A steel fiber distribution quality judgment method and system based on orientation similarity reasoning

Abstract

This invention discloses a method and system for determining the quality of steel fiber distribution based on orientation similarity reasoning, belonging to the field of artificial intelligence reasoning technology. This method for determining the quality of steel fiber distribution based on orientation similarity reasoning acquires steel fiber orientation data corresponding to multiple spatial regions, decouples them by axial orientation state to construct spatial orientation states, performs cross-regional orientation relationship mapping to construct initial orientation similarity relationships, inputs both into a pre-trained orientation similarity reasoning network, obtains enhanced spatial orientation state features through association propagation, and uses an orientation similarity adaptive update module to calculate adjustment amounts and dynamically update the initial orientation similarity relationships to obtain inferred enhanced orientation similarity relationships. This invention maps and outputs the steel fiber distribution quality state through a quality state reasoning module, thereby improving the matching degree between the steel fiber distribution quality state determination result and the actual spatial distribution state.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence reasoning technology, specifically to a method and system for determining the distribution quality of steel fibers based on orientation similarity reasoning. Background Technology

[0002] In complex structural systems with spatial distribution characteristics, the spatial orientation distribution of internal components has a decisive influence on their macroscopic performance. With the maturity of high-precision tomography and three-dimensional digital reconstruction technologies, obtaining the position coordinates and orientation vectors of each spatial unit inside the structure has become a routine technical means, which provides a data foundation for quality judgment based on spatial relationship analysis.

[0003] For example, in the quality evaluation of steel fiber reinforced concrete components, based on the obtained steel fiber-related data, the different spatial regions inside the component are analyzed to determine the distribution state of the steel fibers.

[0004] Based on the above solutions, the limitations of the existing technology include at least the following problems: During the molding process of steel fiber reinforced concrete components, the steel fibers are affected by the flow direction of the slurry, changes in the local flow field, and spatial constraints. Their spatial orientation will form a distribution pattern with regional correlation along the pouring process. The existing technology usually describes the distribution state based on the statistical results of steel fiber orientation, and it mainly establishes the evaluation basis based on the orientation parameters of the overall area or local area.

[0005] However, in the actual forming process, changes in the distribution quality of steel fibers are often accompanied by changes in the orientation relationship between spatial regions. For example, there are phenomena such as a decrease in the consistency of orientation between adjacent regions, deviation of local orientation trends, and differences in the orientation change patterns between regions. Such changes are not simply manifested as changes in the number of fibers or changes in the average angle, but are hidden in the process of changes in the orientation relationship between multiple spatial regions. This makes it difficult for existing technologies to fully reflect the impact of changes in the orientation relationship between different spatial regions on the overall distribution state. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for determining the distribution quality of steel fibers based on orientation similarity reasoning, which solves the problem that existing methods are unable to characterize the impact of changes in orientation correlation between spatial regions on the overall distribution state.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for determining the quality of steel fiber distribution based on orientation similarity inference, comprising the following steps: acquiring steel fiber orientation data corresponding to multiple spatial regions within the target analysis object, and decoupling them through axial orientation state to construct the spatial orientation state of the corresponding spatial region; performing cross-regional orientation relationship mapping on the spatial orientation state to construct an initial orientation similarity relationship; inputting the spatial orientation state and the initial orientation similarity relationship into a pre-trained orientation similarity inference network; propagating the spatial orientation state through an orientation similarity propagation module based on the initial orientation similarity relationship to obtain enhanced spatial orientation state features; calculating the orientation similarity adjustment amount based on the enhanced spatial orientation state features and the initial orientation similarity relationship in an orientation similarity adaptive update module, and dynamically updating the initial orientation similarity relationship to obtain an inferred enhanced orientation similarity relationship; inputting the inferred enhanced orientation similarity relationship and the enhanced spatial orientation state features into a quality state inference module to output the steel fiber distribution quality state.

[0008] Further, the specific steps for constructing the spatial orientation state of each spatial region are as follows: Based on the steel fiber orientation data, obtain the axial direction angle of each steel fiber in each spatial region; perform axial orientation convergence processing on each steel fiber axial direction angle to determine the principal orientation angle of each spatial region; extract the angular change between the principal orientation angle of each spatial region and the principal orientation angle of the previous spatial region along the casting direction, and use it as the orientation evolution change; determine the orientation discrete value of each spatial region based on the deviation statistical aggregation processing of each steel fiber axial direction angle and the principal orientation angle; combine the principal orientation angle, orientation evolution change, and orientation discrete value into the spatial orientation state.

[0009] Furthermore, the specific steps for constructing the initial orientation similarity relationship are as follows: For the principal orientation angles of different spatial regions, the directional orientation similarity between different spatial regions is extracted through axial angle relationship transformation; for the orientation evolution changes of different spatial regions, the orientation evolution similarity is determined through evolution trend homology analysis; the spatial positions of different spatial regions are read from the steel fiber orientation data, and the spatial correlation is determined through spatial distance attenuation mapping; the directional orientation similarity, orientation evolution similarity, and spatial correlation are correlated and aggregated to obtain the initial orientation similarity between different spatial regions, and combined into the initial orientation similarity relationship.

[0010] Furthermore, the specific steps for obtaining enhanced spatial orientation state features are as follows: Based on the initial orientation similarity relationship, other spatial regions that meet the orientation association conditions with each spatial region are selected to form a spatial orientation state association range; based on the spatial orientation state association range, the orientation state association propagation weight is determined according to the initial orientation similarity; based on the state propagation weight, the spatial orientation states corresponding to other spatial regions within the spatial orientation state association range are selectively associated; based on the selective association results corresponding to each spatial region, the spatial orientation states are jointly updated to obtain enhanced spatial orientation state features.

[0011] Furthermore, the specific steps to obtain the reasoning enhancement orientation similarity relationship are as follows:

[0012] The initial orientation similarity between different spatial regions is used as the current relation weight; the current state similarity is determined based on the enhanced spatial orientation state features of any two spatial regions; the difference between the current state similarity and the current relation weight is analyzed to generate orientation similarity residuals; the current relation weights are adjusted according to the orientation similarity residuals to obtain the adjusted relation weights; the adjusted relation weights are subjected to symmetric constraints, and the relation weights after symmetric constraints are used as the input for the next propagation; after iterative propagation and adjustment, the finally obtained relation weights are used as the reasoning enhanced orientation similarity relation.

[0013] Furthermore, the specific steps of symmetric constraint processing are as follows: obtain the first-direction relation weight and the second-direction relation weight between different spatial regions in the adjusted relation weights; determine whether there is a difference between the first-direction relation weight and the second-direction relation weight; if there is a difference, perform fusion processing on the first-direction relation weight and the second-direction relation weight.

[0014] Furthermore, the specific steps for outputting the quality status of steel fiber distribution are as follows: Based on the reasoning reinforcement orientation similarity relationship, determine the correlation contribution of different spatial regions; based on the correlation contribution, perform quality status correlation characterization on the reinforcement spatial orientation status characteristics corresponding to different spatial regions to obtain quality status reasoning features; based on the quality status reasoning features, determine the multidimensional quality evaluation results; based on the multidimensional quality evaluation results, perform quality status mapping to determine the quality status of steel fiber distribution.

[0015] Further, the pre-training steps of the orientation similarity inference network are as follows: Obtain a sample set of steel fiber components, including the corresponding spatial orientation state, initial orientation similarity relationship, multidimensional quality evaluation labels, and corresponding quality state category labels; determine the predicted quality evaluation result for each steel fiber component sample based on the spatial orientation state and initial orientation similarity relationship; construct a first loss based on the difference between the predicted multidimensional quality evaluation result and the multidimensional quality evaluation label; extract the sample orientation similarity residual and sample orientation evolution change generated by each steel fiber component sample in each update process, and construct a second loss; extract the predicted quality state category of each steel fiber component sample, construct a classification loss with the corresponding quality state category label, and weight and sum the first and second losses to obtain the total loss; adjust the network parameters of the orientation similarity inference network through gradient backpropagation until the total loss meets the convergence condition, thus completing the pre-training.

[0016] Furthermore, the specific steps for constructing the second loss are as follows: obtain the sample orientation evolution change amount of each spatial region of each steel fiber component sample, as well as the sample orientation similarity residual generated in each round of updates; for any two spatial regions of each steel fiber component sample, analyze the degree of difference between the corresponding sample orientation evolution change amounts, and determine the evolutionary homology coefficient; based on the sample orientation similarity residual, combined with the evolutionary homology coefficient, calculate the sub-constraint loss, and sum them to obtain the second loss.

[0017] A steel fiber distribution quality determination system based on orientation similarity reasoning includes: a data acquisition unit for acquiring steel fiber orientation data corresponding to multiple spatial regions within the target analysis object; a spatial orientation state construction unit for constructing spatial orientation states; a cross-regional relationship mapping unit for constructing initial orientation similarity relationships; an orientation similarity reasoning unit, including an association propagation subunit for propagating spatial orientation states based on the initial orientation similarity relationships to obtain enhanced spatial orientation state features; a reasoning enhancement subunit for calculating orientation similarity adjustment amounts based on enhanced spatial orientation state features and the initial orientation similarity relationships, and dynamically updating the initial orientation similarity relationships to obtain inferred enhanced orientation similarity relationships; and a quality state reasoning output subunit for outputting the steel fiber distribution quality state.

[0018] The present invention has the following beneficial effects:

[0019] (1) The steel fiber distribution quality judgment method based on orientation similarity reasoning constructs a spatial orientation state and converts the steel fiber orientation data into a structured representation of the spatial region. On this basis, the initial orientation similarity relationship is generated by the comprehensive weighted fusion of orientation similarity, evolutionary orientation similarity and spatial correlation. It not only considers the similarity of the current orientation direction between regions, but also takes into account whether the evolution trend of fiber orientation along the spatial path is close, as well as the influence of spatial distance on the degree of correlation between regions. Thus, a more accurate description of the orientation correlation between regions can be obtained in the initial stage.

[0020] (2) The steel fiber distribution quality judgment method based on orientation similarity reasoning selects the association propagation through the orientation similarity propagation module, so that the spatial region information with high orientation similarity is preferentially fused, avoiding the interference of weakly associated regions on feature extraction. At the same time, the orientation similarity adaptive update module dynamically adjusts the relation weights during the reasoning process, so that the relation weights between regions gradually converge to the state that reflects the true orientation association. The two modules work together alternately, so that the enhanced spatial orientation state features and the orientation similarity relationship between regions promote each other in the iteration, which can more accurately capture the change law of steel fiber orientation along the spatial direction and improve the adaptability to complex orientation patterns.

[0021] (3) The steel fiber distribution quality determination method based on orientation similarity reasoning outputs multi-dimensional quality evaluation results based on the quality state reasoning features through the quality state reasoning module, and determines the steel fiber distribution quality state by combining the state mapping, so that the steel fiber distribution quality determination can comprehensively consider the orientation relationship changes between different spatial regions. At the same time, in the orientation similarity reasoning network pre-training process, the network parameters are jointly optimized by multiple types of loss, so that the network can learn the correspondence between spatial orientation state, orientation similarity relationship and steel fiber distribution quality state, thereby improving the reliability of the output corresponding steel fiber distribution quality state.

[0022] (4) The steel fiber distribution quality judgment system based on orientation similarity reasoning collects steel fiber orientation data of each spatial region through the data acquisition unit. The spatial orientation state construction unit converts the data into a structured state representation containing the main orientation angle, orientation evolution change amount and orientation discrete value. The cross-regional relationship mapping unit constructs the initial orientation similarity relationship according to the state characteristics of each region. The association propagation subunit and the reasoning enhancement subunit alternately perform information propagation and relationship update, so that the similarity relationship between regions gradually converges to a stable reasoning result in the iteration. The quality state reasoning output subunit finally outputs the steel fiber distribution quality state, so that the judgment result is consistent with the actual steel fiber distribution quality.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for determining the distribution quality of steel fibers based on orientation similarity reasoning according to the present invention.

[0025] Figure 2 This is a flowchart illustrating the specific steps involved in constructing the initial orientation similarity relationship in a steel fiber distribution quality determination method based on orientation similarity reasoning according to the present invention.

[0026] Figure 3 This is a block diagram of a steel fiber distribution quality determination system based on orientation similarity reasoning according to the present invention. Detailed Implementation

[0027] Please see Figure 1 The present invention provides a technical solution: a method for determining the distribution quality of steel fibers based on orientation similarity reasoning, comprising the following steps: obtaining steel fiber orientation data corresponding to multiple spatial regions inside the target analysis object (such as steel fiber reinforced concrete components), and constructing the spatial orientation state of the corresponding spatial regions after axial orientation state decoupling processing;

[0028] Perform cross-regional orientation relationship mapping on spatial orientation states to construct initial orientation similarity relationships;

[0029] The spatial orientation state and the initial orientation similarity relationship are input into a pre-trained orientation similarity inference network;

[0030] By using the orientation similarity propagation module in the orientation similarity inference network, the spatial orientation states of different spatial regions are correlated and propagated according to the initial orientation similarity relationship, thereby obtaining the enhanced spatial orientation state features of each spatial region.

[0031] In the orientation similarity adaptive update module of the orientation similarity inference network, the orientation similarity adjustment amount is calculated based on the orientation state features of the enhanced space and the initial orientation similarity relationship, and the initial orientation similarity relationship is dynamically updated based on the orientation similarity adjustment amount to obtain the inference enhanced orientation similarity relationship.

[0032] The enhanced orientation similarity relationship and enhanced spatial orientation state features are input into the quality state reasoning module in the orientation similarity reasoning network, and the output is the quality state of steel fiber distribution.

[0033] Among them, the steel fiber orientation data is the set of the axial direction angles of multiple steel fibers in each spatial region, as well as the spatial position of each spatial region;

[0034] The orientation similarity reasoning network consists of three parts: orientation similarity propagation module, orientation similarity adaptive update module, and quality state reasoning module. The orientation similarity reasoning network uses spatial regions as relation reasoning nodes, spatial orientation states as node features, and orientation similarity relationships between spatial regions as relationships between nodes. It achieves the determination of the quality state of steel fiber distribution through relation propagation, relation update, and state reasoning.

[0035] The orientation similarity propagation module is used to propagate information based on the orientation similarity relationship between regions;

[0036] The orientation similarity adaptive update module is used to dynamically optimize the weights of the relationships between regions during the inference process;

[0037] The quality state reasoning module is used to adaptively aggregate information from multiple spatial region nodes based on optimized relation weights, enhanced spatial orientation state features, and regional contribution weights obtained through learnable attention mapping, and output the quality state of steel fiber distribution.

[0038] This network establishes mapping rules between local orientation features and overall distribution quality state through multi-layer relational reasoning.

[0039] The axial orientation state decoupling process maps the axial direction angle to a double-angle space, eliminating the ambiguity caused by the equivalence of opposite directions, and enabling different axial arrangement states of different steel fibers to be converted into a unified directional representation.

[0040] Specifically, the steps for constructing the spatial orientation state of each spatial region are as follows:

[0041] Based on the steel fiber orientation data for each spatial region, the axial direction angle of each steel fiber in each spatial region is obtained, specifically as follows:

[0042] The axial orientation angle of each steel fiber in each spatial region is the orientation angle of each steel fiber axis relative to the casting direction in that spatial region. It is obtained by XCT scanning or image analysis and is represented according to the 180° axial direction equivalence of the steel fiber.

[0043] The axial orientation angles of each steel fiber are converged through axial orientation processing to determine the principal orientation angles of each spatial region, specifically as follows:

[0044] For the For each spatial region, a double-angle spatial mapping is performed on the axial direction angle of each steel fiber within the spatial region to obtain the direction cosine value and direction sine value corresponding to each direction angle. The cosine values ​​of all directions within the same spatial region are summed and averaged to obtain the cosine mean of the spatial region. The sine values ​​of all directions within the same spatial region are summed and averaged to obtain the sine mean of the spatial region.

[0045] By performing a double-angle spatial inverse mapping on the cosine mean and sine mean, the principal orientation angle of the spatial region can be obtained. :

[0046] ;

[0047] in, The cosine mean of this spatial region. This is the sinusoidal mean of the spatial region;

[0048] Extract the principal orientation angle of each spatial region and the previous spatial region along the pouring direction. The change in angle between the principal orientation angles of -1 is used as the change in orientation evolution in each spatial region, specifically as follows:

[0049] The change in orientation characterizes the spatial evolution trend of steel fiber orientation as it changes from one region to the next along the casting direction.

[0050] For the spatial regions ( Orientation evolution change The extraction rule is to map the two principal orientation angles to a double-angle space and then extract the minimum rotation angle.

[0051] ;

[0052] in, A positive value indicates that the main orientation angle rotates clockwise along the pouring direction, while a negative value indicates that it rotates counterclockwise.

[0053] For the spatial region ranked first along the pouring direction ( Since there is no preceding spatial region, its orientation evolution change is set to zero, i.e. This ensures accurate characterization of the continuous variation trend of steel fiber orientation between adjacent spatial regions;

[0054] Based on the deviation statistical aggregation processing of the axial direction angle and principal orientation angle of each steel fiber, the orientation discrete value of each spatial region is determined, specifically as follows:

[0055] Orientation discrete value characterizes the degree of deviation of the orientation of each steel fiber in the spatial region from the overall arrangement orientation of the region, reflecting the orderliness of the fiber arrangement within the region;

[0056] The specific process of deviation statistical aggregation is as follows: the degree of axial deviation of the axial direction angle of each steel fiber in the same spatial region relative to the main orientation angle of that spatial region is periodically equivalently transformed to obtain the axial direction deviation corresponding to each steel fiber;

[0057] The axial deviation is normalized and converted into a dimensionless deviation value. The normalized deviation is then statistically aggregated, and the orientation dispersion value is determined based on the root mean square dispersion of the normalized deviation. The orientation dispersion value increases as the steel fiber orientation concentration decreases, and the orientation dispersion value ranges from 0 to 1.

[0058] The normalization process involves proportionally converting the axial deviation relative to the maximum permissible axial deviation (90 degrees) so that the normalized axial deviation is between 0 and 1.

[0059] Statistical aggregation specifically involves taking the square root of the squared average of the normalized directional deviations corresponding to each steel fiber.

[0060] The principal orientation angle, orientation evolution change, and orientation discrete value of each spatial region are combined to form the spatial orientation state corresponding to each spatial region. Specifically, it is as follows:

[0061] The spatial orientation state is formed by combining the double-angle direction cosine component, double-angle direction sine component, orientation evolution change, and orientation discrete value corresponding to the principal orientation angle of each spatial region. .

[0062] Specifically, such as Figure 2 As shown, the specific steps for constructing the initial orientation similarity relationship are as follows:

[0063] For the principal orientation angles of different spatial regions, the directional orientation similarity between different spatial regions is extracted through axial angle relationship transformation. Specifically:

[0064] Transform the principal orientation angles corresponding to any two spatial regions to the double-angle representation space to obtain the double-angle direction representation values ​​corresponding to the two spatial regions.

[0065] Calculate the angle difference between two double-angle direction representation values, and perform linear normalization mapping based on the cosine value of the angle difference to transform the cosine value from the interval [-1, 1] to the interval [0, 1] to obtain the directional orientation similarity between the two spatial regions;

[0066] When the principal orientation angles of two spatial regions are the same, or when they satisfy the 180° equivalence relationship of the steel fiber axial direction, the orientation similarity takes the maximum value of 1. As the axial difference between the principal orientation angles of the two spatial regions increases, the orientation correlation characterization gradually decreases.

[0067] When the principal orientation angles of two spatial regions differ by 90°, the orientation similarity takes the minimum value of 0.

[0068] By analyzing the homology of evolutionary trends in the changes in orientation across different spatial regions, the similarity of orientation evolution among these regions is determined. Specifically:

[0069] Obtain any two spatial regions and their corresponding orientation evolution changes, and calculate the difference between the orientation evolution changes of the two spatial regions;

[0070] The cosine value of the difference is used to determine the degree of consistency between the orientation evolution trends of two spatial regions, and the cosine value is transformed from the interval [-1, 1] to the interval [0, 1] by linear normalization to obtain the orientation evolution similarity.

[0071] When the amount of orientation evolution change in two spatial regions is the same, the orientation evolution similarity takes the maximum value of 1, indicating that the two spatial regions have the same orientation change trend along the casting direction;

[0072] As the difference in orientation evolution between two spatial regions increases, the similarity of orientation evolution gradually decreases, which is used to characterize the degree of homology between the two spatial regions in terms of orientation change direction and magnitude.

[0073] The spatial locations of different spatial regions are read from the steel fiber orientation data, and the spatial correlation between different spatial regions is determined by spatial distance attenuation mapping. Specifically:

[0074] For any two spatial regions and Its spatial correlation degree is calculated using an exponential decay function:

[0075] ;

[0076] in, and The first The and the first The spatial coordinates of each spatial region along the pouring direction For spatial correlation scale parameters, The value of is determined based on the spatial dimensions of the component and the spatial continuity of the steel fiber, and is preferably 1 / 5 to 1 / 2 of the total thickness of the component, more preferably 1 / 3 of the total thickness of the component;

[0077] When the distance between two spatial regions is much smaller than When the spatial correlation degree approaches 1, it indicates that there is a strong spatial correlation constraint between the two spatial regions;

[0078] By aggregating and associating directional similarity, directional evolution similarity, and spatial correlation, initial directional similarity between different spatial regions is obtained and combined into initial directional similarity relationships, specifically as follows:

[0079] The initial orientation similarity between any two spatial regions is obtained by weighted summation of orientation similarity, orientation evolution similarity, and spatial correlation. :

[0080] The weight coefficients corresponding to each parameter in the weighting satisfy a sum of 1, and the weight coefficient of orientation similarity ranges from 0.3 to 0.5, preferably 0.4; the weight coefficient of orientation evolution similarity ranges from 0.2 to 0.4, preferably 0.3; and the weight coefficient of spatial association constraint ranges from 0.2 to 0.4, preferably 0.3.

[0081] Iterate through all the initial orientation similarities between any two spatial regions and combine them into an initial orientation similarity relationship.

[0082] Specifically, the steps to obtain enhanced spatial orientation state features are as follows:

[0083] Based on the initial orientation similarity relationship, other spatial regions that meet the orientation association conditions with each spatial region are selected to form the spatial orientation state association range, which is as follows:

[0084] For any spatial region The spatial region is obtained from the initial orientation similarity relationship. Other spatial regions Initial orientation similarity between ;

[0085] When the initial orientation similarity Greater than the preset association threshold At that time, the spatial area Included in spatial area Corresponding spatial orientation state association range ;

[0086] When the initial orientation similarity Not greater than the preset association threshold At that time, the spatial area will not be included. Included in spatial area The corresponding spatial orientation state association range;

[0087] Among them, the association threshold The value range is 0.3 to 0.6, preferably 0.5;

[0088] By using a screening method, each spatial region establishes state associations only with other spatial regions that have a strong orientation correlation, thus avoiding interference from weakly correlated regions in the spatial orientation state fusion process.

[0089] Based on the spatial orientation state association range, the orientation state association propagation weight is determined according to the initial orientation similarity between corresponding spatial regions, specifically as follows:

[0090] For any spatial region Associate its spatial orientation state with other spatial regions within its range Corresponding initial orientation similarity As a weight in orientation state association propagation ;

[0091] For spatial regions that are not within the scope of spatial orientation state association, the corresponding orientation state association propagation weight is set to 0;

[0092] The orientation state association propagation weight is used to represent the degree of influence of spatial orientation state information between different spatial regions during the propagation process. Its value range is [0, 1]. The larger the value, the higher the degree of orientation association between the two spatial regions and the stronger the influence on the spatial orientation state fusion process.

[0093] Based on the state propagation weights between each spatial region and other spatial regions, the spatial orientation states corresponding to other spatial regions within the spatial orientation state association range are selectively associated, so that the spatial orientation states corresponding to spatial regions with a high orientation similarity relationship with the original spatial region are preferentially fused. Specifically:

[0094] For any spatial region Based on the orientation state, the propagation weights are associated. Other spatial regions within the spatial orientation state association range are sorted, and the spatial orientation states of the corresponding spatial regions are selected to participate in the fusion in order of propagation intensity from high to low.

[0095] When the number of spatial regions within the spatial orientation state association range is greater than the preset maximum number of fusions, only the spatial regions with the highest propagation intensity ranking and the maximum number of fusions are selected for state fusion.

[0096] The maximum number of fusions is introduced to control the maximum number of neighbors participating in the fusion of each spatial region, so as to avoid excessive fusion leading to excessive smoothing of features. Its value ranges from 5 to 10, with 8 being preferred.

[0097] By selecting the association method, spatial regions with high orientation similarity are given priority to participate in the construction of enhanced spatial orientation state features, thereby improving the effective utilization of spatial orientation association information;

[0098] Based on the selective association results corresponding to each spatial region, the spatial orientation state is jointly updated to obtain enhanced spatial orientation state features, specifically as follows:

[0099] The spatial orientation state corresponding to the selected associated spatial region is used as the associated state input. The associated states are weighted and fused according to the orientation state association propagation weight to obtain spatial association fusion features.

[0100] The spatial correlation fusion features are fused with the spatial orientation state of the current spatial region, and enhanced spatial orientation state features are obtained through learnable mapping and nonlinear activation.

[0101] For the Each spatial region has enhanced spatial orientation state characteristics. The acquisition process is represented as follows:

[0102] ;

[0103] ;

[0104] in, For the first The neighborhood information features obtained by a spatial region from its associated spatial regions (i.e., the associated fusion features after propagation, but not yet fused with its own state) are obtained by a spatial region. Indicates the first The spatial orientation state corresponding to each spatial region Representing a spatial region With spatial region The orientation state correlation propagation weights between them Represents the learnable feature map parameters. Indicates the state fusion mapping parameters. This indicates a feature concatenation operation. This represents a non-linear activation function.

[0105] The specific steps to obtain the reasoning-enhanced orientation similarity relationship are as follows:

[0106] The initial orientation similarity between different spatial regions is used as the weight of the current relationship between the corresponding spatial regions, specifically as follows:

[0107] The initial orientation similarity relationship is used as the initial relationship representation for the orientation similarity adaptive update module. For any two spatial regions... and The initial orientation similarity As initial relation weights , Represents the spatial region of the initial reasoning stage. With spatial region The current relationship weight between them;

[0108] Based on the enhanced spatial orientation state features of any two spatial regions, the similarity of the current state is determined, specifically as follows:

[0109] For any two spatial regions and Based on the corresponding enhanced spatial orientation state features and The similarity of the current state is obtained through feature similarity mapping:

[0110] ;

[0111] in, Indicates the first The degree of state correlation between two spatial regions in the feature space during round reasoning;

[0112] Analyze the difference between the current state similarity and the current relationship weight to generate orientation similarity residuals, specifically:

[0113] Calculate the similarity of the current state Weight of the current relationship The difference between them is used as the orientation similarity residual. Orientation similarity residuals are used to represent the direction and degree of deviation between the actual degree of correlation reflected by the current orientation state features of the enhanced space and the current relation weights.

[0114] The current relation weights are adjusted based on the orientation similarity residuals to obtain the adjusted relation weights, specifically:

[0115] The weights of the current relationship are adaptively adjusted based on the orientation similarity residuals:

[0116] ;

[0117] in, For the first The spatial region of the layer With spatial region Relationship weights The step size for adjusting relation weights is used to control the correction ratio of the current state similarity to the current relation weights. Its value ranges from 0.01 to 0.1, preferably 0.05. The relation weight adjustment step size parameter does not participate in the training of network parameters, but is used as a fixed adjustment parameter in the adaptive update process of orientation similarity.

[0118] The adjusted relation weights are subjected to symmetric constraints, and the relation weights after symmetric constraints are used as the input for the next propagation.

[0119] After iterative propagation and adjustment, the final obtained relation weights will be used as the reasoning enhancement orientation similarity relation between the two spatial regions, specifically as follows:

[0120] In the orientation similarity reasoning network, the propagation process of spatial orientation state features and the update process of relation weights are carried out alternately.

[0121] After each round of enhanced spatial orientation state feature propagation is completed, the weights of the relationships between spatial regions are adjusted based on the updated enhanced spatial orientation state features, and the adjusted relationship weights are used in the next round of spatial orientation state propagation.

[0122] After a preset number of layers After rounds of propagation and adjustment, the final obtained relation weights will be used as the corresponding reasoning-enhancing orientation similarity relations. , The number of propagation layers in the orientation similarity reasoning network ranges from 2 to 6, with 4 being the preferred value.

[0123] The specific steps for handling symmetric constraints are as follows:

[0124] Obtain the first-direction relationship weights and second-direction relationship weights between different spatial regions in the adjusted relationship weights, the first-direction relationship weight from the first spatial region to the second spatial region, and the second-direction relationship weight from the second spatial region to the first spatial region, specifically as follows:

[0125] For any two spatial regions and The adjusted relation weights include components in two directions: the first direction relation weight is the spatial region. Pointing to spatial region The relation weights in the second direction are spatial region weights. Pointing to spatial region Relationship weights;

[0126] Ideally, the two should be equal, but since the two directions are updated independently during the adjustment process, slight differences may occur.

[0127] To determine whether there is a difference between the weights of the first-direction relationship and the weights of the second-direction relationship, the specific steps are as follows:

[0128] Calculate the absolute difference between the weights of the two directional relationships. If the absolute difference is greater than a preset difference tolerance threshold... ( If the difference is ), then a difference is determined to exist; otherwise, a difference is determined to exist.

[0129] If discrepancies exist, the weights of the first-direction relationship and the second-direction relationship are merged to ensure that the weights of the relationship between the first spatial region and the second spatial region remain consistent. Specifically:

[0130] The fusion processing method is to take the arithmetic average of the bidirectional relationship weights and use the arithmetic average as the updated relationship weight between the two spatial regions.

[0131] If there is no difference, the first direction relationship weight (or the second direction relationship weight, which are equal) is directly used as the relationship weight after fusion.

[0132] The fused relation weights are used as the relation weights between different spatial regions in the next layer of propagation.

[0133] Specifically, the steps for outputting the quality status of the steel fiber distribution are as follows:

[0134] Based on the reasoning-enhanced similarity relationship, the association contribution of different spatial regions is determined, specifically as follows:

[0135] For any spatial region i, the inference enhancement orientation similarity values ​​between the spatial region and all other spatial regions except itself are summarized. Specifically, the arithmetic mean of the inference enhancement orientation similarity between the spatial region and all other spatial regions is calculated to obtain the association contribution of the spatial region.

[0136] Association contribution is used to characterize the degree of association center of the corresponding spatial region in the overall spatial orientation association structure;

[0137] Based on the correlation contribution, the enhanced spatial orientation state features corresponding to different spatial regions are characterized by quality state correlation to obtain quality state inference features, which are as follows:

[0138] Based on the correlation contribution of each spatial region and the enhanced spatial orientation state characteristics of the corresponding spatial region, the attention weights for quality state inference of each spatial region are determined through learnable attention mapping.

[0139] In this embodiment, the learnable attention mapping process includes two steps: attention score calculation and attention weight normalization.

[0140] For any spatial region, the correlation contribution degree corresponding to the spatial region is combined with the enhanced spatial orientation state features, and the attention score corresponding to the spatial region is obtained through feature mapping with learnable parameters. The learnable parameters are updated through gradient backpropagation during the pre-training process of the orientation similarity inference network, so that the attention score can represent the comprehensive contribution of the orientation state of the corresponding spatial region and its spatial correlation features to the determination of the steel fiber distribution quality.

[0141] The attention scores corresponding to each spatial region are normalized, and the normalized results are used as the attention weights for quality state inference of the corresponding spatial region.

[0142] In one implementation, the normalization process uses Softmax normalization, so that the mass state reasoning attention weights corresponding to each spatial region are all between 0 and 1, and the sum of the mass state reasoning attention weights corresponding to all spatial regions in the same steel fiber component is 1.

[0143] Based on the attention weight of quality state reasoning, the enhanced spatial orientation state features corresponding to each spatial region are weighted and aggregated to obtain component-level quality state reasoning features.

[0144] Based on the characteristics of quality state reasoning, a multidimensional quality evaluation result corresponding to the steel fiber distribution state is determined through state mapping, specifically as follows:

[0145] The quality state reasoning module is configured with three independent learnable output branches. Each output branch incorporates component-level quality state reasoning features through linear mapping and nonlinear normalization. Mapped to three multidimensional quality evaluation parameters;

[0146] During the pre-training process, the three output branches are supervised by the corresponding orientation continuous similarity label, local anomaly label and overall evolution trend label, respectively. The trainable parameters in the output branches are optimized by loss, so that the three output branches learn the mapping relationship between the component-level quality state inference features and the corresponding quality evaluation parameters.

[0147] The first output branch outputs the orientation continuous similarity evaluation value. Its value range is [0, 1]. The closer it is to 1, the more it indicates that the orientation similarity between spatial regions maintains a continuous and stable evolution along the pouring direction. The closer it is to 0, the more it indicates that the orientation similarity changes discontinuously along the pouring direction.

[0148] The second output branch outputs the local anomaly evaluation value. Its value range is [0, 1]. The closer it is to 1, the more likely there is a local abnormal region, that is, the orientation similarity between a certain spatial region and the neighboring spatial regions has a sudden decrease. The closer it is to 0, the more likely the orientation relationship between each spatial region remains normal.

[0149] The overall evolution trend evaluation value of the third output branch output Its value range is [0, 1]. The closer it is to 1, the more it indicates that the orientation similarity along the pouring direction shows a continuous decreasing trend. The closer it is to 0, the more it indicates that the orientation similarity along the pouring direction remains relatively stable.

[0150] Based on the multidimensional quality evaluation results, a quality state mapping is performed to determine the quality state of the steel fiber distribution, specifically as follows:

[0151] The orientation continuity similarity evaluation value, local anomaly evaluation value, and overall evolution trend evaluation value are input into the state classification mapping layer. The state classification mapping layer receives the predicted orientation continuity similarity evaluation value, predicted local anomaly evaluation value, and predicted overall evolution trend evaluation value output by the quality state inference module as input.

[0152] A joint feature mapping is performed on the three predicted evaluation values ​​to establish the correspondence between the multidimensional quality evaluation results and the quality status category of steel fiber distribution, and the corresponding predicted quality status category is output.

[0153] The state classification mapping layer is trained under supervision by the quality state category labels in the training samples, so that the predicted quality state category gradually approximates the true quality state category of the corresponding training sample.

[0154] The quality status of steel fiber distribution includes continuous stable state, local abnormal state, overall degradation state, and composite abnormal state;

[0155] Among them, the continuous stable state is used to characterize the orientation similarity relationship between each spatial region, which maintains a continuous and stable evolution along the pouring direction;

[0156] Local anomalies are used to characterize a sudden decrease in the orientation similarity between a local spatial region and its neighboring spatial regions.

[0157] The overall degradation state is used to characterize the continuous decay trend of the orientation similarity relationship between spatial regions along the pouring direction.

[0158] Composite anomalous states are used to characterize the simultaneous existence of local orientation mutation anomalies and overall orientation evolution decay phenomena.

[0159] Specifically, the pre-training steps for the orientation similarity inference network are as follows:

[0160] Obtain a sample set of steel fiber reinforced concrete (SFR) components, including the corresponding spatial orientation state of the samples, the initial orientation similarity relationship of the samples, multidimensional quality evaluation labels, and the corresponding quality status category labels, specifically as follows:

[0161] The multidimensional quality evaluation labels include three dimensions: orientation continuity and similarity labels, local anomaly labels, and overall evolution trend labels. These are used to characterize the spatial orientation continuity, the degree of local orientation anomaly, and the overall orientation evolution trend along the casting direction of the corresponding steel fiber component samples, respectively.

[0162] In this embodiment, the multidimensional quality evaluation label is calculated based on the actual spatial distribution of steel fibers in the corresponding steel fiber component sample;

[0163] Specifically, based on the spatial distribution data of steel fibers, the main orientation angle, orientation evolution change, orientation discrete value, and orientation similarity relationship between spatial regions are calculated for each spatial region according to the above method. Based on the above calculation results, three quality evaluation labels are constructed.

[0164] The method for constructing the orientation continuous similarity label is as follows: based on the calculation results of the orientation similarity between all adjacent spatial regions along the casting direction, the orientation similarity corresponding to all adjacent spatial regions is statistically averaged, and the statistical average result is normalized to the range of 0 to 1, which is used as the orientation continuous similarity label of the corresponding steel fiber component sample.

[0165] The local anomaly label is constructed as follows: based on the orientation similarity calculation results between each spatial region and its associated neighboring spatial regions, the degree of abrupt decrease in orientation similarity between the local spatial region and the neighboring spatial region is statistically analyzed, and normalization is performed according to the degree of local anomaly to obtain the local anomaly label of the corresponding steel fiber component sample.

[0166] The overall evolution trend label is constructed as follows: based on the changes in the orientation evolution change amount along the casting direction in each spatial region, the overall trend of the orientation evolution change amount is statistically analyzed, and the statistical results are normalized to the range of 0 to 1, which serves as the overall evolution trend label for the corresponding steel fiber component sample.

[0167] The aforementioned orientation-continuous similarity labels, local anomaly labels, and overall evolution trend labels together constitute the multidimensional quality evaluation labels for the corresponding steel fiber component samples, and serve as the supervision labels for the three prediction output branches of the quality state reasoning module.

[0168] In this embodiment, based on the actual spatial distribution of steel fibers in the corresponding steel fiber component sample, the quality status of the training sample is labeled by comprehensively considering the orientation continuity state, local orientation anomaly state, and overall orientation evolution state between spatial regions, and the corresponding quality status category label is obtained.

[0169] When the spatial orientation of steel fibers remains continuous and stable, and there are no local orientation anomalies or overall evolution decay, it is marked as a continuous and stable state.

[0170] When there is an orientation deviation of steel fiber between a local spatial region and its neighboring spatial regions, it is marked as a local abnormal state;

[0171] When the spatial orientation of steel fibers shows a continuous changing trend along the casting direction, resulting in a decrease in the overall orientation continuity, it is marked as an overall degradation state.

[0172] When both local orientation anomalies and continuous global orientation degradation exist simultaneously, it is marked as a composite anomaly state.

[0173] The aforementioned quality state category labels serve as the classification labels corresponding to the supervised training of the state classification mapping layer;

[0174] The total number of training samples should be at least 100, preferably 500 to 1000;

[0175] When the number of samples is insufficient, data augmentation methods are used to expand the sample set. Data augmentation methods include geometric transformation and perturbation enhancement. Specifically, when performing geometric transformations such as rotation around the casting direction to keep the spatial evolution law of steel fibers unchanged, the corresponding axial direction angle and spatial position relationship of steel fibers are adjusted simultaneously.

[0176] When performing geometric transformations that change the spatial region arrangement order, such as mirroring and flipping along the pouring direction, the corresponding steel fiber axial direction angle, spatial position relationship, orientation evolution change amount, and quality status category label affected by the geometric transformation are adjusted synchronously according to the change in spatial region arrangement.

[0177] When injecting random noise, the noise amplitude is limited to a range that does not change the original quality status category of the corresponding steel fiber component sample;

[0178] Based on the spatial orientation state and the similarity relationship of the initial orientation of the samples, the predicted quality evaluation results of each steel fiber component sample are determined, specifically as follows:

[0179] Based on the spatial orientation state and initial orientation similarity relationship of each steel fiber component sample, the sample is input into the orientation similarity inference network. After being processed by the propagation module, adaptive update module and quality state inference module, the predicted multidimensional quality evaluation results corresponding to each sample are output, including the predicted orientation continuous similarity evaluation value, the predicted local anomaly evaluation value and the predicted overall evolution trend evaluation value.

[0180] The first loss is constructed based on the difference between the predicted multidimensional quality evaluation results and the multidimensional quality evaluation labels of each steel fiber component sample. Specifically:

[0181] The first loss uses the mean squared error method to calculate the deviation between the predicted quality evaluation parameters and the true quality evaluation labels, and averages it over all training samples.

[0182] Extract the sample orientation similarity residual and sample orientation evolution change generated in each round of update for each steel fiber component sample, and construct a second loss. The second loss is used to constrain the negative adjustment of the relation weight according to the degree of homology of orientation evolution between different spatial regions, so that the orientation similarity relationship between spatial regions with similar orientation evolution trends remains stable in the inference update process.

[0183] Extract the predicted quality state category of each steel fiber component sample, construct a classification loss with the corresponding quality state category label, and then sum the total loss with the first loss and the second loss. Adjust the network parameters of the orientation similarity inference network through gradient backpropagation until the total loss meets the convergence condition, thus completing the pre-training. Specifically:

[0184] The predicted multidimensional quality evaluation results are input into the state classification mapping layer. The state classification mapping layer performs joint feature mapping based on the predicted orientation continuous similarity evaluation value, the predicted local anomaly evaluation value, and the predicted overall evolution trend evaluation value to establish the classification mapping relationship between the predicted multidimensional quality evaluation results and the quality state category, and outputs the corresponding predicted quality state category.

[0185] Construct a classification loss based on the classification error between the predicted quality state category and the corresponding quality state category label;

[0186] The total loss is obtained by weighted summing the first loss, the classification loss, and the second loss:

[0187] ;

[0188] in, For the total loss, The first loss, For classifying losses, This is the second loss. These are the weight coefficients corresponding to the classification loss. The weighting coefficients corresponding to the second loss;

[0189] in, The value range is 0.1 to 1.0, preferably 0.5; The value range is 0.1 to 0.5, preferably 0.2;

[0190] Furthermore, during the pre-training process, the multidimensional quality evaluation label is used to supervise the predicted multidimensional quality evaluation results output by the quality state inference module, and the quality state category label is used to supervise the predicted quality state category output by the state classification mapping layer.

[0191] The first loss, classification loss, and second loss together constitute the total training loss of the network. The trainable parameters in the orientation similarity propagation module, orientation similarity adaptive update module, quality state inference module, and state classification mapping layer are jointly optimized through gradient backpropagation. This enables the network to simultaneously learn the spatial orientation relationship of steel fibers, the spatial evolution law, and the mapping relationship between the quality state of steel fiber distribution, until the total loss meets the preset convergence condition, thus completing the pre-training of the orientation similarity inference network.

[0192] The convergence condition is that the change in total loss over 10 consecutive iterations is less than a preset convergence threshold. , The range of values ​​is Preferred ;

[0193] For the relation weights in the orientation similarity adaptive update module, the relation weights are intermediate states that are dynamically generated based on the spatial region state during the forward propagation of the current training sample and updated during the relation inference process. They are not saved as fixed network parameters shared across training samples.

[0194] The evolution of sample orientation in each spatial region of each steel fiber component sample is obtained, as well as the sample orientation similarity residuals generated in each update process, specifically as follows:

[0195] For each training sample of a steel fiber component, the change in sample orientation of each spatial region is obtained in the manner described above; at the same time, the residual of sample orientation similarity between each spatial region is obtained in each round of update of the orientation similarity adaptive update module in the manner described above.

[0196] For any two spatial regions of each steel fiber component sample, analyze the degree of difference between the corresponding sample orientation evolution changes, and determine the evolutionary homology coefficient between the two spatial regions, specifically as follows:

[0197] For any two spatial regions in the same training sample, calculate the difference in the change of orientation of the two samples, input the difference into the cosine function, calculate half of the sum of the cosine of the difference and 1, and use it as the evolutionary homology coefficient between the two spatial regions.

[0198] When the orientation evolution changes of two spatial regions are completely consistent, the evolutionary homology coefficient is 1; when the orientation evolution changes of two spatial regions differ by 180 degrees, the evolutionary homology coefficient is 0.

[0199] The closer the evolutionary homology coefficient is to 1, the higher the degree of homology in the orientation evolution trends of the two spatial regions, that is, the more consistent the change trends of the two along the casting direction.

[0200] Based on the sample orientation similarity residuals and combined with the evolutionary homology coefficient, the sub-constraint losses corresponding to the two spatial regions are calculated, and the summation is used to obtain the second loss, which is as follows:

[0201] In each round of updates, for any two spatial regions in each training sample, the sample orientation similarity residual between the two spatial regions generated in that round is obtained;

[0202] Multiply the evolutionary homology coefficient by the orientation similarity residual, take the negative value of the product, and when the negative value is greater than zero, take the negative value as the sub-constraint loss of the two spatial regions in this round;

[0203] When the negative value is not greater than zero, the loss of the sub-constraint is zero;

[0204] The second loss is obtained by summing the sub-constraint losses corresponding to all training samples, all update rounds, and all spatial regions, and then dividing by the product of the total number of training samples, the total number of update rounds, and the square of the total number of spatial regions.

[0205] When the evolutionary homology coefficient is large, a strong constraint is imposed on the negative adjustment of the orientation similarity residual to reduce the degree to which the relationship weight between spatial regions with similar orientation evolutionary trends is weakened.

[0206] As the evolutionary homology coefficient decreases, the constraint effect weakens accordingly, allowing the weights of the relationships between corresponding spatial regions to be adjusted primarily based on the similarity of the current state.

[0207] Please see Figure 3 This invention provides a technical solution: a steel fiber distribution quality determination system based on orientation similarity inference, comprising: a data acquisition unit for acquiring steel fiber orientation data corresponding to multiple spatial regions within the target analysis object; a spatial orientation state construction unit for decoupling the steel fiber orientation data by axial orientation state to construct the spatial orientation state of the corresponding spatial region; a cross-regional relationship mapping unit for performing cross-regional orientation relationship mapping on the spatial orientation state to construct an initial orientation similarity relationship; an orientation similarity inference unit for inputting the spatial orientation state and the initial orientation similarity relationship into a pre-trained orientation similarity inference network; an association propagation subunit for propagating the spatial orientation state through the orientation similarity propagation module based on the initial orientation similarity relationship to obtain enhanced spatial orientation state features; an inference enhancement subunit for calculating the orientation similarity adjustment amount based on the enhanced spatial orientation state features and the initial orientation similarity relationship in the orientation similarity adaptive update module, and dynamically updating the initial orientation similarity relationship to obtain an inferred enhanced orientation similarity relationship; and a quality state inference output subunit for inputting the inferred enhanced orientation similarity relationship and the enhanced spatial orientation state features into the quality state inference module to output the steel fiber distribution quality state.

[0208] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0209] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for determining the quality of steel fiber distribution based on orientation similarity reasoning, characterized in that, Includes the following steps: The steel fiber orientation data corresponding to multiple spatial regions within the target analysis object are obtained, and after axial orientation state decoupling processing, the spatial orientation state of the corresponding spatial regions is constructed, specifically as follows: Based on the steel fiber orientation data, obtain the axial direction angle of each steel fiber in each spatial region; The axial orientation angles of each steel fiber are converged through axial orientation processing to determine the principal orientation angles of each spatial region; Extract the angular change between the principal orientation angle of each spatial region and the principal orientation angle of the previous spatial region along the pouring direction, and use it as the orientation evolution change. Based on the deviation statistical aggregation processing of the axial direction angle and main orientation angle of each steel fiber, the orientation discrete value of each spatial region is determined; The main orientation angle, the change in orientation evolution, and the discrete value of orientation are combined to form the spatial orientation state; The spatial orientation states are mapped across regions to construct initial orientation similarity relationships. The specific steps are as follows: For the principal orientation angles of different spatial regions, the directional orientation similarity between different spatial regions is extracted through axial angle relationship transformation; By analyzing the homology of evolutionary trends, the similarity of orientation evolution was determined for the changes in orientation in different spatial regions. The spatial locations of different spatial regions are read from the steel fiber orientation data, and the spatial correlation is determined by spatial distance attenuation mapping. By associating and aggregating directional similarity, orientation evolution similarity, and spatial correlation, the initial orientation similarity between different spatial regions is obtained and combined into an initial orientation similarity relationship. The spatial orientation state and the initial orientation similarity relationship are input into a pre-trained orientation similarity inference network; The spatial orientation state is associated and propagated based on the initial orientation similarity relationship through the orientation similarity propagation module, thereby obtaining enhanced spatial orientation state features. In the orientation similarity adaptive update module, the orientation similarity adjustment amount is calculated based on the enhanced spatial orientation state features and the initial orientation similarity relationship, and the initial orientation similarity relationship is dynamically updated to obtain the inference enhanced orientation similarity relationship. The specific steps are as follows: The initial orientation similarity between different spatial regions is used as the weight of the current relationship. Determine the similarity of the current state based on the enhanced spatial orientation state features of any two spatial regions. Analyze the difference between the current state similarity and the current relationship weight to generate orientation similarity residuals; The current relation weights are adjusted based on the orientation similarity residuals to obtain the adjusted relation weights; The adjusted relation weights are subjected to symmetric constraints, and the relation weights after symmetric constraints are used as the input for the next propagation. After iterative propagation and adjustment, the final relationship weights are used as the reasoning enhancement orientation similarity relationships; The enhanced orientation similarity relationship and enhanced spatial orientation state features are input into the quality state reasoning module, and the steel fiber distribution quality state is output.

2. The method for determining the quality of steel fiber distribution based on orientation similarity reasoning according to claim 1, characterized in that, The specific steps to obtain enhanced spatial orientation state features are as follows: Based on the initial orientation similarity relationship, other spatial regions that meet the orientation association conditions with each spatial region are selected to form the spatial orientation state association range; Based on the spatial orientation state association range, the orientation state association propagation weight is determined according to the initial orientation similarity. Based on the state propagation weight, select and associate spatial orientation states corresponding to other spatial regions within the spatial orientation state association range; Based on the selective association results corresponding to each spatial region, the spatial orientation state is jointly updated to obtain enhanced spatial orientation state features.

3. The method for determining the quality of steel fiber distribution based on orientation similarity reasoning according to claim 1, characterized in that, The specific steps for handling symmetric constraints are as follows: Obtain the first-direction relation weights and second-direction relation weights between different spatial regions in the adjusted relation weights; Determine whether there is a difference between the weights of the first directional relationship and the weights of the second directional relationship; If there are differences, the weights of the first-direction relationship and the weights of the second-direction relationship are merged.

4. The method for determining the quality of steel fiber distribution based on orientation similarity reasoning according to claim 1, characterized in that, The specific steps for determining the quality state of the steel fiber distribution are as follows: Based on the reasoning-enhanced orientation similarity relationship, the correlation contribution of different spatial regions is determined; Based on the correlation contribution, the quality state correlation characterization is performed on the enhanced spatial orientation state features corresponding to different spatial regions to obtain quality state inference features. Based on the characteristics of quality state reasoning, determine the multidimensional quality evaluation results; The quality status is mapped based on the multidimensional quality evaluation results to determine the quality status of steel fiber distribution.

5. The method for determining the quality of steel fiber distribution based on orientation similarity reasoning according to claim 1, characterized in that, The pre-training steps for the orientation similarity inference network are as follows: Obtain a sample set of steel fiber components, including the corresponding spatial orientation state of the samples, the initial orientation similarity relationship of the samples, the multidimensional quality evaluation labels, and the corresponding quality state category labels; Based on the spatial orientation state of the samples and the similarity relationship of the initial orientation of the samples, the predicted quality evaluation results of each steel fiber component sample are determined. The first loss is constructed based on the difference between the predicted multidimensional quality assessment results and the multidimensional quality assessment labels. Extract the sample orientation similarity residuals and sample orientation evolution changes generated in each steel fiber component sample during each update process, and construct a second loss. Extract the predicted quality state category of each steel fiber component sample, construct a classification loss with the corresponding quality state category label, and sum the total loss with the first loss and the second loss. Then, adjust the network parameters of the orientation similarity inference network through gradient backpropagation until the total loss meets the convergence condition, thus completing the pre-training.

6. The method for determining the quality of steel fiber distribution based on orientation similarity reasoning according to claim 5, characterized in that, The specific steps for constructing the second loss are as follows: The evolution of sample orientation in each spatial region of each steel fiber component sample is obtained, as well as the sample orientation similarity residuals generated in each update process; For any two spatial regions of each steel fiber component sample, analyze the degree of difference between the corresponding sample orientation evolution changes and determine the evolution homology coefficient; Based on the sample orientation similarity residuals, combined with the evolutionary homology coefficients, the sub-constraint loss is calculated, and the summation is used to obtain the second loss.

7. A steel fiber distribution quality determination system based on orientation similarity reasoning, employing the steel fiber distribution quality determination method based on orientation similarity reasoning as described in any one of claims 1-6, characterized in that, include: The data acquisition unit is used to acquire steel fiber orientation data corresponding to multiple spatial regions inside the target analysis object; Spatial orientation state construction unit, used to construct spatial orientation states; Cross-regional relationship mapping unit, used to construct initial orientation similarity relationships; The orientation similarity reasoning unit includes an association propagation subunit, which is used to propagate spatial orientation states based on the initial orientation similarity relationship to obtain enhanced spatial orientation state features; The reasoning enhancement subunit is used to calculate the orientation similarity adjustment amount based on the orientation state features of the enhancement space and the initial orientation similarity relationship, and to dynamically update the initial orientation similarity relationship to obtain the reasoning enhancement orientation similarity relationship; The quality state inference output subunit is used to output the quality state of the steel fiber distribution.

Images