A visual detection method for assembling state of a balance based on multi-view image fusion

Abstract

This invention relates to the field of machine vision and intelligent manufacturing inspection technology, and discloses a visual inspection method for the assembly state of a balance wheel based on multi-view image fusion. The method involves acquiring multi-view images of the balance wheel, extracting rigid body edge feature sets using a multi-scale edge detection operator, generating a spatial compensation matrix based on the linear expansion coefficient and projective geometric invariance, and establishing a three-dimensional spatial coordinate system. It also involves calculating the inner product of multi-view gradient vectors to generate a mask scalar that shields preset spatial points, outputting the target topology network; extracting the non-rigid body displacement time series and obtaining feature accelerations through second derivative calculation, and calculating the deformation compensation vector using the stress relaxation time constant; superimposing the deformation compensation vector onto the three-dimensional spatial coordinate system to drive displacement mapping, generating a three-dimensional model of the target assembly, and outputting defect and assembly tolerance parameters.

Description

Technical Field

[0001] This invention relates to the field of machine vision and intelligent manufacturing inspection technology, and in particular to a visual inspection method for the assembly status of a balance wheel based on multi-view image fusion. Background Technology

[0002] In existing technologies for visual inspection of precision mechanical assembly, static two-dimensional image comparison algorithms or deep learning networks are typically used for extracting edge features of parts. However, during the high-frequency assembly of a balance wheel, the ambient temperature drift and electromagnetic scattering from the balance wheel surface interfere with each other. Because the visual processing chain lacks physical reference constraints, static two-dimensional image comparison algorithms are prone to misidentifying spatially phase-lag virtual images caused by light source projection as physical entity boundaries, or introducing thermodynamic deformation errors in the feature matching stage. This results in initial perceptual bias in the static feature extraction layer of the visual inspection system.

[0003] The initial perception bias of the static feature extraction layer, when transmitted from the vision inspection system to the 3D assembly quality assessment stage, can only rely on the static pixel coordinates under discrete timestamps. When the balance wheel experiences non-rigid high-frequency vibration due to assembly impact, the static pixel coordinates forcibly extract transient displacement. Due to the lack of physical damping constraints in the underlying data link, the static pixel coordinate system directly superimposes the high-frequency vibration displacement with the preceding initial perception bias. The superposition of the high-frequency vibration displacement and the initial perception bias disrupts the reconstruction benchmark of the 3D physical space, leading to a measurement offset of the assembly boundary, ultimately causing the assembly line to output incorrect rejection actions or issue abnormal torque compensation commands. Summary of the Invention

[0004] This invention provides a visual inspection method for the assembly status of a balance wheel based on multi-view image fusion, in order to solve the technical problem that existing visual inspection systems suffer from static optical perception deviations caused by environmental temperature drift and projected virtual images. When these static optical perception deviations are transmitted to the dynamic assembly process, they are superimposed with non-rigid high-frequency vibration displacements, causing static and dynamic errors to couple, ultimately leading to deviations in the reconstruction measurement of the target assembly 3D model and abnormal output of defect detection parameters.

[0005] In view of the above problems, the present invention provides a visual detection method for the assembly state of a balance wheel based on multi-view image fusion, comprising the following steps: Acquire multi-view images of the balance wheel containing rigid and non-rigid regions, call a multi-scale edge detection operator to perform feature purification calculations on the multi-view images, output an enhanced feature map, and extract the edge feature set of the rigid region in the enhanced feature map; Based on the preset linear expansion coefficient and the projective geometric invariance of the edge feature set, the three-dimensional relative pose of the multi-view image acquisition device is calculated, a spatial compensation matrix is ​​generated, and a three-dimensional spatial coordinate system is established using the spatial compensation matrix. The enhanced feature map is converted into gradient field data and mapped to the three-dimensional spatial coordinate system; the gradient vectors of the same spatial point under different viewpoints are extracted, the inner product of the gradient vectors is calculated, and a mask scalar is generated based on the inner product; The mask scalar is applied to the three-dimensional spatial coordinate system, and spatial points in the three-dimensional spatial coordinate system with the mask scalar value set to a first preset value are masked, and the target topology network is output. Within a preset time window, extract the displacement time series of the non-rigid body region in the target topology network; The characteristic acceleration is obtained by taking the second derivative of the displacement-time series with respect to time. Obtain the stress relaxation time constant of the material corresponding to the non-rigid region, use the stress relaxation time constant as the constraint boundary, and calculate the deformation compensation vector in combination with the characteristic acceleration. The deformation compensation vector is superimposed onto the three-dimensional spatial coordinate system, driving the spatial points of the non-rigid body region in the three-dimensional spatial coordinate system to perform displacement mapping, generating a target assembly three-dimensional model, and outputting defect and assembly tolerance parameters based on the target assembly three-dimensional model.

[0006] A visual inspection system for balance wheel assembly status based on multi-view image fusion includes a memory and a processor. The memory stores computer instructions, and the processor executes the computer instructions to implement the aforementioned visual inspection method for balance wheel assembly status based on multi-view image fusion.

[0007] The technical solution provided in this application has at least the following technical effects: By calling a multi-scale edge detection operator to extract enhanced feature maps and generating a spatial compensation matrix based on the linear expansion coefficient and projective geometric invariance, the external parameter calibration error of multi-view image acquisition equipment caused by environmental temperature drift is eliminated, thereby improving the accuracy of establishing a three-dimensional spatial coordinate system.

[0008] By calculating the inner product of gradient vectors under different viewpoints and issuing mask scalars based on the inner product results to shield spatial points, the shadow virtual images generated by the light source projection are intercepted and removed in the underlying data link, thus avoiding the impact of optical interference on the reconstruction accuracy of the target topology network.

[0009] By extracting characteristic acceleration and combining it with the stress relaxation time constant to generate a deformation compensation vector, the high-frequency vibration displacement caused by assembly impact is converged by utilizing the intrinsic damping characteristics of non-rigid materials. This breaks the superposition and coupling of dynamic vibration error and static perception deviation, ensuring the accuracy of the final balance wheel defect and assembly tolerance parameter output. Attached Figure Description

[0010] Figure 1A flowchart of a visual detection method for the assembly state of a balance wheel based on multi-view image fusion is provided in an embodiment of the present invention; Figure 2 This is an architecture diagram of a visual inspection system for the assembly status of a balance wheel based on multi-view image fusion, provided in an embodiment of the present invention. Detailed Implementation

[0011] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0012] Example: Please refer to Figure 1 and Figure 2 A visual inspection system for the assembly status of a balance wheel based on multi-view image fusion is configured with an industrial control host that executes calculation instructions, a light source array deployed at the target inspection station, and multi-view image acquisition devices arranged around the target inspection station. The industrial control host internally compiles and instantiates a core processing engine. The multi-view image acquisition devices acquire multi-view images of the balance wheel, including both rigid and non-rigid regions. After discretizing the multi-view images, the multi-view image acquisition devices transmit them to the core processing engine. Upon receiving multi-view image input, the core processing engine initiates a cross-physics data transfer procedure. This procedure sequentially transfers data sequences in memory according to image feature extraction logic, geometric alignment logic, optical image removal logic, and dynamic deduction logic.

[0013] With the initiation of the cross-physics data transfer process, the data processing flow enters the feature extraction stage. The core processing engine calls the multi-scale edge detection operator to perform feature purification calculations on the multi-view images, outputting an enhanced feature map focused on the balance wheel assembly boundary. After outputting the enhanced feature map, the data transfer enters the geometric alignment stage. The core processing engine extracts the edge feature set of the rigid body region from the enhanced feature map, and reads the preset linear expansion coefficient. Based on the projective geometric invariance of the linear expansion coefficient and the edge feature set, the core processing engine calculates the three-dimensional relative pose of the multi-view image acquisition device. After obtaining the three-dimensional relative pose, the core processing engine generates a spatial compensation matrix. Using the spatial compensation matrix, the core processing engine performs coordinate mapping to establish a three-dimensional spatial coordinate system.

[0014] After establishing a three-dimensional spatial coordinate system, the enhanced feature map is transformed into gradient field data and mapped to the three-dimensional spatial coordinate system. For the same spatial point, the core processing engine extracts gradient vectors from different viewpoints and calculates the inner product. The mask scalar is generated by the core processing engine and acts on the three-dimensional spatial coordinate system to output the target topology network.

[0015] After the target topology network is output, the dynamics derivation phase is triggered. Within a preset time window, the displacement time series of non-rigid regions in the target topology network is extracted by a high-frequency sampling register and written to the calculus unit. The displacement time series undergoes a second derivative operation with respect to time to generate characteristic accelerations. Combining the pre-read stress relaxation time constant and characteristic accelerations, the calculus unit calculates and generates a deformation compensation vector. The deformation compensation vector is superimposed onto the three-dimensional coordinate system, driving the spatial points of the non-rigid regions in the three-dimensional coordinate system to perform displacement mapping, generating a target assembly three-dimensional model. The industrial control host extracts and outputs defect and assembly tolerance parameters based on the target assembly three-dimensional model.

[0016] For multi-view images input from multi-view image acquisition devices, the core processing engine activates multi-scale edge detection operators. By activating Gaussian pyramid downsampling operators and gradient magnitude calculation operators in parallel, multi-scale pixel-level differencing and smoothing calculations are performed on the multi-view images, outputting an enhanced feature map that filters out background noise. After the enhanced feature map is generated, the core processing engine scans the enhanced feature map and extracts the edge feature set of rigid body regions from it.

[0017] After the edge feature set is extracted, the core processing engine establishes a projective basis constraint mechanism. The core processing engine extracts the pre-defined linear expansion coefficient from memory; this coefficient reflects the lattice size shift of the target detection station under temperature perturbation conditions. The core processing engine establishes the linear expansion coefficient as a tolerance boundary, which defines the extreme value range of coordinate perturbation in three-dimensional space. After establishing the tolerance boundary, the conic section features in the edge feature set are extracted and set as a rigid reference. The conic section features are initialized as a second-order symmetric tensor matrix within the core processing engine.

[0018] Within the tolerance boundary, constraint equations are established for the second-order symmetric tensor matrices captured by the multi-view image acquisition device at different viewpoints. These constraint equations invoke projective geometry theorems, defining the absolute invariants of spatial conic sections as the ratio of the square of the sum of the elements on the main diagonal of the tensor matrix to the determinant of the tensor matrix. The constraint equations also establish an equality relationship between the absolute invariants of the second-order symmetric tensor matrices generated by mapping the same conic section at different visual angles. Furthermore, the constraint equations algebraically correlate the theoretical projective coordinates of three-dimensional physical points with the acquired coordinates incorporating thermodynamic distortion parameters. After establishing the constraint equations, the core processing engine constructs a least-squares objective function and executes a reverse iterative solution algorithm. This algorithm uses the initial calibration extrinsic parameter matrix of the multi-view image acquisition device as the starting point to perform gradient descent operations to obtain the extrinsic parameter deviation tensor. The extrinsic parameter deviation tensor quantifies the camera's spatial position offset. Based on this deviation tensor, a spatial compensation matrix is ​​generated. This matrix is ​​applied to the original calibration matrix of the multi-view image acquisition device to eliminate thermodynamic displacement errors and construct a three-dimensional spatial coordinate system.

[0019] After the 3D spatial coordinate system is constructed, the enhanced feature map is received by the core processing engine and pixel-level differencing is performed. The core processing engine then uses orthogonal first-order derivative operators to perform spatial convolution on the enhanced feature map, transforming it into gradient field data containing horizontal and vertical gradient components. The core processing engine then invokes the 3D spatial coordinate system and writes the gradient field data into the node array of the 3D spatial coordinate system according to the inverse perspective projection mapping rule. After mapping, the core processing engine addresses the same spatial point in the 3D spatial coordinate system and extracts the first gradient vector from the first viewpoint and the second gradient vector from the second viewpoint.

[0020] In the hardware arithmetic logic unit (ALU), the extracted gradient vectors are subjected to inner product operations. The ALU extracts the horizontal and vertical gradient components of the first and second gradient vectors respectively, and executes multiply-add instructions for the corresponding components to calculate and output the scalar product. When the same spatial point in the 3D coordinate system belongs to the surface of a physical entity, the gradient vectors captured from different viewpoints exhibit acute angles, and the inner product operation outputs a result greater than 0. When the same spatial point belongs to the shadow image caused by the projection of a light source array, the gradient vectors captured from different viewpoints exhibit orthogonal or divergent directions, and the inner product operation outputs a result less than or equal to 0. The core processing engine assigns a mask scalar based on the inner product operation result: when the inner product is greater than 0, the mask scalar is assigned a reserved value; when the inner product is less than or equal to 0, the mask scalar is assigned a first preset value. The assigned mask scalar is applied to the 3D coordinate system, and the core processing engine issues a data filtering instruction, which blocks spatial points with the mask scalar set to the first preset value for invalidation processing. Invalidation removes artifact data at the memory level, and the target topology network is constructed and output by the core processing engine based on the set of spatial points assigned the retained values.

[0021] After the target topology network is output, within a preset time window, the displacement time series of non-rigid regions in the target topology network is continuously extracted by a high-frequency sampling register according to the underlying hardware clock cycle. The core processing engine extracts the transient displacement of the non-rigid region corresponding to the current moment from the displacement time series. The transient displacement is fed into the calculus unit. The calculus unit extracts the transient displacement difference between adjacent sampling points according to the step size of the underlying hardware clock cycle, and performs a first-order central difference discretization operation to obtain the first derivative; the calculus unit synthesizes the transient displacements of three consecutive sampling points, and uses the ratio of the displacement difference to the square of the clock cycle to perform a second-order central difference discretization operation to obtain the characteristic acceleration.

[0022] After completing the differentiation operation, the calculus unit performs a hardware multiplication of the characteristic acceleration and the square of the stress relaxation time constant, obtaining the undamped deformation displacement vector representing the foundation oscillation displacement in the non-rigid body region. Simultaneously, the calculus unit reads the absolute value of the first derivative. A preset reference length parameter is introduced, triggering multipliers and dividers. The calculus unit then performs dimensionless processing on the product of the absolute value of the first derivative and the stress relaxation time constant using the reference length parameter. Based on the dimensionless product result, the calculus unit introduces a second weighting coefficient configured by the system for algebraic scaling, generating a decay weight. The decay weight is configured by hardware logic to decrease as the absolute value of the first derivative increases. After the decay weight is generated, the core processing engine extracts the preset first weighting coefficient. Using the first weighting coefficient and the attenuation weight, the core processing engine issues a weighted calculation instruction to the undamped deformation displacement vector. The calculus unit converts the damped displacement dissipation ratio to offset the basic oscillation displacement, and outputs the deformation compensation vector.

[0023] The formula for calculating the deformation compensation vector is: in, For deformation compensation vector, For transient displacement, For characteristic acceleration, The first derivative, Let be the stress relaxation time constant. As the reference length parameter, As the first weighting coefficient, This is the second weighting coefficient. This is a preset minimum constant.

[0024] The calculus unit extracts the system's preset stress relaxation time constant, first weighting coefficient, and second weighting coefficient. The stress relaxation time constant characterizes the physical period of the exponential decay of internal stress with time in the non-rigid material of the balance wheel under constant strain, and is locked within a millisecond-level numerical range. The first weighting coefficient is configured as a mapping constant characterizing the system's environmental stiffness, scaling the undamped deformation displacement vector to a micrometer-level physical measurement space. The second weighting coefficient is configured as a mapping constant characterizing damping sensitivity, controlling the gradient descent rate of the decay curve. The calculus unit substitutes the absolute value of the first derivative, characteristic acceleration, stress relaxation time constant, reference length parameter, first weighting coefficient, and second weighting coefficient into the calculation formula for the deformation compensation vector.

[0025] When the target detection station experiences a physical impact, causing high-frequency vibrations in the non-rigid region, the absolute value of the first derivative tends towards positive infinity. When the absolute value of the first derivative approaches positive infinity, the stress relaxation time constant and the reference length parameter are positive constants, and the denominator term tends towards positive infinity. Positive infinity triggers the underlying floating-point numerical truncation mechanism, causing the attenuation weight to converge and be assigned a value of 0. When the attenuation weight is 0, the algebraic multiplication mechanism of the characteristic acceleration is blocked. The calculus unit cuts off the compensation flow channel and outputs a deformation compensation vector with a value of 0. This deformation compensation vector with a value of 0 blocks the spatial superposition operation of the high-frequency vibration displacement noise.

[0026] When the non-rigid region reaches the steady-state reversal point of the physical oscillation envelope at the peak or trough, the absolute value of the first derivative approaches 0. As the absolute value of the first derivative approaches 0, the product term also approaches 0. A preset minimum constant ensures that the denominator term is extremely close to the reference length parameter, and the calculus unit outputs a decay weight approaching 1. This decay weight approaching 1 removes the amplitude restriction on the undamped deformation displacement vector, and the product term formed by the characteristic acceleration and the square of the stress relaxation time constant is directly retained. The calculus unit retains the elastic displacement compensation obtained based on the characteristic acceleration. The calculus unit maps the retained elastic displacement compensation to the deformation compensation vector, ensuring the spatial reconstruction accuracy of the balance wheel's three-dimensional model.

[0027] The core processing engine acquires a preset minimum constant and a second weighting coefficient. The hardware multiplier within the calculus unit receives the absolute value of the first derivative, the stress relaxation time constant, and the second weighting coefficient; the hardware multiplier performs a series multiplication operation to output the product term. The calculus unit reads the reference length parameter; the hardware adder performs an accumulation operation on the reference length parameter, the product term, and the preset minimum constant to output the denominator term. The hardware divider calculates the ratio of the reference length parameter to the denominator term and outputs the decay weight.

[0028] A preset minimum constant is injected into the hardware adder as a fixed offset. This preset minimum constant is configured as a positive real number constant calibrated according to the lower limit of floating-point arithmetic precision. When the absolute value of the first derivative is equal to 0, the product term is calculated as 0. The preset minimum constant ensures that the denominator is always greater than 0, and this constant greater than 0 denominator prevents floating-point arithmetic overflow triggered when executing a division instruction. After generating the attenuation weights, the calculus unit uses the first weight coefficient and the attenuation weights to perform multiplication calculations on the undamped deformation displacement vector. The calculus unit outputs the deformation compensation vector.

[0029] The microcontroller unit (MCU) executes a lookup table mapping procedure in the read-only memory (ROM). The mapping curve for the attenuation weights is discretized into a one-dimensional mapping array in offline mode, and this array is burned into the ROM address space of the MCU. The MCU converts the absolute value of the first derivative into an address pointer, and then performs an address lookup in the ROM based on the pointer. The MCU extracts the upper and lower limit attenuation weights for the corresponding interval. The arithmetic logic unit (ALU) performs arithmetic linear interpolation calculations and outputs the attenuation weights.

[0030] The industrial control host is equipped with a central processing unit (CPU) array, a graphics processing unit (GPU), video memory (VRAM), and an instruction register set. Multi-view images are written to the VRAM. A streaming multiprocessor calls a multi-scale edge detection operator to process the multi-view images and generate enhanced feature maps. The streaming multiprocessor extracts the gradient vectors of the same spatial point from the VRAM under different views, allocates the gradient vectors to the floating-point core to execute vector inner product instructions, and the floating-point core calculates the mask scalar and overwrites it to the physical address of the VRAM. The CPU array extracts the linear expansion coefficient and conic section features from the instruction register set. The CPU array attaches constraint equations in hardware threads, and the CPU array calls an algebraic logic unit to perform matrix inverse iterative calculations. The algebraic logic unit outputs a spatial compensation matrix.

[0031] The compiler front-end parses logical actions into machine code streams, which are then loaded into the instruction cache of the central processing unit (CPU) array. When establishing a three-dimensional coordinate system, the arithmetic logic unit (ALU) reads the machine code and pushes the tensor data of the spatial compensation matrix onto the memory stack. The machine code triggers coordinate mapping transformation instructions. When performing a masking action, the condition that the mask scalar equals the first preset value triggers a hardware conditional jump branch instruction, and the program counter skips the data storage addressing cycle of the current spatial point. When performing a weighted calculation action, the floating-point multiplier receives the physical level signals corresponding to the first weighting coefficient, the attenuation weight, and the undamped deformation displacement vector. The floating-point multiplier outputs the deformation compensation vector electrical signal through a shift register and an adder gate array.

[0032] Computer-readable storage media include non-volatile solid-state drives (SSDs), electrically erasable programmable read-only memory (EEPROM), and static random access memory (SRAM). Non-volatile SSDs are used to store multi-scale edge detection operator code and the underlying execution instruction set. EEPROM stores the binary truth values ​​of the first weighting coefficient, the second weighting coefficient, the stress relaxation time constant, and the linear expansion coefficient. SRAM provides a data buffer. Actions performed by the arithmetic logic unit (ALU) change the level toggling state of the flip-flop circuits within the SRAM, which in turn drive the spatial points of the 3D model of the balance wheel to complete the displacement mapping.

[0033] 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; and these 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 visual inspection method for the assembly state of a balance wheel based on multi-view image fusion, characterized in that, Includes the following steps: Acquire multi-view images of the balance wheel containing rigid and non-rigid regions, call a multi-scale edge detection operator to perform feature purification calculations on the multi-view images, output an enhanced feature map, and extract the edge feature set of the rigid region in the enhanced feature map; Based on the preset linear expansion coefficient and the projective geometric invariance of the edge feature set, the three-dimensional relative pose of the multi-view image acquisition device is calculated, a spatial compensation matrix is ​​generated, and a three-dimensional spatial coordinate system is established using the spatial compensation matrix. The enhanced feature map is converted into gradient field data and mapped to the three-dimensional spatial coordinate system; Extract the gradient vectors of the same spatial point under different viewpoints, calculate the inner product of the gradient vectors, and generate a mask scalar based on the inner product; The mask scalar is applied to the three-dimensional spatial coordinate system, and spatial points in the three-dimensional spatial coordinate system with the mask scalar value set to a first preset value are masked, and the target topology network is output. Within a preset time window, extract the displacement time series of the non-rigid body region in the target topology network; The characteristic acceleration is obtained by taking the second derivative of the displacement-time series with respect to time. Obtain the stress relaxation time constant of the material corresponding to the non-rigid region, use the stress relaxation time constant as the constraint boundary, and calculate the deformation compensation vector in combination with the characteristic acceleration. The deformation compensation vector is superimposed onto the three-dimensional spatial coordinate system, driving the spatial points of the non-rigid body region in the three-dimensional spatial coordinate system to perform displacement mapping, generating a target assembly three-dimensional model, and outputting defect and assembly tolerance parameters based on the target assembly three-dimensional model.

2. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 1, characterized in that, The calculation of the three-dimensional relative pose of the multi-view image acquisition device based on the preset linear dilation coefficient and the projective geometric invariance of the edge feature set, and the generation of the spatial compensation matrix, specifically includes: The linear expansion coefficient is used as the tolerance boundary; Using the conic curve features in the edge feature set as a reference, within the tolerance boundary, a constraint equation is established to ensure that the projective invariants are equal under different viewpoints. Solve the constraint equations to obtain the extrinsic deviation tensor of the multi-view image acquisition device, and generate the spatial compensation matrix accordingly.

3. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 1, characterized in that, The process of extracting the gradient vectors of the same spatial point from different viewpoints, calculating the inner product of the gradient vectors, and generating a mask scalar based on the inner product specifically includes: Extract the gradient vector of the same spatial point from different viewpoints; Calculate the inner product of the gradient vectors; When the inner product is greater than zero, the mask scalar is assigned a reserved value; When the inner product is less than or equal to zero, the mask scalar is assigned the first preset value.

4. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 1, characterized in that, The step of taking the second derivative of the displacement-time series with respect to time to obtain the characteristic acceleration specifically includes: Extract the transient displacement of the non-rigid body region corresponding to the current moment from the displacement time series; The first derivative is obtained by taking the first derivative of the transient displacement with respect to time. The characteristic acceleration is obtained by taking the second derivative of the transient displacement with respect to time.

5. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 4, characterized in that, The calculation logic for using the stress relaxation time constant as a constraint boundary and combining it with the characteristic acceleration to calculate the deformation compensation vector includes: Multiply the characteristic acceleration by the square of the stress relaxation time constant to obtain the undamped deformation displacement vector; The product of the first derivative and the stress relaxation time constant is made dimensionless using a preset reference length parameter. Based on the product after dimensionless processing, a decay weight is generated by scaling by introducing a preset second weight coefficient, which decreases as the absolute value of the first derivative increases. Using a preset first weighting coefficient and the attenuation weight, the undamped deformation displacement vector is weighted and calculated to output the deformation compensation vector.

6. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 5, characterized in that, The formula for calculating the deformation compensation vector is: in, The deformation compensation vector is... For the transient displacement, For the characteristic acceleration, Let be the first derivative. The stress relaxation time constant is... The reference length parameter is... The first weighting coefficient, This is the second weighting coefficient. This is a preset minimum constant.

7. The visual inspection method for balance wheel assembly state based on multi-view image fusion according to claim 5, characterized in that, The steps of generating the attenuation weight and outputting the deformation compensation vector include: Obtain a preset minimum constant, and multiply the absolute value of the first derivative, the stress relaxation time constant, and the second weighting coefficient together to obtain the product term; The denominator term is obtained by adding the reference length parameter, the product term, and the preset minimum constant. Calculate the ratio of the baseline length parameter to the denominator term to generate the attenuation weight; The deformation compensation vector is output by multiplying the undamped deformation displacement vector using the first weighting coefficient and the attenuation weight.

8. A visual inspection system for the assembly status of a balance wheel based on multi-view image fusion, characterized in that, It includes a memory and a processor, the memory being used to store computer instructions, and the processor being used to execute the computer instructions to implement the visual inspection method for balance wheel assembly state based on multi-view image fusion as described in any one of claims 1 to 7.

Images