Track traffic protection area engineering instrument detection method based on vision and collaborative filtering

Abstract

The invention discloses a rail transit protection area engineering instrument detection method based on vision and collaborative filtering. The method comprises the steps that S1, vision, sensing, history and other multi-source heterogeneous data are collected and normalized; s2, calculating the average error rate of each data source in real time, and dynamically updating the reliability weight of each data source according to the average error rate; s3, multiplying the normalized original feature value by the corresponding real-time reliability weight, and constructing a dynamically weighted data feature matrix; s4, applying a pre-training model based on a collaborative filtering matrix decomposition principle to decompose the weighting matrix, and generating a standardized instrument state feature vector fusing the potential association of the multi-source data and the real-time reliability of the multi-source data; and S5, inputting the feature vector into a risk assessment model for analysis, and comparing the feature vector with a dynamic risk threshold to realize accurate early warning. According to the invention, the accuracy and robustness of detection are obviously improved, and the crossing from passive monitoring to active early warning is realized.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology for rail transit protection zones, specifically relating to a method for detecting engineering machinery in rail transit protection zones based on vision and collaborative filtering. This method enables real-time and accurate monitoring of the operational status and potential risks of engineering machinery within rail transit protection zones, providing technical support for the safe operation of rail transit. Background Technology

[0002] Safety management of railway lines, especially the safety control of rail transit protection zones, has become a core aspect of ensuring the safe operation of railway lines. The activities of various engineering machinery within the protection zones, such as the unauthorized operation or abnormal condition of excavators, drilling rigs, and cranes, are one of the main hidden dangers that could lead to railway safety accidents.

[0003] Currently, monitoring of engineering equipment within protected areas mainly relies on manual inspections and single-sensor monitoring. However, these traditional methods suffer from three major drawbacks: First, manual inspections are inefficient and costly, and are limited by manpower and inspection routes, making it difficult to achieve full-time, full-area coverage of the vast protected area without blind spots, which can easily lead to monitoring blind spots and delays.

[0004] Secondly, single-sensor monitoring, whether it is a simple visual camera or a vibration or sound sensor, is highly susceptible to interference from complex environments. For example, weather conditions such as rain, snow, fog, haze, and low light at night can severely affect the accuracy of visual recognition; while environmental noise such as surrounding road traffic and nearby construction may interfere with vibration or acoustic sensors, resulting in a high rate of missed detections and false detections.

[0005] Third, existing technologies generally lack the ability for real-time analysis and proactive early warning, and most remain at the stage of retrospective analysis after the fact. They are unable to predict and intervene in potential risks in advance, making it difficult to effectively deal with sudden risks and resulting in a lag in security assurance. Summary of the Invention

[0006] This invention primarily addresses the technical problems of existing technologies, such as low efficiency of manual inspections, susceptibility to interference from single monitoring methods, and lack of proactive early warning capabilities. It provides a vision-based and collaborative filtering-based method for detecting engineering equipment in rail transit protection zones, which can integrate multi-source heterogeneous data, dynamically assess data reliability, and deeply mine potential correlations between data, thereby achieving intelligent safety management of rail transit protection zones.

[0007] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering, comprising the following steps: S1: Data Acquisition and Normalization: Real-time acquisition of multi-source data within the rail transit protection zone, and processing it into structured multi-dimensional raw features; the multi-source data includes at least visual features from visual camera data sources, sensor features from infrared, vibration and acoustic sensor data sources, and historical features retrieved from historical databases; the multi-dimensional raw features are normalized so that their values ​​are uniformly mapped to a preset numerical range. Visual features are obtained through the following method: High-definition intelligent cameras are deployed in key monitoring areas of the rail transit protection zone (such as construction boundaries and equipment parking areas) to collect dynamic image data of engineering machinery in real time at a preset frame rate. Standardized preprocessing is performed on the raw images: Gaussian filtering is used to remove noise interference, contrast enhancement algorithms are used to improve image details, and finally, size normalization processing is performed (uniformly adjusted to 640×480 pixels) to provide high-quality input for subsequent recognition tasks.

[0008] Then, lightweight object detection algorithms (such as YOLOv5s and a lightweight version of Faster R-CNN) are used to analyze the preprocessed images to accurately identify the type of engineering machinery (such as excavators and drilling rigs), its spatial coordinates, and its operating status (such as working / stationary). During the model training phase, a transfer learning strategy is employed. A hybrid training set (containing 50,000+ labeled samples) is constructed based on public datasets and field samples collected in protected areas. Data augmentation (rotation, scaling, and lighting simulation) is used to improve the model's generalization ability, ensuring an accuracy rate of ≥95%.

[0009] S2: Dynamic weight calculation: By comparing the detection results of each data source with the periodically acquired true values, the average error rate e of each data source within a preset sliding window (e.g., 100 times) is calculated and updated in real time. k According to the average error rate e k The following formula is used to dynamically calculate the reliability weight w of each data source at the current moment. k :w k =(1 / e k ) / (Σ m=1 n (1 / e m ), where n is the total number of data sources; This solution addresses the reliability differences of different data sources in complex environments (such as decreased camera recognition rate in rainy weather) by designing a weight adjustment mechanism based on real-time error feedback. A sliding window is used to calculate the recent error rate of each data source, dynamically allocating fusion weights to ensure that high-reliability data (such as visual data from sunny days) dominates. In the presence of backup data sources, if the error rate of a certain data source is e... k When the value is greater than 0.3, the backup data source switching mechanism is automatically triggered.

[0010] S3: Construct the data feature matrix: Multiply the normalized original feature values ​​obtained in step S1 by the real-time reliability weight w calculated in step S2 for the corresponding source data. k We obtain the weighted feature values; we collect the sample vectors composed of feature values ​​collected at several times to construct a dynamically weighted data feature matrix R; the dimension of the data feature matrix is ​​m×n, where m is the number of samples (i.e., the monitored engineering machinery related events or state samples) and n is the feature dimension (i.e., the various attributes describing the samples, such as machinery type, vibration frequency, number of historical failures, etc.). This matrix intuitively presents the original feature distribution of multi-source data.

[0011] S4: Data Feature Matrix Decomposition: Using a feature transformation model, the data feature matrix R obtained in step S3 is decomposed to obtain a standardized user latent vector matrix U that integrates the potential correlations of multi-source data and their real-time reliability. new U new The dimension of U is m×k, where k is the dimension of the latent space, obtained through training. new The i-th row represents the feature vector of the i-th sample (such as the state of the equipment at a certain moment) in the k-dimensional latent space, reflecting the latent attributes of the sample; Matrix factorization algorithms can uncover implicit correlations between multi-source data (visual, sensor, historical data) (such as the potential relationship between vibration frequency and equipment load), while removing outliers and redundant information from the original data, reducing noise interference, and generating standardized equipment state feature vectors. This provides high-quality input for subsequent risk assessment and ultimately improves the accuracy of equipment inspection in rail transit protection zones.

[0012] S5: Risk Assessment and Early Warning: The user latent vector U generated in step S4... new The data is input into the risk assessment model to calculate the final comprehensive risk value. The comprehensive risk value is then compared with a preset risk threshold. If the comprehensive risk value is greater than the risk threshold, an early warning mechanism is triggered.

[0013] The risk assessment model uses a weighted summation method to calculate the comprehensive risk value (covering six risk indicators, including abnormal equipment operation and out-of-bounds operations). When the risk value exceeds the preset threshold, the system automatically triggers a multi-level early warning mechanism (local audible and visual alarm + remote platform push + SMS notification to the responsible person), and simultaneously pushes risk location and handling suggestions.

[0014] Preferably, the feature transformation model described in step S4 is obtained and applied as follows: a) During the offline training phase, the data feature matrix R constructed based on historical data is used... t Perform matrix factorization to train and solidify a latent feature vector matrix V, R tAfter decomposition, we obtain the user latent vector matrix U and the feature latent vector matrix V. The training objective of matrix decomposition is to minimize the following loss function L: L=Σ (i,j)∈S (R t (i,j)–R t '(i,j))²+λ(Σ i ||U i ||²+Σ j ||V j ||²); Among them, R t (i,j) represents the predicted value of the j-th feature of the original i-th sample, R t '(i,j) represents the predicted value of the j-th feature of the i-th sample after reconstruction, R t '(i,j)=U i ·V j U i For the i-th row of U, V j T For the j-th row of V j The transpose of V j The eigenvector representing the j-th original feature (such as vibration sensor data) in the k-dimensional latent space reflects the latent meaning of the feature; the core objective of decomposition is to make the reconstruction matrix R t 'Approximate the original data feature matrix as closely as possible;' The first term of the loss function L is the reconstruction error, ensuring that R t The first term minimizes the deviation from the effective data points (set S) in R; the second term is a regularization term (λ=0.01), which avoids overfitting and enhances the model's adaptability to new data by limiting the vector magnitudes of U and V. The training process ensures the accuracy and generalization ability of the decomposition results.

[0015] b) In the online inference phase, i.e., step S4, the feature latent vector matrix V, which was fixed in step a), is loaded, and the data feature matrix R constructed based on the real-time collected data is decomposed to obtain the real-time user latent vector matrix U. new .

[0016] Preferably, the risk assessment model in step S5 is a Long Short-Term Memory (LSTM) network model, which assesses the risk through the user's latent vector matrix U. new We analyze the time series data to predict potential equipment operation risks.

[0017] This solution utilizes Long Short-Term Memory (LSTM) networks to mine the temporal patterns of equipment operation. By training the model with historical data, it can predict the operating status of the equipment in the next hour (such as predicting abnormal swing of a crane boom). This model reduces the prediction error by 40% compared to the traditional ARIMA model and can detect potential faults 15-30 minutes in advance.

[0018] The core update formula for the LSTM model is: Input gate: i t =σ(W xi x t +W hi h t-1 +b i ); Forgotten Gate: f t =σ(W xf x t +W hf h t-1 +b f ); Cell state: c t =f t ·c t-1 +t t ·tanh(W xc x t +W hc h t-1 +b c ); Output gate: o t =σ(W xo x t +W ho h t-1 +b o ); Hidden state: h t =o t ·tanh(c c ); Where x t Let h be the device state vector at time t. t-1 Let be the hidden state at time t-1, W be the weight matrix, b be the bias term, and σ be the sigmoid activation function.

[0019] Preferably, the risk threshold in step S5 is a dynamic early warning threshold T, which is calculated using the following formula: T = αT0 + (1-α)T h ; Wherein, T0 is the preset regional benchmark threshold based on the risk level zoning of the protected area (core area / buffer zone / impact zone) and the type of equipment (large machinery / small tools); T hα is the dynamic historical threshold calculated based on historical accident data from the past three years; α is the weighting coefficient that is adaptively adjusted based on real-time environmental parameters (such as rainfall and train traffic density).

[0020] The dynamic threshold adjustment model structure is as follows: The model is constructed using a "three-dimensional + dynamic parameter" approach, and the formula serves as the final integration tool for the rules across these dimensions. 1) The basic layer is divided based on region and equipment type: The protected area is divided into core area, buffer zone, and impact zone according to risk level, and the equipment is divided into large machinery and small tools according to risk degree, forming 6 combined scenarios (3 areas × 2 equipment types, core / buffer / impact + large / small). Preset a differentiated baseline threshold T0 for each scenario (e.g., large machinery in the core area T0=0.5, small tools in the affected area T0=0.9).

[0021] The dynamic layer makes corrections based on historical data and the real-time environment: For each scenario, T is calculated based on historical accident data from the past three years (such as the distribution of past accident risk values ​​in that scenario). h The initial value; The weighting coefficient α is adjusted based on real-time environmental parameters (rainfall, train density, etc.). For example, when rainfall is ≥50mm, α is reduced by 0.2, allowing T to... h (As the weight increases, the threshold becomes more sensitive).

[0022] The output layer integrates the formulas to make the threshold effective: The formula T = αT0 + (1-α)T is used. h Calculate the final threshold, T0, T in different scenarios. h The parameters α and β are different, enabling three-dimensional dynamic adjustment of "region + equipment + environment". By automatically optimizing the early warning threshold, the false alarm rate can be reduced by more than 50%.

[0023] Preferably, the normalization process in step S1 refers to linearly mapping the values ​​of all original features to an interval.

[0024] Preferably, the true value mentioned in step S2 is obtained by periodically sampling in the following manner: a) The system backend records the raw data and preliminary test results from each data source; b) Randomly select samples from the recorded data within each preset period; c) The extracted samples are pushed to the manual review interface, where maintenance personnel view and judge the samples and provide the true judgment result as the true value.

[0025] The substantial effects of this invention are: 1. Significantly Improved Detection Accuracy and Robustness: This invention uniquely combines a dynamic weighting mechanism with a collaborative filtering model. By evaluating the reliability of each data source in real time and weighting the input features, it effectively suppresses the interference of adverse factors such as severe weather and environmental noise on a single data source, ensuring the quality of the model input. Furthermore, through collaborative filtering, it deeply mines the inherent correlations between features from different data sources, forming complementary verification. Compared to single visual recognition or single sensor monitoring, the detection accuracy can be improved to over 95%, while the false negative rate and false positive rate are reduced by 80% and 50%, respectively.

[0026] 2. Achieved a leap from passive monitoring to proactive early warning: This invention can not only identify the current state, but also combine time series analysis models such as LSTM to predict the trend of the fused standardized state vector. It can detect potential equipment abnormalities or illegal operation trends 15-30 minutes in advance, moving the safety management checkpoint forward, changing post-event handling to pre-event prevention, and improving the early warning response speed to the second level.

[0027] 3. Optimized Operational Efficiency and Cost-Effectiveness: By achieving automated real-time monitoring across the entire area and around the clock, management personnel are freed from heavy and inefficient manual inspections, increasing detection efficiency by over 60% and reducing labor costs by 70%. Simultaneously, accurate early warnings significantly reduce unnecessary attendance due to false alarms and effectively prevent line delays and even safety accidents caused by equipment failures or violations of operating procedures, demonstrating extremely high economic and promotional value. Attached Figure Description

[0028] Figure 1 This is a flowchart of a method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering, according to the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0030] Example: This example illustrates a vision-based and collaborative filtering-based method for detecting engineering equipment in rail transit protection zones. Figure 1 As shown, it includes the following steps: S1: Data Acquisition and Normalization: In this step, the system first collects multi-source heterogeneous data in real time through diversified front-end devices deployed at key nodes of the rail transit protection zone (such as construction boundaries, equipment parking areas, and areas adjacent to the line).

[0031] The multi-source data includes at least: a) Visual features: Dynamic image data acquired in real time by high-definition intelligent cameras at a preset frame rate. This raw image data is processed by backend target detection algorithms (e.g., lightweight YOLOv5s or Faster R-CNN models) to extract structured visual features, such as: machinery type (excavator, drilling rig, crane, etc.), precise spatial coordinates (x, y, z), and operating status (working / stationary). b) Sensor features: Non-visual data acquired in real time by a sensor network. This network may include: vibration sensors for detecting specific vibration frequencies and amplitudes on the ground caused by large machinery operations; acoustic sensors for collecting and identifying unique acoustic signals emitted by different machinery during operations; and other types of sensors such as infrared sensors. This data is processed into quantified sensor features. c) Historical features: Historical data related to the current monitoring target retrieved from the backend operation and maintenance database. For example, the historical failure rate of this type of equipment, the types and frequencies of past violations in this area, etc. This data constitutes historical features.

[0032] Because the features from different sources have vastly different dimensions and numerical ranges (e.g., coordinate values ​​may be in the thousands, while state values ​​may only be 0 or 1), all original features must be normalized to ensure the fairness and accuracy of subsequent weighting and model calculations. In this embodiment, the normalization process refers to mapping all the values ​​of the original features to a preset numerical range, such as an interval, through a linear mapping (e.g., the max-min-value normalization method).

[0033] S2: Dynamic Weight Calculation: This step is crucial for the invention to achieve adaptability to environmental changes. Its core idea is that the reliability of any data source is not constant, and the system should be able to dynamically assess and quantify this reliability.

[0034] Specifically, the system uses a feedback mechanism to calculate the dynamic reliability weight w of each data source. k The core of this mechanism lies in comparing the detection results from the data source with the actual values.

[0035] The true value is obtained through a cost-controlled periodic sampling method, specifically: a) The system backend continuously records the raw data (such as video clips) and preliminary detection results reported by various data sources. b) Within each preset period (e.g., every 8 hours), the system extracts a portion of samples from a massive historical record according to certain rules (e.g., random or focused on high-risk events). c) The extracted samples (including videos, sensor maps, and the system's preliminary judgment) are pushed to a manual review interface for operation and maintenance personnel to view and judge, providing a final, authoritative judgment result, which is then used as the true value.

[0036] Once the true values ​​are obtained, the system can calculate the average error rate e for each data source (visual camera, vibration sensor, etc.) within a preset sliding window (e.g., the most recent 100 calibrated detection events). k Subsequently, based on the average error rate e k The following formula is used to dynamically calculate the reliability weight w of each data source at the current moment. k :w k =(1 / e k ) / (Σ m=1 n (1 / e m The formula is denoted as n, where n is the total number of data sources. The physical meaning of this formula is that the lower the recent average error rate of a data source, the higher its reliability, and therefore it is given a higher weight. The sum of the weights of all data sources is equal to 1.

[0037] S3: Constructing the Data Feature Matrix: This step is another core innovation that distinguishes this invention from existing technologies. It creatively uses the dynamic weights calculated in the previous step to construct the input matrix of the collaborative filtering model. Specifically, the system multiplies the normalized original feature values ​​obtained in step S1 by the real-time reliability weight w calculated in step S2 for the corresponding source data source. k We obtain the weighted eigenvalues.

[0038] For example, if the normalized value of the visual feature device state in S1 is 1.0, and the camera data source is in good weather at this moment, the reliability weight w calculated in S2... vision If the value is 0.9, then the final weighted eigenvalue is 1.0 × 0.9 = 0.9. Conversely, if the fog causes w to... vision When the value is reduced to 0.2, the weighted eigenvalue is only 0.2. In this way, the real-time reliability information of the data source is cleverly encoded into the eigenvalue itself.

[0039] Subsequently, the sample vectors composed of these weighted feature values ​​collected at several time points are aggregated row by row to construct a dynamically weighted, high-dimensional, and typically sparse data feature matrix R. Each row of this matrix represents an observation sample, and each column represents a weighted feature dimension.

[0040] S4: Data Feature Matrix Decomposition: This step aims to use a machine learning model to deeply mine the complex correlations hidden in the weighted feature matrix R constructed in the previous step. The system uses a pre-trained feature transformation model to perform matrix decomposition, or decomposition, on the data feature matrix R constructed in real time in step S3. This decomposition is performed rapidly during the inference phase. The ultimate goal is to obtain a standardized, low-dimensional, dense user latent vector matrix U that integrates the potential correlations of multi-source data (e.g., a specific vibration pattern always appears simultaneously with the visual features of the excavator) and its real-time reliability. new The user here refers to the sample, U. new Each row is a standardized instrument state feature vector.

[0041] The acquisition and application of feature transformation models is a two-stage process involving offline training and online inference. a) During the offline training phase, the system uses massive amounts of historical data and constructs a historical weighted feature matrix R through steps S1-S3. t Then, by adjusting R t Matrix factorization is performed to train and solidify a latent feature vector matrix V. The training objective of matrix factorization is to minimize the following loss function L: L = Σ (i,j)∈S (R t (i,j)–R t '(i,j))²+λ(Σ i ||U i ||²+Σ j ||V j ||²); where the first term is the reconstruction error term, used to ensure that the matrix after decomposition and reconstruction is as consistent as possible with the original matrix; the second term is the regularization term, used to prevent the model from overfitting. Through repeated iterations using optimization algorithms such as stochastic gradient descent (SGD), a fixed V matrix is ​​finally obtained, which contains the universal patterns in all historical data.

[0042] b) During the online inference phase, i.e., step S4 executed during real-time detection, the system loads the feature latent vector matrix V fixed in step a) and performs rapid decomposition calculation on the weighted feature matrix R constructed based on real-time collected data, thereby efficiently obtaining the real-time user latent vector matrix U. new .

[0043] S5: Risk Assessment and Early Warning: This step is the final stage in realizing the application value of this invention. The system will use the user latent vector U generated in step S4, which has extremely high information density, to... new Each line is input into a preset risk assessment model to calculate the final comprehensive risk value.

[0044] The risk assessment model is a Long Short-Term Memory (LSTM) network model. Because U new It is a state vector on a continuous time series. The LSTM model can learn its temporal dependencies very effectively. It can not only assess the current risk, but also predict the risk evolution trend in the next short period of time (e.g., 15-30 minutes), such as predicting that the equipment will soon experience abnormal swings, thus achieving true forward-looking early warning.

[0045] Finally, the calculated comprehensive risk value is compared with the preset risk threshold. This risk threshold is not a fixed value, but a dynamic early warning threshold T, which is calculated using the following formula: T = αT0 + (1-α)T h ; Wherein, T0 is a pre-set regional benchmark threshold based on the risk level of the protected area (core area / buffer zone / impact area) and the type of equipment (large / small); T h The threshold is a dynamic historical threshold calculated based on historical accident data from the past three years in this scenario; while α is a weighting coefficient that is adaptively adjusted based on real-time environmental parameters (such as rainfall and train traffic density). For example, during heavy rain, α is lowered, making the threshold more inclined to adopt a stricter historical threshold T. h When the overall risk value exceeds the dynamic warning threshold T, the system automatically triggers a multi-level warning mechanism, such as: local audible and visual alarms, remote platform push of warning details, and sending alarm SMS messages containing risk location and handling suggestions to the mobile phones of relevant responsible persons.

[0046] This solution breaks through the traditional single detection mode, pioneering a coupled architecture of visual recognition and collaborative filtering: visual recognition provides intuitive features such as device appearance and location, while collaborative filtering uncovers implicit correlations in multi-source data (such as the correlation between vibration frequency and device load), forming a complementary verification. This fusion mechanism improves detection accuracy by 23% compared to single visual recognition and 35% compared to single sensor monitoring.

[0047] This method can be seamlessly integrated into existing rail transit operation and maintenance management platforms, providing operators with full-process support from equipment monitoring and risk warning to emergency response, significantly reducing the incidence of safety accidents within protected areas. A pilot test on a subway line verified that after applying this invention, equipment violations within the protected area decreased by 92%, and line delays caused by equipment failures were eliminated, demonstrating its high potential for widespread adoption.

[0048] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0049] Although this document uses various terms extensively, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of the invention; interpreting them as any additional limitation would contradict the spirit of the invention.

Claims

1. A method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering, characterized in that, Includes the following steps: S1: Data Acquisition and Normalization: Real-time acquisition of multi-source data within the rail transit protection zone, and processing of it into structured multi-dimensional raw features; the multi-source data includes at least visual features from visual camera data sources, sensor features from vibration and acoustic sensor data sources, and historical features retrieved from historical databases; the multi-dimensional raw features are normalized so that their values ​​are uniformly mapped to a preset numerical range. S2: Dynamic weight calculation: By comparing the detection results of each data source with the periodically acquired true values, the average error rate e of each data source within a preset sliding window is calculated and updated in real time. k According to the average error rate e k The following formula is used to dynamically calculate the reliability weight w of each data source at the current moment. k :w k =(1 / e k ) / (Σ m=1 n (1 / e m ), where n is the total number of data sources; S3: Construct the data feature matrix: Multiply the normalized original feature values ​​obtained in step S1 by the real-time reliability weight w calculated in step S2 for the corresponding source data. k We obtain the weighted feature values; we then aggregate the sample vectors composed of the feature values ​​collected at several time points to construct a dynamically weighted data feature matrix R. S4: Data Feature Matrix Decomposition: Using the feature transformation model, the data feature matrix R obtained in step S3 is decomposed to obtain the user latent vector matrix U. new ; S5: Risk Assessment and Early Warning: The user latent vector U generated in step S4... new The data is input into the risk assessment model to calculate the final comprehensive risk value. The comprehensive risk value is compared with a preset risk threshold. When the comprehensive risk value is greater than the risk threshold, an early warning mechanism is triggered.

2. The method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering according to claim 1, characterized in that, The feature transformation model described in step S4 is obtained and applied as follows: a) During the offline training phase, the data feature matrix R constructed based on historical data is used... t Perform matrix factorization to train and solidify a latent feature vector matrix V, R t After decomposition, we obtain the user latent vector matrix U and the feature latent vector matrix V. The training objective of matrix decomposition is to minimize the following loss function L: L=Σ (i,j)∈S (R t (i,j)–R t '(i,j))²+λ(Σ i ||U i ||²+Σ j ||V j ||²); Among them, R t (i,j) represents the predicted value of the j-th feature of the original i-th sample, R t '(i,j) represents the predicted value of the j-th feature of the i-th sample after reconstruction, R t '(i,j)=U i ·V j U i For the i-th row of U, V j T Let V be the transpose of the j-th row, S be the set of valid data points, and λ be the regularization coefficient. b) In the online inference phase, i.e., step S4, the feature latent vector matrix V, which was fixed in step a), is loaded, and the data feature matrix R constructed based on the real-time collected data is decomposed to obtain the real-time user latent vector matrix U. new .

3. The method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering according to claim 1, characterized in that, The risk assessment model in step S5 is a Long Short-Term Memory (LSTM) network model, which uses the user's hidden vector matrix U... new We analyze the time series data to predict potential equipment operation risks.

4. A method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering according to claim 1 or 3, characterized in that, The risk threshold in step S5 is a dynamic early warning threshold T, which is calculated using the following formula: T=αT0+(1-α)T h ; Wherein, T0 is the regional benchmark threshold preset according to the risk level zoning of the protected area and the type of equipment; T h α is the dynamic historical threshold calculated based on historical accident data from the past three years; α is the weighting coefficient that is adaptively adjusted based on real-time environmental parameters.

5. The method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering according to claim 1, characterized in that, The normalization process in step S1 refers to linearly mapping all the original feature values ​​to an interval.

6. A method for detecting engineering equipment in rail transit protection zones based on vision and collaborative filtering according to claim 1 or 5, characterized in that, The true value mentioned in step S2 is obtained through periodic sampling in the following manner: a) The system backend records the raw data and preliminary test results from each data source; b) Randomly select samples from the recorded data within each preset period; c) The extracted samples are pushed to the manual review interface, where maintenance personnel view and judge the samples and provide the true judgment result as the true value.

Images