Concrete setting quality detection method and system based on time-series sensing data

Abstract

This invention provides a method and system for detecting the solidification quality of concrete based on time-series sensing data. The method includes: preprocessing the data to generate an initial feature sequence and mapping it to a matrix of query Q, key K, and value V; for each query, identifying the corresponding solidification stage based on its temporal position, and using learnable parameters associated with the stage to modulate and generate a query vector focused on a specific stage; calculating the similarity between the query and all key vectors, and after temporal smoothing, constructing relevant sparse key-value pairs to generate attention weights, and weighted summing of a subset of sparse values ​​to obtain an attention output containing global temporal dependencies; fusing this global attention vector, the initial feature vector, and the local convolutional context vector to form a multi-scale fusion feature that combines global, local, and original information; and inputting the features into a prediction network to achieve classification or regression evaluation of the concrete solidification quality.

Description

Technical Field

[0001] This application belongs to the field of quality inspection, and in particular relates to a method and system for testing the solidification quality of concrete based on time-series sensor data. Background Technology

[0002] The setting and hardening of concrete involves cement hydration, temperature changes, humidity evolution, and the formation of its internal microstructure. Current concrete quality testing methods, such as compressive strength tests, rebound hammer tests, or ultrasonic tests, cannot accurately reflect the full evolution of the concrete's internal state in real time and continuously. By deploying multiple sensors inside and on the surface of the concrete, multi-source time-series monitoring data reflecting the setting process can be acquired. However, deep learning models analyzing this time-series data face the problem of gradient vanishing or exploding when processing extremely long sequences, making it difficult to establish dependencies between distant time points. The Transformer model, based on a self-attention mechanism, can detect global dependencies by calculating the weights of the relationships between all time points in the sequence. However, when applying the standard self-attention mechanism to the concrete curing quality inspection scenario, the concrete curing process can last for days or even weeks, resulting in long data sequences. The dot product operation between the query matrix and the key matrix in the standard self-attention mechanism has quadratic time and space complexity, leading to high computational costs for model training and inference. Furthermore, the modeling process ignores domain knowledge related to the phased characteristics of the concrete curing process, and the indiscriminate attention calculation method may fail to detect core information at specific stages or miss the detection of minor local changes or abrupt changes. Therefore, how to design an intelligent detection method that can incorporate domain knowledge and take into account both global and local information, adapting to the temporal characteristics of the concrete curing process, is a crucial problem that urgently needs to be solved in the current technological field. Summary of the Invention

[0003] To address the problems of high computational cost in training and inference of existing technology models, failure to incorporate domain knowledge, and difficulty in balancing global and local information.

[0004] In the first aspect, the present invention proposes a method for detecting the solidification quality of concrete based on time-series sensing data, comprising:

[0005] Using multi-source time-series monitoring data of the concrete solidification process as input, preprocessing is performed to generate an initial feature sequence, and feature transformation is used to generate a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation;

[0006] For each query vector, the corresponding concrete solidification stage is identified based on its position in the time sequence. The query vector is modulated using learnable parameters associated with the stage to generate a staged query vector. The similarity score between the staged query vector and all key vectors is calculated to obtain an original similarity score vector. This original similarity score vector is then processed using temporal smoothing to incorporate the influence of neighboring time points. Based on the smoothed scores, a sparse subset of key values ​​relevant to the current query is selectively constructed, and attention weights are generated. These attention weights are used to perform a weighted summation of the sparse subset of values ​​to obtain an attention output vector containing global temporal dependencies.

[0007] The attention output vector, initial feature vector, and context vector extracted through local convolution are fused at the feature level to construct a multi-scale fused feature that combines global, local, and original information. The multi-scale fused feature is then fed into the prediction network to output classification or regression results on the concrete curing quality, thereby realizing intelligent evaluation and monitoring of the curing process.

[0008] In another aspect, the present invention also proposes a concrete solidification quality detection system based on time-series sensor data, comprising the following modules:

[0009] The generation module is used to take multi-source time-series monitoring data of the concrete solidification process as input, perform preprocessing to generate an initial feature sequence, and generate a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation through feature transformation;

[0010] The summation module identifies the corresponding concrete solidification stage for each query vector based on its position in the time sequence. It modulates the query vector using learnable parameters associated with the stage to generate a staged query vector. The module calculates the similarity score between the staged query vector and all key vectors to obtain an original similarity score vector. It then applies temporal smoothing to this original similarity score vector to fuse the influence of neighboring time points. Based on the smoothed scores, it selectively constructs a sparse subset of key values ​​relevant to the current query and generates attention weights. These attention weights are used to perform a weighted summation of the sparse subset to obtain an attention output vector containing global temporal dependencies.

[0011] The output module is used to fuse the attention output vector, the initial feature vector, and the context vector extracted through local convolution at the feature level to construct a multi-scale fused feature that combines global, local, and original information; the multi-scale fused feature is fed into the prediction network to output the classification or regression results on the concrete solidification quality, thereby realizing intelligent evaluation and monitoring of the solidification process.

[0012] This invention associates attention computation with the stages of concrete solidification, enabling the model to focus on analyzing information associated with each stage. By constructing a sparse attention computation set and utilizing temporal smoothing, the computational burden on the model for processing long-sequence data is reduced, and the model's resistance to noise and minor perturbations in the monitoring data is enhanced. Multi-scale features are constructed by fusing global temporal dependencies, local contextual information, and original features. This allows for the detection of long-range and short-range dependencies determining solidification quality from multi-source time-series data, improving the reliability of the concrete solidification quality detection model. Attached Figure Description

[0013] Figure 1 This is a flowchart of the first embodiment. Detailed Implementation

[0014] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0016] In the first embodiment, the present invention proposes a method for detecting the solidification quality of concrete based on time-series sensor data, such as... Figure 1 As shown, it includes:

[0017] S1, taking multi-source time-series monitoring data of the concrete solidification process as input, performs preprocessing to generate an initial feature sequence, and generates a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation through feature transformation;

[0018] Suppose the input multi-source time-series monitoring data is an N×T×D tensor, where N is the number of samples, T is the time step, and D is the dimension of the monitoring features. The StandardScaler function from the sklearn library is used to perform Z-score standardization on each feature dimension, as shown in the formula: , where μ is the mean and σ is the standard deviation. After processing, the initial feature sequence X is obtained. Three linear transformation layers are defined, namely the torch.nn.Linear module in PyTorch, to transform the feature sequence of dimension X. The initial feature sequence X is multiplied by the weight matrix respectively. , , Generate query matrix Key matrix Value matrix .

[0019] In an optional embodiment, the step of preprocessing multi-source time-series monitoring data of the concrete solidification process as input to generate an initial feature sequence includes:

[0020] For each type of monitoring data, including temperature, humidity, and acoustic emission signal amplitude, the minimum value of the data type is subtracted from the value of each data point, and then divided by the difference between the maximum and minimum values ​​of the data type, thereby linearly mapping the monitoring data of different dimensions to the [0,1] interval.

[0021] A time window and a sliding step are set, and the normalized multi-source time series data are slid segmented. The data in each window is used as a feature vector of a time step, thereby constructing the initial feature sequence.

[0022] The preferred normalization method is minimum-maximum scaling, i.e., using the formula The data is linearly mapped to the interval [0,1]. As an alternative implementation, when the data distribution approximates a Gaussian distribution or outliers exist, Z-score normalization can be used, i.e., by using the formula... Transform the data into a distribution with a mean of 0 and a standard deviation of 1. Statistical data used to calculate min, max, or μ, σ should be extracted only from the training set and applied to the training, validation, and test sets to avoid data leakage.

[0023] A sliding window technique was employed. The preferred length of the time window is 60 time points, which is sufficient to capture representative local changes. The preferred sliding step size is 10 time points, with a 50-time-point overlap between adjacent windows. This overlap helps the model learn smooth and continuous temporal evolution features. With each slide, all data from all sources across the 60 time points within the window are integrated into a flattened feature vector, forming a complete initial feature sequence.

[0024] Optionally, to enhance the model's temporal awareness, positional encoding, such as sinusoidal positional encoding or learnable positional encoding, can be superimposed on the initial feature sequence to obtain a position-enhanced feature sequence for use in subsequent steps.

[0025] S2, For each query vector, identify the corresponding concrete solidification stage based on the position of the query vector in the time sequence; modulate the query vector using learnable parameters associated with the stage to generate a staged query vector; calculate the similarity score between the staged query vector and all key vectors to obtain an original similarity score vector, and use temporal smoothing to fuse the influence of neighboring time points on the original similarity score vector. Based on the smoothed scores, selectively construct a sparse key value subset related to the current query and generate attention weights; the attention weights are used to perform weighted summation on the sparse value subset to obtain an attention output vector containing global temporal dependencies.

[0026] The entire solidification time T is divided into S preset stages, such as initial, middle, and final stages. For the query vector at the i-th time point in the sequence... Stage Index Through formula The calculation shows that, among which The length of each stage, T / S, is given by floor, which is a floor function that rounds down.

[0027] Create a learnable stage parameter matrix P with dimension . This can be achieved through torch.nn.Embedding(S, ) Implementation. Based on the stage index obtained in the previous step. Find the corresponding parameter vector from matrix P. The query vector is processed through the Hadamard product, i.e., element-wise multiplication. Modulation is performed to generate staged query vectors. .

[0028] Compute staged query vector The dot product similarity with the entire key matrix K is then scaled to obtain a fractional vector. The calculation formula is: ,in Let be the dimension of the key vector. The resulting score sequence... A one-dimensional convolutional layer, such as torch.nn.Conv1d, with a kernel size of 3, stride of 1, and padding of 1, is fed into the vector to achieve temporal smoothing. Based on the smoothed scores, the torch.topk function is used for each query vector. The indices of the K highest-scoring key vectors are selected, where K is a constant much smaller than the total sequence length T. A sparse subset of key values ​​is constructed based on the selected indices. The Softmax function is applied only to the K selected smoothed scores to generate sparse attention weights.

[0029] Based on the sparse attention weights generated in the previous step The value vector in the corresponding sparse value subset The weighted summation is calculated using the following formula: , where j belongs to the top-K index set, to obtain the attention output vector.

[0030] In an optional embodiment, modulating the query vector using learnable parameters associated with the stage to generate a staged query vector includes:

[0031] Based on the concrete hydration reaction mechanism, the total monitoring time is pre-divided into three time stages: initial setting, intermediate setting, and hardening.

[0032] Initialize a learnable gating vector with the same dimension as the query vector for each of the three time stages mentioned above;

[0033] Based on the timestamp index of the query vector in the initial feature sequence, determine the time stage to which the vector belongs, and obtain the gate vector corresponding to the stage;

[0034] The staged query vector is obtained by multiplying each element in the query vector with the corresponding element in the gating vector.

[0035] The stage division is a fixed division based on prior knowledge. For example, for a 72-hour monitoring process, 0-8 hours can be divided into the initial solidification stage, 8-24 hours into the intermediate solidification stage, and 24-72 hours into the hardening stage. Three learnable parameter vectors are defined for each of these three stages: , , .

[0036] In an alternative embodiment, the stage division is learnable. Instead of fixed time boundaries, a small neural network, such as a single-layer perceptron, can be used. This network takes normalized timestamps as input and outputs a gating vector of the same dimension as the query vector. The model can learn a smooth transition from one stage to another, rather than relying on hard divisions. Assume the query vector... With a dimension of 128, the corresponding gate vector obtained is: Then, by multiplying element by element, i.e., the Hadamard product, the operation... Generate staged query vectors During training, the model learns gating vectors. The value of allows for the amplification or suppression of different feature dimensions of the query vector at different stages.

[0037] The single-layer perceptron is a feedforward neural network. The network input is a normalized timestamp. Let be a scalar. The network structure includes a mapping of the scalar input to... A fully connected layer for a dimensional vector, where This represents the dimension of the query vector. The calculation process for this layer can be derived from the formula... It is represented as, where W is a dimension of [ The learnable weight matrix is ​​defined as [1, 2], where b is a matrix of dimension 1. The learnable bias vector. The network output is a gating vector of the same dimension as the query vector. .

[0038] In an optional embodiment, calculating the similarity score between the staged query vector and all key vectors to obtain an original similarity score vector, and then applying temporal smoothing processing to the original similarity score vector to fuse the influence of neighboring time points, includes:

[0039] Calculate the dot product between the staged query vector and all N key vectors in the key matrix K to obtain an original similarity score vector of length N;

[0040] Define a one-dimensional Gaussian kernel function as a temporal smoothing convolution kernel;

[0041] The one-dimensional Gaussian kernel function is used to perform a one-dimensional convolution operation on the original similarity score vector of length N, updating the score value of each position to the Gaussian weighted average of the position score value and the scores of the neighboring positions, generating a smoothed score vector.

[0042] Assuming a staged query vector The dimension is The dimension of the key matrix K is [T, ]. Through calculation transpose of K Matrix multiplication yields an original similarity score vector with dimensions [1, T]. .

[0043] Preferably, using a one-dimensional Gaussian kernel with a size of 3 and a standard deviation of 1.0, and normalizing the weights to approximately [0.24, 0.52, 0.24], a smooth weighted average can be achieved. Alternatively, a moving average kernel with a size of 3 and weights of [1 / 3, 1 / 3, 1 / 3] can be used. For The j-th score in the sequence is smoothed by taking its own score as a weighted average of the scores of its immediate and neighboring positions. For sequence boundaries, a mirror-filling strategy is used for calculation. This smoothing process suppresses abnormally high scores, ensuring that the attention distribution focuses on regions that consistently exhibit high correlation over time.

[0044] In an optional embodiment, the selective construction of a sparse subset of key values ​​relevant to the current query based on the smoothed scores, and the generation of attention weights, includes:

[0045] Set the sampling size K to an integer, and identify and record the K highest scores and their corresponding indices in the smoothed score vector;

[0046] Based on the sampling quantity K indices, the corresponding key vectors are extracted from the key matrix K to form a sparse key matrix, and the corresponding value vectors are extracted from the value matrix V to form a sparse value matrix.

[0047] The highest scores of the K samples are normalized using the Softmax function to generate the attention weights.

[0048] For a smooth fractional vector of length T Perform the Top-K operation. The value of K is a key hyperparameter; it can be a fixed integer, such as 25, suitable for scenarios with a fixed sequence length. K can also be set as a proportion of the total length T, for example, K = floor(0.1 × T). This operation returns the vector of highest-scoring values. and the corresponding index vector .

[0049] use Perform an indexing operation to extract the corresponding rows from the complete key matrix K and value matrix V, generating a sparse key matrix. Dimension [K, ] and sparse value matrix Dimension [K, ].Will The vector input to the Softmax function transforms the K highest scores into a probability distribution that sums to 1, i.e., attention weights, which have a dimension of [1,K].

[0050] Optionally, a mask vector with the same dimension as the original similarity score vector is generated. The mask values ​​at the corresponding positions of the K indices are set to 0, and the mask values ​​at the remaining positions are set to negative infinity. The smoothed score vector is added to the mask vector, and then the Softmax function is applied for normalization so that the attention weights of the unselected key vectors approach 0, thereby generating sparse attention weights. This can avoid the problem that unselected features will not be optimized during the entire training process.

[0051] S3, the attention output vector, the initial feature vector, and the context vector extracted through local convolution are fused at the feature level to construct a multi-scale fused feature that combines global, local, and original information; the multi-scale fused feature is fed into the prediction network to output the classification or regression results on the concrete solidification quality, thereby realizing intelligent evaluation and monitoring of the solidification process.

[0052] A one-dimensional convolutional neural network layer, such as torch.nn.Conv1d, is used to convolve the initial feature sequence X to extract the local context vector. The global attention output vector, the initial feature vector X, and the local context vector are then concatenated along the feature dimension to obtain a multi-scale fused feature.

[0053] The multi-scale fused features are input into a multilayer perceptron (MLP) consisting of multiple fully connected layers (torch.nn.Linear) and the ReLU activation function (torch.nn.ReLU). For classification tasks, the softmax activation function is used in the last layer to output the class probability. For regression tasks, no activation function is used in the last layer, and the predicted value is output.

[0054] In an optional embodiment, the feature-level fusion of the attention output vector, the initial feature vector, and the context vector extracted through local convolution to construct a multi-scale fused feature that combines global, local, and original information includes:

[0055] For the initial feature vector corresponding to the attention output vector at the same time step, apply a one-dimensional convolutional layer to extract the local context vector;

[0056] In terms of feature dimension, the attention output vector, the initial feature vector, and the local context vector are sequentially concatenated to form the multi-scale fusion feature.

[0057] To detect local temporal dependencies, a one-dimensional convolution is applied to the initial feature sequence with superimposed positional encodings. The kernel size is preferably 5, the stride is 1, and padding is used to ensure the output sequence length is the same as the input. The convolution operation considers a neighboring window centered at the current time step to extract the context vector encoding the local pattern. .

[0058] The one-dimensional convolutional layer is a neural network layer used to process sequential data. The network input is an initial feature sequence with added positional encoding, and the dimension is [T, ], where T is the sequence length, The feature dimension is [T, 64]. This network layer contains one or more one-dimensional convolutional kernels that slide with a stride of 1 along the temporal dimension of the sequence. The parameters of this layer include the number of output channels, i.e., the output feature dimension, preferably 64; the convolutional kernel size, preferably 5; and padding to maintain the same length for both the input and output sequences. The network output is a sequence of local context vectors C, with dimensions [T, 64]. Each vector in the sequence... Each of them encodes the local neighborhood information corresponding to the input position i.

[0059] Feature fusion can be achieved in various ways, a preferred embodiment being vector concatenation: assuming the attention output... Dimensions are 128, initial features Dimensions are 64, local context If the dimension is 64, then the features are fused. The dimension is 256. As an alternative implementation, if the three dimensions are the same, element-wise summation can be used, saving parameters. A more advanced alternative implementation employs a gated fusion mechanism, utilizing learnable gating units to determine the contribution weight of each feature. , where g is a learnable gating vector.

[0060] In an optional embodiment, feeding the multi-scale fused features into the prediction network and outputting classification or regression results regarding the concrete curing quality includes:

[0061] The multi-scale fused features are input into the first fully connected layer, the output dimension is 256, and the modified linear unit activation function is used for transformation.

[0062] The output of the first fully connected layer is input to the second fully connected layer. The output dimension is 3, which corresponds to the three solidification quality categories of "normal", "under-set" and "over-set".

[0063] The output of the second fully connected layer is normalized using the Softmax function to obtain the predicted probability distribution representing the three solidification quality categories, which is used as the classification detection result.

[0064] In order to extract a global representation from the fused features of the entire time series with dimensions [sequence length, 256], global average pooling is used. It averages the entire sequence over the time dimension to obtain a vector with dimensions [1, 256].

[0065] The global representation is fed into the MLP prediction head, which contains one or more hidden layers. For example, a 256-dimensional input passes through a fully connected layer with a 128-dimensional output and a ReLU activation function before being connected to the output layer. To prevent overfitting, a Dropout layer can be added after the fully connected layer. For classification tasks, such as determining normal, under-coagulated, and over-coagulated states, the output layer has three nodes and outputs probabilities for each category using the Softmax function. For regression tasks, such as predicting the specific value of 28-day compressive strength, the output layer has only one node and typically does not use or uses a linear activation function to output the predicted value.

[0066] The prediction network is a neural network consisting of a global average pooling layer and a multilayer perceptron prediction head cascaded together. The network input is sequential multi-scale fused features with dimensions [T, ... ], where T is the sequence length, To fuse feature dimensions, the network structure consists of two main parts. The first part is a global average pooling layer, which averages the input features over the time dimension T to generate a feature with dimension T. The network consists of two parts: the first is a single feature vector; the second part is the multilayer perceptron prediction head, which receives the feature vector and processes it through a series of fully connected layers, including one or more hidden layers, such as a fully connected layer followed by a ReLU activation function, and then an output layer. The network output depends on the specific task; for classification tasks, the number of nodes in the output layer is equal to the number of classes. And apply the Softmax function, the output dimension is The probability vector is calculated using the following formula: , where z is the Logits output of the output layer; if it is a regression task, the number of nodes in the output layer is 1 and no activation function is used, outputting a scalar prediction value.

[0067] In a second embodiment, the present invention also proposes a concrete curing quality detection system based on time-series sensor data, comprising the following modules:

[0068] The generation module is used to take multi-source time-series monitoring data of the concrete solidification process as input, perform preprocessing to generate an initial feature sequence, and generate a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation through feature transformation;

[0069] The summation module identifies the corresponding concrete solidification stage for each query vector based on its position in the time sequence. It modulates the query vector using learnable parameters associated with the stage to generate a staged query vector. The module calculates the similarity score between the staged query vector and all key vectors to obtain an original similarity score vector. It then applies temporal smoothing to this original similarity score vector to fuse the influence of neighboring time points. Based on the smoothed scores, it selectively constructs a sparse subset of key values ​​relevant to the current query and generates attention weights. These attention weights are used to perform a weighted summation of the sparse subset to obtain an attention output vector containing global temporal dependencies.

[0070] The output module is used to fuse the attention output vector, the initial feature vector, and the context vector extracted through local convolution at the feature level to construct a multi-scale fused feature that combines global, local, and original information; the multi-scale fused feature is fed into the prediction network to output the classification or regression results on the concrete solidification quality, thereby realizing intelligent evaluation and monitoring of the solidification process.

[0071] In an optional embodiment, the step of preprocessing multi-source time-series monitoring data of the concrete solidification process as input to generate an initial feature sequence includes:

[0072] For each type of monitoring data, including temperature, humidity, and acoustic emission signal amplitude, the minimum value of the data type is subtracted from the value of each data point, and then divided by the difference between the maximum and minimum values ​​of the data type, thereby linearly mapping the monitoring data of different dimensions to the [0,1] interval.

[0073] A time window and a sliding step are set, and the normalized multi-source time series data are slid segmented. The data in each window is used as a feature vector of a time step, thereby constructing the initial feature sequence.

[0074] In an optional embodiment, modulating the query vector using learnable parameters associated with the stage to generate a staged query vector includes:

[0075] Based on the concrete hydration reaction mechanism, the total monitoring time is pre-divided into three time stages: initial setting, intermediate setting, and hardening.

[0076] Initialize a learnable gating vector with the same dimension as the query vector for each of the three time stages mentioned above;

[0077] Based on the timestamp index of the query vector in the initial feature sequence, determine the time stage to which the vector belongs, and obtain the gate vector corresponding to the stage;

[0078] The staged query vector is obtained by multiplying each element in the query vector with the corresponding element in the gating vector.

[0079] In an optional embodiment, calculating the similarity score between the staged query vector and all key vectors to obtain an original similarity score vector, and then applying temporal smoothing processing to the original similarity score vector to fuse the influence of neighboring time points, includes:

[0080] Calculate the dot product between the staged query vector and all N key vectors in the key matrix K to obtain an original similarity score vector of length N;

[0081] Define a one-dimensional Gaussian kernel function as a temporal smoothing convolution kernel;

[0082] The one-dimensional Gaussian kernel function is used to perform a one-dimensional convolution operation on the original similarity score vector of length N, updating the score value of each position to the Gaussian weighted average of the position score value and the scores of the neighboring positions, generating a smoothed score vector.

[0083] In an optional embodiment, the selective construction of a sparse subset of key values ​​relevant to the current query based on the smoothed scores, and the generation of attention weights, includes:

[0084] Set the sampling size K to an integer, and identify and record the K highest scores and their corresponding indices in the smoothed score vector;

[0085] Based on the sampling quantity K indices, the corresponding key vectors are extracted from the key matrix K to form a sparse key matrix, and the corresponding value vectors are extracted from the value matrix V to form a sparse value matrix.

[0086] The highest scores of the K samples are normalized using the Softmax function to generate the attention weights.

[0087] In an optional embodiment, the feature-level fusion of the attention output vector, the initial feature vector, and the context vector extracted through local convolution to construct a multi-scale fused feature that combines global, local, and original information includes:

[0088] For the initial feature vector corresponding to the attention output vector at the same time step, apply a one-dimensional convolutional layer to extract the local context vector;

[0089] In terms of feature dimension, the attention output vector, the initial feature vector, and the local context vector are sequentially concatenated to form the multi-scale fusion feature.

[0090] In an optional embodiment, feeding the multi-scale fused features into the prediction network and outputting classification or regression results regarding the concrete curing quality includes:

[0091] The multi-scale fused features are input into the first fully connected layer, the output dimension is 256, and the modified linear unit activation function is used for transformation.

[0092] The output of the first fully connected layer is input to the second fully connected layer. The output dimension is 3, which corresponds to the three solidification quality categories of "normal", "under-set" and "over-set".

[0093] The output of the second fully connected layer is normalized using the Softmax function to obtain the predicted probability distribution representing the three solidification quality categories, which is used as the classification detection result.

[0094] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0095] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0096] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0097] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0098] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for detecting the setting quality of concrete based on time-series sensing data, characterized by, Includes the following steps: Using multi-source time-series monitoring data of the concrete solidification process as input, preprocessing is performed to generate an initial feature sequence, and feature transformation is used to generate a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation; For each query vector, the corresponding concrete solidification stage is identified based on the position of the query vector in the time sequence; the query vector is modulated using learnable parameters associated with the stage to generate a staged query vector. The similarity score between the staged query vector and all key vectors is calculated to obtain the original similarity score vector. The original similarity score vector is then processed by temporal smoothing to integrate the influence of adjacent time points. Based on the smoothed scores, a sparse subset of key values ​​related to the current query is selectively constructed, and attention weights are generated. The attention weights are used to perform a weighted summation on a subset of sparse values ​​to obtain an attention output vector that contains global temporal dependencies. The attention output vector, initial feature vector, and context vector extracted through local convolution are fused at the feature level to construct a multi-scale fused feature that combines global, local, and original information. The multi-scale fused feature is then fed into the prediction network to output classification or regression results on the concrete curing quality, thereby realizing intelligent evaluation and monitoring of the curing process.

2. The method of claim 1, wherein, The process of preprocessing multi-source time-series monitoring data of the concrete solidification process to generate an initial feature sequence includes: For each type of monitoring data, including temperature, humidity, and acoustic emission signal amplitude, the minimum value of the class of data is subtracted from the value of each data point, and then divided by the difference between the maximum and minimum values ​​of the class of data, so as to linearly map the monitoring data of different dimensions to the interval [0,1]. A time window and a sliding step are set, and the normalized multi-source time series data are slid segmented. The data in each window is used as a feature vector of a time step, thereby constructing the initial feature sequence.

3. The method of claim 2, wherein, The step of modulating the query vector using learnable parameters associated with the stage to generate a staged query vector includes: Based on the concrete hydration reaction mechanism, the total monitoring time is pre-divided into three time stages: initial setting, intermediate setting, and hardening. Initialize a learnable gating vector with the same dimension as the query vector for each of the three time stages mentioned above; Based on the timestamp index of the query vector in the initial feature sequence, determine the time stage to which the vector belongs, and obtain the gate vector corresponding to the stage; The staged query vector is obtained by multiplying each element in the query vector with the corresponding element in the gating vector.

4. The method according to claim 1, characterized in that, The calculation of the similarity score between the staged query vector and all key vectors yields an original similarity score vector. The original similarity score vector is then processed using temporal smoothing to fuse the influence of neighboring time points, including: Calculate the dot product between the staged query vector and all N key vectors in the key matrix K to obtain an original similarity score vector of length N; Define a one-dimensional Gaussian kernel function as a temporal smoothing convolution kernel; The one-dimensional Gaussian kernel function is used to perform a one-dimensional convolution operation on the original similarity score vector of length N, updating the score value of each position to the Gaussian weighted average of the position score value and the scores of the neighboring positions, generating a smoothed score vector.

5. The method according to claim 1, characterized in that, The process involves selectively constructing a sparse subset of key values ​​relevant to the current query based on the smoothed scores, and generating attention weights, including: Set the sampling size K to an integer, and identify and record the K highest scores and their corresponding indices in the smoothed score vector; Based on the sampling quantity K indices, the corresponding key vectors are extracted from the key matrix K to form a sparse key matrix, and the corresponding value vectors are extracted from the value matrix V to form a sparse value matrix. The highest scores of the sampled quantity K are normalized using the Softmax function to generate the attention weights.

6. The method according to claim 4, characterized in that, The process of fusing the attention output vector, the initial feature vector, and the context vector extracted through local convolution at the feature level to construct a multi-scale fused feature that combines global, local, and original information includes: For the initial feature vector corresponding to the attention output vector at the same time step, apply a one-dimensional convolutional layer to extract the local context vector; In terms of feature dimension, the attention output vector, the initial feature vector, and the local context vector are sequentially concatenated to form the multi-scale fusion feature.

7. The method according to claim 1, characterized in that, The process of feeding the multi-scale fused features into the prediction network and outputting classification or regression results regarding the concrete curing quality includes: The multi-scale fused features are input into the first fully connected layer, the output dimension is 256, and the modified linear unit activation function is used for transformation. The output of the first fully connected layer is input to the second fully connected layer. The output dimension is 3, which corresponds to the three solidification quality categories of "normal", "under-set" and "over-set". The output of the second fully connected layer is normalized using the Softmax function to obtain the predicted probability distribution representing the three solidification quality categories, which is used as the classification detection result.

8. A concrete solidification quality detection system based on time-series sensor data, characterized in that, Includes the following modules: The generation module is used to take multi-source time-series monitoring data of the concrete solidification process as input, perform preprocessing to generate an initial feature sequence, and generate a query matrix Q, a key matrix K, and a value matrix V for self-attention calculation through feature transformation; The summation module is used to identify the corresponding concrete curing stage for each query vector based on its position in the time sequence. The query vector is modulated using learnable parameters associated with the stage to generate a staged query vector; The similarity score between the staged query vector and all key vectors is calculated to obtain the original similarity score vector. The original similarity score vector is then processed by temporal smoothing to integrate the influence of adjacent time points. Based on the smoothed scores, a sparse subset of key values ​​related to the current query is selectively constructed, and attention weights are generated. The attention weights are used to perform a weighted summation on a subset of sparse values ​​to obtain an attention output vector that contains global temporal dependencies. The output module is used to fuse the attention output vector, the initial feature vector, and the context vector extracted through local convolution at the feature level to construct a multi-scale fused feature that combines global, local, and original information; the multi-scale fused feature is fed into the prediction network to output the classification or regression results on the concrete solidification quality, thereby realizing intelligent evaluation and monitoring of the solidification process.

9. The system according to claim 8, characterized in that, The process of preprocessing multi-source time-series monitoring data of the concrete solidification process to generate an initial feature sequence includes: For each type of monitoring data, including temperature, humidity, and acoustic emission signal amplitude, the minimum value of the class of data is subtracted from the value of each data point, and then divided by the difference between the maximum and minimum values ​​of the class of data, so as to linearly map the monitoring data of different dimensions to the interval [0,1]. A time window and a sliding step are set, and the normalized multi-source time series data are slid segmented. The data in each window is used as a feature vector of a time step, thereby constructing the initial feature sequence.

10. The system according to claim 8, characterized in that, The step of modulating the query vector using learnable parameters associated with the stage to generate a staged query vector includes: Based on the concrete hydration reaction mechanism, the total monitoring time is pre-divided into three time stages: initial setting, intermediate setting, and hardening. Initialize a learnable gating vector with the same dimension as the query vector for each of the three time stages mentioned above; Based on the timestamp index of the query vector in the initial feature sequence, determine the time stage to which the vector belongs, and obtain the gate vector corresponding to the stage; The staged query vector is obtained by multiplying each element in the query vector with the corresponding element in the gating vector.

Images