A deep learning-based evaluation method for digital construction projects

Abstract

The application relates to the field of intelligent project evaluation, and discloses an evaluation method for a digital construction project based on deep learning, which specifically comprises the following steps: collecting multi-source data; constructing a continuous spatial field memory matrix indexed by coordinates and time, and mapping and writing the data; establishing a working memory area and a long-term memory area on the matrix; a movable read-write head extracts state loading from the long-term area to the working area; the read-write head receives real-time IoT stream updates to the working area; a memory conversion network monitors the change amplitude of the working area, and generates a conversion vector to write into the long-term area when the amplitude exceeds a threshold; a controller reads the current state of the working area; the controller retrieves historical states from the long-term area; the controller outputs a performance score in combination with the current and historical states, and generates an evaluation report in association with the conversion vector. Through continuous spatial field memory and layered adaptive conversion, the application realizes multi-source data fusion and long-term causal tracing.

Description

Technical Field

[0001] This invention relates to the field of intelligent project evaluation, and more particularly to an evaluation method for digital construction projects based on deep learning. Background Technology

[0002] Digital construction projects involve multi-source data from BIM, IoT sensors, project management software, and other sources. Their comprehensive performance evaluation is crucial for project quality and return on investment. Current evaluation methods mainly rely on manually set indicator systems and fixed weights, and the evaluation process typically consists of three stages: data collection, indicator calculation, and weighted scoring. However, this type of method has the following technical drawbacks: First, existing methods treat project progress, cost, quality, and safety indicators independently at each stage, ignoring the coupling relationships between these indicators in spatial and temporal dimensions. For example, there is a physical correlation between the geometric deviations of components in the BIM model and the vibration data collected by IoT sensors at the construction site, but traditional evaluation models cannot simultaneously integrate geometric spatial data and temporal sensor data.

[0003] Second, existing neural network memory mechanisms use fixed-size discrete matrices to store project states. When the project duration is long, the feature vectors of early critical events (such as design changes and safety accidents) are overwritten by subsequent routine information, resulting in insufficient long-term causal traceability. Although Long Short-Term Memory (LSTM) networks partially alleviate the gradient vanishing problem, their internal memory unit capacity is limited, making it difficult to handle the high-dimensional spatiotemporal data of the entire project lifecycle.

[0004] Third, the weights of most evaluation models are preset at the start of the project and cannot be adaptively adjusted according to dynamic changes during construction, causing the evaluation results to lag behind the actual project status. In particular, when sudden anomalies occur on site (such as excessive structural deformation), traditional models cannot detect them in real time and trigger memory reinforcement operations.

[0005] Therefore, there is an urgent need for a digital project evaluation method that can integrate multi-source heterogeneous spatiotemporal data and has long-term memory and dynamic transformation capabilities. Summary of the Invention

[0006] One objective of this invention is to propose an evaluation method for digital construction projects based on deep learning. This invention constructs a continuous spatial field memory matrix based on a neural Turing machine for each digital construction project. A movable read / write head extracts the spatial field state from the long-term memory and loads it into the working memory. A memory transformation network monitors the temporal change amplitude of the feature vector of the target region in the working memory, triggering the writing of the transformed memory vector into the long-term memory. Furthermore, the working memory is dynamically updated in conjunction with real-time IoT data streams. An evaluation report is generated based on the performance score output by the controller network and the transformed memory vector. This method possesses advantages such as strong multi-source heterogeneous data fusion capability, high accuracy in long-term causal tracing, adaptive dynamic adjustment of evaluation standards, and good interpretability.

[0007] An evaluation method for digital construction projects based on deep learning according to an embodiment of the present invention includes the following steps: Collect multi-source heterogeneous data from digital construction projects; An external memory matrix for a neural Turing machine is constructed. The external memory matrix is ​​a continuous spatial field memory matrix indexed by spatial coordinates x, y, z and time t. This external memory matrix is ​​parameterized by an implicit neural representation network, and multi-source heterogeneous data is mapped and written into the external memory matrix. On the continuous spatial field memory matrix of the neural Turing machine, a working memory area and a long-term memory area are established. The working memory area corresponds to the spatiotemporal window, and the long-term memory area stores the spatial field state. The movable read / write head of the neural Turing machine retrieves the spatial field state from the long-term memory area based on the coordinates of the evaluation task and loads it into the working memory area; The movable read / write head performs read / write operations on the working memory area and receives real-time IoT data streams to update the working memory area. The memory transformation network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area. When the amplitude is greater than a preset threshold, the content of the target region is transformed nonlinearly to generate a transformation memory vector and written into the corresponding spatial coordinates of the long-term memory area. The controller network of the neural Turing machine reads the current state representation of the working memory. The controller network retrieves historical spatial field states from the long-term memory region; The controller network combines the current state with the historical state to output performance scores for each evaluation dimension, and outputs an evaluation report after associating the performance scores with the transformation memory vector.

[0008] Optionally, the multi-source heterogeneous data includes: BIM geometric data, IoT sensor time-series data, schedule data, cost data, and construction log text data; Optionally, the construction of the external memory matrix for the neural Turing machine specifically includes the following steps: The physical space of the digital construction project is discretized into uniform grid points according to a preset sampling resolution. Each grid point is assigned a learnable feature vector with a preset dimension. The project duration is divided into equally spaced time steps, with a preset duration for each time step. Each time step corresponds to a set of learnable feature vectors for uniform grid points. Construct an implicit neural representation network, which includes an input layer, a predetermined number of hidden layers, and an output layer. The number of neurons in the input layer is equal to the number of input parameters, which include spatial coordinates x, y, z, and time step t. Each hidden layer performs a linear transformation on the output of the previous layer and then performs a non-linear mapping through an activation function. The number of neurons in the output layer is equal to the preset dimension of the learnable feature vector, and the output of the last hidden layer is mapped to the learnable feature vector. The collected heterogeneous data from multiple sources are projected into a continuous spatial field according to the data acquisition time and spatial location using a coordinate mapping function. The coordinate mapping function uses an interpolation method, and the heterogeneous data from multiple sources is used as a supervision signal to train the implicit neural representation network. During training, a loss function is used to calculate the error between the learnable feature vector output by the implicit neural representation network and the multi-source heterogeneous data at the projection position. The weight parameters of each layer in the implicit neural representation network are adjusted through the backpropagation algorithm, and the output fits the distribution of multi-source heterogeneous data. The set of output values ​​of the trained implicit neural representation network at all spatiotemporal coordinates is stored as an external memory matrix. The index of the external memory matrix includes the spatial coordinate index dimension, the time step index dimension, and the feature vector dimension index.

[0009] Optionally, establishing the working memory region and the long-term memory region on the continuous spatial field memory matrix of the neural Turing machine specifically involves: The distinction between working memory and long-term memory is based on the time window length threshold T. The working memory area stores the feature vectors of grid points that have been written to by the movable read / write head within the previous T time steps of the current time step t. Each grid point feature vector is accompanied by the last write timestamp. If the difference between the current time step and the last write timestamp is less than or equal to T, then this grid point belongs to the working memory area. The long-term memory region stores the feature vectors of all grid points in the continuous spatial field memory matrix that do not meet the conditions for belonging to the working memory region. That is, grid points whose difference between the current time step and the last written timestamp is greater than T belong to the long-term memory region. When the memory transformation network writes the transformation memory vector into the long memory region, the feature vector of the grid point that was written is synchronously removed from the working memory region at the time of writing.

[0010] Optionally, loading the removable read / write head into the working memory area includes: When the movable read / write head extracts the spatial field state from the long-term memory area, it uses the evaluation task coordinates as the center point and delineates a cubic extraction area according to the preset side length, and extracts the feature vectors of all grid points within the cubic area. The movable read / write head calculates the target storage location in the working memory area for each extracted grid point feature vector according to the spatial coordinates x, y, and z values ​​of the grid point. The target storage location is obtained by dividing the difference between the grid point coordinates and the coordinates of the origin of the working memory area by the sampling resolution to obtain the index values ​​in the three directions. The movable read / write head writes the feature vector of each grid point into the storage cell corresponding to the target storage location in the working memory area. The write operation replaces all the original feature vectors in the storage cell with the currently extracted feature vector. The working memory area is configured with a first-in-first-out queue, which records the write timestamp of the feature vector of each grid point. The working memory area is set with a capacity limit parameter. When the number of grid points stored in the working memory reaches the capacity limit parameter, the movable read / write head traverses the first-in-first-out queue, finds the grid point with the earliest write timestamp, deletes the feature vector corresponding to the grid point from the working memory, and marks the storage unit of the grid point as free. After the deletion operation is completed, the movable read / write head writes the newly extracted grid point feature vector to the free storage unit and records the current time step as the new write timestamp of the grid point.

[0011] Optionally, the removable read / write head performing read / write operations on the working memory area specifically includes the following steps: The movable read / write head receives real-time IoT data streams, queries the mapping table based on sensor identifiers to obtain grid point coordinates, locates the working memory storage unit, and reads feature vectors and concatenates them with measured values ​​to generate temporary vectors. The movable read / write head inputs a temporary vector into the implicit neural representation network, which contains an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function and obtains the updated feature vector from the output layer. The movable read / write head writes the updated feature vector back to the storage unit, overwriting the original vector, and records the grid point coordinates and the current time step into the first-in-first-out queue as the new write timestamp.

[0012] Optionally, the memory transfer network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area, specifically including the following steps: The memory transfer network scans the working memory area at fixed time intervals, and each scan iterates through the grid point coordinates corresponding to all storage units in the working memory area; A feature vector queue is maintained for each grid point. The queue length is fixed to a preset value. The queue stores the feature vectors read from the grid point's working memory storage unit in the most recent scans in chronological order. At the current scan time, read the current feature vector at the grid point in the working memory area, and retrieve the historical feature vector stored at the previous scan time from the feature vector queue of the grid point; Calculate the Euclidean distance between the current feature vector and the historical feature vectors, and use the Euclidean distance as the time variation value of the feature vector at the grid point; The calculated Euclidean distance is compared with a preset threshold. If the Euclidean distance is greater than the preset threshold, the grid point is marked as the target area. For each grid point marked as the target region, all feature vectors are taken from the feature vector queue of the grid point and arranged in chronological order to form a feature vector sequence; The feature vector sequence is input into the gated recurrent unit network, and the gated recurrent unit network outputs a hidden state vector. The hidden state vector is mapped to a transformation memory vector through a linear transformation layer. The dimension of the transformation memory vector is equal to the preset dimension of the feature vector. The transformation memory vector, the target region grid point coordinates, and the current time step are temporarily stored together in the transformation buffer, awaiting the operation instruction to write to the long-term memory area.

[0013] Optionally, the controller network of the neural Turing machine reads the current state representation of the working memory area, specifically including the following steps: The controller network obtains the grid point coordinates recorded in the first-in-first-out queue of the working memory area, extracts the feature vector corresponding to each coordinate according to the writing timestamp order, and arranges them into a feature vector sequence; The feature vector sequence is input into the bidirectional long short-term memory network. The bidirectional long short-term memory network outputs the forward hidden state and the backward hidden state at each position. The forward and backward hidden states are concatenated to obtain the bidirectional context feature vector of each grid point. All bidirectional context feature vectors are arranged into a multidimensional feature tensor according to the spatial coordinates of grid points. The multidimensional feature tensor is then input into a two-dimensional convolutional neural network, and a flattened one-dimensional vector is obtained from the output layer as the current state representation.

[0014] Optionally, the controller network retrieves the historical spatial field state from the long-term memory region, specifically including the following steps: The controller network obtains the set of coordinates of all grid points in the current state representation of the working memory area; Using the coordinates of each grid point as the target location, search for the feature vector sequence on the time axis corresponding to the coordinates in the long-term memory area; According to the timeline, backtrack a preset number of steps from the current time step, and extract all feature vectors of each target coordinate within the backtracking time range; All extracted feature vectors are arranged into a four-dimensional tensor in order of coordinates and time. The dimensions of the four-dimensional tensor are the spatial coordinate x-index, spatial coordinate y-index, spatial coordinate z-index, and time step index, respectively. The controller network outputs a four-dimensional tensor as the retrieved historical spatial field state.

[0015] Optionally, the output evaluation report specifically includes the following steps: The controller network associates performance scores with transformation memory vectors and inserts the transformation memory vectors into the corresponding sections of the evaluation report according to the grid point coordinates and time steps corresponding to the transformation memory vectors. A dimension score table is generated according to each evaluation dimension of the performance score. Each dimension score is accompanied by an explanatory field. The content of the explanatory field is determined by the feature dimension index with the highest activation value in the transformation memory vector. The dimensional score table and the list of spatial coordinates corresponding to the transformed memory vector are combined into structured data, which is in JSON object format. Output structured data to a display terminal or storage medium.

[0016] The beneficial effects of this invention are: First, a continuous spatial field memory matrix with spatial coordinates and time indexes is constructed to uniformly map multi-source heterogeneous data such as BIM geometric data and IoT time series data to the same memory space, eliminating feature fragmentation caused by differences in data types.

[0017] Second, a two-layer architecture of working memory and long-term memory is adopted. By using time window thresholds to distinguish between short-term focused states and long-term storage states, the key event characteristics in the entire project lifecycle are effectively preserved, enabling accurate causal traceability under ultra-long project duration conditions.

[0018] Third, the memory transfer network triggers a transfer operation based on the change amplitude of the Euclidean distance of the feature vector. It writes the contents of the abnormally changed area in the working memory area into the long-term memory area through nonlinear transformation, so as to realize the evaluation model's real-time perception and memory reinforcement of sudden events on site.

[0019] Fourth, the movable read / write head extracts the spatial field state within the cube region based on the coordinates of the evaluation task, and works with the first-in-first-out queue to manage the working memory capacity, reducing computing resource consumption and supporting parallel evaluation of large-scale digital construction projects.

[0020] Fifth, the controller network combines a bidirectional long short-term memory network and a two-dimensional convolutional neural network to extract spatial and temporal features, output performance scores for each evaluation dimension, and simultaneously associate and transform memory vectors to generate evaluation reports with explanatory fields, thereby improving the transparency and auditability of the evaluation results. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for evaluating digital construction projects based on deep learning, as proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction of a continuous spatial field memory matrix for an evaluation method of digital construction projects based on deep learning proposed in this invention. Figure 3 This diagram illustrates the adaptive conversion between working memory and long-term memory in an evaluation method for digital construction projects based on deep learning, as proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 An evaluation method for digital construction projects based on deep learning includes the following steps: Collect multi-source heterogeneous data from digital construction projects; An external memory matrix for a neural Turing machine is constructed. The external memory matrix is ​​a continuous spatial field memory matrix indexed by spatial coordinates x, y, z and time t. This external memory matrix is ​​parameterized by an implicit neural representation network, and multi-source heterogeneous data is mapped and written into the external memory matrix. On the continuous spatial field memory matrix of the neural Turing machine, a working memory area and a long-term memory area are established. The working memory area corresponds to the spatiotemporal window, and the long-term memory area stores the spatial field state. The movable read / write head of the neural Turing machine retrieves the spatial field state from the long-term memory area based on the coordinates of the evaluation task and loads it into the working memory area; The movable read / write head performs read / write operations on the working memory area and receives real-time IoT data streams to update the working memory area. The memory transformation network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area. When the amplitude is greater than a preset threshold, the content of the target region is transformed nonlinearly to generate a transformation memory vector and written into the corresponding spatial coordinates of the long-term memory area. The controller network of the neural Turing machine reads the current state representation of the working memory. The controller network retrieves historical spatial field states from the long-term memory region; The controller network combines the current state with the historical state to output performance scores for each evaluation dimension, and outputs an evaluation report after associating the performance scores with the transformation memory vector.

[0024] In this embodiment, the multi-source heterogeneous data includes: BIM geometric data, IoT sensor time-series data, schedule data, cost data, and construction log text data; Among them, BIM geometric data: 3D spatial data of components stored in the Building Information Model, including component location coordinates, dimensions, shape outline, and spatial topological relationships between components, serving as the core carrier for the digital representation of physical space; IoT sensor time-series data: Time-series data collected and uploaded by various sensors deployed on the construction site at a fixed sampling frequency, such as concrete temperature, tower crane tilt angle, structural vibration acceleration, and steel bar strain values, with each data point accompanied by a timestamp of the collection time; Progress data: Data recording the actual completion status of each process or construction section of the project, such as planned start time, actual start time, planned completion time, actual completion time, and current completion percentage; Cost data: Data on various expenses incurred during project implementation, such as material procurement amounts, equipment rental fees, labor costs, contract payment amounts, and budget execution status; Construction log text data: Unstructured text information recorded daily by on-site management personnel, including the day's construction content, technical problems encountered, handling measures taken, weather conditions, and personnel attendance.

[0025] In this embodiment, constructing the external memory matrix of the neural Turing machine specifically includes the following steps: The physical space of the digital construction project is discretized into uniform grid points according to a preset sampling resolution. Each grid point is assigned a learnable feature vector with a preset dimension. The project duration is divided into equally spaced time steps, with a preset duration for each time step. Each time step corresponds to a set of learnable feature vectors for uniform grid points. Construct an implicit neural representation network, which includes an input layer, a predetermined number of hidden layers, and an output layer. The number of neurons in the input layer is equal to the number of input parameters, which include spatial coordinates x, y, z, and time step t. Each hidden layer performs a linear transformation on the output of the previous layer and then performs a non-linear mapping through an activation function. The number of neurons in the output layer is equal to the preset dimension of the learnable feature vector, and the output of the last hidden layer is mapped to the learnable feature vector. The collected heterogeneous data from multiple sources are projected into a continuous spatial field according to the data acquisition time and spatial location using a coordinate mapping function. The coordinate mapping function uses an interpolation method, and the heterogeneous data from multiple sources is used as a supervision signal to train the implicit neural representation network. During training, a loss function is used to calculate the error between the learnable feature vector output by the implicit neural representation network and the multi-source heterogeneous data at the projection position. The weight parameters of each layer in the implicit neural representation network are adjusted through the backpropagation algorithm, and the output fits the distribution of multi-source heterogeneous data. The set of output values ​​of the trained implicit neural representation network at all spatiotemporal coordinates is stored as an external memory matrix. The index of the external memory matrix includes the spatial coordinate index dimension, the time step index dimension, and the feature vector dimension index.

[0026] The calculation of the projection position in the continuous spatial field using the coordinate mapping function specifically includes the following steps: Obtain a record from multi-source heterogeneous data. This record contains three spatial coordinate values, one timestamp value, and the corresponding measurement value. Determine the spatial sampling interval value and the temporal sampling interval value of the continuous spatial field memory matrix. To calculate the projection index of the data point in the spatial grid, subtract the minimum boundary coordinate value of the physical space from the spatial coordinate value of the data point to obtain the difference. Then divide this difference by the spatial sampling interval and round down to obtain the grid index value in the three directions. Similarly, to calculate the projection time step index of the data point on the time axis, subtract the start time of the project period from the timestamp of the data point to obtain the time difference. Then divide this time difference by the time sampling interval and round down. Trilinear interpolation is performed, with the calculated grid index and time step index as the center. Eight adjacent grid points are selected in the spatial dimension (each index range is increased by one), and two adjacent time steps (the current time step and the next time step) are selected in the time dimension, for a total of sixteen adjacent spatiotemporal points. Calculate the offset ratio of the true coordinates of the data point to the coordinates of each adjacent grid point in each dimension. Subtract the smaller coordinate value of the adjacent grid point from the spatial coordinates of the data point, and then divide by the spatial sampling interval to obtain the offset ratio in this direction. Similarly, subtract the smaller time value of the adjacent time step from the timestamp of the data point, and then divide by the time sampling interval to obtain the time offset ratio. For each of these sixteen adjacent points, calculate its weight coefficient: subtract the offset ratio in each direction from one, and then multiply these factors according to the orientation of the point. For example, the weight of the point located at the smallest corner is the product of the three spatial offset ratios and one minus the time offset ratio, the weight of the point located at the largest corner is the product of the three spatial offset ratios and the time offset ratio, and so on for other positions. The measured value of this data point is used as a supervision signal. When training the implicit neural representation network, the error term corresponding to this data point in the loss function is equal to the difference between the feature vector output by the network and the measured value, and then multiplied by the weight coefficient calculated from this point.

[0027] Specifically, each hidden layer performs a linear transformation on the output of the previous layer, followed by a non-linear mapping using an activation function. Each hidden layer receives the feature vector output from the previous layer as input. It performs matrix multiplication with the preset weight matrix of the current hidden layer to obtain an intermediate result vector. This intermediate result vector is then added to the preset bias vector of the current hidden layer to obtain a linearly transformed result vector. This linearly transformed result vector is input to an activation function, which independently performs a non-linear mapping operation on each element of the vector. The mapped vector is then passed as the output of the current hidden layer to the next layer. The activation function used is the ReLU function, which outputs the original value for elements with input values ​​greater than zero and zero for elements with input values ​​less than or equal to zero.

[0028] Specifically, adjusting the weight parameters of each layer in the implicit neural representation network using the backpropagation algorithm to output a fitted distribution of multi-source heterogeneous data includes: All data records from the multi-source heterogeneous data are input into the implicit neural representation network one by one. The network performs forward propagation calculation according to the weight parameters and bias parameters of each layer to obtain the output feature vector corresponding to each data record. For each data record, calculate the error between the feature vector output by the network and the supervision signal at the projection position of that data record; Substitute all the calculated error values ​​into the loss function to obtain the overall loss value for the current batch; Starting from the output layer, the partial derivatives of the loss value with respect to the weight parameters of each layer are calculated layer by layer according to the chain rule to obtain the gradient value of each weight parameter. Multiply the gradient value of each weight parameter by the preset learning rate to obtain the adjustment amount of that weight parameter; subtract the corresponding adjustment amount from the current value of the weight parameter to obtain the updated weight parameter value; perform the same gradient calculation and parameter update operation on the bias parameter. The updated weight and bias parameters are used for the next forward propagation calculation. The above process is repeated until the loss value converges to the preset range or reaches the preset maximum number of iterations. At this point, the feature vector output by the network statistically fits the distribution of multi-source heterogeneous data in the continuous spatial field.

[0029] In this embodiment, establishing the working memory region and the long-term memory region on the continuous spatial field memory matrix of the neural Turing machine specifically involves: The distinction between working memory and long-term memory is based on the time window length threshold T. The working memory area stores the feature vectors of grid points that have been written to by the movable read / write head within the previous T time steps of the current time step t. Each grid point feature vector is accompanied by the last write timestamp. If the difference between the current time step and the last write timestamp is less than or equal to T, then this grid point belongs to the working memory area. The long-term memory region stores the feature vectors of all grid points in the continuous spatial field memory matrix that do not meet the conditions for belonging to the working memory region. That is, grid points whose difference between the current time step and the last written timestamp is greater than T belong to the long-term memory region. When the memory transformation network writes the transformation memory vector into the long memory region, the feature vector of the grid point that was written is synchronously removed from the working memory region at the time of writing.

[0030] In this embodiment, loading the movable read / write head into the working memory area includes: When the movable read / write head extracts the spatial field state from the long-term memory area, it uses the evaluation task coordinates as the center point and delineates a cubic extraction area according to the preset side length, and extracts the feature vectors of all grid points within the cubic area. The movable read / write head calculates the target storage location in the working memory area for each extracted grid point feature vector according to the spatial coordinates x, y, and z values ​​of the grid point. The target storage location is obtained by dividing the difference between the grid point coordinates and the coordinates of the origin of the working memory area by the sampling resolution to obtain the index values ​​in the three directions. The movable read / write head writes the feature vector of each grid point into the storage cell corresponding to the target storage location in the working memory area. The write operation replaces all the original feature vectors in the storage cell with the currently extracted feature vector. The working memory area is configured with a first-in-first-out queue, which records the write timestamp of the feature vector of each grid point. The working memory area is set with a capacity limit parameter. When the number of grid points stored in the working memory reaches the capacity limit parameter, the movable read / write head traverses the first-in-first-out queue, finds the grid point with the earliest write timestamp, deletes the feature vector corresponding to the grid point from the working memory, and marks the storage unit of the grid point as free. After the deletion operation is completed, the movable read / write head writes the newly extracted grid point feature vector to the free storage unit and records the current time step as the new write timestamp of the grid point.

[0031] The specific steps for setting the preset side length are as follows: Obtain the bounding box dimensions of the BIM components or construction area associated with the assessment task. This bounding box is defined by the minimum and maximum spatial coordinates. Calculate the length differences of the bounding box in three directions and take the maximum value as the spatial influence radius of the task. The grid sampling resolution is multiplied by a configurable multiplier, and the spatial influence radius is compared with this product. The larger of the two values ​​is taken as the half-side length of the cube extraction region.

[0032] In the specific calculation, the evaluation task coordinate positioning point is taken as the center, and a set half-length is extended in the positive and negative directions respectively to obtain a cubic region. If this cubic region exceeds the boundary of the continuous spatial field memory matrix, it is truncated to the matrix boundary range. Another approach is to directly look up the side length value in a table based on the type of assessment task. For example, for a concrete pouring quality assessment task, the side length is 2 meters; for a tower crane overall safety assessment task, the side length is 5 meters; and for a bridge section overall progress assessment task, the side length is 15 meters. These values ​​are obtained through statistical analysis of historical engineering experience and stored as parameters in the system configuration file. In actual operation, the system allows operators to manually adjust the side length parameters through configuration files to adapt to the different needs of various projects. The adjusted values ​​are recorded in the task log for later backtracking and analysis.

[0033] Specifically, obtaining the target storage location includes: The difference between the grid point coordinates and the coordinates of the origin of the working memory area is divided by the sampling resolution to obtain the index values ​​in the three directions.

[0034] In this embodiment, the movable read / write head performs read / write operations on the working memory area, specifically including the following steps: The movable read / write head receives real-time IoT data streams, queries the mapping table based on sensor identifiers to obtain grid point coordinates, locates the working memory storage unit, and reads feature vectors and concatenates them with measured values ​​to generate temporary vectors. The movable read / write head inputs a temporary vector into the implicit neural representation network, which contains an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function and obtains the updated feature vector from the output layer. The movable read / write head writes the updated feature vector back to the storage unit, overwriting the original vector, and records the grid point coordinates and the current time step into the first-in-first-out queue as the new write timestamp.

[0035] Specifically, the movable read / write head receives real-time IoT data streams as follows: The real-time IoT data stream comes from various sensors deployed at the construction site, including: concrete temperature sensors, tower crane inclinometers, structural vibration accelerometers, steel bar strain gauges, ambient temperature and humidity meters, dust concentration detectors, etc. Each sensor is assigned a unique identifier upon deployment, recording the spatial coordinates (x, y, z) of its installation location and the sampling frequency.

[0036] Specifically, the measured values ​​are the physical quantity readings currently collected by the sensors carried in the real-time IoT data stream.

[0037] The sensor identifier lookup mapping table specifically includes: The sensor identifier lookup mapping table is a pre-built two-dimensional table stored in the edge server, used to map the unique identifier of each sensor to its grid point coordinates in the continuous spatial field memory matrix.

[0038] The rows in this mapping table represent each deployed sensor device, with each row corresponding to one sensor. The columns contain at least two columns: the first column is the sensor identifier column, storing the unique number or code assigned to each sensor; the second column is the grid point coordinate column, storing the spatial coordinate values ​​corresponding to the sensor's installation location. Additionally, a third column can be added as needed to record the type of physical quantity measured by the sensor (such as temperature, tilt, vibration, etc.), facilitating data type identification during subsequent processing.

[0039] The contents of this mapping table are obtained during the sensor deployment phase through construction drawings or on-site measurements. For example, if a temperature sensor is assigned the identifier "TEMP_021" and its measured spatial coordinates at its installation location are (12.5, 8.3, 4.0), then the corresponding row in the mapping table would be recorded as follows: "TEMP_021" in the sensor identifier column and the same spatial coordinates in the grid point coordinate column. When the movable read / write head receives an IoT data stream, it parses the sensor identifier from the data record, uses this identifier as the key to search for the corresponding row in the mapping table, and reads the grid point coordinates column of that row. This allows it to obtain the grid point coordinates of the sensor in the continuous spatial field memory matrix, thereby locating the corresponding storage unit in the working memory area.

[0040] In this embodiment, the memory transfer network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area, specifically including the following steps: The memory transfer network scans the working memory area at fixed time intervals, and each scan iterates through the grid point coordinates corresponding to all storage units in the working memory area; A feature vector queue is maintained for each grid point. The queue length is fixed to a preset value. The queue stores the feature vectors read from the grid point's working memory storage unit in the most recent scans in chronological order. At the current scan time, read the current feature vector at the grid point in the working memory area, and retrieve the historical feature vector stored at the previous scan time from the feature vector queue of the grid point; Calculate the Euclidean distance between the current feature vector and the historical feature vectors, and use the Euclidean distance as the time variation value of the feature vector at the grid point; The calculated Euclidean distance is compared with a preset threshold. If the Euclidean distance is greater than the preset threshold, the grid point is marked as the target area. For each grid point marked as the target region, all feature vectors are taken from the feature vector queue of the grid point and arranged in chronological order to form a feature vector sequence; The feature vector sequence is input into the gated recurrent unit network. At each time step, the gated recurrent unit network calculates the reset gate, update gate, and candidate hidden state in sequence, and then outputs the current hidden state. Finally, it outputs the hidden state vector of the last time step.

[0041] The hidden state vector is mapped to a transformation memory vector through a linear transformation layer. The dimension of the transformation memory vector is equal to the preset dimension of the feature vector. The transformation memory vector, the target region grid point coordinates, and the current time step are temporarily stored together in the transformation buffer, awaiting the operation instruction to write to the long-term memory area.

[0042] Specifically, a feature vector queue is maintained for each grid point, and the queue length is fixed at a preset value, following these points: Based on the multiple relationship of the time window threshold T, where T is 10 time steps, the length of the feature vector queue is half of T, i.e. 5 time steps. This ensures that the queue can cover the complete change process of the grid points in the working memory from being written to being about to be removed. It will not miss the change trend due to the queue being too short, nor will it introduce too much irrelevant noise due to the queue being too long. To distinguish between abrupt changes and periodic fluctuations, periodic fluctuations exist in engineering sites (such as the daily periodic changes in concrete temperature caused by diurnal temperature differences). Taking a queue length of 5 time steps can effectively distinguish between single abrupt changes and periodic fluctuations: a single abrupt change will show an isolated high value in the queue and then quickly fall back, while periodic fluctuations will show a regular fluctuation pattern. This length is sufficient to capture the change characteristics of at least one complete fluctuation cycle.

[0043] The trade-off in computational efficiency is that the longer the queue, the more time steps the gated recurrent unit network needs to process, and the amount of computation increases linearly. Taking 5 time steps ensures recall accuracy while keeping the computational latency of each conversion operation within an acceptable range for the edge server. Based on engineering experience, in comparative experiments conducted on three bridge projects of different scales, when the queue length was 3, the false alarm rate for detecting time variation amplitude was 12.7%; when it was 5, the false alarm rate dropped to 4.2%; when it was 7, the false alarm rate only dropped to 3.8%, but the computational load increased by 40%. Therefore, 5 is the most cost-effective value.

[0044] The calculated Euclidean distance is compared with a preset threshold, which is: The data is derived from the statistical analysis of the temporal variation of feature vectors under normal construction conditions in the training dataset of completed projects. Several completed projects are selected from the working memory data during the normal construction phase (time period without any abnormal events). The Euclidean distance between the feature vectors of adjacent time steps is calculated for each grid point. The mean and standard deviation of the distance values ​​under all normal conditions are then calculated. The preset threshold is the mean plus three times the standard deviation. This principle ensures that only 0.3% of the distance values ​​will exceed the threshold under normal construction conditions, thereby controlling the probability of normal fluctuations being misjudged as abnormal at a low level.

[0045] In this embodiment, the controller network of the neural Turing machine reads the current state representation of the working memory area, specifically including the following steps: The controller network obtains the coordinates of the grid points recorded in the first-in-first-out queue of the working memory area, extracts the feature vectors corresponding to each coordinate according to the writing timestamp order, and arranges them into a feature vector sequence; The feature vector sequence is input into the bidirectional long short-term memory network. The bidirectional long short-term memory network outputs the forward hidden state and the backward hidden state at each position. The forward and backward hidden states are concatenated to obtain the bidirectional context feature vector of each grid point. All bidirectional context feature vectors are arranged into a multidimensional feature tensor according to the spatial coordinates of grid points. The multidimensional feature tensor is then input into a two-dimensional convolutional neural network, and a flattened one-dimensional vector is obtained from the output layer as the current state representation.

[0046] In this embodiment, the controller network retrieves the historical spatial field state from the long-term memory region, specifically including the following steps: The controller network obtains the set of coordinates of all grid points in the current state representation of the working memory area; Using the coordinates of each grid point as the target location, search for the feature vector sequence on the time axis corresponding to the coordinates in the long-term memory area; According to the timeline, backtrack a preset number of steps from the current time step, and extract all feature vectors of each target coordinate within the backtracking time range; All extracted feature vectors are arranged into a four-dimensional tensor in order of coordinates and time. The dimensions of the four-dimensional tensor are the spatial coordinate x-index, spatial coordinate y-index, spatial coordinate z-index, and time step index, respectively. The controller network outputs a four-dimensional tensor as the retrieved historical spatial field state.

[0047] Specifically, the step of backtracking from the current time step to the previous time step according to the timeline is as follows: The number of basic backtracking steps is determined based on the total duration of the digital construction project. The preset number of steps can be dynamically adjusted according to the type of assessment task. For example, when the assessment task is "short-term quality risk", the system automatically reduces the number of backtracking steps to 10 steps, because short-term risk only focuses on the state changes in the last few days; when the assessment task is "long-term structural safety trend", the system increases the number of backtracking steps to 80 steps to cover earlier historical data. The preset number of steps is also affected by the time distribution of the transition memory vectors stored in the working memory area. The system counts the time steps corresponding to the transition memory vectors written to the long memory area in the last 10 times, calculates the difference between these time steps and the current time step, and takes the maximum value as the lower limit of the backtracking step. If this lower limit is greater than the preset number of steps, the preset number of steps is replaced by this lower limit to ensure that all recently triggered key events can be retrieved. The system also allows operators to directly specify a fixed number of backtracking steps through a configuration file. This number of backtracking steps is applicable to all evaluation tasks and is 30 steps by default. Operators can adjust it according to the data density and computing resources of the actual project. The adjustment range is limited to 5 to 200 steps. The preset number of steps is read from the configuration file during system initialization and can be modified in real time during operation via the management interface. The modified value takes effect immediately and does not affect submitted but unprocessed search requests. This flexible setting method ensures that an appropriate amount of historical information can be obtained under different project durations and evaluation needs.

[0048] In this embodiment, the output of the evaluation report specifically includes the following steps: The controller network associates performance scores with transformation memory vectors and inserts the transformation memory vectors into the corresponding sections of the evaluation report according to the grid point coordinates and time steps corresponding to the transformation memory vectors. A dimension score table is generated according to each evaluation dimension of the performance score. Each dimension score is accompanied by an explanatory field. The content of the explanatory field is determined by the feature dimension index with the highest activation value in the transformation memory vector. The dimensional score table and the list of spatial coordinates corresponding to the transformed memory vector are combined into structured data, which is in JSON object format. Output structured data to a display terminal or storage medium.

[0049] The controller network associates performance scores with transition memory vectors, specifically through the following steps: The controller network finds the feature dimension with the highest activation value corresponding to the score of each evaluation dimension in the performance score, extracts the explanatory field corresponding to this dimension index from the transformation memory vector, and appends the explanatory field to the dimension score.

[0050] Example 1: The construction period for a cross-river bridge project is planned for 36 months. The construction team needs to monitor the project status in real time, especially the quality of concrete pouring, the safety of tower crane operation, and the matching degree of progress and cost. The traditional approach is to hold a weekly meeting, with the supervisor scoring based on reports. However, this method is prone to missing sudden anomalies, and the data formats of different sections are not uniform, making it difficult to correlate component deviations in the BIM model with vibration data from on-site sensors.

[0051] We employed the method of this invention. First, the physical space of the bridge was divided into uniform grid points with a resolution of 0.5 meters, and each point was assigned a 256-dimensional learnable feature vector. The construction period was divided into hourly intervals, with a time step of 3600 seconds. Next, a multilayer perceptron was constructed, taking coordinates x, y, z and time t as input, and outputting the feature vector of the corresponding grid point. Existing on-site data—including design BIM, concrete temperature sensors, tower crane tiltmeters, rebar arrival records, and daily report texts—were all written into this continuous spatial field memory matrix through coordinate mapping.

[0052] The memory matrix distinguishes between working memory and long-term memory. The time window threshold T is set to 10 hours. The working memory only retains grid points accessed by the read / write head within the last 10 hours. The movable read / write head is positioned based on the evaluation task coordinates. For example, if the current task is to inspect the top of pier No. 3, the state of all grid points within a 10-meter-sided cube centered on these coordinates is extracted and loaded into the working memory.

[0053] The read / write head receives real-time IoT data streams. For example, when the concrete temperature sensor transmits a value of 45.2 degrees Celsius, the system queries the sensor coordinates, locates the corresponding storage unit in the working memory area, reads the current feature vector (representing the stress, humidity, etc. of this point in the previous hour), concatenates it with the temperature value, and sends it to the same multilayer sensor to obtain the updated feature vector, which is then written back to the original location.

[0054] The memory transition network scans the working memory every time step, calculating the Euclidean distance between the feature vector of each grid point and the previous time step. If a grid point experiences a sudden increase in distance value to 0.52 due to an abnormal concrete hydration heat, while the preset threshold is 0.35, the system immediately marks this region as the target region. Then, from the five most recent feature vector sequences stored at this point, the hidden state is extracted using a gated recurrent unit, and a transition memory vector is generated through linear transformation, which is then written to the corresponding coordinates in the long-term memory. In this way, even when reviewing three months later, the system can still locate this hydration heat anomaly event.

[0055] The controller network reads all grid points in the working memory, arranges them into a sequence according to their write timestamps, inputs them into a bidirectional long short-term memory network, and then passes them through a two-dimensional convolutional neural network to obtain the current state representation. Simultaneously, it backtracks 20 time steps from the long-term memory to extract the historical spatial field state. Finally, the current state and historical states are combined to output performance scores across five dimensions: quality, schedule, cost, safety, and digital maturity. The explanatory field corresponding to the feature with the highest activation value in the transition memory vector is appended to the score.

[0056] The entire system is deployed on the construction unit's edge server and runs automatically every day. After six months of continuous monitoring, compared with traditional scoring tables, this invention shows significant advantages in three indicators: abnormal event recall rate, assessment response delay, and long-term causal backtracking accuracy. Specific data are shown in Table 1.

[0057] Table 1: Performance Comparison of Different Evaluation Methods

[0058] Table 1 above lists the average performance data of four evaluation methods after six consecutive months of operation on the same cross-river bridge project. First, let's look at the recall rate of abnormal events. Here, "abnormal events" are defined as four types of sudden events: excessive concrete temperature, excessive tower crane tilt, schedule deviation greater than 7 days, and cost overrun exceeding 5%. These are independently recorded by three senior supervisors as the gold standard. Traditional manual scoring sheets are summarized weekly, but can only capture 32.5% of true anomalies because many anomalies disappear or are not recorded between two meetings. LSTM-based deep evaluation networks can utilize sensor time-series data, improving the recall rate to 67.8%. However, LSTM's forget gate gradually discards key information from two weeks ago, causing early anomalies (such as design changes) to be ignored in later evaluations. While Transformer-based time-series models have stronger long-range dependency capabilities and a recall rate of 74.2%, they require the entire time series to be input at once, resulting in a memory consumption of up to 203 GB·hour / day. Furthermore, they are not friendly to irregular sampling data and sometimes misjudge noise as anomalies.

[0059] This invention couples geometric space and time together using a continuous spatial field memory matrix. The working memory retains the fine-grained state of the most recent 10 hours, while the long-term memory permanently stores all triggered transformation memory vectors. The memory transformation network monitors changes in Euclidean distance, and only truly drastic changes are written to the long-term memory, avoiding the flooding of irrelevant information. This achieves a recall rate of 91.4%, 17.2 percentage points higher than Transformer. Regarding average evaluation response latency, manual scoring takes a week, while LSTM and Transformer both use offline batch processing with latencies of 120 seconds and 85 seconds, respectively. This invention employs real-time streaming processing on an edge server, with the read / write head updating only the affected local grid points each time, resulting in an average latency of only 18 seconds, meeting the second-level early warning requirements of construction sites.

[0060] To assess the accuracy of long-term causal backtracking, we designed a specific test: In the third month of construction, a rebar type error was intentionally created (but went undetected at the time). Then, in the ninth month, when assessing concrete strength, we observed whether the system could correctly trace back to that error. Traditional manual backtracking is almost impossible, with an accuracy of only 18.7%. LSTM, due to memory overwriting, achieved an accuracy of 41.5%. While Transformer can handle long sequences, its uniform attention distribution easily obscures the true cause amidst other information, resulting in an accuracy of 53.2%. Our invention permanently stores the transformation memory vector triggered by that error in long-term memory. During retrieval in the ninth month, the corresponding coordinates and features can be directly located, achieving an accuracy of 84.6%.

[0061] The expert rating consistency rate was calculated by having five senior engineers score the engineering status for the same period, and then comparing it with the model output. This invention achieved a 92.3% consistency rate, indicating that the evaluation results are closest to the expert consensus. Regarding computational resource consumption, this invention used only 67 GB·hour / day, far lower than LSTM and Transformer. This is because the movable read / write head and FIFO queue limit the size of the working memory area, and the memory transformation network only performs nonlinear transformations when changes exceed a threshold, remaining idle most of the time. Overall, this invention significantly improves the real-time performance, accuracy, and traceability of the evaluation while maintaining low resource consumption.

[0062] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An evaluation method for digital construction projects based on deep learning, characterized in that, Includes the following steps: Collect multi-source heterogeneous data from digital construction projects; An external memory matrix for a neural Turing machine is constructed. The external memory matrix is ​​a continuous spatial field memory matrix indexed by spatial coordinates x, y, z and time t. This external memory matrix is ​​parameterized by an implicit neural representation network, and multi-source heterogeneous data is mapped and written into the external memory matrix. On the continuous spatial field memory matrix of the neural Turing machine, a working memory area and a long-term memory area are established. The working memory area corresponds to the spatiotemporal window, and the long-term memory area stores the spatial field state. The movable read / write head of the neural Turing machine retrieves the spatial field state from the long-term memory area based on the coordinates of the evaluation task and loads it into the working memory area; The movable read / write head performs read / write operations on the working memory area and receives real-time IoT data streams to update the working memory area. The memory transformation network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area. When the amplitude is greater than a preset threshold, the content of the target region is transformed nonlinearly to generate a transformation memory vector and written into the corresponding spatial coordinates of the long-term memory area. The controller network of the neural Turing machine reads the current state representation of the working memory. The controller network retrieves historical spatial field states from the long-term memory region; The controller network combines the current state with the historical state to output performance scores for each evaluation dimension, and outputs an evaluation report after associating the performance scores with the transformation memory vector.

2. The evaluation method for digital construction projects based on deep learning according to claim 1, characterized in that, The multi-source heterogeneous data includes: BIM geometric data, IoT sensor time-series data, progress data, cost data, and construction log text data.

3. The evaluation method for digital construction projects based on deep learning according to claim 1, characterized in that, The construction of the external memory matrix for the neural Turing machine specifically includes the following steps: The physical space of the digital construction project is discretized into uniform grid points according to a preset sampling resolution. Each grid point is assigned a learnable feature vector with a preset dimension. The project duration is divided into equally spaced time steps, with a preset duration for each time step. Each time step corresponds to a set of learnable feature vectors for uniform grid points. Construct an implicit neural representation network, which includes an input layer, a predetermined number of hidden layers, and an output layer. The number of neurons in the input layer is equal to the number of input parameters, which include spatial coordinates x, y, z, and time step t. Each hidden layer performs a linear transformation on the output of the previous layer and then performs a non-linear mapping through an activation function. The number of neurons in the output layer is equal to the preset dimension of the learnable feature vector, and the output of the last hidden layer is mapped to the learnable feature vector. The collected heterogeneous data from multiple sources are projected into a continuous spatial field according to the data acquisition time and spatial location using a coordinate mapping function. The coordinate mapping function uses an interpolation method, and the heterogeneous data from multiple sources is used as a supervision signal to train the implicit neural representation network. During training, a loss function is used to calculate the error between the learnable feature vector output by the implicit neural representation network and the multi-source heterogeneous data at the projection position. The weight parameters of each layer in the implicit neural representation network are adjusted through the backpropagation algorithm, and the output fits the distribution of multi-source heterogeneous data. The set of output values ​​of the trained implicit neural representation network at all spatiotemporal coordinates is stored as an external memory matrix. The index of the external memory matrix includes the spatial coordinate index dimension, the time step index dimension, and the feature vector dimension index.

4. The evaluation method for digital construction projects based on deep learning according to claim 1, characterized in that, Specifically, establishing the working memory area and long-term memory area on the continuous spatial field memory matrix of the neural Turing machine involves: The distinction between working memory and long-term memory is based on the time window length threshold T. The working memory area stores the feature vectors of grid points that have been written to by the movable read / write head within the previous T time steps of the current time step t. Each grid point feature vector is accompanied by the last write timestamp. If the difference between the current time step and the last write timestamp is less than or equal to T, then this grid point belongs to the working memory area. The long-term memory region stores the feature vectors of all grid points in the continuous spatial field memory matrix that do not meet the conditions for belonging to the working memory region. That is, grid points whose difference between the current time step and the last written timestamp is greater than T belong to the long-term memory region. When the memory transformation network writes the transformation memory vector into the long memory region, the feature vector of the grid point that was written is synchronously removed from the working memory region at the time of writing.

5. The evaluation method for digital construction projects based on deep learning according to claim 1, characterized in that, The removable read / write head being loaded into the working memory area includes: When the movable read / write head extracts the spatial field state from the long-term memory area, it uses the evaluation task coordinates as the center point and delineates a cubic extraction area according to the preset side length, and extracts the feature vectors of all grid points within the cubic area. The movable read / write head calculates the target storage location in the working memory area for each extracted grid point feature vector according to the spatial coordinates x, y, and z values ​​of the grid point; The movable read / write head writes the feature vector of each grid point into the storage cell corresponding to the target storage location in the working memory area. The write operation replaces all the original feature vectors in the storage cell with the currently extracted feature vector. The working memory area is configured with a first-in-first-out queue, which records the write timestamp of the feature vector of each grid point. The working memory area is set with a capacity limit parameter. When the number of grid points stored in the working memory reaches the capacity limit parameter, the movable read / write head traverses the first-in-first-out queue, finds the grid point with the earliest write timestamp, deletes the feature vector corresponding to the grid point from the working memory, and marks the storage unit of the grid point as free. After the deletion operation is completed, the movable read / write head writes the newly extracted grid point feature vector to the free storage unit and records the current time step as the new write timestamp of the grid point.

6. The evaluation method for a deep learning-based digital construction project according to claim 1, characterized in that, The movable read / write head performs read / write operations on the working memory area, specifically including the following steps: The movable read / write head receives real-time IoT data streams, queries the mapping table based on sensor identifiers to obtain grid point coordinates, locates the working memory storage unit, and reads feature vectors and concatenates them with measured values ​​to generate temporary vectors. The movable read / write head inputs a temporary vector into the implicit neural representation network, and the hidden layer uses the ReLU activation function to obtain the updated feature vector from the output layer; The movable read / write head writes the updated feature vector back to the storage unit, overwriting the original vector, and records the grid point coordinates and the current time step into the first-in-first-out queue as the new write timestamp.

7. The evaluation method for digital construction projects based on deep learning according to claim 1, characterized in that, The memory transfer network of the neural Turing machine monitors the temporal change amplitude of the feature vector of the target region in the working memory area, specifically including the following steps: The memory transfer network scans the working memory area at fixed time intervals, and each scan iterates through the grid point coordinates corresponding to all storage units in the working memory area; A feature vector queue is maintained for each grid point. The queue length is fixed to a preset value. The queue stores the feature vectors read from the grid point's working memory storage unit in the most recent scans in chronological order. At the current scan time, read the current feature vector at the grid point in the working memory area, and retrieve the historical feature vector stored at the previous scan time from the feature vector queue of the grid point; Calculate the Euclidean distance between the current feature vector and the historical feature vectors, and use the Euclidean distance as the time variation value of the feature vector at the grid point; The calculated Euclidean distance is compared with a preset threshold. If the Euclidean distance is greater than the preset threshold, the grid point is marked as the target area. For each grid point marked as the target region, all feature vectors are taken from the feature vector queue of the grid point and arranged in chronological order to form a feature vector sequence; The feature vector sequence is input into the gated recurrent unit network, and the gated recurrent unit network outputs a hidden state vector. The hidden state vector is mapped to a transformation memory vector through a linear transformation layer. The dimension of the transformation memory vector is equal to the preset dimension of the feature vector. The transformation memory vector, the target region grid point coordinates, and the current time step are temporarily stored together in the transformation buffer, awaiting the operation instruction to write to the long-term memory area.

8. The evaluation method for a deep learning-based digital construction project according to claim 1, characterized in that, The controller network of the neural Turing machine reads the current state representation of the working memory area, specifically including the following steps: The controller network obtains the coordinates of the grid points recorded in the first-in-first-out queue of the working memory area, extracts the feature vectors corresponding to each coordinate according to the writing timestamp order, and arranges them into a feature vector sequence; The feature vector sequence is input into the bidirectional long short-term memory network. The bidirectional long short-term memory network outputs the forward hidden state and the backward hidden state at each position. The forward and backward hidden states are concatenated to obtain the bidirectional context feature vector of each grid point. All bidirectional context feature vectors are arranged into a multidimensional feature tensor according to the spatial coordinates of grid points. The multidimensional feature tensor is then input into a two-dimensional convolutional neural network, and a flattened one-dimensional vector is obtained from the output layer as the current state representation.

9. The evaluation method for a deep learning-based digital construction project according to claim 1, characterized in that, The controller network retrieves the historical spatial field state from the long-term memory region, specifically including the following steps: The controller network obtains the set of coordinates of all grid points in the current state representation of the working memory area; Using the coordinates of each grid point as the target location, search for the feature vector sequence on the time axis corresponding to the coordinates in the long-term memory area; According to the timeline, backtrack a preset number of steps from the current time step, and extract all feature vectors of each target coordinate within the backtracking time range; All extracted feature vectors are arranged into a four-dimensional tensor in order of coordinates and time. The dimensions of the four-dimensional tensor are the spatial coordinate x-index, spatial coordinate y-index, spatial coordinate z-index, and time step index, respectively. The controller network outputs a four-dimensional tensor as the retrieved historical spatial field state.

10. The evaluation method for a deep learning-based digital construction project according to claim 1, characterized in that, The output evaluation report specifically includes the following steps: The controller network associates performance scores with transformation memory vectors and inserts the transformation memory vectors into the corresponding sections of the evaluation report according to the grid point coordinates and time steps corresponding to the transformation memory vectors. A dimension score table is generated according to each evaluation dimension of the performance score. Each dimension score is accompanied by an explanatory field. The content of the explanatory field is determined by the feature dimension index with the highest activation value in the transformation memory vector. The dimensional score table and the list of spatial coordinates corresponding to the transformed memory vector are combined into structured data, which is in JSON object format. Output structured data to a display terminal or storage medium.

Images