A method, system and electronic device for predicting the autogenous shrinkage of ultra-high performance concrete

Abstract

The present application belongs to the field of concrete performance prediction and data processing, and particularly relates to a method and system for predicting the self-shrinkage of ultra-high performance concrete and an electronic device, aiming to solve the problem that the existing method relies on experiments and is difficult to jointly represent the nonlinear coupling correlation between material component parameters and curing parameters. The present application comprises: deleting abnormal samples and standardizing the concrete samples, constructing a prediction network connected in series by a one-dimensional convolution network, a bidirectional long short-term memory network and a multi-head attention mechanism, training the prediction parameters according to the prediction error, inputting the to-be-measured parameters into the prediction network after standardization, and performing anti-standardization on the network output to obtain the self-shrinkage prediction value. The present application can jointly extract local correlation features, bidirectional sequence features and weighted sequence features, and realize the prediction of the self-shrinkage value of ultra-high performance concrete.

Description

Technical Field

[0001] This invention belongs to the field of concrete performance prediction and data processing technology, specifically involving a method, system and electronic equipment for predicting the autogenous shrinkage of ultra-high performance concrete. Background Technology

[0002] Ultra-high performance concrete (UHVPC) typically features a high amount of cementitious materials and a low water-cement ratio, making it prone to autogenous shrinkage deformation in the early stages of hardening. Autogenous shrinkage deformation can generate tensile stress within the concrete and increase the risk of early microcracks and cracking. Therefore, predicting the autogenous shrinkage value of UHVPC under different material compositions and curing conditions is of great significance for mix design and autogenous shrinkage performance analysis.

[0003] The current method for obtaining the autogenous shrinkage value of ultra-high performance concrete mainly relies on specimen preparation, curing, and autogenous shrinkage testing. This method requires a certain curing period to obtain measured data, and when the material composition or curing conditions change, the test usually needs to be repeated. Therefore, it suffers from problems such as long test cycles, high data acquisition costs, and difficulty in comprehensively covering different parameter combinations under various mix proportions and curing conditions.

[0004] Predicting autogenous shrinkage using empirical or statistical regression relationships can reduce the need for repeated experiments to some extent. However, the autogenous shrinkage of ultra-high performance concrete is influenced by multiple parameters, including cementitious materials, aggregates, mixing water, admixtures, fibers, and curing age. Complex nonlinear coupling relationships may also exist between material component parameters and between material component parameters and curing age. Some empirical or single-data prediction models have limited ability to jointly express local correlations between adjacent input features, correlations between two directions of input feature sequences, and differences in contributions from different sequence positions. Therefore, the adaptability of the prediction results still has room for improvement when material composition and curing conditions change.

[0005] Furthermore, different material composition parameters, curing parameters, and autogenous shrinkage values ​​have different dimensions and numerical ranges. Abnormal samples and different data scales may affect the training process of prediction parameters. How to perform consistent data processing on concrete samples and jointly extract local correlation features, bidirectional sequence features, and weighted sequence features from standardized input features to obtain autogenous shrinkage prediction values ​​that can be used for the tested material composition parameters and curing parameters is a technical problem that needs to be solved. Summary of the Invention

[0006] This invention aims to address the problems of existing methods for predicting the autogenous shrinkage value of ultra-high performance concrete, such as reliance on experiments, difficulty in jointly characterizing the nonlinear coupling relationships between different material component parameters and between material component parameters and curing parameters, and the potential impact of abnormal samples and dimensional differences on the training of prediction networks. The invention provides a method, system, and electronic device for predicting the autogenous shrinkage of ultra-high performance concrete.

[0007] In a first aspect, this invention proposes a method for predicting the autogenous shrinkage of ultra-high performance concrete, comprising: Concrete samples were collected, and material composition parameters and curing parameters were used as input features. The measured autogenous shrinkage value was used as the output value. The input features and the output value were standardized to obtain standardized input features, standardized measured values ​​and standardized parameters. A prediction network is constructed by sequentially connecting a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism. The standardized input features are convolved and nonlinearly mapped by the one-dimensional convolutional network to obtain a local feature sequence. The local feature sequence is encoded by the bidirectional long short-term memory network to obtain bidirectional sequence features. The bidirectional sequence features are weighted by the multi-head attention mechanism to obtain weighted sequence features. Standardized predicted values ​​are generated using the weighted sequence features, and the prediction network is trained based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters. The component parameters and curing parameters of the material to be tested are converted into predictive input features using the standardized parameters. The predictive input features are then input into the predictive network using the predicted parameters. The output of the predictive network is de-standardized to obtain the self-shrinkage predicted value.

[0008] Furthermore, each concrete test record corresponds to one concrete sample, and the material composition parameters include at least the amount of cement, silica fume, fly ash, sand, mixing water, superplasticizer, superabsorbent polymer, expansion agent, shrinkage retardant, and steel fiber. The curing parameters include at least the curing age. The material composition parameters and curing parameters of the same concrete sample are arranged in a preset order, so that each material composition parameter and curing parameter corresponds to a sequence position, forming an input feature sequence corresponding to a measured autogenous shrinkage value.

[0009] Furthermore, the standardized input features, the standardized measured values, and the standardized parameters are calculated as follows: Delete samples whose absolute value of the difference between any input feature or output value and the mean of the same term, divided by the standard deviation of the same term, exceeds a preset outlier threshold; Calculate the mean and standard deviation of each input feature and output value of the retained sample, and use the mean and standard deviation as the standardization parameters; Subtract the mean of the same item from each value of the retained sample and divide by the standard deviation of the same item to obtain the standardized input feature and the standardized measured value; The standardized input features and standardized measured values ​​of the same sample are combined to form a standardized sample, which is then randomly divided into a training set and a test set according to a preset ratio. The prediction network is trained using the training set, and the prediction network using the prediction parameters is tested using the test set.

[0010] Furthermore, the specific processing method for the one-dimensional convolutional network is as follows: The standardized input features are arranged into an input feature sequence according to a preset order, and then input into multiple sequentially connected one-dimensional convolutional layers; Each one-dimensional convolutional layer performs a one-dimensional convolution on the input to obtain a channel response, performs batch normalization on the channel response, and sets negative values ​​to zero. Except for the last one-dimensional convolutional layer, the processing result of each one-dimensional convolutional layer is input into the next one-dimensional convolutional layer, and the processing result of the last one-dimensional convolutional layer is used as the local feature sequence.

[0011] Furthermore, the specific processing method of the bidirectional long short-term memory network is as follows: The positive long short-term memory state is updated from the first item to the last item along the local feature sequence, and the negative long short-term memory state is updated from the last item to the first item. By splicing the forward long short-term memory state and the reverse long short-term memory state at the same sequence position, and performing random deactivation on the splicing result, the bidirectional sequence feature is obtained.

[0012] Furthermore, the weighted sequence features and the standardized predicted values ​​are calculated as follows: The bidirectional sequence features are subjected to query projection, key projection, and value projection within multiple attention heads to obtain a query matrix, a key matrix, and a value matrix. Multiply the query matrix of each attention head by the transpose of the key matrix and divide by the square root of the dimension of the key vector. Apply a normalized exponential function to the division result to obtain the attention weight. The value matrix is ​​weighted by the attention weights, the weighted results of each attention head are concatenated and linear mapping is performed to obtain the weighted sequence features; global average pooling is performed on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector; the fixed-length feature vector is sequentially input into the first fully connected layer and the second fully connected layer to obtain the standardized prediction value.

[0013] Furthermore, the prediction parameters are calculated as follows: The standardized samples in the training set are used as training samples, and the standardized samples in the test set are used as early cessation monitoring samples. The loss value is obtained by calculating the squared mean of the differences between the standardized predicted values ​​and the standardized measured values ​​of the training samples. Calculate the gradient of the loss value. When the gradient norm exceeds a preset clipping threshold, scale the gradient according to the ratio of the preset clipping threshold to the gradient norm. An adaptive moment estimation optimizer is used to update the parameters of the prediction network with scaled gradients or gradients that do not exceed the preset pruning threshold; After each round of parameter updates, the monitoring loss is calculated based on the early stop monitoring samples. When the monitoring loss does not decrease for a preset number of consecutive rounds, the update stops, and the prediction network parameters with the minimum monitoring loss are used as the prediction parameters.

[0014] Furthermore, the calculation method for the predicted input features and the self-shrinking predicted value is as follows: The standardized parameters include the mean and standard deviation of each input feature and output value; The predicted input features are obtained by subtracting the mean of the same input feature from each of the material composition parameters and maintenance parameters and then dividing by the standard deviation of the same input feature. The self-shrinking prediction value is obtained by multiplying the standardized prediction value output by the standard deviation of the output value and adding the mean of the output value.

[0015] In a second aspect, the present invention provides a system for predicting the autogenous shrinkage of ultra-high performance concrete, comprising: The data processing module is configured to collect concrete samples, use material composition parameters and curing parameters as input features, use measured autogenous shrinkage value as output value, standardize the input features and the output value, and obtain standardized input features, standardized measured values ​​and standardized parameters. The prediction network module is configured to construct a prediction network consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence; the one-dimensional convolutional network convolves nonlinearly to map the standardized input features to obtain a local feature sequence; the bidirectional long short-term memory network encodes the local feature sequence to obtain bidirectional sequence features; and the multi-head attention mechanism weights the bidirectional sequence features to obtain weighted sequence features. The training module is configured to generate standardized predicted values ​​using the weighted sequence features, and train the prediction network based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters. The prediction module is configured to convert the component parameters and curing parameters of the material to be tested into prediction input features using the standardized parameters, input the prediction input features into the prediction network using the prediction parameters, de-standardize the output of the prediction network, and obtain the self-shrinkage prediction value.

[0016] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method for predicting the autogenous shrinkage of ultra-high performance concrete as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are: This invention performs abnormal sample deletion and standardization on the material composition parameters, curing parameters and measured autogenous shrinkage values ​​in concrete samples, and saves the standardized parameters corresponding to each input feature and output value. Therefore, it can reduce the impact of abnormal data and different dimensions on the training of the prediction network, and make the training stage and the prediction stage use the same data scale and field order.

[0018] The prediction network of this invention sequentially connects a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism. The one-dimensional convolutional network extracts local correlation features from the standardized input features, the bidirectional long short-term memory network encodes local feature sequences from two opposite directions, and the multi-head attention mechanism forms attention weights based on the correlation between different sequence positions. Therefore, it can jointly characterize the local correlation, bidirectional sequence correlation, and contribution differences between different sequence positions between material composition parameters and maintenance parameters.

[0019] This invention trains a prediction network based on the mean square error between standardized predicted values ​​and standardized measured values, and prunes the gradient proportionally when the gradient norm exceeds a preset pruning threshold. Therefore, it can limit the impact of abnormal gradients on the update process of trainable parameters and reduce the possibility of excessively large gradient values ​​during training.

[0020] This invention uses an adaptive moment estimation optimizer to update trainable parameters and implements early stopping based on the monitoring loss corresponding to the early stopping monitoring sample. The prediction network parameters at the lowest monitoring loss are used as prediction parameters. Therefore, prediction parameters that are adapted to the early stopping monitoring sample can be retained during training, reducing the situation where the monitoring loss increases due to continued training.

[0021] In this invention, the component parameters and curing parameters of the material to be tested are converted into predictive input features using the same standardized parameters as those in the training phase. Furthermore, the standardized predicted values ​​output by the prediction network are de-standardized. Therefore, the network output can be restored to the dimensions corresponding to the measured self-shrinkage values, directly forming self-shrinkage predicted values ​​corresponding to the component parameters and curing parameters of the material to be tested. Attached Figure Description

[0022] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of the method for predicting the autogenous shrinkage of ultra-high performance concrete in an embodiment of the present invention; Figure 2 This is a schematic diagram of a prediction network structure in an embodiment of the present invention, consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence. Figure 3 This is a scatter plot comparing the predicted and measured self-shrinkage values ​​of the training and test sets in this embodiment of the invention. Figure 4 This is a comparison chart of the prediction performance of different models in the embodiments of the present invention; Figure 4 (a) is a comparison of the mean absolute error of different models, (b) is a comparison of the mean square error of different models, (c) is a comparison of the coefficient of determination of different models, and (d) is a comparison of the root mean square error of different models. Figure 5 This is a line graph comparing the measured self-shrinkage value and the predicted self-shrinkage value in an embodiment of the present invention; Figure 5 (a) in the figure is a comparison of the measured shrinkage values ​​and the predicted shrinkage values ​​of the training set, and (b) is a comparison of the measured shrinkage values ​​and the predicted shrinkage values ​​of the test set. Figure 6 This is a structural block diagram of the ultra-high performance concrete self-shrinkage prediction system in an embodiment of the present invention. Detailed Implementation

[0023] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] See Figure 1 The first embodiment of this invention proposes a method for predicting the autogenous shrinkage of ultra-high performance concrete, comprising the following steps: S10: Collect concrete samples, use material composition parameters and curing parameters as input features, use the measured autogenous shrinkage value as the output value, standardize the input features and the output value to obtain standardized input features, standardized measured values ​​and standardized parameters; S20: Construct a prediction network consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence; S30: Use a one-dimensional convolutional network to convolve and nonlinearly map the standardized input features to obtain a local feature sequence; S40: Encode the local feature sequence using a bidirectional long short-term memory network to obtain bidirectional sequence features; S50: The bidirectional sequence features are weighted using a multi-head attention mechanism to obtain weighted sequence features; S60: Generate standardized predicted values ​​using the weighted sequence features, and train the prediction network based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters; S70: Convert the component parameters and curing parameters of the material to be tested into predictive input features using the standardized parameters, input the predictive input features into the predictive network using the predicted parameters, destandardize the output of the predictive network, and obtain the self-shrinkage predicted value.

[0026] To more clearly explain the method for predicting the autogenous shrinkage of ultra-high performance concrete of the present invention, the following is in conjunction with... Figure 1 Each step is described in detail.

[0027] S10: Collect concrete samples, use material composition parameters and curing parameters as input features, use measured autogenous shrinkage value as output value, standardize the input features and the output value to obtain standardized input features, standardized measured values ​​and standardized parameters.

[0028] In this embodiment, the material composition parameters include the amount of cement, silica fume, fly ash, sand, mixing water, superplasticizer, superabsorbent polymer, expansion agent, shrinkage retardant, and steel fiber. The curing parameters include the curing age.

[0029] In this embodiment, step S10 includes: Step S11: Each concrete test record corresponds to a concrete sample. The material composition parameters include at least the amount of cement, silica fume, fly ash, sand, mixing water, superplasticizer, superabsorbent polymer, expansion agent, shrinkage retardant, and steel fiber. The curing parameters include at least the curing age. The material composition parameters and curing parameters of the same concrete sample are arranged in a preset order so that each material composition parameter and curing parameter corresponds to a sequence position, forming an input feature sequence corresponding to a measured autogenous shrinkage value. Step S12: Delete samples whose absolute value of the difference between any input feature or output value and the mean of the same term, divided by the standard deviation of the same term, exceeds a preset anomaly threshold. Step S13: Calculate the mean and standard deviation of each input feature and output value of the retained sample, and use the mean and standard deviation as the standardization parameters; Step S14: Subtract the mean of the same item from each value of the retained sample and divide by the standard deviation of the same item to obtain the standardized input features and the standardized measured values; Step S15: Combine the standardized input features and standardized measured values ​​of the same sample into a standardized sample, and randomly divide it into a training set and a test set according to a preset ratio. Train the prediction network with the training set and test the prediction network using the prediction parameters with the test set.

[0030] Concrete samples are linked by sample identifiers. Material composition parameters, curing parameters, and measured autogenous shrinkage values ​​within the same concrete sample correspond to the same mix proportion and the same measurement conditions. In this embodiment, the amounts of cement, silica fume, fly ash, sand, mixing water, superplasticizer, superabsorbent polymer, expanding agent, shrinkage retarder, and steel fiber are treated as independent material composition parameters, and the curing age is used as a curing parameter.

[0031] The preset order in step S11 remains consistent throughout the training and prediction processes. In this embodiment, the order is as follows: cement dosage, silica fume dosage, fly ash dosage, sand dosage, mixing water dosage, superplasticizer dosage, superabsorbent polymer dosage, expanding agent dosage, shrinkage retarder dosage, steel fiber dosage, and curing age. This preset order corresponds one-to-one with the field indices in the standardized parameters and is saved along with the prediction parameters. In this embodiment, the input feature sequence includes ten material component parameters and one curing age, corresponding to eleven sequence positions. The number of sequence positions described above is used to illustrate the specific input structure used in this embodiment and does not constitute a limitation on other material component parameters or curing parameters.

[0032] During data reading, if a field is complete but its order differs from that used during training, it should be rearranged according to the field index. If necessary input features, maintenance parameters, or measured shrinkage values ​​are missing, the corresponding sample should be marked as invalid and not included in the prediction network training. For data recorded using different units of measurement, they should first be converted to the same unit before anomaly identification and standardization.

[0033] Step S12 adopts The criteria identify outlier samples. Specifically, the mean and standard deviation of each input feature and output value are calculated separately. Then, the absolute value of the difference between each sample data point and the mean of the same feature is calculated, and this absolute value is divided by the standard deviation of the same feature. When the result is greater than 3, the sample is considered outlier in the corresponding data item, and the concrete sample containing the outlier data item is deleted; when the result is not greater than 3, the concrete sample is retained. If the standard deviation of a data item is 0, the division is not performed, but the source of the data item is reviewed and supplemented with concrete samples with different values.

[0034] Step S13 involves recalculating the mean and standard deviation of each input feature and output value in the retained concrete samples after deleting the abnormal samples. The recalculated mean and standard deviation are used as the standardized parameters. The initial mean and standard deviation used in the anomaly identification stage are not directly used for subsequent predictions.

[0035] Step S14 uses a standard score method for standardization. For each input feature, the mean of the corresponding input feature is subtracted from the input feature value, and then divided by the standard deviation of the corresponding input feature to obtain the standardized input feature. For the output value, the mean of the output value is subtracted from the measured autogenous shrinkage value, and then divided by the standard deviation of the output value to obtain the standardized measured value. By processing different data items separately, the differences in dimensions and numerical ranges between material usage, curing age, and measured autogenous shrinkage value can be eliminated.

[0036] Step S15 combines the standardized input features and standardized measured values ​​of the same concrete sample into a standardized sample record. In this embodiment, the standardized samples are randomly divided into training and test sets at a ratio of 75% and 25%, respectively. Related records belonging to the same specimen or the same mix proportion sequence are preferably grouped into the same data subset to avoid highly correlated data from simultaneously entering the training and test sets. The test set does not participate in the backpropagation of training loss or the gradient update of trainable parameters; its monitoring loss is used for early stopping judgment and prediction parameter selection.

[0037] When partitioning the dataset, a fixed random number seed can be used, and the data subset corresponding to each sample identifier can be recorded to reproduce the same data partitioning results during repeated training. Standardization parameters, dataset partitioning results, input field order, and unit conversion relationships are recorded as version identifiers, and a correspondence is established with subsequent prediction parameters.

[0038] In this embodiment, other anomaly thresholds and data partitioning ratios can also be set based on the sample size and data distribution. When using other settings, the anomaly handling rules, field order, units of measurement, and standardization parameters used in the same training, testing, and prediction should remain consistent.

[0039] S20: Construct a prediction network consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence.

[0040] In this embodiment, step S20 includes: Step S21: Set up a one-dimensional convolutional network for receiving the standardized input features; Step S22: Connect the output of the one-dimensional convolutional network to the input of the bidirectional long short-term memory network; Step S23: Connect the output of the bidirectional long short-term memory network to the input of the multi-head attention mechanism, connect the output of the multi-head attention mechanism to the input of the global average pooling layer, and input the output of the global average pooling layer into the first fully connected layer and the second fully connected layer in sequence.

[0041] See Figure 2 The feature propagation directions in the prediction network are, in sequence, a one-dimensional convolutional network, a bidirectional long short-term memory network, a multi-head attention mechanism, a global average pooling layer, a first fully connected layer, and a second fully connected layer. The one-dimensional convolutional network receives standardized input features arranged in a preset order and outputs a local feature sequence. The bidirectional long short-term memory network encodes the local feature sequence from two opposite directions and outputs bidirectional sequence features. The multi-head attention mechanism weights the bidirectional sequence features based on the correlation between different sequence positions and outputs weighted sequence features. The global average pooling layer performs average pooling along the sequence position dimension to obtain a fixed-length feature vector. The first and second fully connected layers map the fixed-length feature vector to standardized predicted values.

[0042] The standardized input features of each concrete sample are arranged into a single-channel input sequence according to the preset order determined in step S11. During batch training, multiple input sequences are combined into an input tensor, but each input sequence still maintains its independent sample identifier and field order.

[0043] In one specific implementation, the one-dimensional convolutional network includes three cascaded one-dimensional convolutional layers, with output channels of 128, 64, and 64 respectively. Each one-dimensional convolutional layer is followed by batch normalization and a non-linear mapping that sets negative values ​​to 0. The bidirectional long short-term memory network adopts a single-layer bidirectional structure, with a hidden state size of 128 in one direction. The multi-head attention mechanism uses four attention heads, each outputting a 256-dimensional feature corresponding to a sequence position. A global average pooling layer averages the features at each sequence position along the sequence position dimension to obtain a 256-dimensional fixed-length feature vector. A first fully connected layer maps the 256-dimensional fixed-length feature vector to a 32-dimensional feature, and a second fully connected layer maps the 32-dimensional feature to a 1-dimensional standardized prediction value. The number of layers, channels, hidden state size, number of attention heads, and dimension of the fully connected layer are used to describe a specific network structure and do not constitute a limitation on the prediction network structure.

[0044] The output of a one-dimensional convolutional network preserves the positional order of the original input sequence and feeds the multi-channel response at each position into a bidirectional long short-term memory network. If the number of output channels of the one-dimensional convolutional network differs from the input size of the bidirectional long short-term memory network, the feature dimension can be unified through a linear mapping. This linear mapping only adjusts the feature dimension at each sequence position, without changing the sample correspondence or sequence position.

[0045] The bidirectional states output by the bidirectional long short-term memory network are arranged according to their original sequence positions and then fed into a multi-head attention mechanism to obtain weighted sequence features. Global average pooling is performed on the weighted sequence features along the sequence position dimension to compress the features corresponding to different sequence positions into fixed-length feature vectors. The fixed-length feature vectors are then passed through the first fully connected layer and the second fully connected layer in sequence, ultimately generating a standardized prediction value for each concrete sample.

[0046] Before training begins, the trainable parameters of the one-dimensional convolutional network, bidirectional long short-term memory network, multi-head attention mechanism, first fully connected layer, and second fully connected layer are initialized. The bidirectional long short-term memory network initializes the hidden state and memory state to 0 at the beginning of each independent sample sequence. Batch normalization updates statistics based on the training batch during the training phase, and uses the statistics saved after training is completed during the prediction phase.

[0047] In this embodiment, each concrete test record forms an independent concrete sample. The material composition parameters and curing parameters of the same concrete sample are arranged in a preset order as an input feature sequence, and correspond to a measured autogenous shrinkage value. Test records with the same mix proportion at different curing ages are treated as different concrete samples, and multiple test records are not grouped into the same input sequence according to curing age. The same data organization method is used in the training and prediction phases.

[0048] If the number or meaning of the input feature fields changes, the input field mapping and the input size of the prediction network should be adjusted accordingly, and the network should be retrained to obtain the corresponding prediction parameters. Input data with inconsistent dimensions should not be directly fed into the original prediction network. If the intermediate results of the network are non-numerical, infinite, or have mismatched dimensions, the corresponding batch processing should be stopped, the abnormal location and input sample identifier should be recorded, and the abnormal intermediate results should not be passed into subsequent networks.

[0049] In this embodiment, other numbers of one-dimensional convolutional layers, other hidden state sizes, or other numbers of attention heads can also be used, as long as the one-dimensional convolutional network, bidirectional long short-term memory network, and multi-head attention mechanism are connected in sequence and respectively complete local feature extraction, bidirectional state encoding, and sequence feature weighting.

[0050] S30: The standardized input features are obtained by convolutional nonlinear mapping using a one-dimensional convolutional network.

[0051] In this embodiment, the plurality of cascaded one-dimensional convolutional layers includes three one-dimensional convolutional layers.

[0052] In this embodiment, step S30 includes: Step S31: Arrange the standardized input features into an input feature sequence according to a preset order, and input multiple sequentially connected one-dimensional convolutional layers; Step S32: Each one-dimensional convolutional layer performs a one-dimensional convolution on the input to obtain the channel response, performs batch normalization on the channel response and sets negative values ​​to zero; Step S33: Except for the last one-dimensional convolutional layer, input the processing result of each one-dimensional convolutional layer into the next one-dimensional convolutional layer, and use the processing result of the last one-dimensional convolutional layer as the local feature sequence.

[0053] Step S31 further includes: Step S311: The first one-dimensional convolutional layer maps one input channel to 128 channel responses; In step S312, the second one-dimensional convolutional layer maps the 128-channel response to a 64-channel response; In step S313, the third one-dimensional convolutional layer maps the 64-channel response to the 64-channel response and uses the output of the third one-dimensional convolutional layer as the local feature sequence.

[0054] Step S31 receives the standardized input features formed in step S10. Before inputting, the number of fields, their meanings, and their order are checked, and then the standardized input features are arranged into a single-channel input sequence. Each position in this input sequence corresponds to a material composition parameter or curing parameter, and the mapping relationship between that position and the input field is preserved.

[0055] The first one-dimensional convolutional layer moves the convolutional kernel along the input sequence direction, allowing multiple normalized input features from adjacent positions to jointly participate in the formation of the channel response. Each convolutional kernel forms one output channel, and 128 convolutional kernels form 128 channel responses. The convolutional kernels and biases are updated during training based on the error between the normalized predicted values ​​and the normalized measured values, without pre-specifying the contribution relationships between the various material component parameters.

[0056] After the first one-dimensional convolutional layer generates channel responses, batch normalization is performed on each channel response to ensure that the different channel responses in the same training batch are within a numerical range that facilitates further training. Then, negative values ​​in the batch normalization result are set to 0. The processed 128 channel responses are used as the input to the second one-dimensional convolutional layer.

[0057] The second one-dimensional convolutional layer performs another one-dimensional convolution on the 128-channel response output of the first one-dimensional convolutional layer, mapping it to a 64-channel response. The output of the second one-dimensional convolutional layer is also processed sequentially through batch normalization and zeroing of negative values, and then used as the input of the third one-dimensional convolutional layer.

[0058] The third one-dimensional convolutional layer further maps the 64-channel response output of the second one-dimensional convolutional layer into a 64-channel response. After batch normalization and nonlinear mapping, the output of the third one-dimensional convolutional layer is not fed into the next convolutional layer, but is arranged into a local feature sequence according to the original sequence position and sent to step S40.

[0059] Each one-dimensional convolutional layer can use a convolutional kernel of length 3, a stride of 1, and zero padding at both ends of the sequence to ensure that the length of the output sequence of each layer is consistent with the length of the input sequence. Zero padding at both ends is only used to handle convolutional boundaries and does not indicate the addition of new material composition parameters or change the meaning of the input feature fields.

[0060] During the training phase, batch normalization calculates the corresponding statistics based on the training batches and saves the moving statistics updated with training. In the prediction phase, statistics are not recalculated based on individual test samples; instead, the batch normalized statistics saved in the prediction parameters are used to ensure consistency between the training and prediction phases. To prevent division by zero during batch normalization, a positive stability constant can be added to the standard deviation processing. This processing is only used to ensure numerical stability and does not change the source or subsequent path of the channel response.

[0061] When the number of valid samples in the training batch is insufficient, the training batch can be expanded or a saved moving average statistic can be used. If there are missing values ​​in the input sequence, convolution is not performed; instead, a missing value flag for the input field is returned. If the standardized input features exceed the main distribution range of the training samples, convolution can continue, and an out-of-range flag is written to the execution record for verification of the prediction results.

[0062] In this embodiment, different numbers of one-dimensional convolutional layers, different kernel lengths, and different numbers of channels can also be used to complete local feature extraction. When using other structures, the output of the last one-dimensional convolutional layer is still fed into the bidirectional long short-term memory network as a local feature sequence.

[0063] S40: Encode the local feature sequence using a bidirectional long short-term memory network to obtain bidirectional sequence features.

[0064] In this embodiment, step S40 includes: Step S41: Update the positive long short-term memory state from the first item to the last item along the local feature sequence, and update the reverse long short-term memory state from the last item to the first item. Step S42: Concatenate the forward long short-term memory state and the reverse long short-term memory state at the same sequence position, and perform random deactivation on the concatenation result to obtain the bidirectional sequence feature.

[0065] Step S40 receives the local feature sequence formed in step S30. Each position in the local feature sequence has a corresponding multi-channel local feature, and the relationship between that position and the original input feature or maintenance age record is preserved.

[0066] The forward long short-term memory network (FSMB) starts from the first position of the local feature sequence and reads the local features at each position sequentially. At each position, the FMB updates its state based on the local features at the current position, the hidden state of the previous position, and the memory state of the previous position. The input gate controls the extent to which the local features of the current position are written into the memory state, the forget gate controls the extent to which the memory state of the previous position is retained, and the output gate controls the extent to which the hidden state of the current position is output.

[0067] The reverse long short-term memory network starts from the last position of the local feature sequence, reads the local features of each position in reverse order, and updates the reverse hidden state and reverse memory state based on the local features of the current position and the previous state in the backpropagation direction. The forward long short-term memory network and the reverse long short-term memory network each have independent network parameters and do not share the state of the previous position.

[0068] The initial hidden state and initial memory state in both directions are set to 0, and are initialized separately for each independent concrete sample. The forward long short-term memory network encodes the association between the input fields before the current position and the current position, while the backward long short-term memory network encodes the association between the input fields after the current position and the current position, thereby obtaining bidirectional association information of each sequence position relative to the preceding and following parts of the sequence.

[0069] At each sequence position, the forward hidden state and the backward hidden state are concatenated along the feature dimension. When using a specific structure with a one-way hidden state size of 128, the concatenation result at each position has 256-dimensional features. The concatenation results at each position are still arranged in the original order of the local feature sequence, without changing the position index.

[0070] During the training phase, the connection results are randomly deactivated. In one specific implementation, the random deactivation ratio is 0.6, meaning that some feature elements are randomly masked in each training iteration, and the retained feature elements are scaled accordingly. Random deactivation is only used during the training phase; during the prediction phase, random deactivation is disabled, and the complete connection results are used. The features at each position after random deactivation are combined to form bidirectional sequence features, which are then fed into step S50.

[0071] In this embodiment, the input feature sequences of each concrete sample have the same number of sequence positions, so there is no need to set padding positions to unify the sequence length during batch training. If a concrete sample lacks any necessary input field, the corresponding concrete sample is marked as an invalid sample and is not sent to the prediction network.

[0072] If a non-numerical or numerical overflow occurs during the state update process, the state transmission for the corresponding batch is stopped, the abnormal location and sample identifier are recorded, and the input data, prediction network parameters, and gradients are reviewed in step S60. Abnormal states are not further fed into the multi-head attention mechanism.

[0073] In this embodiment, different hidden state sizes, different random deactivation ratios, or multi-layer bidirectional long short-term memory networks can also be used. When using other implementation methods, the states are still updated along two opposite directions of the local feature sequence, and the forward and reverse states at the same sequence position are combined into bidirectional sequence features.

[0074] S50: The bidirectional sequence features are weighted using a multi-head attention mechanism to obtain weighted sequence features.

[0075] In this embodiment, step S50 includes: Step S51: Perform query projection, key projection and value projection on the bidirectional sequence features in multiple attention heads respectively to obtain query matrix, key matrix and value matrix; Step S52: Multiply the query matrix of each attention head by the transpose of the key matrix and divide by the square root of the dimension of the key vector. Apply the normalized exponential function to the division result to obtain the attention weight. Step S53: Weight the value matrix with the attention weights, concatenate the weighted results of each attention head, and perform a linear mapping to obtain the weighted sequence features; Step S54: Perform global average pooling on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector; Step S55: Input the fixed-length feature vector into the first fully connected layer and the second fully connected layer in sequence to obtain the standardized prediction value; Step S50 receives the bidirectional sequence features output from step S40. Each position in the bidirectional sequence features has a feature vector formed by the forward and reverse states. The multi-head attention mechanism forms attention weights based on the projection parameters obtained during training and the current bidirectional sequence features, without pre-specifying the importance of a certain material component parameter or a certain maintenance age.

[0076] For each attention head, the bidirectional sequence features are processed using a query projection matrix, a key projection matrix, and a value projection matrix, respectively, to obtain a query matrix, a key matrix, and a value matrix. The query matrix represents the information used to retrieve features from other positions at each sequence position, the key matrix represents the information used for query matching at each sequence position, and the value matrix represents the feature content involved in the combination during attention weighting. Each projection matrix is ​​a trainable parameter of the prediction network and is updated in step S60 based on the prediction error.

[0077] Each attention point is weighted according to the following formula: ; Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; The transpose of the key matrix; The key vector dimension is represented; the normalized exponent operation is used to convert the correlation values ​​between different sequence positions into attention weights.

[0078] Specifically, the query matrix is ​​first multiplied by the transpose of the key matrix to obtain the association value between different sequence positions; then the association value is divided by the square root of the key vector dimension; subsequently, the scaled association value is normalized exponentially to make the attention weights corresponding to the same query position all between 0 and 1 and the sum of the weights is 1; finally, the value matrix is ​​weighted using the attention weights to obtain the weighted result of the corresponding attention head.

[0079] Before performing the normalized exponentiation, the maximum value can be subtracted from the associated values ​​corresponding to the same query position, and then the exponentiation operation can be performed on the processed associated values. This process limits the numerical range of the exponentiation operation without changing the relative size between the associated values.

[0080] In this embodiment, the input feature sequences of each concrete sample have the same number of sequence positions, and no padding positions are set for batch alignment; the multi-head attention mechanism calculates attention weights for each valid sequence position.

[0081] In one specific implementation, the multi-head attention mechanism includes four attention heads. Each attention head uses an independent query projection matrix, key projection matrix, and value projection matrix to form associations between sequence positions from different feature subspaces. After each attention head generates a weighted result, all weighted results are connected along the feature dimension, and then a weighted sequence feature is formed through linear mapping.

[0082] The weighted sequence features retain the feature representations corresponding to each sequence position and are input into a global average pooling layer. The global average pooling layer averages the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector; the fixed-length feature vector passes through a first fully connected layer and a second fully connected layer in sequence to obtain the standardized prediction value used in step S60.

[0083] Attention weights can be saved as a readable intermediate result. The saved content may include the sample identifier, attention head number, query position, weighted position, and corresponding attention weight, used to verify whether the multi-head attention mechanism has properly assigned weights to valid positions. This record does not participate in denormalization, nor does it change the self-shrinking prediction value.

[0084] In this embodiment, other numbers of attention heads or other connectable projection dimensions can also be used. In other implementations, each attention head still forms a query matrix, a key matrix, and a value matrix, respectively. Normalized attention weights are formed based on the correlation between the query matrix and the key matrix, and the attention weights are used to weight the value matrix.

[0085] S60: Generate standardized predicted values ​​using the weighted sequence features, and train the prediction network based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters.

[0086] In this embodiment, step S60 includes: Step S61: Use the standardized samples in the training set as training samples and the standardized samples in the test set as early cessation monitoring samples. Step S62: Calculate the squared mean of the difference between the standardized predicted value and the standardized measured value of the training sample to obtain the loss value; Step S63: Calculate the gradient of the loss value. When the gradient norm exceeds a preset clipping threshold, scale the gradient according to the ratio of the preset clipping threshold to the gradient norm. Step S64: An adaptive moment estimation optimizer is used to update the parameters of the prediction network with the scaled gradient or the gradient that does not exceed the preset pruning threshold. Step S641: Apply exponential weighting to the gradient of the current iteration to obtain the first-order moment estimate and the second-order moment estimate; Step S642: Correct the deviation of the first-order moment estimate and the second-order moment estimate according to the current iteration number; Step S643: Divide the first-order moment estimate after bias correction by the sum of the square root of the second-order moment estimate after bias correction and the positive stability constant to obtain the parameter adjustment amount; Step S644: Scale the parameter adjustment amount with the learning rate, and subtract the scaled parameter adjustment amount from the currently trainable parameters; Step S65: After each round of parameter update, calculate the monitoring loss based on the early stop monitoring sample. When the monitoring loss does not decrease for a preset number of consecutive rounds, stop updating and use the prediction network parameter with the minimum monitoring loss as the prediction parameter.

[0087] Step S61 uses the standardized samples in the training set formed in step S10 as training samples and the standardized samples in the test set as early stopping monitoring samples. Both the training samples and the early stopping monitoring samples use the same set of standardized parameters formed in step S13. The early stopping monitoring samples do not participate in the backpropagation of the training loss and gradient update; their monitoring loss is used for early stopping judgment and prediction parameter selection.

[0088] The standardized input features of the training samples are sequentially passed through a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism to form weighted sequence features. Global average pooling is performed on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector. This fixed-length feature vector is then sequentially input into the first fully connected layer and the second fully connected layer to obtain standardized predicted values. Step S62 calculates the difference between the standardized predicted value and the corresponding standardized measured value for each training sample, squares the difference, and averages the squared results in the same training batch. The result is used as the training loss in the form of mean squared error.

[0089] Step S63 calculates the gradient of the training loss relative to each trainable parameter through backpropagation, and calculates the overall norm of all gradients. In this embodiment, the preset clipping threshold can be set to 1.0. When the gradient norm is not greater than 1.0, the original gradient is maintained; when the gradient norm is greater than 1.0, each gradient is multiplied by the ratio of 1.0 to the current gradient norm, so that the clipped gradient norm is 1.0. When the gradient norm is 0, the zero gradient is directly retained.

[0090] Step S64 updates the trainable parameters using the adaptive moment estimator optimizer. In each iteration, the first-order moment estimates and second-order moment estimates are updated based on the current gradient and the state retained from previous iterations. The first-order moment estimates are used to represent the exponentially weighted average of the gradients, and the second-order moment estimates are used to represent the exponentially weighted average of the squared gradients.

[0091] In one specific implementation, the first-order moment decay coefficient is set to 0.9, and the second-order moment decay coefficient is set to 0.999. Since the first-order moment estimates and second-order moment estimates start from 0 in the early stage of training, step S642 corrects the deviation of the two according to the current iteration number to reduce the influence of the initial values ​​on the parameter update.

[0092] Step S643 uses the bias-corrected first-order moment estimate as the parameter adjustment direction and the square root of the bias-corrected second-order moment estimate to adjust the update scale of different parameters. A positive stability constant is added to the denominator to avoid division by zero when the second-order moment estimate is 0. In one specific embodiment, the positive stability constant is set to 10. -8 The first-order moment decay coefficient is 0.9, the second-order moment decay coefficient is 0.999, and the positive stability constant is 10. -8 All use the default values ​​of the Adam optimizer in the PyTorch framework.

[0093] Step S644 uses the learning rate to scale the parameter adjustment and subtracts the scaled parameter adjustment from the current trainable parameters. In one specific implementation, the learning rate is 0.001. The updated trainable parameters are used for the next batch of training until the early stopping condition or the maximum number of training epochs is reached.

[0094] After each training round, forward computation is performed using the early stopping monitoring samples to obtain the monitoring loss, but backpropagation or gradient updates are not performed based on the monitoring loss calculated from the early stopping monitoring samples. During training, the current minimum monitoring loss, corresponding predicted network parameters, and the number of consecutive epochs without improvement are saved. When the monitoring loss in the current round is lower than the saved minimum monitoring loss, the minimum monitoring loss and predicted network parameters are updated, and the number of consecutive epochs without improvement is reset to zero; otherwise, the number of consecutive epochs without improvement is incremented by 1.

[0095] In one specific implementation, the maximum number of training epochs is set to 500, the training batch size is set to 32, and training stops when the monitoring loss does not decrease after 50 consecutive epochs. After training stops, the predicted network parameters saved when the monitoring loss is at its lowest are read, instead of directly using the predicted network parameters from the last epoch.

[0096] During training, random deactivation, second-norm regularization, and weight decay can also be used to suppress overfitting. Random deactivation is used for the bidirectional sequence features formed in step S40, while second-norm regularization or weight decay is used to limit the excessive increase of trainable parameters. The above processing belongs to the training implementation method of the prediction network and does not change the relationship between the standardized predicted values ​​and the standardized measured values ​​established through the mean square error to form the training objective.

[0097] If the training loss, monitoring loss, or gradient becomes non-finite, the parameters for the current batch are not updated, and the network is restored to the most recently valid predicted parameters. If the same anomaly occurs in multiple consecutive batches, training is stopped, and the input data, standardized parameters, and learning rate are reviewed.

[0098] The final predicted parameters include at least the parameters of the one-dimensional convolutional network, the batch normalization parameters, the bidirectional long short-term memory network parameters, the projection parameters of the multi-head attention mechanism, the parameters of the first fully connected layer, and the parameters of the second fully connected layer. The predicted parameters, along with the normalized parameters, the order of the input fields, and the network structure version, are stored together for use in step S70.

[0099] During training, the test set formed in step S10 is used as the early stopping monitoring sample. The test set does not participate in the backpropagation of training loss or the gradient update of trainable parameters; its monitoring loss is used for early stopping judgment and prediction parameter selection. After training stops, the prediction parameters saved when the monitoring loss is at its lowest are used to process the test set to obtain the predicted value of autogenous shrinkage for each concrete sample in the test set, and the corresponding prediction evaluation index is calculated.

[0100] The difference between the predicted and measured shrinkage values ​​can be evaluated using mean square error, root mean square error, mean absolute error, and coefficient of determination. Specifically, smaller mean square error, root mean square error, and mean absolute error indicate smaller prediction errors; a coefficient of determination closer to 1 indicates a higher degree of fit between the predicted and measured shrinkage values.

[0101] See Figure 3 The x-axis represents the measured shrinkage value, and the y-axis represents the predicted shrinkage value. The solid line in the middle indicates the correspondence between the predicted and measured shrinkage values, while the dashed lines on either side represent the ±10% error range. Most scatter points in both the training and test sets are distributed along the solid line in the middle and concentrated around the ±10% error range, indicating that the predicted shrinkage value changes accordingly with the increase or decrease of the measured shrinkage value. For samples with large or small shrinkage values, although there are a few scatter points deviating from the error range, the scatter points in both the training and test sets do not consistently concentrate on the same side of the solid line, indicating that the prediction network does not exhibit a systematic bias of consistently overestimating or underestimating the measured shrinkage value.

[0102] See Figure 4 , Figure 4 Figures (a), (b), (c), and (d) show the mean absolute error, mean square error, coefficient of determination, and root mean square error for different models, respectively. Figure 4 As can be seen from (a), (b), and (d) in the figure, the mean absolute error, mean square error, and root mean square error of the prediction network used in this embodiment are all lower than those of the comparison models on both the training and test sets; Figure 4 As can be seen from (c), the determination coefficient of the prediction network on both the training and test sets is higher than that of the comparative models and closer to 1. The error index on the test set is higher than that on the training set, and the determination coefficient is slightly lower than that on the training set. However, the index changes of different models on the training and test sets are basically consistent, indicating that under the data partitioning method and training parameter settings adopted in this embodiment, the prediction network formed by sequentially connecting a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism has better evaluation metrics on both the training and test sets than the comparative models.

[0103] See Figure 5 The red curve represents the measured auto-shrinkage value, and the blue curve represents the predicted auto-shrinkage value. Figure 5 (a) in the training set shows that the positive and negative changes, peak positions and valley positions of the measured self-shrinkage value and the predicted self-shrinkage value are generally consistent at different sample positions, indicating that the prediction network can learn the corresponding relationship between the self-shrinkage value and the sample based on the material composition parameters and maintenance parameters in the training samples. Figure 5In the test set corresponding to (b), even when the self-shrinkage value of the test sample undergoes continuous changes or local abrupt changes, the predicted self-shrinkage value can still follow the direction of change of the measured self-shrinkage value; although there are some deviations in the positions of individual peaks and valleys, the two curves do not show continuous separation. This indicates that the prediction network can maintain its ability to characterize the trend of self-shrinkage value changes across samples corresponding to different material composition parameters and curing parameters.

[0104] Under the same training set, test set, input features, and data preprocessing conditions, the prediction network can also be compared with extreme gradient boosting models, adaptive boosting models, gradient boosting decision tree models, random forest models, and lightweight gradient boosting models.

[0105] In one set of implementation data, the coefficient of determination for the training set was 0.9915, and the coefficient of determination for the test set was 0.9783. These values ​​are used to illustrate the training and testing status of the prediction network under the corresponding implementation data and do not constitute a limitation on the prediction results for other datasets.

[0106] To verify the rationality of the training parameter settings, training and validation were conducted using both the original parameter configuration and commonly used parameter configurations under the same concrete samples, input features, data preprocessing methods, and prediction network structure. In the original parameter configuration, the random deactivation ratio was 0.6, the global gradient norm pruning threshold was 1.0, and the number of early stopping rounds was 50. The resulting coefficient of determination (COP) for the training set was 0.9915, and for the test set it was 0.9783. In the commonly used parameter configuration, the random deactivation ratio was 0.2, the global gradient norm pruning threshold was 5.0, and the number of early stopping rounds was 20. The resulting COP for the training set was 0.9826, and for the test set it was 0.9641.

[0107] Both sets of parameters were able to complete the prediction network training normally, and the commonly used parameter configurations were reasonable. However, their training set and test set determination coefficients were slightly lower than the original parameter configurations. Given the relatively limited amount of data on the autogenous shrinkage of ultra-high performance concrete, this embodiment uses a random inactivation ratio of 0.6, a global gradient norm pruning threshold of 1.0, and an early stopping waiting round of 50 to reduce the possibility of overfitting during training and to obtain higher training set and test set determination coefficients in this embodiment. The above parameter configurations and evaluation results are only used to illustrate specific training implementation methods and do not constitute a limitation on the scope of protection of this invention.

[0108] S70: Convert the component parameters and curing parameters of the material to be tested into predictive input features using the standardized parameters, input the predictive input features into the predictive network using the predicted parameters, destandardize the output of the predictive network, and obtain the self-shrinkage predicted value.

[0109] In this embodiment, the standardization parameters include the mean and standard deviation of each input feature and output value.

[0110] In this embodiment, step S70 includes: Step S71: Subtract the mean of the same input feature from each of the material composition parameters and maintenance parameters to obtain the predicted input features by dividing by the standard deviation of the same input features. Step S72: Multiply the standardized predicted value output by the prediction network by the standard deviation of the output value and add the mean of the output value to obtain the self-shrinking predicted value.

[0111] Before performing predictions, the standardized parameters, input field order, and network structure bound to the prediction parameter version are read. In this embodiment, the parameters of the material components to be tested include cement dosage, silica fume dosage, fly ash dosage, sand dosage, mixing water dosage, superplasticizer dosage, superabsorbent polymer dosage, expanding agent dosage, shrinkage retarder dosage, and steel fiber dosage. The curing parameters include curing age. The field composition, arrangement order, and units of measurement of the data to be tested are consistent with those in the training phase.

[0112] Step S71 processes the various component parameters and curing parameters of the material to be tested sequentially. For each parameter, the mean of the corresponding input feature in the standardized parameters is first subtracted, and then divided by the standard deviation of the corresponding input feature. The processing results are arranged in the preset order determined in step S11 to form the predicted input features.

[0113] A prediction network using prediction parameters receives the prediction input features. During prediction, the initial hidden state and memory state of the bidirectional long short-term memory network are set to 0, batch normalization uses the statistics saved after training, and random deactivation is disabled. The prediction input features are sequentially passed through a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism to form local feature sequences, bidirectional sequence features, and weighted sequence features. Global average pooling is performed on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector, which is then sequentially input into the first fully connected layer and the second fully connected layer to obtain standardized prediction values.

[0114] Step S72 performs destandardization on the standardized predicted values. Specifically, the standardized predicted value is multiplied by the standard deviation of the output value, and then added to the mean of the output value to obtain the self-shrinking predicted value restored to the dimensions of the measured self-shrinking value. The mean and standard deviation of the output value must be consistent with the dataset version used when forming the current prediction parameters; if the versions are inconsistent, prediction is stopped and a parameter mismatch flag is returned.

[0115] When it is necessary to predict the autogenous shrinkage value of the same material component at multiple curing ages, each curing age can be combined with the parameters of the same material component to form multiple sets of predictive input features, thereby obtaining multiple autogenous shrinkage prediction values, which are then arranged according to the curing age to form an autogenous shrinkage prediction sequence. This process involves repeatedly calling the same prediction workflow and does not change the input-output relationship of a single prediction.

[0116] If the input to be tested lacks a necessary field, the missing field is not replaced with 0, but a field missing flag is returned; if the units of measurement are different, the conversion is completed first according to the preset unit conversion relationship; if the input to be tested exceeds the value range of the corresponding input feature in the training data, the conversion is still performed according to the corresponding standardized parameters, and an out-of-range flag is recorded to indicate that the prediction belongs to the extrapolation result outside the distribution of the training data.

[0117] If any non-finite values ​​appear in the standardized predicted input features, local feature sequences, bidirectional sequence features, weighted sequence features, or standardized predicted values, the destandardization process is stopped, and an anomaly flag is output. Anomaly samples are not merged with the self-shrinking predicted values ​​of normal samples in the output.

[0118] To verify the prediction process, a run log can be generated after each prediction. The run log should include at least the sample identifier, prediction parameter version, normalized parameter version, input field order, predicted input features, attention weight summary, normalized predicted value, shrinking predicted value, and anomaly flag. The run log does not participate in prediction parameter updates, nor does it alter the shrinking predicted value.

[0119] In this embodiment, multiple sets of material composition parameters and curing parameters can be received in batches. During batch prediction, each set of inputs and outputs maintains the same sample identifier to avoid mismatch between the test data and the self-shrinkage prediction value caused by batch rearrangement.

[0120] After obtaining the predicted autogenous shrinkage value, correlation analysis can be performed based on concrete samples. Specifically, Pearson correlation coefficients are calculated between different input features and between input features and measured autogenous shrinkage values, and a correlation heatmap is used to represent the degree of linear association between the corresponding data; a violin plot is used to represent the data distribution range, concentration, and dispersion of each input feature. If any data item has the same value in all samples, the correlation coefficient between that data item and other data items is not calculated, and it is marked as having no effective correlation coefficient.

[0121] It can also calculate the marginal contribution of each input feature to a single self-shrinking prediction value, average the absolute values ​​of the marginal contributions of the same input feature across all samples, and rank the input features by importance based on the average value. A positive marginal contribution indicates that the corresponding input feature increases the current self-shrinking prediction value relative to the baseline prediction value, while a negative marginal contribution indicates that the corresponding input feature decreases the current self-shrinking prediction value relative to the baseline prediction value.

[0122] Furthermore, while keeping other input features according to the concrete sample distribution, one input feature can be changed sequentially, and the corresponding autoshrinkage prediction values ​​can be summarized to form a partial dependency of that input feature. This dependency represents the marginal trend between the change of that input feature and the autoshrinkage prediction value. The above correlation analysis and prediction interpretation results are used to assist in the analysis of the prediction results, but they do not participate in the update of prediction parameters, nor do they change the autoshrinkage prediction value obtained in step S70.

[0123] The first embodiment will be described in detail with specific examples. The following values ​​are for illustrative purposes only and do not represent actual experimental results, optimal mix proportions, or optimal training parameters. For ease of explanation, only one sample from the constructed data and some intermediate results are listed.

[0124] In step S10, the concrete sample uses eleven input features: cement content, silica fume content, fly ash content, sand content, mixing water content, superplasticizer content, superabsorbent polymer content, expansion agent content, shrinkage retarder content, steel fiber content, and curing age. For ease of explanation, only six of these input features—cement content, sand content, mixing water content, superplasticizer content, steel fiber content, and curing age—are selected below to illustrate the standardization process. The original six inputs for a constructed sample are 1000, 1000, 180, 25, 120, and 7, respectively. The example units for the first five quantities are kg / m³. 3 The example unit for maintenance age is days; the remaining input features are standardized in the same way.

[0125] Assuming the mean values ​​of the above 6 input features are 900, 950, 170, 20, 100 and 14 respectively, and the standard deviations are 100, 100, 20, 10, 20 and 7 respectively, then after subtracting the mean from each feature and dividing by the standard deviation, the resulting standardized input features are 1, 0.5, 0.5, 0.5, 1 and -1 respectively.

[0126] Assume the measured shrinkage value of the constructed training sample is The average output value is The standard deviation of the output value is If the measured value is 1, then the standardized value is 1. This record is only a constructed sample used to illustrate the processing procedure; in actual training, the training set is composed of multiple concrete samples.

[0127] In step S20, all eleven standardized input features, including the six input features mentioned above, are arranged into a single-channel input sequence according to a preset order. The example prediction network uses three one-dimensional convolutional layers with 128, 64, and 64 channels respectively; the bidirectional long short-term memory network uses a single-layer bidirectional structure; and the multi-head attention mechanism uses four attention heads.

[0128] In step S30, the first one-dimensional convolutional layer maps the single-channel input sequence to 128-channel responses, the second one-dimensional convolutional layer maps the 128-channel responses to 64-channel responses, and the third one-dimensional convolutional layer maps the 64-channel responses to 64-channel responses. The output of each one-dimensional convolutional layer is sequentially processed by batch normalization and setting negative values ​​to 0. The output of the third one-dimensional convolutional layer forms a local feature sequence.

[0129] In step S40, the forward Long Short-Term Memory (LSTM) network updates its state from the first position to the last position of the local feature sequence, while the backward LSM network updates its state from the last position to the first position. Taking one sequence position as an example, assuming the forward state portion has values ​​of 0.2 and 0.3, and the backward state portion has values ​​of 0.4 and 0.1, the connection results are 0.2, 0.3, 0.4, and 0.1 respectively. During the training phase, the connection results are randomly deactivated, while the prediction phase uses the complete connection results.

[0130] In step S50, taking one attention head as an example, assuming that after the three valid positions are calculated and normalized using the query matrix and key matrix, the attention weights are 0.665, 0.245, and 0.090 respectively, and the sum of the three attention weights is 1. This attention head uses the above attention weights to weight the value features of the three positions. The remaining attention heads generate weighted results respectively. The weighted results of each attention head are connected and linearly mapped to obtain weighted sequence features. Global average pooling is performed on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector. The fixed-length feature vector is then sequentially input into the first fully connected layer and the second fully connected layer to obtain standardized prediction values.

[0131] In step S60, assuming the standardized predicted value of the current constructed sample is 0.8 and the standardized measured value is 1, the square of the difference between the two is 0.04, which is averaged with the squared errors of other samples in the same training batch to form the training loss. Assuming the current gradient norm is 8 and the preset clipping threshold is 1.0, each gradient is multiplied by the ratio of 1.0 to 8, so that the clipped gradient norm is 1.0. The adaptive moment estimation optimizer then updates the trainable parameters based on the clipped gradients. Training continues until the loss has not decreased for 50 consecutive rounds, and the predicted parameters saved when the loss was at its lowest are read.

[0132] In step S70, it is assumed that the standardized prediction value of the prediction network for the output of the test data is 0.8, and the mean of the output value is... The standard deviation of the output value is Then, multiplying 0.8 by the standard deviation of the output value and adding it to the mean of the output value yields an exemplary self-shrinking prediction value. .

[0133] Thus, starting from the example material composition parameters and maintenance parameters, the following steps are completed in sequence: standardization, one-dimensional convolution, bidirectional state encoding, multi-head attention weighting, global average pooling, fully connected mapping, prediction network training, prediction output, and destandardization, ultimately forming the self-shrinking prediction value corresponding to step S70.

[0134] See Figure 6 The second embodiment of the present invention proposes an ultra-high performance concrete autogenous shrinkage prediction system. The system includes: a data processing module configured to collect concrete samples, use material composition parameters and curing parameters as input features, use measured autogenous shrinkage values ​​as output values, and standardize the input features and the output values ​​to obtain standardized input features, standardized measured values ​​and standardized parameters. The prediction network module is configured to construct a prediction network consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence; the one-dimensional convolutional network convolves nonlinearly to map the standardized input features to obtain a local feature sequence; the bidirectional long short-term memory network encodes the local feature sequence to obtain bidirectional sequence features; and the multi-head attention mechanism weights the bidirectional sequence features to obtain weighted sequence features. The training module is configured to generate standardized predicted values ​​using the weighted sequence features, and train the prediction network based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters. The prediction module is configured to convert the component parameters and curing parameters of the material to be tested into prediction input features using the standardized parameters, input the prediction input features into the prediction network using the prediction parameters, de-standardize the output of the prediction network, and obtain the self-shrinkage prediction value.

[0135] Each module is a logical module divided according to its processing function. It can be executed by the same processor or by multiple processors that communicate with each other. The module division method does not change the data processing order, processing basis, and input-output relationship of the method described in the first embodiment.

[0136] The third embodiment of the present invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the method for predicting the autogenous shrinkage of ultra-high performance concrete as described in the first embodiment.

[0137] The memory can be used to store concrete samples, input field order, standardized parameters, prediction parameters, prediction input features, self-shrinkage prediction values, and running records. After the processor reads the computer program from the memory, it sequentially executes sample standardization, prediction network construction, local feature sequence generation, bidirectional sequence feature generation, weighted sequence feature generation, global average pooling, fully connected mapping, prediction network training, test data standardization, and prediction result de-standardization.

[0138] The electronic device can receive material composition parameters, curing parameters, and measured autogenous shrinkage values ​​through a data interface, and output predicted autogenous shrinkage values ​​through a display interface, file interface, or data communication interface. Prediction network training and prediction can be performed by the same electronic device, or the prediction parameters can be generated by the training device and then configured with the prediction parameters and their associated standardized parameters into the electronic device used for prediction.

[0139] When training and prediction devices are separated, the prediction device verifies the prediction parameter version, normalized parameter version, network structure, and input field order before performing prediction. If the verification is consistent, normalization, prediction network processing, and denormalization are performed; if the verification is inconsistent, prediction is not performed to avoid erroneous results due to mismatches in normalized parameters or input field order.

[0140] The processor can be a general-purpose processor, a graphics processor, or other data processor capable of executing the computer program. The memory can be volatile or non-volatile, or it can be composed of a combination of different types of memory cells. The specific hardware type of the electronic device does not change the processing order in which it performs the method described in the first embodiment.

[0141] It should be noted that the module names mentioned above are only used to distinguish different processing functions and do not constitute a limitation on the physical structure or deployment location of the modules. The same module can be divided into multiple processing units, and multiple modules can also be executed by the same processing unit. Unless otherwise expressly defined, the term "comprising" means that it may also include other components not expressly listed, but which do not impede the implementation of the described technical solution.

[0142] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0143] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0144] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the autogenous shrinkage of ultra-high performance concrete, characterized in that, include: Concrete samples were collected, and material composition parameters and curing parameters were used as input features. The measured autogenous shrinkage value was used as the output value. The input features and the output value were standardized to obtain standardized input features, standardized measured values ​​and standardized parameters. A prediction network is constructed by sequentially connecting a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism. The standardized input features are convolved and nonlinearly mapped by the one-dimensional convolutional network to obtain a local feature sequence. The local feature sequence is encoded by the bidirectional long short-term memory network to obtain bidirectional sequence features. The bidirectional sequence features are weighted by the multi-head attention mechanism to obtain weighted sequence features. Standardized predicted values ​​are generated using the weighted sequence features, and the prediction network is trained based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters. The component parameters and curing parameters of the material to be tested are converted into predictive input features using the standardized parameters. The predictive input features are then input into the predictive network using the predicted parameters. The output of the predictive network is de-standardized to obtain the self-shrinkage predicted value.

2. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, Each concrete test record corresponds to one concrete sample. The material composition parameters include at least the amount of cement, silica fume, fly ash, sand, mixing water, superplasticizer, superabsorbent polymer, expansion agent, shrinkage retardant, and steel fiber. The curing parameters include at least the curing age. The material composition parameters and curing parameters of the same concrete sample are arranged in a preset order, so that each material composition parameter and curing parameter corresponds to a sequence position, forming an input feature sequence corresponding to a measured autogenous shrinkage value.

3. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, The standardized input features, the standardized measured values, and the standardized parameters are calculated as follows: Delete samples whose absolute value of the difference between any input feature or output value and the mean of the same term, divided by the standard deviation of the same term, exceeds a preset outlier threshold; Calculate the mean and standard deviation of each input feature and output value of the retained sample, and use the mean and standard deviation as the standardization parameters; Subtract the mean of the same item from each value of the retained sample and divide by the standard deviation of the same item to obtain the standardized input feature and the standardized measured value; The standardized input features and standardized measured values ​​of the same sample are combined to form a standardized sample, which is then randomly divided into a training set and a test set according to a preset ratio. The prediction network is trained using the training set, and the prediction network using the prediction parameters is tested using the test set.

4. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, The specific processing method for the one-dimensional convolutional network is as follows: The standardized input features are arranged into an input feature sequence according to a preset order, and then input into multiple sequentially connected one-dimensional convolutional layers; Each one-dimensional convolutional layer performs a one-dimensional convolution on the input to obtain a channel response, performs batch normalization on the channel response, and sets negative values ​​to zero. Except for the last one-dimensional convolutional layer, the processing result of each one-dimensional convolutional layer is input into the next one-dimensional convolutional layer, and the processing result of the last one-dimensional convolutional layer is used as the local feature sequence.

5. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, The specific processing method of the bidirectional long short-term memory network is as follows: The positive long short-term memory state is updated from the first item to the last item along the local feature sequence, and the negative long short-term memory state is updated from the last item to the first item. By splicing the forward long short-term memory state and the reverse long short-term memory state at the same sequence position, and performing random deactivation on the splicing result, the bidirectional sequence feature is obtained.

6. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, The weighted sequence features and the standardized predicted values ​​are calculated as follows: The bidirectional sequence features are subjected to query projection, key projection, and value projection within multiple attention heads to obtain a query matrix, a key matrix, and a value matrix. Multiply the query matrix of each attention head by the transpose of the key matrix and divide by the square root of the dimension of the key vector. Apply a normalized exponential function to the division result to obtain the attention weight. The value matrix is ​​weighted by the attention weights, the weighted results of each attention head are concatenated and a linear mapping is performed to obtain the weighted sequence features; Global average pooling is performed on the weighted sequence features along the sequence position dimension to obtain a fixed-length feature vector; The fixed-length feature vector is sequentially input into the first fully connected layer and the second fully connected layer to obtain the standardized prediction value.

7. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 3, characterized in that, The prediction parameters are calculated as follows: The standardized samples in the training set are used as training samples, and the standardized samples in the test set are used as early cessation monitoring samples. The loss value is obtained by calculating the squared mean of the differences between the standardized predicted values ​​and the standardized measured values ​​of the training samples. Calculate the gradient of the loss value. When the gradient norm exceeds a preset clipping threshold, scale the gradient according to the ratio of the preset clipping threshold to the gradient norm. An adaptive moment estimation optimizer is used to update the parameters of the prediction network with scaled gradients or gradients that do not exceed the preset pruning threshold; After each round of parameter updates, the monitoring loss is calculated based on the early stop monitoring samples. When the monitoring loss does not decrease for a preset number of consecutive rounds, the update stops, and the prediction network parameters with the minimum monitoring loss are used as the prediction parameters.

8. The method for predicting the autogenous shrinkage of ultra-high performance concrete according to claim 1, characterized in that, The calculation methods for the predicted input features and the self-shrinking predicted values ​​are as follows: The standardized parameters include the mean and standard deviation of each input feature and output value; The predicted input features are obtained by subtracting the mean of the same input feature from each of the material composition parameters and maintenance parameters and then dividing by the standard deviation of the same input feature. The self-shrinking prediction value is obtained by multiplying the standardized prediction value output by the standard deviation of the output value and adding the mean of the output value.

9. A system for predicting the autogenous shrinkage of ultra-high performance concrete, characterized in that, include: The data processing module is configured to collect concrete samples, use material composition parameters and curing parameters as input features, use measured autogenous shrinkage value as output value, standardize the input features and the output value, and obtain standardized input features, standardized measured values ​​and standardized parameters. The prediction network module is configured to construct a prediction network consisting of a one-dimensional convolutional network, a bidirectional long short-term memory network, and a multi-head attention mechanism connected in sequence; the one-dimensional convolutional network convolves nonlinearly to map the standardized input features to obtain a local feature sequence; the bidirectional long short-term memory network encodes the local feature sequence to obtain bidirectional sequence features; and the multi-head attention mechanism weights the bidirectional sequence features to obtain weighted sequence features. The training module is configured to generate standardized predicted values ​​using the weighted sequence features, and train the prediction network based on the error between the standardized predicted values ​​and the standardized measured values ​​to obtain prediction parameters. The prediction module is configured to convert the component parameters and curing parameters of the material to be tested into prediction input features using the standardized parameters, input the prediction input features into the prediction network using the prediction parameters, de-standardize the output of the prediction network, and obtain the self-shrinkage prediction value.

10. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method for predicting the autogenous shrinkage of ultra-high performance concrete as described in any one of claims 1 to 8.

Images