow disturbance sequence is translated based on the time offset. After the translation operation is completed, the contact pressure distribution map is repeatedly arranged according to a fixed time window to form a spatial map sequence with the same length as the airflow disturbance sequence. The spatial map remains unchanged within each time window. The time-domain aligned spatial map sequence and the translated airflow disturbance sequence are output, and the two together constitute the aligned composite feature sequence. Expand all element values ​​of the spatial map row by row into a one-dimensional spatial feature vector, and then perform a Cartesian product operation on the one-dimensional spatial feature vector and the airflow disturbance value to generate a two-dimensional fusion matrix; stack the two-dimensional fusion matrices generated at all time points in chronological order along the third dimension to obtain a three-dimensional composite feature tensor. The three-dimensional composite feature tensor is used for classification mapping calculation based on a pre-generated behavior classification template library. The element-wise product of the composite feature tensor and each template tensor is calculated sequentially and then summed to obtain the matching score of the classification template. The classification label corresponding to the classification template with the highest matching score is taken as the output, and the classification label is the classification label of abnormal intrusion behavior.

3. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 2, characterized in that, The process of generating a two-dimensional fusion matrix includes the following steps: The pressure value of each grid cell in the spatial map is extracted. All pressure values ​​are extracted in ascending order of row index and column index to form the original row vector. The sum of all pressure values ​​in the original row vector is calculated. Each pressure value in the original row vector is divided by the sum to obtain the normalized row vector. Then, the coordinate set of local anomaly regions is extracted. The grid cell index covered by each positioning coordinate interval in the coordinate set of local anomaly regions is used as a mask. The component at the corresponding index position in the normalized row vector is multiplied by a weighting coefficient of two, and the components at other positions are multiplied by a weighting coefficient of one to obtain the weighted feature vector. Perform a Cartesian product operation between the weighted eigenvector and the airflow disturbance value at the current time point in the translated airflow disturbance sequence; For each row of the two-dimensional fusion matrix, extract the four neighboring grid cells of the corresponding grid cell in the original spatial map, obtain the corresponding components of the four neighboring grid cells in the weighted eigenvector, and calculate the arithmetic mean of the four components; add the element of the first column of the row to the arithmetic mean and divide by two to obtain the updated value of the first column element; add the element of the second column of the row to the square of the arithmetic mean and divide by two to obtain the updated value of the second column element; after all rows are updated, output the two-dimensional fusion matrix.

4. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 3, characterized in that, The process of calculating the arithmetic mean of four components includes the following steps: For the current grid cell, read the row and column offsets in four directions (horizontal left neighbor, horizontal right neighbor, vertical top neighbor, and vertical bottom neighbor) from the pre-stored neighborhood offset record table; add the row index and column index of the current grid cell to the four offsets respectively to generate four neighborhood index pairs; if the row index of any neighborhood index pair is less than one row or greater than the maximum number of rows, or the column index is less than one column or greater than the maximum number of columns, then replace the neighborhood index pair with the index pair of the current grid cell itself. The four neighborhood index pairs output are used as lookup addresses in sequence to extract the component values ​​at the corresponding positions from the weighted feature vector. During extraction, the storage order of the weighted feature vector is followed. The storage order is a one-dimensional arrangement with the row index increasing and the column index increasing. After converting each neighborhood index pair into a linear index, the component values ​​are extracted to obtain four neighborhood component values. The four neighbor component values ​​are summed sequentially to obtain the sum; the sum is divided by four to obtain the arithmetic mean; at the same time, the effective count of the four neighbors of the current grid cell is recorded. If there is a case where the boundary is replaced by its own index, the arithmetic mean is still calculated as four values ​​without changing the denominator; the arithmetic mean is output as the neighborhood modulation value of the current row.

5. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 4, characterized in that, The process of converting each neighborhood index pair into a linear index and then retrieving the component value includes the following steps: Obtain the number of rows and columns of the original spatial map corresponding to the weighted feature vector, and use the number of rows as the step size parameter; for each neighborhood index pair, subtract one from its row index, multiply by the step size parameter, and add its column index to obtain the linear index value of the index pair, and store the linear index value in the linear index cache sequence. Sort the four linear index values ​​in the output linear index cache sequence in ascending order according to their numerical values ​​to obtain the sorted linear index sequence. Then, traverse the sorted linear index sequence. If two adjacent index values ​​are the same, delete the duplicate and record the neighborhood index pair corresponding to the duplicate position. Use the sequence after deleting duplicates as the unique linear index sequence. Using each linear index value in the output unique linear index sequence as the access address, retrieve the corresponding component value from the contiguous storage space of the weighted feature vector in sequence; if there are duplicate indexes that have been deleted, copy the retrieved corresponding component value to the original duplicate position, arrange the four component values ​​in the order of the original four neighborhood index pairs, and output the four neighborhood component values.

6. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 5, characterized in that, The process of obtaining the linear index value of an index pair includes the following steps: Extract the row index and column index values ​​of each neighborhood index pair, subtract one from the row index value to obtain the zero-based row number, and subtract one from the column index value to obtain the zero-based column number; at the same time, read the partition coefficient corresponding to the neighborhood index pair from the pre-stored door partition weight table; the partition coefficient is pre-calibrated according to different areas of the door hinge side, lock side, upper and lower parts, and the value range is zero to one; Multiply the output zero-based row number by the step size parameter to obtain the row offset; then add the row offset to the zero-based column number to obtain the basic linear index; then multiply the basic linear index by the partition coefficient to obtain the weighted linear index; at the same time, calculate the door edge distance value between the row and column of the neighborhood index, take the minimum value of the edge distance values ​​in the row direction and column direction as the edge distance, divide the edge distance by the maximum edge distance of the door and round it to obtain the edge decay value; The weighted linear index is added to the edge decay value to obtain the final linear index value. The linear index value is modulo the total length of the weighted feature vector. If the linear index value is greater than or equal to the total length, the total length is subtracted to ensure that the index does not go out of bounds. The final adjusted linear index value is stored in the linear index cache sequence, and the row index and column index of the original neighborhood index pair corresponding to the linear index value are recorded at the same time.

7. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 6, characterized in that, The process of calculating the distance between the row and column edges of a neighbor index includes the following steps: Take the row index of the neighbor index pair, calculate the distance from the row index to the first row (or the row index minus one) and the distance to the maximum number of rows (or the maximum number of rows minus the row index), and take the smaller of the two values ​​as the row edge distance; similarly, take the column index, calculate the distance to the first column and the distance to the maximum number of columns, and take the smaller of the two values ​​as the column edge distance. The row stiffness coefficient and column stiffness coefficient corresponding to the neighborhood index pair are read from the pre-stored door stiffness distribution table. The row stiffness coefficient increases linearly along the row direction from the door hinge side to the lock side, and the column stiffness coefficient decreases linearly along the column direction from the top to the bottom of the door. The row edge distance is multiplied by the row stiffness coefficient to obtain the weighted row edge distance. The column edge distance is multiplied by the column stiffness coefficient to obtain the weighted column edge distance. Add the weighted row edge distance to the weighted column edge distance to get the weighted edge distance sum; then extract the partition coefficient corresponding to the neighborhood index pair, and multiply the weighted edge distance sum by the reciprocal of the partition coefficient to get the final edge distance value.

8. The method for identifying abnormal intrusion behavior of anti-theft doors based on multimodal perception as described in claim 1, characterized in that, The method for identifying abnormal intrusion behavior of security doors based on multimodal perception also includes collecting tensile length data sets of strain fiber bundles embedded in different areas of the door body, performing differential comparison operations on the tensile length data sets, and outputting a deformation gradient matrix; then, based on the position of elements in the deformation gradient matrix that exceed a preset threshold, extracting the coordinate set of local abnormal areas.

9. The method for identifying abnormal intrusion behavior of a security door based on multimodal perception as described in claim 8, characterized in that, The obtained set of coordinates of local abnormal areas is spatially superimposed with the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel. The result of the superposition operation generates the fracture mode vector of the conductive paint mesh. Then, the fracture mode vector is subjected to pressure inversion calculation to output the contact pressure distribution map of the intrusion action.

10. A multimodal perception-based system for identifying abnormal intrusion behavior in security doors, used to implement the multimodal perception-based method for identifying abnormal intrusion behavior in security doors as described in any one of claims 1 to 9, characterized in that, include: The abnormal area confirmation module is used to collect tensile length data sets of strain fiber bundles embedded in different areas of the door body, perform differential comparison operations on the tensile length data sets, and output the deformation gradient matrix; then, based on the position of the elements in the deformation gradient matrix that exceed the preset threshold, the coordinate set of the local abnormal area is extracted. The overlay operation module is used to perform spatial overlay operation on the obtained set of coordinates of local abnormal areas and the breakpoint distribution of the conductive paint mesh pre-printed on the inner layer of the door panel; the overlay operation result generates the fracture mode vector of the conductive paint mesh, and then performs pressure inversion calculation on the fracture mode vector to output the contact pressure distribution map of the intrusion action. The temporal alignment module is used to perform temporal alignment operations on the obtained contact pressure distribution map and the airflow disturbance sequence collected by the micro piezoelectric film pre-placed in the gap between the door leaf and the door frame. The aligned composite feature sequence is then used to perform feature fusion calculation to obtain a composite feature tensor. Finally, the composite feature tensor is used to perform classification mapping calculation to output the classification label of abnormal intrusion behavior.

Images