Method and system for automatically generating supervision log for construction supervision

Abstract

The present invention relates to the field of data processing, and in particular to a method and system for automatically generating a supervision log for construction supervision. The automatic generation system for a supervision log for construction supervision comprises: a supervision data acquisition module, a text data construction module, a text feature extraction module, a supervision entity label output module and a supervision log generation module. The present invention acquires voice data and supervision image data, generates supervision entity labels based on the voice data and the supervision image data, and then dynamically selects a log template according to the supervision entity label, and finally fills the selected log template with the supervision entity label to construct a supervision log. The dynamic log template can be adapted to different supervision log situations, which is convenient for subsequent users to view the supervision log; and in the second entity extraction model, the temporal feature extraction of the supervision image is strengthened based on the text feature, so that the recognition of the supervision image pays more attention to the content corresponding to the voice data.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method and system for automatically generating a supervision log for construction supervision. Background Art

[0002] In the process of construction project management, the supervision log is an important tool for project supervisors to record the daily work conditions, quality inspections, safety conditions, etc. of the construction site. The supervision log is not only the basis for monitoring the progress, quality and safety of the project, but also a reference for the completion acceptance and responsibility tracing of the project in the future. However, the existing supervision log template is usually a fixed format, requiring supervisors to fill it out according to the preset items and steps. Although the design of this fixed template can standardize the recording process, it also has many shortcomings.

[0003] First, the actual situation of construction projects is complex and changeable, with different stages and types of construction content, environmental conditions and safety requirements. Fixed-format log templates cannot flexibly adapt to these changes, causing supervisors to often have to make additional adjustments or add supplementary information outside the template when filling out the log, which increases the workload and is prone to omissions or inaccurate records. Summary of the invention

[0004] In order to solve the problem that fixed-format log templates cannot flexibly adapt to changes, the present invention provides a method and system for automatically generating supervision logs for construction supervision, which obtains voice data and supervision image data, generates supervision entity labels based on the voice data and supervision image data, and then dynamically selects log templates according to the supervision entity labels. Finally, the selected log templates are filled with supervision entity labels to construct supervision logs. The dynamic log template can be adapted to different supervision log situations, making it convenient for subsequent users to view the supervision logs. In addition, in the second entity extraction model, the temporal feature extraction of supervision images is enhanced based on text features, so that the recognition of supervision images pays more attention to the content corresponding to the voice data, thereby improving the accuracy of the output of supervision entity labels.

[0005] The present invention provides a method for automatically generating a supervision log for construction supervision, comprising:

[0006] Responding to user operations, obtaining supervision data, the supervision data including voice data and supervision image data, and the supervision image data including a plurality of frames of supervision images input by the user and sorted by time;

[0007] Convert the voice data in the supervision data into text data through voice recognition technology;

[0008] Perform named entity extraction on the text data through the first entity extraction model, output the supervision entity label and the corresponding supervision entity label vector, and concatenate the supervision entity label vectors from top to bottom in the order of appearance to construct text features;

[0009] The text features and the supervision image data are sent to the second entity extraction model for processing, and the supervision entity label and the corresponding supervision entity label vector are output. In the process of processing the text features and the supervision image data, the second entity extraction model strengthens the temporal feature extraction of the supervision image based on the text features;

[0010] The supervision entity label vectors output by all the second entity extraction models are spliced ​​to construct the supervision analysis data, and the supervision analysis data is sent to the log template selection model for processing to output the log template label, which is used to select the corresponding log template; the corresponding log template is selected through the log template label, and the supervision entity labels output by all the second entity extraction models are used to fill the selected log template to construct the supervision log.

[0011] As a preferred aspect, the first entity extraction model is established based on the BiLSTM-CRF model, including a preprocessing layer, a word embedding layer, a BiLSTM layer and a CRF layer;

[0012] The text data is processed by the first entity extraction model to output the supervision entity label and the corresponding supervision entity label vector, which specifically includes the following steps:

[0013] The text data is input into the preprocessing layer for processing, specifically performing word segmentation and deleting stop words; the words output by the preprocessing layer are embedded in words, and the word embedding operation adopts the Word2vec model to construct several word vectors; all the word vectors are then input into the BiLSTM layer in order, and finally processed through the CRF layer to output the supervision entity label, and the word vector corresponding to the supervision entity label is used as the supervision entity label vector corresponding to the supervision entity label.

[0014] As a preferred aspect, the second entity extraction model includes a temporal feature extraction layer, a multi-task recognition layer and a supervision entity label output layer, wherein the temporal feature extraction layer is established based on the ConvLSTM model, and is used to extract temporal features based on text features and supervision image data, during which the temporal feature extraction of the supervision image is enhanced based on the text features to construct image features; the multi-task recognition layer includes N multi-task recognition units, each of which performs a corresponding task recognition and generates a supervision entity label; the supervision entity label output layer is used to output the supervision entity label, and the word vector corresponding to the supervision entity label is used as the supervision entity label vector corresponding to the supervision entity label.

[0015] As a preferred aspect, temporal feature extraction is performed based on text features and supervision image data, during which the temporal feature extraction of the supervision image is enhanced based on the text features, specifically including the following steps:

[0016] The temporal feature extraction layer is established based on the ConvLSTM model. The supervision images in the supervision image data are sequentially input into the temporal feature extraction layer to realize temporal feature extraction. The difference from the ConvLSTM model is that the weight matrix in the input gate of the temporal feature extraction layer is a self-attention weight matrix constructed based on text features and supervision images.

[0017] The self-attention weight matrix is ​​constructed based on text features and supervision images, which specifically includes the following steps: multiplying the supervision images with the key weight matrix to construct the supervision image key matrix K, multiplying the text features with the query weight matrix to construct the supervision image query matrix Q, and calculating the self-attention weight matrix ATT=softmax(QK T / (D) 0.5 ), where softmax is the normalized exponential function, T is the matrix transpose operation, and D is the dimension of the supervision image.

[0018] As a preferred aspect, training the first entity extraction model specifically includes the following steps:

[0019] Obtain several text training samples, annotate the text training samples with supervision entity labels, form all annotated text training samples into a text training set, and then send the text training set to the first entity extraction model with parameter initialization for training to determine whether the training conditions are met. If the training conditions are met, output the trained first entity extraction model; otherwise, continue to train the first entity extraction model through the text training set.

[0020] As a preferred aspect, training the second entity extraction model specifically includes the following steps:

[0021] A number of image training samples are obtained, which include text features and supervision image data, and the image training samples are labeled with supervision entity labels; all labeled image training samples are combined into an image training set, and a second entity extraction model with initialized parameters is trained with the image training set to determine whether the training conditions are met. If the training conditions are met, the trained second entity extraction model is output; otherwise, the second entity extraction model is continued to be trained with the image training set.

[0022] As a preferred aspect, the log template selection model is established based on the BP neural network model, and the training of the log template selection model includes the following steps:

[0023] Obtain several log template selection training samples, which include several supervision entity label vectors, and label the log template selection training samples through log template labels; combine all labeled log template selection training samples into a log template selection training set, and then send the log template selection training set to the log template selection model with parameter initialization for training to determine whether the training conditions are met. If so, output the trained log template selection model; otherwise, continue to train the log template selection model through the log template selection training set.

[0024] As a preferred aspect, after building the supervision log, the log template is also adjusted. The specific steps are as follows:

[0025] Get user operations, adjust the supervision log in response to the user operations, generate a new log template copy for the adjusted supervision log, and each time an adjustment is made, if there is an existing log template copy, do not add a new log template copy; otherwise, add a new log template copy; at the same time, count the number of times the log template copy is used, and whenever the corresponding log template is selected through the log template label, the log template corresponding to the log template label and the one with the highest number of uses among several log template copies will be selected as the log template.

[0026] The present invention also provides a supervision log automatic generation system for construction supervision, comprising:

[0027] A supervision data acquisition module, used to acquire supervision data in response to user operations, the supervision data including voice data and supervision image data, and the supervision image data including a number of supervision image frames input by the user and sorted by time;

[0028] A text data construction module is used to convert the voice data in the supervision data into text data through voice recognition technology;

[0029] A text feature extraction module is used to extract named entities from text data through a first entity extraction model, output a supervision entity label and a corresponding supervision entity label vector, and concatenate the supervision entity label vectors from top to bottom in order of appearance to construct text features;

[0030] A supervision entity label output module is used to send text features and supervision image data to the second entity extraction model for processing, and output the supervision entity label and the corresponding supervision entity label vector. In the process of processing the text features and supervision image data, the second entity extraction model strengthens the temporal feature extraction of the supervision image based on the text features;

[0031] The supervision log generation module is used to splice the supervision entity label vectors output by all the second entity extraction models, construct the supervision analysis data, and send the supervision analysis data to the log template selection model for processing, and output the log template label, which is used to select the corresponding log template; select the corresponding log template through the log template label, and fill the selected log template with the supervision entity labels output by all the second entity extraction models to construct the supervision log.

[0032] The present invention has the following advantages:

[0033] 1. The present invention obtains voice data and supervision image data, generates supervision entity labels based on the voice data and supervision image data, and then dynamically selects a log template according to the supervision entity label. Finally, the supervision entity label is filled in the selected log template to construct a supervision log. The dynamic log template can be adapted to different supervision log situations, which is convenient for subsequent users to view the supervision log; and in the second entity extraction model, the temporal feature extraction of the supervision image is enhanced based on the text feature, so that the recognition of the supervision image pays more attention to the content corresponding to the voice data, thereby improving the accuracy of the supervision entity label output.

[0034] 2. The present invention calculates the self-attention weight matrix through text features and supervision images, and retains the features of the voice data in the supervision image through the calculated self-attention weight matrix in the input gate, which can make more attention be paid to the content corresponding to the voice data in the process of temporal feature extraction, improve the accuracy of supervision entity label extraction, and then improve the accuracy of supervision log generation.

[0035] 3. The present invention also adaptively selects the log template according to the user's usage habits. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a schematic diagram of the structure of a system for automatically generating a supervision log for construction supervision used in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0038] Embodiment 1, a method for automatically generating a supervision log for construction supervision, comprising:

[0039] Respond to user operations to obtain supervision data. It should be noted that during the construction process, supervisors will conduct supervision work through surveillance cameras and on-site observations. Supervisors can input voice data and image data as supervision data to automatically generate logs. That is, supervision data includes voice data and supervision image data, and supervision image data includes several frames of supervision images input by the user in order of time. Supervision images can also be obtained through videos obtained by surveillance cameras.

[0040] The voice data in the supervision data is converted into text data through voice recognition technology. It should be noted that the voice recognition technology used in this embodiment is the wav2vec 2.0 model. The wav2vec 2.0 model is a deep learning model built based on the Transformer model, which can realize voice recognition. The specific operation refers to the existing technology and will not be repeated here. In addition, this application does not limit the specific implementation means of the voice recognition technology.

[0041] The text data is subjected to named entity extraction through the first entity extraction model, and the supervision entity label and the corresponding supervision entity label vector are output. The supervision entity label includes the area name, equipment name, building material name, construction personnel name and construction operation name, etc., which are used for the subsequent generation of log content. For example, "underground foundation concrete pouring" is the area name, and concrete pouring is the construction operation name; the supervision entity label vectors are spliced ​​from top to bottom in the order of appearance to construct text features;

[0042] The text features and the supervision image data are sent to the second entity extraction model for processing, and the supervision entity label and the corresponding supervision entity label vector are output. It should be noted that the supervision entity label output here has the same form as the supervision entity label output by the first entity extraction model, and the supervision entity label output by the second entity extraction model can supplement and correct the supervision entity label output by the first entity extraction model; in the process of processing the text features and the supervision image data, the second entity extraction model will strengthen the temporal feature extraction of the supervision image based on the text features, so that the recognition of the supervision image pays more attention to the content corresponding to the voice data, thereby improving the accuracy of the supervision entity label output;

[0043] The supervision entity label vectors output by all the second entity extraction models are spliced ​​to construct the supervision analysis data, and the supervision analysis data is sent to the log template selection model for processing, and the log template label is output. The log template label is used to select the corresponding log template. The log template includes the positions corresponding to various supervision entity labels. The log template is set in advance by the user, and the log template label corresponds to the log template one by one; the corresponding log template is selected through the log template label, and the supervision entity labels output by all the second entity extraction models are filled in the selected log template to construct the supervision log;

[0044] This application obtains voice data and supervision image data, generates supervision entity labels based on the voice data and supervision image data, and then dynamically selects a log template based on the supervision entity label. Finally, the supervision entity label is filled in the selected log template to construct a supervision log. The dynamic log template can be adapted to different supervision log situations, making it convenient for subsequent users to view the supervision log. In the second entity extraction model, the temporal feature extraction of the supervision image will be enhanced based on the text features, so that the recognition of the supervision image pays more attention to the content corresponding to the voice data, thereby improving the accuracy of the supervision entity label output.

[0045] The first entity extraction model is built based on the BiLSTM-CRF model, including a preprocessing layer, a word embedding layer, a BiLSTM layer, and a CRF layer;

[0046] The text data is processed by the first entity extraction model to output the supervision entity label and the corresponding supervision entity label vector, which specifically includes the following steps:

[0047] The text data is input into the preprocessing layer for processing, specifically performing word segmentation and deleting stop words. Stop words refer to words such as modal particles, function words and punctuation marks. The words output by the preprocessing layer are embedded in words using the Word2vec model to construct several word vectors. All word vectors are then input into the BiLSTM layer in order, and finally processed through the CRF layer to output the supervision entity label, and the word vector corresponding to the supervision entity label is used as the supervision entity label vector corresponding to the supervision entity label.

[0048] The second entity extraction model includes a temporal feature extraction layer, a multi-task recognition layer and a supervision entity label output layer. The temporal feature extraction layer is established based on the ConvLSTM model, and is used to extract temporal features based on text features and supervision image data. During this period, the temporal feature extraction of the supervision image is strengthened based on the text features to construct image features; the multi-task recognition layer includes N multi-task recognition units, each of which performs corresponding task recognition. Since the supervision entity label includes multiple categories, such as person names, area names, safety type names and construction progress, for example, "underground foundation concrete pouring 60%", 60% of which is the construction progress, task recognition is performed for different categories of supervision entity labels. The specific N value is designed by the developer according to user needs to generate supervision entity labels; the supervision entity label output layer is used to output the supervision entity label, and the word vector corresponding to the supervision entity label is used as the supervision entity label vector corresponding to the supervision entity label.

[0049] Temporal feature extraction is performed based on text features and supervision image data, and the temporal feature extraction of supervision images is enhanced based on text features. Specifically, the steps include:

[0050] The temporal feature extraction layer is established based on the ConvLSTM model. With reference to the existing ConvLSTM model, the supervision images in the supervision image data are sequentially input into the temporal feature extraction layer to realize temporal feature extraction. The difference from the ConvLSTM model is that the weight matrix in the input gate of the temporal feature extraction layer is a self-attention weight matrix constructed based on text features and supervision images.

[0051] The self-attention weight matrix is ​​constructed based on text features and supervision images, which specifically includes the following steps: multiplying the supervision images with the key weight matrix to construct the supervision image key matrix K, multiplying the text features with the query weight matrix to construct the supervision image query matrix Q, and calculating the self-attention weight matrix ATT=softmax(QK T / (D) 0.5 ), where softmax is a normalized exponential function, T is a matrix transpose operation, and D is the dimension of the supervision image. The self-attention weight matrix is ​​calculated through text features and supervision images, and the features of the voice data in the supervision image are retained by the calculated self-attention weight matrix in the input gate. This allows more attention to be paid to the content corresponding to the voice data during the temporal feature extraction process, thereby improving the accuracy of supervision entity label extraction and thus improving the accuracy of supervision log generation. Both the key weight matrix and the query weight matrix are set based on the Transformer model.

[0052] Training the first entity extraction model specifically includes the following steps:

[0053] Obtain several text training samples. The text training samples can be text data of actual speech conversion or supervision logs actually obtained. Label the text training samples with supervision entity labels. The labeling process can be to manually label the words in the text data. All labeled text training samples are combined into a text training set. The text training set is then sent to the first entity extraction model with parameter initialization for training to determine whether the training conditions are met. The training condition can be that the supervision entity label output by the first entity extraction model reaches a certain accuracy rate. If the training conditions are met, the trained first entity extraction model is output; otherwise, the first entity extraction model continues to be trained through the text training set.

[0054] Training the second entity extraction model specifically includes the following steps:

[0055] A number of image training samples are obtained, and the image training samples include text features and supervision image data. The text features are the actual conversion of voice data, and the supervision image data is the actual supervision image set corresponding to the text features. The image training samples are labeled with supervision entity labels; all labeled image training samples are combined into an image training set, and the second entity extraction model with initialized parameters is trained with the image training set to determine whether the training conditions are met. If the training conditions are met, the trained second entity extraction model is output; otherwise, the second entity extraction model is continued to be trained with the image training set.

[0056] The log template selection model is established based on the BP neural network model. The training of the log template selection model includes the following steps:

[0057] Obtain several log template selection training samples, which include several supervision entity label vectors, and annotate the log template selection training samples through log template labels. It should be noted that when training the log template selection model, several supervision logs are first obtained, and the corresponding supervision entity labels are manually extracted based on the supervision logs and the supervision entity label vectors corresponding to the supervision entity labels are determined, and the annotated log template labels are determined based on the log templates corresponding to the supervision logs; all annotated log template selection training samples are combined into a log template selection training set, and then the log template selection training set is sent to the log template selection model with parameter initialization for training to determine whether the training conditions are met. The training conditions can be that the log template labels output by the log template selection model reach a certain accuracy rate. If the training conditions are met, the trained log template selection model is output; otherwise, the log template selection model continues to be trained through the log template selection training set.

[0058] In order to adapt to the user's usage habits, after building the supervision log, the log template is also adjusted. The specific steps are as follows:

[0059] Get user operations, which can be addition and deletion operations on the generated supervision log. In response to user operations, adjust the supervision log, generate a new log template copy for the adjusted supervision log, and for each adjustment, if there is an existing log template copy, do not add the log template copy; otherwise, add a new log template copy; at the same time, count the number of times the log template copy is used, and whenever the corresponding log template is selected through the log template label, the log template corresponding to the log template label and the log template copies with the highest number of uses are selected as the log template; in this way, the log template can be adaptively selected according to the user's usage habits.

[0060] Example 2, a system for automatically generating a supervision log for construction supervision, such as Figure 1 As shown, including:

[0061] The supervision data acquisition module is used to respond to user operations to obtain supervision data. It should be noted that during the construction process, the supervisor will conduct supervision work through monitoring cameras and field observations. The supervisor can input voice data and image data as supervision data to automatically generate logs. That is, the supervision data includes voice data and supervision image data, and the supervision image data includes several frames of supervision images input by the user in order of time. The supervision images can also be obtained through videos obtained by monitoring cameras;

[0062] The text data construction module is used to convert the voice data in the supervision data into text data through the voice recognition technology. It should be noted that the voice recognition technology used in this embodiment is the wav2vec 2.0 model. The wav2vec 2.0 model is a deep learning model built based on the Transformer model, which can realize voice recognition. The specific operation refers to the existing technology and will not be repeated here. In addition, this application does not limit the specific implementation means of the voice recognition technology.

[0063] The text feature extraction module is used to extract named entities from the text data through the first entity extraction model, and output the supervision entity label and the corresponding supervision entity label vector. The supervision entity label includes the area name, equipment name, building material name, construction personnel name and construction operation name, etc., which are used to generate subsequent log content. For example, "underground foundation concrete pouring", the underground foundation is the area name, and concrete pouring is the construction operation name; the supervision entity label vectors are spliced ​​from top to bottom in the order of appearance to construct text features;

[0064] The supervision entity label output module is used to send the text features and supervision image data to the second entity extraction model for processing, and output the supervision entity label and the corresponding supervision entity label vector. It should be noted that the supervision entity label output here has the same form as the supervision entity label output by the first entity extraction model, and the supervision entity label output by the second entity extraction model can supplement and correct the supervision entity label output by the first entity extraction model; in the process of processing text features and supervision image data, the second entity extraction model will strengthen the temporal feature extraction of the supervision image based on the text features, so that the recognition of the supervision image pays more attention to the content corresponding to the voice data, thereby improving the accuracy of the supervision entity label output;

[0065] The supervision log generation module is used to splice the supervision entity label vectors output by all the second entity extraction models, construct the supervision analysis data, and send the supervision analysis data to the log template selection model for processing, and output the log template label. The log template label is used to select the corresponding log template. The log template includes the positions corresponding to various supervision entity labels. The log template is set in advance by the user, and the log template label corresponds to the log template one by one; the corresponding log template is selected through the log template label, and the supervision entity labels output by all the second entity extraction models are filled in the selected log template to construct the supervision log.

[0066] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A method for automatically generating a supervision log for construction supervision, characterized in that: include: Responding to user operations, obtaining supervision data, the supervision data including voice data and supervision image data, and the supervision image data including a plurality of frames of supervision images input by the user and sorted by time; Convert the voice data in the supervision data into text data through voice recognition technology; Perform named entity extraction on the text data through the first entity extraction model, output the supervision entity label and the corresponding supervision entity label vector, and concatenate the supervision entity label vectors from top to bottom in the order of appearance to construct text features; The text features and the supervision image data are sent to the second entity extraction model for processing, and the supervision entity label and the corresponding supervision entity label vector are output. In the process of processing the text features and the supervision image data, the second entity extraction model strengthens the temporal feature extraction of the supervision image based on the text features; The supervision entity label vectors output by all the second entity extraction models are spliced ​​to construct the supervision analysis data, and the supervision analysis data is sent to the log template selection model for processing, and the log template label is output, and the log template label is used to select the corresponding log template; Select the corresponding log template through the log template tag, and fill the selected log template with the supervision entity tags output by all the second entity extraction models to construct the supervision log; The first entity extraction model is built based on the BiLSTM-CRF model, including a preprocessing layer, a word embedding layer, a BiLSTM layer, and a CRF layer; The text data is processed by the first entity extraction model to output the supervision entity label and the corresponding supervision entity label vector, which specifically includes the following steps: The text data is input into the preprocessing layer for processing, specifically, word segmentation and stop word deletion; the words output by the preprocessing layer are embedded, and the word embedding operation adopts the Word2vec model to construct several word vectors; all word vectors are then input into the BiLSTM layer in order, and finally processed through the CRF layer to output the supervision entity label, and the word vector corresponding to the supervision entity label is used as the supervision entity label vector corresponding to the supervision entity label; The second entity extraction model includes a temporal feature extraction layer, a multi-task recognition layer and a supervision entity label output layer. The temporal feature extraction layer is established based on the ConvLSTM model and is used to extract temporal features based on text features and supervision image data. During this period, the temporal feature extraction of the supervision image is enhanced based on the text features to construct image features. The multi-task recognition layer includes N multi-task recognition units, each of which performs corresponding task recognition and generates a supervision entity label; the supervision entity label output layer is used to output the supervision entity label, and use the word vector corresponding to the supervision entity label as the supervision entity label vector corresponding to the supervision entity label; Temporal feature extraction is performed based on text features and supervision image data, and the temporal feature extraction of supervision images is enhanced based on text features. Specifically, the steps include: The temporal feature extraction layer is established based on the ConvLSTM model. The supervision images in the supervision image data are sequentially input into the temporal feature extraction layer to realize temporal feature extraction. The difference from the ConvLSTM model is that the weight matrix in the input gate of the temporal feature extraction layer is a self-attention weight matrix constructed based on text features and supervision images. The self-attention weight matrix is ​​constructed based on text features and supervision images, which specifically includes the following steps: multiplying the supervision images with the key weight matrix to construct the supervision image key matrix K, multiplying the text features with the query weight matrix to construct the supervision image query matrix Q, and calculating the self-attention weight matrix ATT=softmax(QK T / (D) 0.5 ), where softmax is the normalized exponential function, T is the matrix transpose operation, and D is the dimension of the supervision image; The log template selection model is established based on the BP neural network model. The training of the log template selection model includes the following steps: Obtain several log template selection training samples, which include several supervision entity label vectors, and label the log template selection training samples through log template labels; combine all labeled log template selection training samples into a log template selection training set, and then send the log template selection training set to the log template selection model with parameter initialization for training to determine whether the training conditions are met. If so, output the trained log template selection model; otherwise, continue to train the log template selection model through the log template selection training set.

2. The method for automatically generating a supervision log for construction supervision according to claim 1, characterized in that: Training the first entity extraction model specifically includes the following steps: Obtain several text training samples, annotate the text training samples with supervision entity labels, form all annotated text training samples into a text training set, and then send the text training set to the first entity extraction model with parameter initialization for training to determine whether the training conditions are met. If the training conditions are met, output the trained first entity extraction model; otherwise, continue to train the first entity extraction model through the text training set.

3. The method for automatically generating a supervision log for construction supervision according to claim 2, characterized in that: The training of the second entity extraction model specifically includes the following steps: A number of image training samples are obtained, which include text features and supervision image data, and the image training samples are labeled with supervision entity labels; all labeled image training samples are combined into an image training set, and a second entity extraction model with initialized parameters is trained with the image training set to determine whether the training conditions are met. If the training conditions are met, the trained second entity extraction model is output; otherwise, the second entity extraction model is continued to be trained with the image training set.

4. The method for automatically generating a supervision log for construction supervision according to claim 3, characterized in that: After building the supervision log, the log template needs to be adjusted. The specific steps are as follows: Obtain user operations, adjust the supervision log in response to the user operations, generate a new log template copy for the adjusted supervision log, and for each adjustment, if there is an existing log template copy, do not add a new log template copy; Otherwise, add a new copy of the log template; At the same time, the number of times the log template copy is used is counted. Whenever the corresponding log template is selected through the log template label, the log template corresponding to the log template label and the log template copies with the highest number of uses are selected as the log template.

5. The automatic generation system of supervision log for construction supervision is characterized by: The system applies the method for automatically generating a supervision log for construction supervision according to any one of claims 1 to 4, comprising: A supervision data acquisition module, used to acquire supervision data in response to user operations, the supervision data including voice data and supervision image data, and the supervision image data including a number of supervision image frames input by the user and sorted by time; A text data construction module is used to convert the voice data in the supervision data into text data through voice recognition technology; A text feature extraction module is used to extract named entities from text data through a first entity extraction model, output a supervision entity label and a corresponding supervision entity label vector, and concatenate the supervision entity label vectors from top to bottom in order of appearance to construct text features; A supervision entity label output module is used to send text features and supervision image data to the second entity extraction model for processing, and output the supervision entity label and the corresponding supervision entity label vector. In the process of processing the text features and supervision image data, the second entity extraction model strengthens the temporal feature extraction of the supervision image based on the text features; The supervision log generation module is used to splice the supervision entity label vectors output by all the second entity extraction models, construct the supervision analysis data, and send the supervision analysis data to the log template selection model for processing, and output the log template label, which is used to select the corresponding log template; select the corresponding log template through the log template label, and fill the selected log template with the supervision entity labels output by all the second entity extraction models to construct the supervision log.

Images