Intelligent mechanical part defect detection system based on deep learning

Abstract

The invention belongs to the technical field of intelligent defect detection, and relates to an intelligent mechanical part defect detection system based on deep learning. The system is provided with a parameter matrix acquisition module, a mechanical part quantum state data output module, a fractional dimension convolution feature extraction module, a mechanical part defect manifold coordinate extraction module, a manifold transposition sub-calculation module and a mechanical part defect confidence output module. The method solves the problems that traditional mechanical part defect detection is generally identified according to professional detection personnel in the field, the subjectivity of manual detection is high, the false detection rate of a detection result is high, and the detection efficiency and the detection precision of manual detection are low; the possibility that the mechanical part has defects can be evaluated more comprehensively and reliably, manual intervention is reduced, and the detection efficiency and consistency are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent defect detection, and relates to an intelligent defect detection system for mechanical parts based on deep learning. Background Technique

[0002] Mechanical equipment is mostly used to replace manual labor in high-intensity, repetitive, and relatively dangerous work to liberate humans. Therefore, the working environment of mechanical equipment is often relatively harsh. Mechanical equipment is often used to replace humans in working environments with strong alkalis, strong acids, high temperatures, and high humidity.

[0003] During long-term operation in a harsh environment, the metallic component parts of mechanical equipment are prone to deformation and corrosion, resulting in faults and defects that cause the mechanical equipment to malfunction or the operating efficiency to decline. To prevent the mechanical equipment from stopping operation or even being scrapped due to defective mechanical parts, traditional mechanical part defect detection is generally carried out by professional inspectors in the field. This method has a high labor cost, strong subjectivity in manual inspection, a high false detection rate in the detection results, and high requirements for the professionalism of inspectors. The detection efficiency and detection accuracy of manual inspection are low. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides an intelligent defect detection system for mechanical parts based on deep learning, which can effectively solve the problems.

[0005] The present invention is realized through the following technical solutions:

[0006] An intelligent defect detection system for mechanical parts based on deep learning includes: a parameter matrix acquisition module for acquiring the material parameter matrix of mechanical parts; a mechanical part quantum state data output module for outputting the mechanical part quantum state data based on a multi-modal generation model excited by a quantum field; a fractional dimension convolution feature extraction module for performing fractional dimension convolution feature extraction on the mechanical part quantum state data and performing quantum state decoding to obtain the fractal features of the mechanical parts; a mechanical part defect manifold coordinate extraction module for constructing a topological manifold detection network, calculating the manifold extreme points, and extracting the mechanical part defect manifold coordinates from the fractal features of the mechanical parts; a manifold commutator calculation module for calculating the manifold commutator based on the mechanical part defect manifold coordinates; and a mechanical part defect confidence output module for performing non-commutative geometric decision fusion and outputting the mechanical part defect fusion confidence.

[0007] As a further solution, for the multi-modal generation model excited by a quantum field to output the mechanical part quantum state data, the specific analysis process includes:

[0008] (1) Parameter matrix acquisition module:

[0009] Acquire the material parameters of the mechanical parts, specifically including the elastic modulus of the mechanical parts , Poisson's ratio of mechanical parts , Density of mechanical parts , Yield strength of mechanical parts , Fracture toughness of mechanical parts , Melting point of mechanical parts ; Construct a material parameter matrix for mechanical parts :[[]]

[0010] ; (1)[[]]

[0011] Combine the material parameters of mechanical parts with the quantum field, and each material parameter controls a component of the quantum field;

[0012] (2) Quantum state data output module for mechanical parts:

[0013] Obtain the generated network weights stored in the database , and generate quantum states through the path integral of quantum field theory :[[]]

[0014] ; (2)[[]]

[0015] In the formula, is the quantum state function obtained through path integral, describing the quantum state of the system at position x in space, is the normalization constant, is the functional integral measure of the quantum field , is the action of the quantum field stored in the database, is the generation network, and e is the natural constant; represents matrix multiplication; refers to the combination of the material parameter matrix of mechanical parts and the quantum field;

[0016] The quantum state needs to satisfy the material mechanics equation:

[0017] ; (3)[[]]

[0018] In the formula, is the displacement field, is the Lame constant stored in the database, is the field nonlinear coefficient stored in the database, is the displacement field with respect to time 's second derivative, i.e., acceleration;

[0019] Finally, obtain the quantum state data through projective measurement :[[]]

[0020] ; (4)

[0021] In the formula, is the three-dimensional complex-valued data obtained by quantum measurement, where a, b, and c are indices in three-dimensional space, is the quantum state is the conjugate transpose of, is the measurement operator for a, b, and c, is the trace of the matrix, a represents the surface topography, b represents the internal stress, and c represents the crack depth;

[0022] Output quantum state data .

[0023] As a further solution, combine the mechanical part material parameters with the quantum field. The specific analysis process is as follows:

[0024] ; (5)

[0025] In the formula, is the quantum field at space x, is the fusion rule of the mechanical part material parameters and the quantum field, is the k-th component of the quantum field at space x, is the k-th component corresponding to the i-th mechanical part material parameter matrix sample, where i is the mechanical part material parameter matrix sample number and k is the quantum field component number corresponding to the mechanical part material parameter.

[0026] (3) Fractional-dimensional convolution feature extraction module:

[0027] As a further solution, perform fractional-dimensional convolution feature extraction on the mechanical part quantum state data and perform quantum state decoding to obtain the fractal features of the mechanical part. The specific analysis process is as follows: Obtain the convolution kernel , and initialize it using the Mittag-Leffler function; Obtain the fractional order stored in the database;

[0028] Fractional-order differentiation:

[0029] ; (6)

[0030] In the formula, is the result function of the fractional-order differential operation, is the summation index, is the gamma function, is the space interval, is the quantum state data;

[0031] Obtain the initial fractal features, that is, the initial features of the fractal pyramid;

[0032] Fractal pyramid construction:

[0033] ; (7)

[0034] In the formula, is the feature of the th layer of the fractal pyramid, is obtained by performing an operation on and the th convolution kernel , and after transformation, it is the feature of the next layer of the fractal pyramid. N is the number of convolution kernels participating in the calculation in the fractal pyramid construction. is the transformation operator related to the Hausdorff dimension , is the convolution kernel number; represents the convolution operation;

[0035] Obtain the feature of the last layer of the fractal pyramid ; Perform quantum state decoding on the feature of the last layer of the fractal pyramid , decompose the complex feature into physical quantities, and obtain the fractal feature of the mechanical part .

[0036] As a further solution, the quantum state decoding, the specific analysis process is: decompose the complex feature into physical quantities:

[0037] ; (8)

[0038] In the formula, is the real part encoded geometric feature, is the real part, is the imaginary part quantum phase feature, is the imaginary part; obtain the fractal feature of the mechanical part , .

[0039] (4) Mechanical part defect manifold coordinate extraction module:

[0040] As a further solution, construct a topological manifold detection network, calculate the extreme points of the manifold, and extract the mechanical part defect manifold coordinates from the fractal features of the mechanical part. The specific analysis process is: generate the mechanical part defect manifold coordinates of multiple sensors based on X-ray, ultrasonic, and infrared modalities; obtain the weights of the fully connected layer stored in the database, and output the p-direction metric tensor manifold space weight , the p-direction metric tensor manifold space bias ;

[0041] Obtain the initial metric :

[0042] ; (9)

[0043] Among them, is the fractal feature of the mechanical part, is the feature covariance matrix, p is the component number of the metric tensor in the p direction, and q is the component number of the metric tensor in the q direction;

[0044] Construct the differentiable manifold equation:

[0045] ; (10)

[0046] Among them, is the coordinate component in the metric tensor in the p direction, and e is the natural constant;

[0047] Improve the manifold shape through Ricci flow:

[0048] ; (11)

[0049] Among them, represents the Gaussian curvature of the manifold, are the metric tensors in the p and q directions, and t is the time index;

[0050] When the Gaussian curvature stop the iteration; calculate the extreme point m of the manifold. The extreme point of the manifold corresponds to the crack tip or the hole center, and record the coordinates of the extreme point of the manifold as the manifold coordinates of the mechanical part defect.

[0051] As a further solution, calculate the extreme point of the manifold. The specific analysis process is as follows:

[0052] ; (12)

[0053] Among them, represents the set obtained by calculating the points where the gradient of the Gaussian curvature on the manifold with respect to the coordinate x is zero, denoted as the set of extreme points of the manifold;

[0054] Obtain the coordinates of the extreme point m of the manifold as ; Among them, is the x-axis coordinate of the extreme point of the manifold, is the y-axis coordinate of the extreme point of the manifold, is the z-axis coordinate of the extreme point of the manifold, is the defect depth of the extreme point of the manifold.

[0055] (5) Manifold commutator calculation module:

[0056] As a further solution, based on the manifold coordinates of mechanical part defects, calculate the manifold commutator. The specific analysis process is as follows: The manifold coordinates of mechanical part defects specifically include the manifold coordinates of X-ray mechanical part defects , the manifold coordinates of ultrasonic mechanical part defects , the manifold coordinates of infrared modal mechanical part defects ; Obtain the Dirac matrices defined by the standard Clifford algebra stored in the database;

[0057] Encode the manifold coordinates of mechanical part defects as Clifford numbers to generate the Dirac matrix basis, satisfying:

[0058] ; (13)

[0059] Among them, is the Dirac matrix basis of the th dimension, is the Dirac matrix basis of the th dimension, is the Minkowski metric, is the identity matrix, is the dimension number, is the dimension number, and the dimensions are the X-ray dimension, the ultrasonic dimension, and the infrared modal dimension;

[0060] Calculate the manifold commutator , determine the defect dominant direction. A non-zero manifold commutator reveals multimodal data conflicts. The larger the modulus length of the manifold commutator value, the more significant the conflict: Obtain the mapping set of the defect dominant direction - defect weight factor set stored in the database, and based on the determined defect dominant direction, determine the matching defect weight factor set.

[0061] As a further solution, calculate the manifold commutator. The specific analysis process is as follows:

[0062] ; (14)

[0063] In the formula, is the imaginary unit, is the three-dimensional Levi-Civita symbol, is the Dirac matrix basis of the Kth dimension, is the coordinate component of the R direction of the th dimension, is the coordinate component of the r direction of the th dimension, and K is the dimension number;

[0064] If In a three-dimensional spatial orientation even permutation, the three-dimensional Levi-Civita symbol is equal to 1; if In a three-dimensional spatial orientation odd permutation, the three-dimensional Levi-Civita symbol is equal to -1; if In the three-dimensional space, if any two directions are the same, the three-dimensional Levi-Civita symbol is equal to 0.

[0065] (6) Mechanical part defect confidence output module:

[0066] As a further solution, perform non-commutative geometry decision fusion to output the mechanical part defect fusion confidence. The specific analysis process is as follows:

[0067] Perform non-commutative geometry decision fusion, and obtain the mechanical part defect confidence through eigenvalue decomposition:

[0068] ; (15)

[0069] Among them, is the mechanical part defect confidence, is the operation of taking the real part, is the trace operation, is the chiral projection matrix, is the operation of projecting the confidence onto the chiral subspace;

[0070] ; (16)

[0071] ; (17)

[0072] Among them, is the Pauli matrix of the th dimension, is the imaginary unit;

[0073] Based on each mechanical part defect confidence and combined with the matching defect weight factor set, perform weighted summation to obtain the matching mechanical part defect fusion confidence ; Output the mechanical part defect fusion confidence .

[0074] Compared with the prior art, the beneficial effects of the present invention:

[0075] (1) By providing an intelligent detection system for mechanical part defects based on deep learning, the fractional-dimensional convolution feature extraction module uses fractional-dimensional convolution for feature extraction and quantum state decoding. Fractional-dimensional convolution can capture more subtle and complex feature relationships in data. Compared with traditional integer-dimensional convolution, it can extract fractal features at a finer scale, dig deeper into features such as the microstructure of mechanical parts, and improve the feature representation ability. The mechanical part defect manifold coordinate extraction module constructs a topological manifold detection network to calculate the manifold extreme points for extracting defect manifold coordinates. The concept of manifold can describe the internal structure of data from a geometric perspective. By finding the extreme points to determine the defect positions, it can accurately locate the defect positions such as the crack tips and hole centers of mechanical parts, providing accurate position information for subsequent repair and evaluation.

[0076] (2) Through the mechanical part defect confidence output module, the manifold commutator is calculated based on the defect manifold coordinates and non-commutative geometric decision fusion is performed to output the confidence. Non-commutative geometric decision fusion synthesizes various geometric information and decision factors, and can more comprehensively and reliably evaluate the possibility of mechanical parts having defects, providing a more credible basis for quality control and decision-making. With the functions of each module implemented based on deep learning, the entire system has intelligent processing capabilities, automatically running throughout the process from data acquisition, feature extraction to defect location and confidence evaluation, reducing manual intervention, improving detection efficiency and consistency, and meeting the needs of large-scale rapid detection. Brief Description of the Drawings

[0077] Figure 1 It is a schematic diagram of the system module connection of the present invention.

[0078] Figure 2 It is a flowchart of the steps for determining the set of matching defect weight factors.

[0079] Figure 3 It is a flowchart of the steps of the present invention. Detailed Embodiments

[0080] The embodiments of the present invention will be described clearly and completely below with reference to the drawings.

[0081] Please refer to Figure 1 , the embodiments of the present invention provide an intelligent detection technical solution for mechanical part defects based on deep learning: As Figure 3 shown, the intelligent detection system for mechanical part defects based on deep learning includes a parameter matrix acquisition module, a mechanical part quantum state data output module, a fractional-dimensional convolution feature extraction module, a mechanical part defect manifold coordinate extraction module, a manifold commutator calculation module, and a mechanical part defect confidence output module.

[0082] The parameter matrix acquisition module acquires the material parameter matrix of the mechanical part;

[0083] Mechanical part quantum state data output module, based on a multi-modal generation model excited by a quantum field, outputs mechanical part quantum state data.

[0084] (1) Parameter matrix acquisition module:

[0085] The specific analysis process includes: obtaining the mechanical part material parameters, specifically including the elastic modulus of the mechanical part , Poisson's ratio of the mechanical part , density of the mechanical part , yield strength of the mechanical part , fracture toughness of the mechanical part , melting point of the mechanical part ; constructing a mechanical part material parameter matrix :

[0086] ; (1)

[0087] Combining the mechanical part material parameters with the quantum field, each material parameter controls one component of the quantum field;

[0088] (2) Mechanical part quantum state data output module:

[0089] Obtaining the weights of the generation network stored in the database , generating a quantum state through the path integral of quantum field theory :

[0090] ; (2)

[0091] In the formula, is the quantum state function obtained through path integral, describing the quantum state of the system at position x in space, is the normalization constant, is the functional integral measure of the quantum field , is the action of the quantum field stored in the database, is the generation network, and e is the natural constant;

[0092] The quantum state needs to satisfy the material mechanics equation:

[0093] ; (3)

[0094] In the formula, is the displacement field, is the Lame constant stored in the database, is the field nonlinear coefficient stored in the database, is the displacement field with respect to time 's second derivative, that is, acceleration;​​​​

[0095] Finally, quantum state data is obtained through projective measurement :

[0096] ; (4)

[0097] In the formula, is three-dimensional complex-valued data obtained through quantum measurement, where a, b, and c are indices in three-dimensional space is the quantum state of the conjugate transpose is the measurement operator for a, b, and c is the trace of the matrix, a represents the surface topography, b represents the internal stress, and c represents the crack depth

[0098] Output quantum state data .

[0099] Combining the material parameters of mechanical parts with the quantum field, the specific analysis process is as follows

[0100] ; (5)

[0101] In the formula, is the quantum field at position x in space is the fusion rule of the material parameters of mechanical parts and the quantum field is the k-th component of the quantum field at position x in space is the k-th component corresponding to the i-th mechanical part material parameter matrix sample, where i is the mechanical part material parameter matrix sample number and k is the quantum field component number corresponding to the mechanical part material parameter

[0102] Obtain various mechanical part material parameters such as elastic modulus and Poisson's ratio to construct a matrix, comprehensively covering the information of material physical properties, providing a rich data basis for subsequent analysis, being able to more completely characterize the attributes of mechanical parts, and avoiding analysis deviations caused by information loss

[0103] Combining the material parameters with the quantum field, each parameter controls the quantum field component, and using the characteristics of the quantum field to explore the potential connections between material parameters, fusing data from a new perspective, and being able to obtain deeper and more unique data features and information compared with traditional methods

[0104] Based on the multi-modal generation model excited by the quantum field, combining the path integral of quantum field theory to generate quantum states, modeling from the microscopic quantum level, conforming to modern physical theory, making the model's description of the mechanical part state more scientific and accurate, and being able to capture physical phenomena and laws that are difficult to reach by traditional methods

[0105] Three-dimensional complex quantum state data are obtained through projection measurement, corresponding to surface topography, internal stress, crack depth, etc., providing mechanical part status information from multiple dimensions, more comprehensively reflecting the part condition than single-dimensional detection, and facilitating the comprehensive evaluation of part quality.

[0106] (3) Fractional-dimensional convolution feature extraction module

[0107] The fractional-dimensional convolution feature extraction module performs fractional-dimensional convolution feature extraction on the quantum state data of mechanical parts and decodes the quantum state to obtain the fractal features of mechanical parts.

[0108] The specific analysis process is as follows: Obtain the convolution kernel , and initialize it using the Mittag-Leffler function; Obtain the fractional order stored in the database;

[0109] Fractional order differentiation:

[0110] ; (6)

[0111] In the formula, is the result function of the fractional order differential operation, is the summation index, is the gamma function, is the space interval, is the quantum state data;

[0112] Obtain the initial fractal features, that is, the initial features of the fractal pyramid;

[0113] Fractal pyramid construction:

[0114] ; (7)

[0115] In the formula, is the feature of the layer of the fractal pyramid, is obtained by performing an operation on and the th convolution kernel , and after being transformed by to obtain the feature of the next layer of the fractal pyramid. N is the number of convolution kernels participating in the calculation in the fractal pyramid construction, is the transformation operator related to the Hausdorff dimension , is the convolution kernel number;

[0116] Obtain the feature of the last layer of the fractal pyramid ; For the feature of the last layer of the fractal pyramid Perform quantum state decoding, decompose the complex number feature into physical quantities, and obtain the fractal feature of mechanical parts .

[0117] Quantum state decoding, the specific analysis process is: decompose the complex number feature into physical quantities:

[0118] ; (8)

[0119] In the formula, is the real part encoded geometric feature, is the real part, is the imaginary part quantum phase feature, is the imaginary part; obtain the fractal feature of mechanical parts , .

[0120] Fractional dimension convolution and fractional order differentiation break the traditional integer order limit and can capture complex and subtle features in quantum state data more precisely. Quantum state data contains microscopic information of mechanical parts. Fractional order operations can mine features at different scales, such as the irregularity of the microscopic structure of materials, etc., improving the depth and breadth of feature extraction.

[0121] The construction of the fractal pyramid is based on fractal theory and has the characteristics of self-similarity and multi-scale. By iteratively using different convolution kernel operations and combining with the related transformation of the Hausdorff dimension, the features can be analyzed and fused at different scales, comprehensively describing the features of mechanical parts from macro to micro and meeting the analysis requirements of different hierarchical structures of parts.

[0122] Initialize the convolution kernel using the Mittag-Leffler function, introduce the concept of fractional order, and apply the advanced mathematical theory to the construction of the feature extraction model. Compared with the traditional convolution method, new characteristics and capabilities are given to the model. For the complex characteristics of the quantum state data of mechanical parts, quantum state decoding obtains geometric features and quantum phase features by separating the real part and the imaginary part respectively, which fits the essence of the data, enables the model to effectively process and utilize quantum state data, and improves the pertinence and effectiveness of feature extraction.

[0123] The finally obtained fractal feature F contains the real part geometric feature and the imaginary part quantum phase feature. This clear feature representation method provides intuitive and physically meaningful feature information for subsequent defect detection and quality assessment of mechanical parts.

[0124] (4) Mechanical part defect manifold coordinate extraction module:

[0125] The mechanical part defect manifold coordinate extraction module constructs a topological manifold detection network, calculates the extreme points of the manifold, and extracts the mechanical part defect manifold coordinates from the fractal features of the mechanical parts.

[0126] The specific analysis process is as follows: Generate the defect manifold coordinates of mechanical parts for multi-sensors based on X-ray, ultrasonic, and infrared modalities; obtain the weights of the fully connected layer stored in the database , and output the weights of the p-direction metric tensor manifold space , and the bias of the p-direction metric tensor manifold space ;

[0127] Obtain the initial metric :

[0128] ; (9)

[0129] Among them, is the fractal feature of the mechanical part, is the feature covariance matrix, p is the component number of the p-direction metric tensor, and q is the component number of the q-direction metric tensor;

[0130] Construct a differentiable manifold equation:

[0131] ; (10)

[0132] Among them, is the coordinate component in the p-direction metric tensor;

[0133] Improve the manifold shape through Ricci flow:

[0134] ; (11)

[0135] Among them, represents the Gaussian curvature of the manifold, are the p- and q-direction metric tensors, and t is the time index;

[0136] When the Gaussian curvature , stop the iteration; calculate the extreme points m of the manifold. The extreme points of the manifold correspond to the crack tip or the hole center, and record the coordinates of the extreme points of the manifold as the defect manifold coordinates of the mechanical part.

[0137] Calculate the extreme points of the manifold. The specific analysis process is as follows:

[0138] ; (12)

[0139] Among them, represents the set obtained by calculating the points where the gradient of the Gaussian curvature on the manifold with respect to the coordinate x is zero, denoted as the set of extreme points of the manifold;

[0140] Obtain the coordinates of the extreme point m of the manifold as ; Among them, is the x-axis coordinate of the extreme point of the manifold, is the y-axis coordinate of the extreme point of the manifold, is the z-axis coordinate of the manifold extreme point, is the defect depth of the manifold extreme point.

[0141] Generate the manifold coordinates of mechanical part defects based on multi-modalities such as X-ray, ultrasound, and infrared. Different modalities can perceive the part state from multiple angles. X-ray can detect internal structure defects, ultrasound is sensitive to cracks, etc., and infrared can detect defects reflected by temperature anomalies. Multi-modal fusion can integrate the advantages of each modality, reduce the limitations and misjudgments of single-modal detection, and obtain defect information comprehensively and accurately.

[0142] Construct a topological manifold detection network, and use the concept of manifold to describe the features and defect positions of mechanical parts from a geometric perspective. Manifold can depict the internal structure of data, map the fractal features of mechanical parts to the manifold space, discover defects from the changes in geometric structure, provide a new perspective and theoretical framework for defect detection, and more essentially reveal the relationship between defects and part structure.

[0143] Use the weights of the fully connected layer to output the weights and biases in the manifold space, and obtain the initial metric based on the feature covariance matrix. The fully connected layer can effectively transform features and determine the parameters of the manifold space; the initial metric provides a basis for the geometric structure of the manifold space, and reasonable parameter and metric settings enable the model to more accurately describe the manifold characteristics and lay a good foundation for subsequent calculations.

[0144] Construct a differentiable manifold equation and improve the manifold shape through Ricci flow. The differentiable manifold equation defines the mathematical form of the manifold, and Ricci flow dynamically adjusts the manifold according to the Gaussian curvature, making the manifold shape approximate the real part structure and defect distribution. When the Gaussian curvature satisfies certain conditions, stop the iteration to ensure the optimization convergence of the manifold shape and accurately reflect the defect position and characteristics.

[0145] Calculate the extreme points of the manifold to determine the defect positions. The extreme points correspond to key defect parts such as crack tips or hole centers. This method can accurately capture the defect positions, output information including spatial coordinates and defect depths, provide an accurate position basis for mechanical part repair, quality assessment, etc., and is more accurate in positioning than traditional methods.

[0146] Manifold commutator calculation module, based on the manifold coordinates of mechanical part defects, calculate the manifold commutator.

[0147] Mechanical part defect confidence output module, perform non-commutative geometric decision fusion, and output the fused confidence of mechanical part defects.

[0148] (5) Manifold commutator calculation module:

[0149] As Figure 2 shown, determine the set of matching defect weight factors. The specific analysis process is as follows: The manifold coordinates of mechanical part defects specifically include the manifold coordinates of X-ray mechanical part defects 、Ultrasonic mechanical part defect manifold coordinates 、Infrared modal mechanical part defect manifold coordinates ; Obtain the Dirac matrices defined by the standard Clifford algebra stored in the database;

[0150] Encode the mechanical part defect manifold coordinates as Clifford numbers to generate a Dirac matrix basis, satisfying:

[0151] ; (13)

[0152] Wherein, is the Dirac matrix basis of the th dimension, is the Dirac matrix basis of the th dimension, is the Minkowski metric, is the identity matrix, is the dimension number, is the dimension number, and the dimensions are the X-ray dimension, the ultrasonic dimension, and the infrared modal dimension;

[0153] Calculate the manifold commutator , determine the defect dominant direction. A non-zero manifold commutator reveals multimodal data conflicts. The greater the modulus length of the manifold commutator value, the more significant the conflict: Obtain the defect dominant direction - defect weight factor set mapping set stored in the database, and determine the matching defect weight factor set based on the determined defect dominant direction.

[0154] For the manifold commutator, the specific analysis process is:

[0155] ; (14)

[0156] In the formula, is the imaginary unit, is the three-dimensional Levi-Civita symbol, is the Dirac matrix basis of the Kth dimension, is the coordinate component of the R direction of the th dimension, is the coordinate component of the r direction of the th dimension, and K is the dimension number;

[0157] If is arranged in an even permutation in the three-dimensional space direction, the three-dimensional Levi-Civita symbol is equal to 1; if is arranged in an odd permutation in the three-dimensional space direction, the three-dimensional Levi-Civita symbol is equal to -1; if In three-dimensional space, if any two directions are the same, the three-dimensional Levi-Civita symbol is equal to 0.

[0158] (6) Mechanical part defect confidence output module:

[0159] Perform non-commutative geometric decision fusion and output the fused confidence of mechanical part defects. The specific analysis process is as follows:

[0160] Perform non-commutative geometric decision fusion and obtain the confidence of mechanical part defects through eigenvalue decomposition:

[0161] ; (15)

[0162] Where, is the confidence of mechanical part defects, is the operation of taking the real part, is the trace operation, is the chiral projection matrix, is the operation of projecting the confidence onto the chiral subspace;

[0163] ; (16)

[0164] ; (17)

[0165] Where, is the Pauli matrix of the -th dimension, is the imaginary unit;

[0166] Based on the confidence of each mechanical part defect and combined with the matching defect weight factor set, perform weighted summation to obtain the fused confidence of the matching mechanical part defects ; Output the fused confidence of mechanical part defects .

[0167] Synthesize the defect manifold coordinates of multiple modalities such as X-ray, ultrasound, and infrared, and comprehensively consider the information of parts under different detection methods. Different modalities have different sensitivities and emphases on defects. After fusion, the defect characteristics can be captured more completely, reducing the one-sidedness of single-modal detection and improving the accuracy of defect judgment.

[0168] Calculate the manifold commutator to determine the dominant direction of defects and can detect conflicts in multi-modal data. A non-zero commutator indicates that there are differences in the data, and the magnitude of its value reflects the significance of the conflict, which helps to analyze the contradictions between different detection data, clarify the performance of defects in each dimension and the main influencing direction, and provide key information for accurately evaluating defects.

[0169] Using Dirac matrices defined by standard Clifford algebra, the coordinates of the defective manifold are encoded as a matrix basis for generating Clifford numbers. Clifford algebra has a powerful descriptive ability in fields such as geometry and physics. It can uniformly process multi-modal data from both algebraic and geometric perspectives, providing a more abstract and general mathematical framework for defect analysis, enhancing the depth and theoretical nature of the analysis.

[0170] Confidence is obtained through non-commutative geometric decision fusion. Non-commutative geometry has unique advantages in dealing with complex space and data relationships. It can comprehensively consider the non-commutative characteristics and geometric structures of multi-modal data, fuse information from a more fundamental level, and can more reasonably integrate multi-source data compared to traditional methods, making the confidence calculation more scientific.

[0171] The defect confidence is obtained through calculations such as eigenvalue decomposition to quantify the possibility of defects in the parts. The specific numerical values facilitate the operator to intuitively judge the quality status of the parts, providing a clear basis for whether the parts are qualified and whether further inspection or repair is required.

[0172] If the confidence of the mechanical part defect fusion is higher than the threshold of the mechanical part defect fusion stored in the database, and the defect depth of the manifold extreme point is higher than the threshold of the defect depth of the manifold extreme point stored in the database, it indicates that the mechanical part defect is serious, and the detected mechanical part defect is highly credible, and further inspection and repair are required.

Claims

1. An intelligent defect detection system for mechanical parts based on deep learning, characterized in that, Including: Parameter matrix acquisition module: Acquire the material parameter matrix of mechanical parts; Quantum state data output module of mechanical parts: Based on the multi-modal generation model excited by the quantum field, output the quantum state data of mechanical parts; Fractional-dimensional convolution feature extraction module: Perform fractional-dimensional convolution feature extraction on the quantum state data of mechanical parts and perform quantum state decoding to obtain the fractal features of mechanical parts; Defect manifold coordinate extraction module of mechanical parts: Construct a topological manifold detection network, calculate the extreme points of the manifold, and extract the defect manifold coordinates of mechanical parts from the fractal features of mechanical parts; Manifold commutator calculation module: Calculate the manifold commutator based on the defect manifold coordinates of mechanical parts; Defect confidence output module of mechanical parts: Perform non-commutative geometric decision fusion and output the defect fusion confidence of mechanical parts.

2. The intelligent defect detection system for mechanical parts based on deep learning according to claim 1, wherein: The multi-modal generation model excited by the quantum field outputs the quantum state data of mechanical parts. The specific analysis process includes: (1) Acquire the material parameter matrix of mechanical parts: Obtain the material parameters of mechanical parts, specifically including the elastic modulus of mechanical parts , the Poisson's ratio of mechanical parts , the density of mechanical parts , the yield strength of mechanical parts , the fracture toughness of mechanical parts , the melting point of mechanical parts ; Construct a material parameter matrix for mechanical parts : ；（1） Combine the material parameters of mechanical parts with the quantum field, and each material parameter controls one component of the quantum field; (2) Multi-modal generation model excited by the quantum field: Obtain the generated network weights stored in the database , generate quantum states through quantum field theory path integral : ；（2） In the formula, is the quantum state function obtained through path integral, describing the quantum state of the system at space x, is the normalization constant, is the functional integral measure of the quantum field ; is the action of the quantum field stored in the database, is the generation network, and e is the natural constant; represents matrix multiplication; refers to the combination of the mechanical part material parameter matrix and the quantum field; The quantum state needs to satisfy the material mechanics equation: ；（3） In the formula, is the displacement field, is the Lame constant stored in the database, is the field nonlinearity coefficient stored in the database, is the displacement field with respect to time second derivative, i.e., acceleration; (3) Output the quantum state data of mechanical parts: Finally, quantum state data is obtained through projective measurement : ；（4） wherein, is three-dimensional complex-valued data obtained by quantum measurement, a, b, c are indices in three-dimensional space, is the quantum state is the conjugate transpose of, is the measurement operator for a, b, c, is the trace of the matrix, a represents the surface topography, b represents the internal stress, and c represents the crack depth; Output quantum state data .

3. The intelligent mechanical part defect detection system based on deep learning according to claim 2, wherein: The combination of the material parameters of mechanical parts and the quantum field, the specific analysis process is: ；（5） wherein, is the quantum field at space x, is the fusion rule of the mechanical part material parameters and the quantum field, is the k-th component of the quantum field at space x, is the k-th component corresponding to the i-th mechanical part material parameter matrix sample, i is the mechanical part material parameter matrix sample number, and k is the quantum field component number corresponding to the mechanical part material parameter.

4. The intelligent detection system for mechanical part defects based on deep learning according to claim 1, wherein: The fractional-dimensional convolution feature extraction of the quantum state data of mechanical parts and the quantum state decoding to obtain the fractal features of mechanical parts. The specific analysis process is: (1) Fractional-dimensional convolution feature extraction: Obtain the convolutional kernel , and initialize it using the Mittag-Leffler function; Obtain the fractional order stored in the database ; Fractional-order differentiation: ；（6） In the formula, is the result function of the fractional-order differential operation, is the summation index, is the gamma function, is the spatial interval, is the quantum state data; (2) Fractal pyramid construction: Obtain the initial fractal features, that is, the initial features of the fractal pyramid; Fractal pyramid construction: ；（7） In the formula, is the feature of the layer of the fractal pyramid, is obtained by performing an operation on and the th convolution kernel , and is the feature of the next layer of the fractal pyramid obtained after transformation. N is the number of convolution kernels participating in the calculation in the construction of the fractal pyramid. is a transformation operator related to the Hausdorff dimension . is the convolution kernel number; represents the convolution operation; Obtain the features of the last layer of the fractal pyramid ; Perform quantum state decoding on the features of the last layer of the fractal pyramid, decompose the complex features into physical quantities, and obtain the fractal features of mechanical parts .​ 5. The intelligent detection system for mechanical part defects based on deep learning according to claim 4, characterized in that: The quantum state decoding, the specific analysis process is: Decompose the complex number features into physical quantities: ；（8） In the formula, is the real part coding geometric feature, is the real part, is the imaginary part quantum phase feature, is the imaginary part; Obtain the fractal features of mechanical parts , .

6. The intelligent defect detection system for mechanical parts based on deep learning according to claim 1, wherein: The construction of the topological manifold detection network, calculating the extreme points of the manifold, and extracting the defect manifold coordinates of mechanical parts from the fractal features of mechanical parts. The specific analysis process is: (1) Construct a topological manifold monitoring network: Generate the defect manifold coordinates of multi-sensors of mechanical parts based on X-ray, ultrasonic, and infrared modalities; Obtain the fully connected layer weights stored in the database , and output the weights of the p-direction metric tensor manifold space , the bias of the p-direction metric tensor manifold space ; Obtain initial metrics : ； （9） Among them, is the fractal feature of mechanical parts, is the feature covariance matrix, p is the component number of the metric tensor in the p direction, and q is the component number of the metric tensor in the q direction; Construct a differentiable manifold equation: ；（10） Among them, is the coordinate component in the p-direction metric tensor, and e is the natural constant; (2) Manifold optimization and defect localization: Improve the manifold shape through Ricci flow: ；（11） wherein, represents the Gaussian curvature of the manifold, is the metric tensor in the p and q directions, and t is the time index; Stop the iteration when the Gaussian curvature is reached; Calculate the extreme point m of the manifold. The extreme point of the manifold corresponds to the crack tip or the hole center, and record the coordinates of the extreme point of the manifold as the defect manifold coordinates of mechanical parts.

7. The intelligent detection system for mechanical part defects based on deep learning according to claim 6, characterized in that: The calculation of the extreme points of the manifold, the specific analysis process is: ；（12） Among them, represents the set obtained by calculating the points where the gradient of the Gaussian curvature on the manifold with respect to the coordinate x is zero, denoted as the set of manifold extreme points; Obtain the coordinates of the extreme point m of the manifold as ; Among them, is the x-axis coordinate of the manifold extreme point, is the y-axis coordinate of the manifold extreme point, is the z-axis coordinate of the manifold extreme point, is the defect depth of the manifold extreme point.

8. The intelligent defect detection system for mechanical parts based on deep learning according to claim 1, characterized in that: The calculation of the manifold commutator based on the defect manifold coordinates of mechanical parts. The specific analysis process is: The manifold coordinates of mechanical part defects specifically include the X-ray manifold coordinates of mechanical part defects , the ultrasonic manifold coordinates of mechanical part defects , the infrared modal manifold coordinates of mechanical part defects ; Obtain the Dirac matrix defined by the standard Clifford algebra stored in the database; Encode the defect manifold coordinates of mechanical parts as Clifford numbers to generate the Dirac matrix basis, satisfying: ； （13） Among them, is the Dirac matrix basis of the th dimension, is the Dirac matrix basis of the th dimension, is the Minkowski metric, is the identity matrix, is the dimension number, is the dimension number, and the dimensions are the X-ray dimension, the ultrasonic dimension, and the infrared mode dimension; Computational manifold commutator , determine the defect-dominant direction. A non-zero manifold commutator reveals multimodal data conflicts. The magnitude of the manifold commutator value is larger, and the conflict is more significant: Obtain the mapping set of the defect dominant direction - defect weight factor set stored in the database, and determine the matching defect weight factor set based on the determined defect dominant direction.

9. The intelligent mechanical part defect detection system based on deep learning according to claim 8, characterized in that: The calculation of the manifold commutator, the specific analysis process is: ； （14） wherein, is the imaginary unit, is the three-dimensional Levi-Civita symbol, is the Dirac matrix basis in the K-th dimension, is the coordinate component in the R direction in the -th dimension, is the coordinate component in the r direction in the -th dimension, and K is the dimension number; If is arranged in an even permutation in the three-dimensional spatial direction, the three-dimensional Levi-Civita symbol is equal to 1; If is an odd permutation in the three-dimensional spatial directions, the three-dimensional Levi-Civita symbol is equal to -1; If If any two of the directions in the three-dimensional space directions are the same, the three-dimensional Levi-Civita symbol is equal to 0.

10. The intelligent mechanical part defect detection system based on deep learning according to claim 8, characterized in that: The non-commutative geometric decision fusion is performed to output the defect fusion confidence of mechanical parts. The specific analysis process is: (1) Perform non-commutative geometric decision fusion: Perform non-commutative geometric decision fusion and obtain the confidence of mechanical part defects through eigenvalue decomposition: ； （15） Among them, is the confidence level of mechanical part defects, is the operation of taking the real part, is the trace operation, is the chirality projection matrix, is the operation of projecting the confidence level onto the chiral subspace; ；（16） ；（17） Among them, is the Pauli matrix of the th dimension, is the imaginary unit; (2)Output the fused confidence of mechanical part defects: Based on the defect confidence levels of each mechanical part and combined with the matching set of defect weight factors, perform a weighted sum to obtain the fused confidence level of the matching mechanical part defects .

Images