Industrial heritage retention degree studying and judging method and system based on historical image comparison and recognition

Abstract

The invention provides an industrial heritage retention degree studying and judging method and system based on historical image comparison and recognition, and the method comprises the steps: collecting multi-source heterogeneous data of a target region, extracting space-time and semantic feature vectors, and strengthening the feature representation, thereby obtaining a comprehensive feature vector; performing confrontation optimization by using a label predictor and a domain discriminator to enable the comprehensive feature vector to realize feature distribution alignment of an industrial quality inspection source domain and an industrial heritage target domain to obtain a feature matching degree; and analyzing and predicting the functional coefficient by using an LSTM time sequence based on the multi-source heterogeneous data, and outputting a retention degree evaluation report. The method comprises the following steps: firstly, extracting feature vectors in parallel, enhancing feature fusion in combination with a double-attention module, and capturing cross-year image sequence changes; the two-domain feature distribution difference is minimized through adversarial training, and the accuracy of the feature matching degree is improved; and finally, weighting and fusing the feature matching degree and the expert score to calculate an integrity coefficient, and predicting a functional coefficient by using an LSTM time sequence model, thereby improving the scientificity and efficiency of industrial heritage protection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of identifying the degree of heritage preservation, in particular to an industrial heritage preservation degree identification method and system based on historical image comparison. BACKGROUND

[0002] Current industrial heritage preservation degree identification mainly relies on manual field investigation and experience judgment mode. Traditional methods usually involve expert teams comparing historical images of different periods, combining with on-site measurement data to identify component damage conditions such as building structure cracks and equipment aging degree, and using simple image processing techniques such as edge detection and gray scale comparison to extract static features. Some studies attempt to introduce single-scale image analysis or basic time series models, but are mostly limited to static images or single-year data, and lack of spatiotemporal evolution information mining in continuous year image sequences. The fusion application of multi-source heterogeneous data such as images, equipment logs, and environmental monitoring data is still in its infancy, with low data utilization rate, and the identification results are highly dependent on expert subjective experience, making it difficult to balance efficiency and objectivity. SUMMARY

[0003] The present application aims to at least solve the technical problem in the prior art that the identification results are highly dependent on expert subjective experience, and it is difficult to balance efficiency and objectivity. The present application provides an industrial heritage preservation degree identification method based on historical image comparison and recognition.

[0004] To achieve the above-mentioned purpose of the present application, the present application provides an industrial heritage preservation degree identification method based on historical image comparison and recognition, which comprises: S1, collecting multi-source heterogeneous data of a target area and preprocessing; the multi-source heterogeneous data includes a continuous year image sequence; S2, extracting spatiotemporal feature vectors in the continuous year image sequence using a CNN network; S3, extracting semantic feature vectors in the continuous year image sequence using a self-attention mechanism; S4, strengthening feature representation based on the spatiotemporal feature vectors and semantic feature vectors using a space-channel dual attention SA module, and obtaining a comprehensive feature vector; S5, using the adversarial optimization of a label predictor and a domain discriminator to align the feature distribution of the industrial quality inspection source domain and the industrial heritage target domain based on the comprehensive feature vector, and obtaining a feature matching degree; S6, calculating an integrity coefficient based on the feature matching degree and expert score; predicting a functionality coefficient based on the multi-source heterogeneous data through LSTM time series analysis, and outputting a preservation degree evaluation report based on the integrity coefficient and the predicted functionality coefficient.

[0005] In another aspect, the present application also provides an industrial heritage retention degree judgment system based on historical image contrast recognition, the system comprises: a processor; a memory for storing processor executable instructions; wherein the processor is configured to implement the industrial heritage retention degree judgment method based on historical image contrast recognition when executing the executable instructions.

[0006] The present application has the following advantages: the present application effectively solves the pain points of traditional manual judgment relying on subjective experience, low efficiency and insufficient utilization of multi-source data through multi-source heterogeneous data fusion, spatio-temporal-semantic feature double-channel extraction and domain discriminator adversarial optimization technology. Specifically, the spatio-temporal feature vector and the semantic feature vector are extracted in parallel by using CNN and self-attention mechanism, the feature fusion is strengthened by combining the space-channel double-attention module, and the subtle changes of the cross-year image sequence are captured; through the adversarial training of the label predictor and the domain discriminator, the feature distribution difference between the industrial quality inspection source domain and the heritage target domain is minimized, and the feature matching degree accuracy is improved; finally, the integrity coefficient is calculated based on the weighted fusion of the feature matching degree and the expert score, and the functional coefficient is predicted by using the LSTM time series model, forming a dual quantification system of "data driven + experience calibration", so that the retention degree evaluation result has objectivity and interpretability, supporting dynamic repair decision and life cycle management, and significantly improving the scientific nature and efficiency of industrial heritage protection.

[0007] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 is a flowchart of the industrial heritage retention degree judgment method based on historical image contrast recognition of the present application. DETAILED DESCRIPTION

[0009] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar reference numerals refer to the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0010] Example 1 As Figure 1 shown, the industrial heritage retention degree judgment method based on historical image contrast recognition, the method comprises: S1, collect multi-source heterogeneous data of the target area and preprocess; the multi-source heterogeneous data includes continuous-year image sequences; In step S1, first, continuous-year image sequences collected by high-resolution satellite remote sensing images, unmanned aerial vehicle oblique photography data, and ground monitoring cameras are used as core data sources, which need to cover the whole life cycle stage of industrial heritage; at the same time, structured data such as engineering drawings and equipment operation logs in historical archives, and unstructured text data such as on-site investigation records and expert evaluation reports of cultural relics protection units are integrated to form a multi-dimensional data set containing spatial and temporal information, functional attributes, and maintenance records; in the preprocessing stage, histogram equalization is used to enhance the contrast of image data, SIFT algorithm is used for image registration to eliminate geometric distortion, and median filtering is used to remove noise interference; for text data, digital conversion is realized through OCR technology, NLP model is used to extract key entities and semantic relationships, and finally all preprocessed data are stored in a spatial and temporal database.

[0011] S2, use CNN network to extract spatial and temporal feature vectors in continuous-year image sequences; S3, use self-attention mechanism to extract semantic feature vectors in continuous-year image sequences; S4, based on spatial and temporal feature vectors and semantic feature vectors, use spatial-channel dual attention SA module to strengthen feature representation and obtain comprehensive feature vectors; S5, use the adversarial optimization of label predictor and domain discriminator to realize the feature distribution alignment of industrial quality inspection source domain and industrial heritage target domain for comprehensive feature vectors, and obtain feature matching degree; S6, calculate integrity coefficient based on feature matching degree and expert score; predict functional coefficient based on multi-source heterogeneous data through LSTM time series analysis; output retention degree evaluation report based on integrity coefficient and predicted functional coefficient.

[0012] The principle of the industrial heritage retention degree judgment method based on historical image contrast recognition in the embodiment is as follows: through the construction of a multi-source heterogeneous data fusion framework, first, a CNN network is used to perform convolution operation on the image sequence of consecutive years, and the geometric shape, spatial position and other low-level to high-level spatio-temporal features of the industrial heritage component are extracted layer by layer to form a spatio-temporal feature vector; at the same time, a self-attention mechanism is used to capture the semantic association of different time steps in the image sequence, the similarity weight of the Query vector and the Key vector is calculated, the Value vector is weighted and aggregated, and a semantic feature vector reflecting the functional properties and structural state of the industrial heritage is generated. Then, the two kinds of feature vectors are input into a space-channel dual attention SA module, a spatial weight map is generated through global average pooling in the spatial dimension, a channel weight vector is generated through global maximum pooling in the channel dimension, and the two are fused to form a dual attention weight matrix, which dynamically weights and enhances the original features, and finally a comprehensive feature vector containing spatio-temporal evolution information and semantic association information is obtained. On this basis, the label predictor and the domain discriminator are used for adversarial optimization training, so that the distribution of the comprehensive feature vector tends to be consistent between the industrial quality inspection source domain and the heritage target domain, and the feature matching degree is calculated through the binary classification probability of the domain discriminator; the integrity coefficient is calculated by combining the expert score matrix, and the future functional coefficient is predicted by using a double-layer LSTM network combined with a time attention mechanism, and finally the quantitative retention degree level is output based on the weighted sum model to form a comprehensive evaluation result covering the morphological integrity and functional availability.

[0013] As an optional embodiment of the present application, the spatio-temporal feature vector in the image sequence of consecutive years is extracted by using a CNN network in step S2, which includes: S201, standardizing the image sequence of consecutive years; In step S201, it needs to be specified that the standardization of the image sequence of consecutive years mainly includes two steps of size normalization and pixel value standardization. First, the images of different years and different resolutions are uniformly scaled to a preset size (such as 512x512 pixels), and the bilinear interpolation method is used to maintain the smoothness of the image edges and avoid geometric distortion; second, the normalized image is subjected to pixel value standardization, the RGB three-channel pixel values are linearly mapped from the [0, 255] range to the [0, 1] interval, and the global mean and standard deviation are calculated, and the Z-Score standardization processing is performed on each channel, so that the image data conforms to the distribution characteristics of zero mean and unit variance. Through the standardization processing, the feature distribution deviation caused by the differences in shooting equipment and lighting conditions of different years of images can be eliminated, and the stability and robustness of spatio-temporal feature extraction can be improved.

[0014] S202, constructing a CNN network and extracting spatio-temporal information of different scales based on the image sequence of consecutive years after standardization processing; In step S202, it needs to be specified in detail that in the present embodiment, the CNN network adopts a hierarchical architecture, including an input layer, multiple convolutional layers, a pooling layer and a fully connected layer. The input layer receives the standardized continuous-year image sequence, and each image is input as an independent channel; the convolutional layer adopts different sizes of convolutional kernels (such as 3x3, 5x5) for sliding window operation, introduces nonlinearity through a ReLU activation function, and extracts multi-scale spatio-temporal features from edges, textures to components and scenes layer by layer; the pooling layer adopts a maximum pooling strategy, and down-samples each feature map to reduce the parameter amount while retaining significant features; the fully connected layer fuses the multi-scale features, and outputs a spatio-temporal feature vector containing spatio-temporal evolution information. Through this hierarchical extraction and fusion mechanism, the CNN network can effectively capture the morphological changes and spatial displacement features of industrial heritage in different years.

[0015] S203, fuse spatio-temporal information of different scales to obtain a spatio-temporal feature vector.

[0016] In step S203, it needs to be specified in detail that the fusion of spatio-temporal information adopts a composite strategy of multi-scale feature splicing and 1x1 convolutional dimension reduction. First, the different scale feature maps output by each convolutional layer are spliced in the channel to form a multi-dimensional feature tensor containing shallow detail features and deep semantic features; then, a 1x1 convolutional kernel is used for channel compression to reduce parameter redundancy while maintaining feature expression capability, and finally a unified dimension spatio-temporal feature vector is generated. The vector is further processed by a global average pooling operation to eliminate spatial redundancy and ensure the translation invariance of the feature representation.

[0017] As an optional embodiment of the present application, in step S3, the semantic feature vector in the continuous-year image sequence is extracted by using a self-attention mechanism, which includes: S301, segment the preprocessed continuous-year image sequence, and then map the segmented continuous-year image sequence to initial tokens using a learnable linear projection layer; In step S301, the preprocessed continuous-year image sequence is first segmented into fixed-length image segments (such as every 5 years) according to time steps, and each segment is processed as an independent sample; then, a learnable linear projection layer (parameter matrix W ∈ R(CxD), where C is the number of image channels and D is the dimension of the token) is used to map the pixel values of each image segment to D-dimensional initial tokens, forming a token sequence {t1, t2,..., tT} (T is the total number of time steps). The projection layer automatically optimizes the parameters through backpropagation, so that the initial tokens can not only retain the local features of the original image, but also have dimensionality computability.

[0018] S302, calculating the similarity weight between each segmented continuous year image sequence based on the initial token using a self-attention mechanism; The expression for calculating the similarity weight between each segmented continuous year image sequence is: wherein, represents the similarity weight between the i-th image sequence block and the j-th image sequence block in the frame image, represents a normalized exponential function, represents the Query vector corresponding to the i-th image block in the frame image, reflecting the spatiotemporal semantic query requirement of a specific industrial heritage component (such as a chimney) in the frame, represents the dimension (positive integer) of the Key vector, used to scale the dot product result to prevent gradient disappearance (usually 64 / 128 / 256); In step S302, the original similarity score is obtained by calculating the dot product of the Query vector and the Key vector, then scaled by dividing by the square root of the dimension of the Key vector, and finally converted to a probability form of the similarity weight using the normalized exponential function. The weight matrix (W ∈ R(T × T)) reflects the semantic association strength between different time step image blocks, for example, the image block corresponding to the equipment modification in a certain year will form a high weight connection with the image block related to the function in the subsequent year.

[0019] S303, weighting and aggregating the initial token based on the similarity weight to generate a semantic feature vector.

[0020] ​​​​​​​​In step S303, the initial token sequence {t1, t2,..., tT} is weighted and aggregated based on a similarity weight matrix (W e R(T x T)). First, the initial token at each time step is multiplied by the weight vector of the corresponding row to generate a weighted token sequence. Then, all weighted tokens are summed to obtain an aggregated semantic feature vector. This vector is fused with the mean of the initial token sequence through a residual connection, preserving both local features of the original image and semantic associations across time steps. Finally, layer normalization is performed to stabilize the training process and ensure the generalization ability of the semantic feature vector across different image sequences. For example, when processing an image sequence containing device update cycles, the semantic feature vector can automatically focus on the image blocks related to functional changes, ignoring the interference caused by background environment changes.

[0021] As an optional embodiment of the present application, in step S4, the feature representation is enhanced based on the spatio-temporal feature vector and the semantic feature vector using a space-channel dual attention SA module to obtain a comprehensive feature vector, which includes: S401, the spatio-temporal feature vector and the semantic feature vector are concatenated in the channel dimension to obtain a fusion feature tensor; In step S401, when the spatio-temporal feature vector and the semantic feature vector are concatenated in the channel dimension, it is necessary to ensure that the spatial resolutions of the two vectors are consistent. The semantic feature vector is upsampled to the same spatial size (such as 512 x 512 pixels) as the spatio-temporal feature vector through bilinear interpolation, and then a concat operation is performed along the channel dimension to generate a fusion feature tensor F e R(H x W x (C1 + C2)), where H and W are the spatial height and width, and C1 and C2 are the channel numbers of the spatio-temporal feature and the semantic feature, respectively. The tensor retains both the geometric shape information (such as building outlines) and the functional semantic information (such as device status) of the industrial heritage.

[0022] S402, the fusion feature tensor is compressed in the spatial dimension to generate a single-channel spatial weight map; In step S402, when the fusion feature tensor is compressed in the spatial dimension, a global average pooling operation is used to calculate the mean value of each channel along the spatial dimension to generate a single-channel spatial weight map. This weight map reflects the importance of different channel features in the overall spatial distribution by aggregating global spatial information. For example, when there is a local damage area in the industrial heritage image, the spatial weight value of the corresponding channel will be significantly reduced, thereby suppressing the interference of irrelevant areas in subsequent feature enhancement. At the same time, to preserve local spatial details, a global maximum pooling can be used in parallel to generate an auxiliary spatial weight map, and the two are fused through a learnable weight parameter to form a more robust spatial weight representation.

[0023] S403, performing global average pooling on the fused feature tensor to generate a channel weight vector after compressing the spatial dimension; In step S403, when performing global average pooling on the fused feature tensor, the mean value of each channel is calculated along the spatial dimension to generate a channel weight vector. This vector can reflect the importance of different channel features in the overall semantic expression by aggregating global spatial information. For example, when a certain type of feature (such as rust texture) in the industrial heritage image appears in multiple spatial locations, the weight value of the corresponding channel will significantly increase, thereby highlighting the contribution of this type of feature in subsequent feature enhancement. To further improve the discriminability of the channel weight vector, a global maximum pooling can be used to generate an auxiliary channel weight vector in parallel. The two vectors are fused through a learnable weight parameter to form a more robust channel weight representation.

[0024] S404, multiplying the single-channel spatial weight map and the channel weight vector by broadcasting to generate a dual-attention weight matrix; performing feature enhancement by element-wise multiplication of the dual-attention weight matrix and the fused feature tensor, and adding the spatial-temporal feature vector and the semantic feature vector through residual connection to form a style-enhanced feature vector; In step S405, when multiplying the single-channel spatial weight map and the channel weight vector by broadcasting, the spatial weight map and the channel weight vector are first expanded to the same dimension through broadcasting mechanism, and then element-wise multiplication is performed to generate a dual-attention weight matrix. The value of each element in the matrix reflects the joint importance of the corresponding spatial position and channel feature. When performing element-wise multiplication of the dual-attention weight matrix and the fused feature tensor, the original feature value of each spatial position will be dynamically weighted according to its spatial importance and channel importance. For example, the well-preserved area in the industrial heritage image will obtain a higher weight value, while the features of damaged or irrelevant areas will be suppressed. The feature-enhanced tensor is added to the original spatial-temporal feature vector and semantic feature vector through residual connection, which not only preserves the integrity of the original features, but also strengthens the semantic expression of the key areas, finally forming a style-enhanced feature vector. The vector is normalized through layer normalization to stabilize the training process and ensure the generalization ability between different year image sequences.

[0025] S405, performing global average pooling on the style-enhanced feature vector to obtain a comprehensive feature vector, which includes spatial-temporal evolution information and semantic association information.

[0026] In step S405, it needs to be specified that when the style enhancement feature vector is globally averaged, the mean value is calculated along the spatial dimension and the channel dimension, and the multi-dimensional feature tensor is compressed into a fixed-length comprehensive feature vector. This operation aggregates global information, eliminates spatial redundancy and strengthens the semantic association between channels, and finally generates a comprehensive feature representation containing spatio-temporal evolution information (such as the trajectory of the change of the building structure over time) and semantic association information (such as the cross-year correspondence of functional components). For example, when processing image sequences of an industrial plant containing multiple renovations, the comprehensive feature vector can encode both the progressive changes in roof shape (spatio-temporal information) and the type association of equipment updates (semantic information). At the same time, to improve the interpretability of the features, a gradient-weighted class activation mapping (Grad-CAM) can be introduced after pooling to visualize the contribution of key regions in different year images, assisting in verifying the effectiveness of feature extraction.

[0027] As an optional embodiment of the present application, optionally, in step S5, the comprehensive feature vector is subjected to adversarial optimization of the label predictor and the domain discriminator to align the feature distribution of the industrial quality inspection source domain and the industrial heritage target domain, and the feature matching degree is obtained, including: S501, a multi-layer perceptron structure is used to construct a label predictor, which takes the comprehensive feature vector as input and outputs the industrial heritage preservation level; In step S501, it needs to be specified that the multi-layer perceptron structure includes an input layer, a hidden layer and an output layer. The input layer receives the comprehensive feature vector, and its dimension is determined according to the specific length of the comprehensive feature vector; the hidden layer uses a fully connected manner, introduces non-linear transformation ability through a non-linear activation function (such as ReLU), can learn complex spatio-temporal-semantic association patterns in the comprehensive feature vector, and the number of hidden layer nodes can be adjusted according to actual needs to balance the model complexity and generalization ability; the output layer maps the output of the hidden layer to a probability distribution of the industrial heritage preservation level through a Softmax activation function, for example, divided into multiple levels such as "perfectly preserved", "partially damaged" and "severely damaged", each level corresponds to a probability value, and the sum of the probability values is 1, thereby realizing quantitative evaluation of the industrial heritage preservation level. In the training process, the label predictor continuously optimizes the network parameters by minimizing the cross-entropy loss function between the predicted results and the true labels, and improves the prediction accuracy of the industrial heritage preservation level.

[0028] S502, a convolutional neural network structure is used to construct a domain discriminator, which takes the comprehensive feature vector as input and outputs a binary classification probability of whether the sample belongs to the industrial quality inspection source domain or the industrial heritage target domain; In step S502, the domain discriminator of the convolutional neural network structure is composed of multiple convolutional layers, pooling layers and fully connected layers. After the input layer receives the comprehensive feature vector, local feature extraction is first performed through the convolutional layer, and the convolution kernel size and step can be adjusted according to the feature dimension to capture domain-related features of different scales; the pooling layer adopts a maximum pooling or average pooling strategy to reduce the dimension of the feature map output by the convolution, thereby enhancing the translation invariance of the model; then, the multi-dimensional features are mapped to binary classification probabilities through the fully connected layer, and the probability value (range 0~1) of the sample belonging to the industrial quality inspection source domain or the industrial heritage target domain is output. In the training process, the domain discriminator minimizes the binary cross-entropy loss function between the predicted domain label and the real domain label, optimizes the network parameters to improve the domain classification accuracy. At the same time, in order to avoid gradient disappearance, batch normalization operation can be introduced after the convolutional layer to speed up the training convergence and improve the stability of the model. For example, when the input comprehensive feature vector contains device texture differences, the domain discriminator can capture the subtle differences in material reflection characteristics between the source domain and the target domain through the convolution kernel, thereby accurately distinguishing the sample source.

[0029] S503, through the adversarial optimization training, the cross-entropy loss function of the label predictor is minimized, and the confusion loss function of the domain discriminator is maximized, the confusion loss function is gradient reversed through GRL, and the feature distribution of the comprehensive feature vector is aligned between the industrial quality inspection source domain and the industrial heritage target domain; The expression of the cross-entropy loss function is: , wherein, represents the cross-entropy loss function value of the label predictor, which is used to measure the difference between the predicted retention level and the real level, and the smaller the value is, the more accurate the prediction is, represents the total number of samples, that is, the number of image sequences participating in the training in the industrial heritage target domain, represents the number of retention level categories (such as excellent / good / average / poor, corresponding to =4), represents the real label of the th sample in the th category (one-hot encoding, if the sample belongs to the th category, it is 1, otherwise it is 0), represents the predicted probability of the label predictor for the th sample in the th category; The expression of the confusion loss function is: , wherein, represents the confusion loss function value of the domain discriminator, which is maximized when the gradient is reversed through GRL, and the domain discriminator is difficult to distinguish between the source domain and the target domain, The total number of input samples for the domain discriminator (including mixed samples from the industrial quality inspection source domain and the industrial heritage target domain). Indicates the first Domain labels of each sample ( =0 indicates a sample from the source domain. =1 indicates a target domain sample). The domain discriminator represents the first domain. The predicted probability of a sample belonging to the target domain (output by the Sigmoid activation function, with a value range of [0, 1]). In step S503, it is necessary to explain in detail that during the adversarial optimization training process, the label predictor and the domain discriminator form a dynamic game relationship. Specifically, the label predictor continuously optimizes its own parameters by minimizing the cross-entropy loss function to improve the prediction accuracy of the industrial heritage retention level; the domain discriminator optimizes its domain classification ability by minimizing the binary cross-entropy loss function, and simultaneously maximizes the confusion loss function by inverting the sign of the gradient during backpropagation through a gradient reversal layer (GRL). This adversarial mechanism prompts the comprehensive feature vector to gradually eliminate the domain offset between the source and target domains while retaining task-related features (such as retention level semantics). For example, when the source domain samples contain standardized quality inspection equipment and the target domain samples contain old industrial heritage equipment, adversarial training will make the comprehensive feature vector focus more on the functional status of the equipment rather than the material aging differences, thereby improving the cross-domain generalization ability. An alternating optimization strategy is adopted during training: first, the domain discriminator parameters are fixed, and the label predictor parameters are updated to reduce the cross-entropy loss function value; then, the label predictor parameters are fixed, and the domain discriminator parameters are updated through GRL backpropagation to increase the confusion loss function value. Through multiple iterations, the comprehensive feature vector is finally aligned in feature distribution between the industrial quality inspection source domain and the industrial heritage target domain. At this point, the domain discriminator's prediction probability of the sample source is close to 0.5, indicating that the model can no longer effectively distinguish the sample domain, while the label predictor's prediction accuracy of the retention rate of the target domain samples is significantly improved.

[0030] S504. Calculate the feature matching degree based on the binary classification probability output by the domain discriminator. The feature matching degree reflects the alignment between the features of the target domain of industrial heritage and the features of the source domain of industrial quality inspection.

[0031] The expression for feature matching degree is: , in, Indicates feature matching degree. The domain discriminant predicts the probability of a sample in the target domain. This represents the Gaussian kernel bandwidth parameter. This represents the Softmax activation function. The weight matrix of the domain discriminator is represented. represents a comprehensive feature vector, represents a bias vector of the domain discriminator.

[0032] In step S504, it needs to be explained in detail that the calculation of the feature matching degree is realized by quantifying the similarity of the prediction probability distribution of the target domain sample in the domain discriminator and the source domain distribution. Specifically, first, the output probability of the domain discriminator (i.e. the probability that the sample is judged as the target domain) is combined with the Gaussian kernel function to smooth the probability distribution, where the Gaussian kernel bandwidth parameter controls the smoothing degree. Subsequently, the linear combination of the weight matrix of the domain discriminator and the comprehensive feature vector is mapped to the probability distribution through the Softmax activation function, and finally the feature matching degree is generated. The value of this index ranges from 0 to 1, and the closer the value is to 1, the more similar the target domain feature is to the source domain feature, that is, the higher the alignment degree; otherwise, it indicates that the domain shift still exists. For example, when processing the same industrial heritage images taken in different years, if the feature matching degree reaches 0.85 or above, it means that the model has effectively eliminated the domain differences caused by the shooting equipment, lighting conditions or building aging, and the comprehensive feature vector can stably represent the preservation state of the heritage across years. In addition, in order to dynamically monitor the training process, a matching degree change curve can be introduced: if it continues to rise and tends to be stable during the iteration process, it indicates that the adversarial optimization is successful; if it fluctuates or stagnates, the GRL gradient reversal intensity or the domain discriminator structure (such as increasing the depth of convolution layer) needs to be adjusted. Finally, the feature matching degree not only serves as a quantitative indicator for cross-domain alignment, but also can be directly used to select high-quality training samples - only samples with a matching degree above a threshold (such as 0.7) are retained for label predictor update, thereby further improving the generalization performance of the model in the target domain.

[0033] As an optional embodiment of the present application, optionally, in step S6, the integrity coefficient is calculated based on the feature matching degree and the expert score; the functional coefficient is predicted based on the multi-source heterogeneous data through LSTM time series analysis, and the preservation degree evaluation report is output based on the integrity coefficient and the predicted functional coefficient, including: S601, based on the feature matching degree and the expert score matrix, the integrity coefficient is calculated through a weighted fusion model; The expression for calculating the integrity coefficient is: wherein, represents the integrity coefficient, and represents the weighted coefficient, represents the feature matching degree, represents the average expert score; In step S601, it needs to be explained in detail that the calculation of the integrity coefficient is realized by dynamically weighting the objective feature matching degree and the subjective expert score. Specifically, the weighted coefficient and According to the data reliability, when the historical image annotation quality is high and the domain alignment effect is significant, the weight of the feature matching degree can be increased The weight (such as = 0.7, = 0.3) is adjusted to make the feature matching degree dominant in the evaluation; if the expert's judgment on the historical value of the heritage is more authoritative, the weight is adjusted to = 0.4, = 0.6. The mean value E of the expert score is obtained by taking the arithmetic mean of the independent scores of multiple experts on the structural integrity and material degradation of the industrial heritage, and the score range is usually set to [0, 10] points. For example, for a 19th century foundry, if the feature matching degree is calculated as 0.82 (indicating high alignment of cross-domain features), and three experts give scores of 8.5, 8.0, and 9.0 (mean value E = 8.5), then the integrity coefficient I = 0.6 x 0.82 + 0.4 x 8.5 / 10 = 0.492 + 0.34 = 0.832. This coefficient not only quantifies the degree of feature alignment at the technical level, but also incorporates the expert's experience and judgment, especially for heritage cases with non-standard modifications or missing historical data. To improve the adaptability of the model, a dynamic weight adjustment mechanism can be introduced: when the correlation coefficient between the feature matching degree and the expert score is lower than the threshold (such as r < 0.6), the weight redistribution process is automatically triggered, and the optimal , combination is searched through Bayesian optimization to ensure the balance between technical feasibility and historical authenticity. The final integrity coefficient will be one of the core indicators of the preservation degree assessment report, directly reflecting the comprehensive preservation of the physical state and historical value of the heritage.

[0034] S602, based on the spatio-temporal feature vector in the multi-source heterogeneous data, using a double-layer LSTM network, combining a time attention mechanism to weight the key time steps, outputting a predicted functionality coefficient for the next n years; In step S602, it needs to be specified in detail that the double-layer LSTM network can effectively capture the complex spatio-temporal dependence relationship in multi-source heterogeneous data by stacking two long short-term memory units. Specifically, the bottom-layer LSTM first encodes the input spatio-temporal feature vector to extract the dynamic pattern (such as the device aging rate and the structural deformation trend) in the local time window; the upper-layer LSTM further integrates the bottom-layer output to build a global time series representation and capture the long-term evolution law across time periods. The time attention mechanism strengthens the contribution of key historical moments (such as major maintenance events and natural disaster impact periods) to the prediction result by calculating the attention weight of each time step. For example, when analyzing the functional changes of an industrial heritage in the past 50 years, the attention mechanism may assign a weight of 0.3 to the device modification period in 1980, a weight of 0.25 to the earthquake repair period in 2010, and the remaining time steps are allocated according to the remaining weight proportion. Finally, the predicted functionality coefficient is mapped to the interval [0, 1] by a fully connected layer, and the value closer to 1 indicates a higher probability of the heritage maintaining its original functionality in the future n years. For example, if the predicted functionality coefficient of a textile mill in the next 10 years is 0.78, it means that the core functional elements such as production equipment and spatial layout have a high probability of maintaining their usability. To improve the prediction robustness, a Dropout layer (dropout rate 0.2) is introduced during network training to prevent overfitting, and a Huber loss function is used to balance the sensitivity to outliers. In actual deployment, the prediction period n can be dynamically adjusted according to the type of heritage: for fast-iterating light industrial heritage (such as food processing plants), n is 3-5 years; for structurally stable heavy industrial heritage (such as steel plants), n can be extended to 10-15 years. The final generated functionality coefficient will be the quantitative basis for the preservation degree assessment report together with the integrity coefficient.

[0035] S603, calculating a weighted sum based on the integrity coefficient and the predicted functionality coefficient to obtain a preservation degree grade; The expression for calculating the preservation degree grade is: wherein, R represents the preservation degree grade (value range [0, 1]), which quantifies the comprehensive preservation state of the industrial heritage through a weighted sum. The grade division criteria are: excellent (R≥0.85), good (0.70≤R<0.85), general (0.50≤R<0.70), and poor (R<0.50), and α and β represent the weighted coefficients, reflecting the relative importance of the integrity coefficient and the functionality coefficient, α and β represent the weighted coefficients, reflecting the relative importance of the integrity coefficient and the functionality coefficient, α and β represent the weighted coefficients, reflecting the relative importance of the integrity coefficient and the functionality coefficient, The calculation of the retention degree grade is achieved by comprehensively considering the current integrity state and future functional availability of the industrial heritage. Specifically, the weighting coefficients and need to be configured differently according to the type of heritage and protection goals: when the evaluation focuses on the historical authenticity of the building (such as a cultural heritage unit), the integrity coefficient can be set as = 0.6, = 0.4, so that the integrity coefficient dominates the grade division; if the focus is on the continuous use value of the heritage (such as an industrial tourism site), the coefficients are adjusted to = 0.4, = 0.6, to strengthen the influence of the functional coefficient. For example, for a mechanical manufacturing plant built in the early 20th century, if the integrity coefficient I = 0.82 (the structure is well preserved) and the functional coefficient F = 0.75 (the equipment is updated and can still be used for production), when = 0.5, = 0.5, the retention degree grade R = 0.5 x 0.82 + 0.5 x 0.75 = 0.785, which is classified as "good" according to the standard. This quantitative model not only avoids the one-sidedness of a single index, but also adapts to the needs of different protection scenarios through weight adjustment. To improve the flexibility of evaluation, a dynamic weight calibration mechanism can be introduced: when the functional requirements of the heritage change significantly (such as from production to exhibition), the weight optimization process is automatically triggered, and the optimal , combination is searched through genetic algorithm to ensure the consistency of grade division and protection goals. The final retention degree grade will serve as the core basis for heritage protection decisions, such as applying for cultural relic protection funds for "excellent" grade heritage and starting a rescue repair plan for "poor" grade heritage.

[0036] S604, based on the integrity coefficient, the predicted functional coefficient, and the retention degree grade, a retention degree evaluation report is generated.

[0037] In step S604, it is necessary to explain in detail that the generation of the retention assessment report relies on the quantitative results of the integrity coefficient, predicted functionality coefficient, and retention level, and achieves information integration and visualization through a structured template. The report content covers four core modules: First, heritage overview, including basic information such as the name, geographical location, construction date, and historical function of the industrial heritage, combined with historical images and current photos to form a spatiotemporal comparison, intuitively showing the evolution of the physical form of the heritage; Second, analysis of assessment indicators, presenting the specific values ​​of the integrity coefficient, predicted functionality coefficient, and retention level in tabular form, and indicating the weighting configuration scheme (e.g., α=0.6, β=0.4), and simultaneously displaying intermediate calculation results such as feature matching degree and expert score mean, ensuring the traceability of the assessment process; Third, the basis for level determination, based on the R value range (excellent ≥0.85, ...). The report provides clear grading conclusions (Good 0.70-0.85, Average 0.50-0.70, Poor <0.50), explaining the logic behind weight selection in conjunction with heritage type and protection objectives. For example, it emphasizes the rationale for emphasizing the "integrity coefficient as the primary factor" for cultural heritage sites. Fourthly, it offers protection recommendations, proposing differentiated measures based on the grading. For example, it recommends applying for cultural heritage protection funds and conducting regular monitoring for "Excellent" heritage sites, and developing rescue restoration plans and restricting functional changes for "Poor" heritage sites. It also includes a note predicting the probability of maintaining functionality over the next n years corresponding to the functional coefficient (e.g., F=0.78 corresponds to a 78% probability of usable production functions within 10 years). The report includes a dynamic update prompt at the end, clarifying that when the correlation coefficient between feature matching and expert scores falls below a threshold (r<0.6) or when there are significant changes in heritage functional needs, the weights need to be recalibrated and a revised report generated. The final report is output in PDF format, supporting mixed text and graphics layouts and interactive data queries.

[0038] As an optional embodiment of the present invention, the expression for outputting the predicted functionality coefficient for the next n years is optionally: ,in, Indicates the future number Annual predicted functional coefficient This represents the output layer weight matrix. Indicates a historical time step. Indicates the length of the input time window. This represents the time attention weight, calculated using a time attention mechanism. Higher weights are assigned to key time steps (such as the year of major equipment overhaul) to enhance predictive sensitivity. This indicates the hidden state of the second LSTM layer. This represents the output layer bias vector.

[0039] As an optional embodiment of the present invention, the method may further include: S7, visual rendering is performed on the retention degree evaluation report to generate a dynamic heat map and a three-dimensional reconstruction model, and spatial distribution and time evolution characteristics of the industrial heritage are intuitively displayed; In step S7, it needs to be specified that the generation of the dynamic heat map is based on the spatial distribution data of the retention degree level, and the color gradient (such as red-yellow-green) is used to intuitively reflect the preservation state difference of different regions. Specifically, the industrial heritage is divided into regular grid units, and the retention degree level R value of each unit is mapped to a preset color scale: R≥0.85 shows dark green (excellent), 0.70≤R<0.85 shows light green (good), 0.50≤R<0.70 shows yellow (general), and R<0.50 shows red (poor). For example, the heat map of a certain steel plant may show a layered effect with the core production area in green (frequent equipment updates), the auxiliary facility area in yellow (partial aging), and the abandoned warehouse area in red. To enhance the information density, the heat map is superimposed with historical image slices, and users can observe the preservation state changes from 1950 to the present through the time slider. The process of the red area shrinking over time can intuitively verify the effectiveness of the protection measures.

[0040] The construction of the three-dimensional reconstruction model relies on the geometric feature vector in the multi-source heterogeneous data, and uses point cloud registration and texture mapping technology to restore the three-dimensional form of the heritage. First, the SfM (structure from motion) algorithm is used to process historical images of different years to generate a high-precision point cloud model. Second, laser scanning data is combined to complete the occluded area and construct a complete geometric framework. Finally, the texture information in the historical images is mapped to the model surface to form a three-dimensional scene with realistic and time dimension. For example, the reconstruction model of a certain textile mill can simultaneously display the brick wall in 1960 (extracted from black and white images) and the steel structure roof installed in 2020 (obtained from laser scanning), and users can observe the spatio-temporal trajectory of structural changes by rotating the viewing angle. To support interactive analysis, the model is embedded with voxelized encoding of the retention degree level: the R value of each voxel (three-dimensional pixel) is displayed through transparency adjustment, high R value area (such as well-preserved factory building) presents semi-transparent green, and low R value area (such as collapsed chimney) presents semi-transparent red, which helps decision makers quickly locate the protection priorities.

[0041] The dynamic rendering process introduces a timeline control module, with which users can specify the starting year and step size (e.g., every 5 years). The system then automatically generates an animation sequence showing the evolution of the heritage. In the animation, the color changes of the heat map and the shape changes of the 3D model are synchronized. For example, in 1980, the heat map shows an expansion of the red area (corresponding to the collapse of the roof in the 3D model), while in 2010, the heat map shows an increase in the green area (corresponding to the repaired factory in the 3D model). To improve data reliability, the rendering results are accompanied by metadata tags. By clicking on any area, users can view the original data of the integrity coefficient, functionality coefficient, and expert rating average for that location, as well as the shooting time and equipment parameters of the corresponding historical images. The final visualization report supports multi-terminal access, with real-time rendering on the browser side through WebGL technology, allowing decision-makers to perform spatial analysis and time backtracking without the need to install professional software.

[0042] In this embodiment, the above-mentioned shaft control module adopts a layered architecture design: the bottom layer integrates a time step resolver (e.g., converting user input of the starting year and step size into a system-recognizable time sequence format, such as converting "1950 starting, every 5 years" into a time array of [1950, 1955, 1960, …, 2025]), supporting user-defined starting year (e.g., 1950) and time interval (e.g., every 5 or 10 years), generating discrete time sequence nodes by analyzing time parameters; the middle layer deploys a time event trigger, which automatically triggers heat map updates and 3D model shape changes for the corresponding year when the user drags the time slider to a specific node (e.g., the 1980 equipment modification period); the upper layer builds a multi-modal synchronization engine to ensure the spatio-temporal alignment of color gradient (heat map) and structure deformation (3D model), such as when the timeline points to the 2010 earthquake repair period, the red area in the heat map is reduced, and the collapsed factory structure in the 3D model is repaired to the reinforced state. To improve operation smoothness, the module introduces a buffer preload mechanism, which renders the results of the current time node and the previous and next two time steps in advance, reducing the waiting time when the user drags the slider. At the same time, the shaft control module supports time jump operation, users can directly input the target year (e.g., 2020) or click on the marked points on the timeline (e.g., the years corresponding to major historical events), the system quickly locates to the specified time and updates the visualization content. For mobile access scenarios, the module is optimized for touch-friendly design, adjusting the timeline range through two-finger zoom, and long-pressing a time node to pop up an assessment indicator summary (e.g., integrity coefficient = 0.78, functionality coefficient = 0.65) for that year. Finally, the interactive timeline generated by the shaft control module serves as the core navigation tool of the visualization report, allowing decision-makers to intuitively trace the trajectory of the preservation state of the industrial heritage from construction to the present.

[0043] S8, associate the retention degree assessment report with the historical maintenance records, use knowledge graph technology to construct an industrial heritage life cycle graph, and output optimization maintenance suggestions and risk warning indicators.

[0044] In step S8, the static attributes (such as name, geographic location, construction year, heritage type) and dynamic indicators (integrity coefficient, functionality coefficient, retention degree level) of the industrial heritage are first extracted from the retention degree assessment report, and the time-sequenced maintenance events (such as repair time, repair location, repair method, repair cost) are extracted from the historical maintenance records to form a structured data layer. Secondly, define the entity relationship model: take the industrial heritage as the core entity, associate its "contains" sub-components (such as factory building, equipment, chimney), "experiences" maintenance events, "meets" protection standards (such as "Cultural Relics Protection Law"), "belongs to" heritage categories (such as important modern historical sites); the maintenance event entity is further associated with "uses" repair materials, "performs" repair teams, and "follows" technical specifications, etc. For example, in the life cycle graph of a certain steel plant, "blast furnace" as a sub-component entity is connected to the "furnace lining repair" event in 1985 through the "experiences" relationship, and the event is further associated with the "uses" "refractory brick" material entity and the "performs" "third repair team" team entity.

[0045] To enhance the semantic reasoning ability of the graph, introduce ontology library to define the logical rules between entities: such as "when the functionality coefficient <0.5 and no maintenance events are recorded in the past 3 years, trigger 'high risk' warning"; or "if the heritage type is 'protected unit' and the integrity coefficient <0.7, recommend 'apply for special repair funds'". For example, for a certain textile mill, the graph discovers through rule reasoning that the integrity coefficient of its "weaving workshop" sub-component decreased from 0.82 in 2018 to 0.65 in 2023, and there is no maintenance record after 2020, and automatically generates the optimization suggestion "suggest starting structural reinforcement in 2024, budget about 500,000 yuan". At the same time, the graph embeds time and space dimension analysis: arrange the maintenance events according to the time axis, combine with the change trend of the retention degree level, and identify the key intervention nodes (such as the functionality coefficient significantly improves after a major repair).

[0046] The generation of the risk early warning index relies on the association rule mining and anomaly detection algorithm in the graph. On the one hand, frequent item set mining is used to discover high-frequency co-occurrence patterns: such as the association rule "equipment aging (integrity coefficient annual decline > 0.1) → functional coefficient annual decline > 0.15 → failure within 3 years", which triggers an early warning when new data matches the pattern; on the other hand, the isolation forest algorithm is used to detect abnormal maintenance behavior: if the cost of a certain maintenance is significantly higher than the average cost of similar events (such as more than 2 times the median), and is not accompanied by an improvement in the retention level, it is marked as a "potential resource waste" risk. For example, the "machine tool overhaul" event in a machinery plant in 2022 cost 800,000 yuan, but the functional coefficient only increased from 0.68 to 0.71, and the graph automatically generated a warning "suggest reviewing the reasonableness of the maintenance plan".

[0047] The output of the optimized maintenance suggestion adopts a hierarchical recommendation mechanism: the basic layer generates general suggestions based on the retention level (such as "excellent" level heritage suggestion "full detection every 5 years", "poor" level heritage suggestion "immediately restrict use"); the advanced layer customizes the plan based on the heritage type and sub-component characteristics (such as recommending "modular update" for "light industry heritage" and "structural reinforcement first" for "heavy industry heritage"); the intelligent layer dynamically adjusts the suggestion using a reinforcement learning model: taking maintenance cost, functional recovery rate, and time efficiency as optimization objectives, it simulates the long-term effects of different maintenance strategies to output the optimal combination. For example, for the boiler equipment in a power plant, the graph recommends a phased plan of "replace the refractory layer in 2024 (cost 300,000 yuan, functional coefficient increased to 0.85) + upgrade the control system in 2026 (cost 200,000 yuan, functional coefficient increased to 0.92)", with a total cost lower than the one-time full renovation of 550,000 yuan, and a higher functional recovery rate.

[0048] The finally generated industrial heritage life cycle graph supports interactive query: users can ask questions in natural language (such as "which components have a functional coefficient decline of more than 0.2 since 2018?"), and the system returns structured answers based on the semantic analysis capabilities of the graph, and highlights the associated entities and evidence chains. The graph data is stored in RDF format, supporting federated query with external databases (such as the list of cultural relics protection units and the maintenance material price library), ensuring the timeliness and accuracy of the recommendations. For example, when the user queries "the maintenance team qualification of a certain heritage", the graph can link to the national enterprise credit information public system to verify whether the team has "cultural relics protection engineering construction qualification", avoiding invalid recommendations.

[0049] Embodiment 2 An industrial heritage retention level research and judgment system based on historical image contrast recognition, comprising: a processor; a memory for storing processor-executable instructions; The processor is configured to implement the historical image contrast-based identification method for judging the remaining degree of industrial heritage when executing the executable instructions.

[0050] It should be noted that the computer device includes a processor and a memory, and can further include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0051] The processor is configured to implement the historical image contrast-based identification method for judging the remaining degree of industrial heritage when executing the executable instructions.

[0052] The memory is configured to store various types of data to support the operation of the computer device, which can include, for example, instructions for operating any application or method on the computer device, and application-related data; the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0053] The multimedia component can include a screen, such as a touch screen, and an audio component for outputting and / or inputting audio signals; for example, the audio component can include a microphone for receiving external audio signals, and the received audio signals can be further stored in the memory or transmitted through the communication component; the audio component also includes at least one speaker for outputting audio signals.

[0054] The I / O interface provides an interface between the processor and other interface modules, which can be a keyboard, a mouse, a button, etc.; these buttons can be virtual buttons or physical buttons.

[0055] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or a combination of one or more of them, so the corresponding communication component can include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0056] As a preferred scheme of the present embodiment, the computer device can be implemented by one or more Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor or other electronic elements, for executing the above-mentioned industrial heritage retention degree judgment method based on historical image comparison and recognition.

[0057] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for judging the remaining degree of industrial heritage based on historical image contrast recognition, characterized in that, The method comprises: S1, collecting multi-source heterogeneous data of a target area and preprocessing; the multi-source heterogeneous data comprises a continuous year image sequence; S2, extracting a space-time feature vector in the continuous year image sequence by using a CNN network; S3, extracting a semantic feature vector in the continuous year image sequence by using a self-attention mechanism; S4, strengthening feature representation by using a space-channel dual attention SA module based on the space-time feature vector and the semantic feature vector, and obtaining a comprehensive feature vector; S5, implementing feature distribution alignment of an industrial quality inspection source domain and an industrial heritage target domain on the comprehensive feature vector by using adversarial optimization of a label predictor and a domain discriminator, and obtaining a feature matching degree; S6, calculating an integrity coefficient based on the feature matching degree and expert scores, predicting a functional coefficient based on the multi-source heterogeneous data through LSTM time series analysis, and outputting a retention degree evaluation report based on the integrity coefficient and the predicted functional coefficient.

2. The method of claim 1, wherein the historical image pair is obtained from a historical image database.

2. The method of claim 1, wherein the historical image pair is obtained from a historical image database. In step S2, the space-time feature vector in the continuous year image sequence is extracted, comprising: S201, performing standardization processing on the continuous year image sequence; S202, constructing a CNN network, and extracting space-time information of different scales based on the continuous year image sequence after standardization processing; S203, fusing the space-time information of different scales to obtain a space-time feature vector.

3. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 1, characterized in that, In step S3, the semantic feature vector in the continuous year image sequence is extracted, comprising: S301, segmenting the preprocessed continuous year image sequence, and then mapping the segmented continuous year image sequence to an initial token by using a learnable linear projection layer; S302, calculating similarity weight between each segmented continuous year image sequence based on the initial token by using a self-attention mechanism; S303, weighting and aggregating the initial token based on the similarity weight to generate a semantic feature vector.

4. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 1, characterized in that, In step S4, the comprehensive feature vector is obtained, comprising: S401, concatenating the space-time feature vector and the semantic feature vector in a channel dimension to obtain a fusion feature tensor; S402, performing spatial dimension compression on the fusion feature tensor to generate a single-channel spatial weight map; S403, performing global average pooling on the fusion feature tensor to generate a channel weight vector after spatial dimension compression; S404, multiplying the single-channel spatial weight map and the channel weight vector by broadcasting to generate a dual-attention weight matrix; performing feature enhancement by element-wise multiplication of the dual-attention weight matrix and the fusion feature tensor, and superimposing the space-time feature vector and the semantic feature vector through a residual connection to form a style-enhanced feature vector; S405, performing global average pooling on the style-enhanced feature vector to obtain a comprehensive feature vector, wherein the comprehensive feature vector comprises space-time evolution information and semantic association information.

5. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 1, characterized in that, In step S5, the feature matching degree is obtained, comprising: S501, constructing a label predictor by using a multi-layer perception structure, wherein the input of the label predictor is the comprehensive feature vector, and the output of the label predictor is an industrial heritage retention degree grade; S502, constructing a domain discriminator using a convolutional neural network structure, wherein an input of the domain discriminator is the comprehensive feature vector, and an output of the domain discriminator is a binary classification probability of a sample belonging to an industrial quality inspection source domain or an industrial heritage target domain; S503, through adversarial optimization training, minimizing a cross-entropy loss function of the label predictor, and maximizing a confusion loss function of the domain discriminator, wherein the confusion loss function is gradient reversed through GRL, so as to promote the alignment of the feature distribution of the comprehensive feature vector between the industrial quality inspection source domain and the industrial heritage target domain; S504, calculating a feature matching degree based on the binary classification probability output by the domain discriminator, wherein the feature matching degree reflects the alignment degree of the industrial heritage target domain feature and the industrial quality inspection source domain feature.

6. The method of claim 1 or 5, wherein the historical image-based contrast recognition of the industrial heritage preservation degree is determined by using a historical image-based contrast recognition model. The expression for calculating the feature matching degree is: wherein, denotes a feature matching degree, denotes a prediction probability of the domain discriminator on the target domain sample, denotes a Gaussian kernel bandwidth parameter, denotes a Softmax activation function, denotes a weight matrix of the domain discriminator, denotes a comprehensive feature vector, denotes a bias vector of the domain discriminator.

7. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 5, characterized in that, The output of the remaining degree evaluation report in step S6 includes: S601, calculating an integrity coefficient based on the feature matching degree and the expert scoring matrix through a weighted fusion model; S602, outputting a predicted functional coefficient in the next n years based on the spatio-temporal feature vector in the multi-source heterogeneous data using a double-layer LSTM network combined with a time attention mechanism to weight the key time steps; S603, calculating a weighted sum based on the integrity coefficient and the predicted functional coefficient to obtain the remaining degree level; S604, generating a remaining degree evaluation report based on the integrity coefficient, the predicted functional coefficient, and the remaining degree level.

8. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 7, characterized in that, The expression for outputting the predicted functional coefficient in the next n years is: in, Indicates the future number Annual predicted functional coefficient This represents the output layer weight matrix. Indicates a historical time step. Indicates the length of the input time window. This represents the time attention weight, calculated using a time attention mechanism. Higher weights are assigned to key time steps (such as the year of major equipment overhaul) to enhance predictive sensitivity. This indicates the hidden state of the second LSTM layer. This represents the output layer bias vector.

9. The method for assessing the preservation level of industrial heritage based on historical image comparison and recognition as described in claim 1, characterized in that, The method further includes: S7, visualizing and rendering the remaining degree evaluation report to generate a dynamic heat map and a three-dimensional reconstruction model, and intuitively displaying the spatial distribution and time evolution characteristics of the industrial heritage; S8, associating the remaining degree evaluation report with historical maintenance records, constructing an industrial heritage life cycle graph using knowledge graph technology, and outputting optimization maintenance suggestions and risk early warning indicators.

10. An industrial heritage retention degree judgment system based on historical image contrast recognition, characterized in that, The system includes: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the historical image contrast identification-based industrial heritage remaining degree evaluation method of any one of claims 1 to 9 when executing the executable instructions.

Images