AI attitude analysis-based high-altitude operation safety harness dangerous attitude early warning method and system

Abstract

The invention discloses an AI attitude analysis-based high-altitude operation safety harness dangerous attitude early warning method and system, and the method comprises the steps: collecting the real-time acceleration, angular velocity and attitude angle data of a safety harness multi-axis sensor, and carrying out the denoising, time synchronization and standardization preprocessing; inputting a pre-training attitude signal coding model to obtain a real-time attitude feature vector; and calculating the similarity between the vector and a feature vector in a preset dangerous attitude feature library, and if the similarity is greater than a preset threshold value, determining that the attitude is a dangerous attitude, triggering an audible and visual alarm to an operator and sending early warning information to a monitoring terminal. According to the method, accurate real-time recognition and active early warning of dangerous postures are achieved through the AI technology, the initiative and reliability of high-altitude operation safety protection are effectively improved, and the accident risk is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-altitude operation safety, in particular to a high-altitude operation safety harness dangerous posture early warning method and system based on AI posture analysis. BACKGROUND

[0002] High-altitude operation belongs to the field of high-risk operation, and the safety harness is one of the core protective equipment for the operating personnel. The traditional safety harness mainly relies on passive protection, and lacks active identification and early warning capability for dangerous postures of operating personnel. The existing active early warning scheme has obvious deficiencies: some rely on manual monitoring, which has poor real-time performance and limited coverage, and is difficult to cope with complex operation scenarios; some use simple sensor data collection and threshold judgment, without combining artificial intelligence technology to analyze posture data in depth, and the recognition accuracy of complex dangerous postures such as bending imbalance and excessive body tilt is low, which is prone to false positives and false negatives, and cannot timely and effectively avoid safety risks. Therefore, an active early warning method capable of accurately and timely identifying dangerous postures is urgently needed to improve the safety protection level of high-altitude operation. SUMMARY

[0003] The purpose of the present application is to provide a high-altitude operation safety harness dangerous posture early warning method and system based on AI posture analysis.

[0004] In a first aspect, the present application provides a high-altitude operation safety harness dangerous posture early warning method based on AI posture analysis, comprising:

[0005] Collecting real-time posture sensing data of the high-altitude operation safety harness, the real-time posture sensing data including acceleration data, angular velocity data and posture angle data collected by multi-axis sensors deployed on the safety harness;

[0006] Preprocessing the real-time posture sensing data to obtain standardized posture sensing data, the preprocessing including data denoising, time synchronization and standardization processing;

[0007] Inputting the standardized posture sensing data into a pre-trained posture signal coding model to obtain a real-time posture feature vector corresponding to the real-time posture sensing data;

[0008] Based on a pre-set dangerous posture feature library, calculating the similarity between the real-time posture feature vector and each dangerous posture feature vector in the pre-set dangerous posture feature library;

[0009] If the similarity is greater than a pre-set dangerous threshold, it is determined that the current posture is a dangerous posture, and a safety warning is triggered, including issuing an audible and visual alarm to the high-altitude operator and sending warning information to a monitoring terminal.

[0010] In a possible implementation, the posture signal coding model is trained by the following method, comprising:

[0011] obtaining a plurality of posture sensor data combinations and a safety state identifier of each posture sensor data in the plurality of posture sensor data combinations, the posture sensor data combination comprising two posture sensor data with close posture amplitudes;

[0012] obtaining a posture feature vector of each posture sensor data in the plurality of posture sensor data combinations according to a multi-dimensional posture feature extraction network;

[0013] determining a first state feature of each posture sensor data in the plurality of posture sensor data combinations based on the posture feature vector of each posture sensor data in the plurality of posture sensor data combinations, the first state feature of the posture sensor data corresponding to a safety state level of the posture sensor data in the calculation result;

[0014] determining a second state feature of each posture sensor data in the plurality of posture sensor data combinations based on the safety state identifier of each posture sensor data in the plurality of posture sensor data combinations, the second state feature of the posture sensor data corresponding to a pre-set safety state level of the posture sensor data;

[0015] for a reference posture sensor data combination, determining a first deviation degree of the reference posture sensor data combination based on the posture feature vector of at least one posture sensor data in the reference posture sensor data combination; the reference posture sensor data combination is any posture sensor data combination;

[0016] determining a pending safety state of the reference posture sensor data combination, the pending safety state of the reference posture sensor data combination comprising one posture sensor data in the reference posture sensor data combination and a target posture sensor data combination, the target posture sensor data combination being a posture sensor data combination in the plurality of posture sensor data combinations except the reference posture sensor data combination and having a posture difference degree value within a set threshold range;

[0017] determining a second deviation degree of the reference posture sensor data combination based on the posture feature vector of each posture sensor data in the pending safety state of the reference posture sensor data combination;

[0018] The multi-dimensional posture feature extraction network is updated in network parameters based on the first state feature and the second state feature of each posture sensor instance in the posture sensor instance combination and the first deviation degree and the second deviation degree of the posture sensor instance combination, to obtain a posture signal coding model.

[0019] In a possible implementation, the security state identifier of the posture sensor instance is one or more, and determining the second state feature of each posture sensor instance in the posture sensor instance combination based on the security state identifier of each posture sensor instance in the posture sensor instance combination comprises:

[0020] For any security state identifier of each posture sensor instance in the posture sensor instance combination, a preset security state level corresponding to the any security state identifier is determined from a plurality of preset security state levels, and a posture feature of the preset security state level corresponding to the any security state identifier is taken as a posture feature of the any security state identifier.

[0021] For any posture sensor instance in the posture sensor instance combination, a second state feature of the any posture sensor instance is determined based on the posture feature of one or more security state identifiers of the any posture sensor instance.

[0022] In a possible implementation, the security state identifier of the posture sensor instance is one or more, and before the posture feature of the preset security state level corresponding to the any security state identifier is taken as the posture feature of the any security state identifier, the method further comprises:

[0023] An initial state association tensor is obtained, each state unit of a first state dimension of the initial state association tensor corresponds to each preset security state level, and each state unit of a second state dimension of the initial state association tensor corresponds to each preset security state level.

[0024] For any one of the plurality of posture sensing data instances, according to a plurality of safety state identifiers of the any one of the plurality of posture sensing data instances, an effective association value is written at a state dimension intersection unit of a state cell corresponding to a first state dimension and a state cell corresponding to a second state dimension, the state cell corresponding to the first state dimension being a first feature dimension position corresponding to a preset safety state level corresponding to one safety state identifier of the any one of the plurality of posture sensing data instances in the initial state association tensor, and the state cell corresponding to the second state dimension being a second feature dimension position corresponding to a preset safety state level corresponding to other safety state identifiers of the any one of the plurality of posture sensing data instances except the one safety state identifier in the initial state association tensor;

[0025] The effective association values at the state dimension intersection units of each state cell of the first state dimension and each state cell of the second state dimension in the initial state association tensor are superimposed to obtain the state association mapping tensor, and any data in the state association mapping tensor represents a cooperative intensity value of a preset safety state level corresponding to the first state dimension and a preset safety state level corresponding to the second state dimension being observed at the same time, and one safety state identifier is one preset safety state level.

[0026] A first preset safety state level is randomly selected from the preset safety state levels without determined posture features.

[0027] A data slice at a first feature dimension position in an unassigned state is randomly selected from the posture disturbance tensor template as the posture feature of the first preset safety state level.

[0028] Based on each effective association value in the first feature dimension position corresponding to the first preset safety state level in the state association mapping tensor, a second preset safety state level corresponding to the second state dimension and without determined posture features is determined.

[0029] A data slice at a first feature dimension position in an unassigned state with the highest similarity to the posture feature of the first preset label is determined from the posture disturbance tensor template as the posture feature of the second preset safety state level.

[0030] if there is the preset safety state level of the undetermined attitude feature, starting from the first preset safety state level randomly selected from the preset safety state level of the undetermined attitude feature, and circulating until there is no preset safety state level of the undetermined attitude feature, the dimension of the attitude disturbance tensor template is greater than the number of preset safety state levels.

[0031] if there is the preset safety state level of the undetermined attitude feature, starting from the first preset safety state level randomly selected from the preset safety state level of the undetermined attitude feature, and circulating until there is no preset safety state level of the undetermined attitude feature, the dimension of the attitude disturbance tensor template is greater than the number of preset safety state levels.

[0032] In a possible implementation, the first deviation degree of the reference attitude sensor instance combination is determined based on the attitude feature vector of at least one attitude sensor instance in the reference attitude sensor instance combination, including:

[0033] The attitude feature vector of at least one attitude sensor instance in the reference attitude sensor instance combination is binarized to obtain a binary attitude feature vector of at least one attitude sensor instance in the reference attitude sensor instance combination.

[0034] The first deviation degree is determined based on the attitude feature vector and the binary attitude feature vector of at least one attitude sensor instance in the reference attitude sensor instance combination.

[0035] In a possible implementation, the pending safety state of the reference attitude sensor instance combination is determined, including:

[0036] A first attitude difference degree value between one attitude sensor instance in the reference attitude sensor instance combination and one attitude sensor instance in each other attitude sensor instance combination is determined, the each other attitude sensor instance combination being each attitude sensor instance combination in the plurality of attitude sensor instance combinations except the reference attitude sensor instance combination.

[0037] A target attitude sensor instance combination with a first attitude difference degree value less than a first reference attitude difference degree value is determined from the each other attitude sensor instance combination.

[0038] determine the pending safety state of the reference attitude sensor instance combination based on one attitude sensor instance in the target attitude sensor instance combination and the reference attitude sensor instance combination.

[0039] In a possible implementation, the determining the target attitude sensor instance combination with the first attitude difference degree value less than the first reference attitude difference degree value from each of the other attitude sensor instance combinations comprises:

[0040] determining the target attitude sensor instance combination with the first attitude difference degree value less than the first reference attitude difference degree value from preset attitude sensor instance combinations, the preset attitude sensor instance combination being the other attitude sensor instance combination with the first attitude difference degree value not less than the second reference attitude difference degree value in each of the other attitude sensor instance combinations, the first reference attitude difference degree value being greater than the second reference attitude difference value.

[0041] In a possible implementation, the determining the second bias degree of the reference attitude sensor instance combination based on the attitude feature vectors of each attitude sensor instance in the pending safety state of the reference attitude sensor instance combination comprises:

[0042] determining a second attitude difference degree value between one attitude sensor instance and another attitude sensor instance in the reference attitude sensor instance combination based on the attitude feature vector of the one attitude sensor instance and the attitude feature vector of the another attitude sensor instance in the pending safety state of the reference attitude sensor instance combination, determining a third attitude difference degree value between the one attitude sensor instance in the reference attitude sensor instance combination and one attitude sensor instance in the target attitude sensor instance combination based on the attitude feature vector of the one attitude sensor instance in the reference attitude sensor instance combination and the attitude feature vector of the one attitude sensor instance in the target attitude sensor instance combination;

[0043] determining the second bias degree of the reference attitude sensor instance combination based on the second attitude difference degree value and the third attitude difference degree value.

[0044] In a possible implementation, the performing network parameter update on the multi-dimensional attitude feature extraction network based on the first state feature, the second state feature of each attitude sensor instance in the plurality of attitude sensor instance combinations and the first bias degree, the second bias degree of the plurality of attitude sensor instance combinations to obtain the attitude signal coding model comprises:

[0045] determine a posture feature deviation degree based on the first state feature and the second state feature of each posture sensor data in the plurality of posture sensor data combinations;

[0046] determine a posture feature vector deviation degree based on the first deviation degree and the second deviation degree of the plurality of posture sensor data combinations;

[0047] update network parameters of the multi-dimensional posture feature extraction network based on the posture feature deviation degree and the posture feature vector deviation degree, to obtain a posture signal coding model.

[0048] In a second aspect, an embodiment of the present application provides a server system, comprising a server configured to execute the method of the first aspect.

[0049] Compared with the prior art, the present application has the following beneficial effects: the AI posture analysis based high-altitude operation safety harness dangerous posture early warning method and system provided by the present application collects real-time acceleration, angular velocity and posture angle data of a multi-axis sensor of a safety harness, and after denoising, time synchronization and standardization preprocessing, inputs a pre-trained posture signal coding model to obtain a real-time posture feature vector; calculates the similarity of the vector and a feature vector in a preset dangerous posture feature library, and if the similarity is greater than a preset threshold, determines that it is a dangerous posture, triggers an audible and visual alarm for the operator and sends early warning information to a monitoring terminal. The present application realizes accurate and real-time identification and active early warning of dangerous postures through AI technology, effectively improves the initiative and reliability of high-altitude operation safety protection, and reduces the risk of accidents. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0051] Figure 1 The step flowchart of the AI posture analysis based high-altitude operation safety harness dangerous posture early warning method provided by the present application is shown in the following figure:

[0052] Figure 2 The structural schematic block diagram of the computer device provided by the present application is shown in the following figure: DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0054] The specific embodiments of the present application will be described in detail below with reference to the drawings.

[0055] In order to solve the technical problems in the foregoing background art, Figure 1 The flowchart of the high-altitude operation safety harness dangerous posture early warning method based on AI posture analysis provided by the embodiments of the present disclosure is as follows, and the high-altitude operation safety harness dangerous posture early warning method based on AI posture analysis will be described in detail.

[0056] Step S201, real-time posture sensing data of a high-altitude operation safety harness is collected, the real-time posture sensing data including acceleration data, angular velocity data and posture angle data collected by a multi-axis sensor deployed on the safety harness;

[0057] Step S202, the real-time posture sensing data is preprocessed to obtain standardized posture sensing data, the preprocessing including data denoising, time synchronization and standardization processing;

[0058] Step S203, the standardized posture sensing data is input into a pre-trained posture signal coding model to obtain a real-time posture feature vector corresponding to the real-time posture sensing data;

[0059] Step S204, based on a pre-set dangerous posture feature library, the similarity between the real-time posture feature vector and each dangerous posture feature vector in the pre-set dangerous posture feature library is calculated;

[0060] Step S205, if the similarity is greater than a pre-set dangerous threshold, it is determined that the current posture is a dangerous posture, and a safety early warning is triggered, the safety early warning including issuing an audible and visual alarm to a high-altitude operation personnel and sending early warning information to a monitoring terminal.

[0061] In the embodiment of the application, an exemplary building service company undertakes the exterior wall cleaning project of the 30-45 floor of the 5A-level office building "Global Financial Center" in the city center. The operation personnel are equipped with intelligent safety belts integrated with multi-axis sensors, the server is deployed in the construction monitoring room, and the Ubuntu 20.04 operating system and customized monitoring software are carried. The LoRa module communicates with the belt, and the ground monitoring terminal is real-time guarded by the safety officer. The operation personnel Zhang wears an intelligent belt on the 38th floor at 9:15 am on May 20, 2024, and the belt shoulder strap left and right, waistband front side each integrated with 1 MPU9250 nine-axis sensor, left leg belt integrated with 1 MPU6050 six-axis sensor, all sensors collect acceleration, angular velocity and attitude angle data at a sampling frequency of 50 Hz. Each data contains sensor ID, GPS timestamp and three-axis data, and sends 128-byte data packets to the server every 20 ms through the LoRa module. The server LoRa receiving module receives and stores in the Redis cache area in real time, and sets a 10-second sliding window to avoid data loss.

[0062] After the server monitoring software reads the data from the Redis cache area, it first performs data preprocessing to ensure input quality: for high-frequency noise caused by belt friction and tool vibration during high-altitude operation, 3-layer wavelet transform decomposition is performed using db4 wavelet basis, and low-frequency approximation coefficients and high-frequency detail coefficients are extracted. After threshold truncation of high-frequency coefficients, the data is reconstructed to correct the X-axis acceleration spike noise of the shoulder strap sensor; for the attitude angle drift of the magnetometer disturbed by the metal structure of the office building, the Kalman filter is used to fuse the accelerometer and gyroscope data, and the heading angle drift is controlled within 0.5°; for the timestamp deviation of the four sensors caused by the difference in crystal oscillator, the GPS timing module standard time is taken as the reference, the time deviation value of each data packet is calculated, and the data points are linearly interpolated or repeated data is discarded to achieve time synchronization within 1 ms; finally, the Min-Max standardization method is used to call the pre-stored statistical parameters to normalize the acceleration, angular velocity and attitude angle data to the [0, 1] interval, and integrate them into a 50×4×33 multi-dimensional time series matrix, where 50 is the number of time steps, 4 is the number of sensors, and 33 is the total number of features.

[0063] After the preprocessing is completed, the server inputs the standardized posture sensing data into a pre-trained CNN-LSTM fusion posture signal coding model. The model first extracts local features of the sensor and spatial correlation of multiple sensors through 3 convolutional layers. The first layer uses 32 3x3 convolutional kernels to extract single sensor features, the second layer uses 64 3x3 convolutional kernels to strengthen feature expression, and the third layer uses 128 3x3 convolutional kernels to capture the collaborative changes of multiple sensors. The convolutional output is subjected to global average pooling to obtain a spatial feature vector of 4x128. Then the spatial feature vector is input into a 2-layer LSTM network with 64 hidden units in each layer to capture the trend of temporal changes in posture. Finally, a 128-dimensional real-time posture feature vector is output through a fully connected layer and a dropout layer. Under normal posture, the dimension value of the feature vector is in a lower interval, and when the imbalance forward leaning occurs, the dimension value of multiple dimensions significantly increases.

[0064] The server calls a pre-set dangerous posture feature library which stores feature vectors of 10 typical dangerous postures, each containing 1000 verification vectors from accident cases and simulation experiments. The cosine similarity algorithm is used to calculate the similarity between the real-time feature vector and the vectors in the library. The maximum similarity of each dangerous posture is taken as the matching degree. The pre-set dangerous threshold is determined to be 0.75 through ROC curve analysis, ensuring that the recall rate is greater than or equal to 95% and the false positive rate is less than or equal to 5%. When Zhang causes imbalance forward leaning due to stepping on the void, the maximum similarity between the real-time feature vector and the “imbalance forward leaning” vector reaches 0.88, which exceeds the pre-set threshold. The server determines that the current posture is a dangerous posture and immediately triggers a warning: sends instructions to the harness through the LoRa module, activates the harness buzzer to continuously beep at a volume of 80dB, and controls the front 3 red LED lights to flash at a frequency of 1Hz; simultaneously sends structured warning information to the ground monitoring terminal, including the worker ID, name, work location, dangerous posture type, similarity value, warning time, and real-time data screenshot. The monitoring terminal pops up a red warning window and emits a prompt sound. The safety officer confirms Zhang’s status through the intercom and records the event to the MySQL database.

[0065] When Zhang adjusts to a safe posture, the server continuously monitors his posture data and automatically stops the sound and light alarm when the matching degree of all dangerous posture classes is less than 0.6, and sends a “posture recovery safe” prompt to the monitoring terminal. The technical personnel review the warning log every month. If the false positive rate of a certain dangerous posture is high, they will supplement simulation experiment data to retrain the model and update the pre-set dangerous posture feature library to continuously optimize the warning accuracy. The entire process takes the server as the core execution subject, realizes real-time monitoring and warning of dangerous postures in high-altitude work, and effectively reduces the risk of falling.

[0066] In the embodiment of the application, the posture signal coding model is trained in the following manner, which can be implemented through the following examples.

[0067] obtaining a plurality of attitude sensor data instances and a safety state identifier of each attitude sensor data instance in the plurality of attitude sensor data instances, the plurality of attitude sensor data instances including two attitude sensor data instances with close attitude amplitudes;

[0068] obtaining an attitude feature vector of each attitude sensor data instance in the plurality of attitude sensor data instances according to the multi-dimensional attitude feature extraction network;

[0069] determining a first state feature of each attitude sensor data instance in the plurality of attitude sensor data instances based on the attitude feature vector of each attitude sensor data instance, the first state feature of the attitude sensor data corresponding to a safety state level of the attitude sensor data in the calculation result;

[0070] determining a second state feature of each attitude sensor data instance in the plurality of attitude sensor data instances based on the safety state identifier of each attitude sensor data instance, the second state feature of the attitude sensor data corresponding to a pre-set safety state level of the attitude sensor data;

[0071] for a reference attitude sensor data instance combination, determining a first deviation degree of the reference attitude sensor data instance combination based on the attitude feature vector of at least one attitude sensor data instance in the reference attitude sensor data instance combination; the reference attitude sensor data instance combination is any attitude sensor data instance combination;

[0072] determining a pending safety state of the reference attitude sensor data instance combination, the pending safety state of the reference attitude sensor data instance combination including one attitude sensor data instance in the reference attitude sensor data instance combination and a target attitude sensor data instance combination, the target attitude sensor data instance combination being an attitude sensor data instance combination in the plurality of attitude sensor data combinations except the reference attitude sensor data instance combination and having an attitude difference degree value within a set threshold range;

[0073] determining a second deviation degree of the reference attitude sensor data instance combination based on the attitude feature vector of each attitude sensor data instance in the pending safety state of the reference attitude sensor data instance combination;

[0074] updating network parameters of the multi-dimensional attitude feature extraction network based on the first state feature, the second state feature of each attitude sensor data instance in the plurality of attitude sensor data instances, and the first deviation degree, the second deviation degree of the plurality of attitude sensor data instance combinations, to obtain an attitude signal coding model.

[0075] In the embodiment of the present application, the server first retrieves training data from the company aerial work posture database, which contains posture sensing data instances collected in simulated laboratory and real external wall cleaning projects, and pairs the data into multiple posture sensing data instance combinations according to the "posture amplitude proximity" principle, each combination containing two instances with a posture angle difference of no more than 3 degrees, such as "normal working posture with a body-wall angle of 30 degrees" and "slightly unbalanced posture with an angle of 32 degrees", and retrieves the safety state labels annotated by 3 senior safety experts, which are classified into "safe", "low risk" and "high risk", and are labeled in each instance. Then, the server loads the initially constructed CNN-LSTM multi-dimensional posture feature extraction network, inputs the two instances in each combination into the network in turn, extracts multi-sensor spatial correlation features through 3 convolution layers, captures time series trends through 2 LSTM layers, and outputs 128-dimensional posture feature vectors, such as the value of the "trunk inclination correlation dimension" in the feature vector of the normal working instance is 0.3, and the value of the dimension in the slightly unbalanced instance is 0.5. After that, the server performs state estimation based on the extracted feature vectors, divides all feature vectors into 3 clusters corresponding to safety levels through the built-in K-means clustering algorithm, and obtains the first state feature of each instance, such as clustering the feature vector of the above slightly unbalanced instance to the "low risk" cluster, and the first state feature is 0.2 (corresponding to the preset low risk state value mapping); at the same time, the server retrieves the preset state feature mapping table according to the safety state label annotated by the expert, maps the "high risk" label to the value of 0.8, and takes it as the second state feature of the instance. Subsequently, the server randomly selects a combination as a reference posture sensing data instance combination, such as a combination containing "30-degree normal work" and "32-degree slight imbalance", extracts the feature vectors of the two instances in the combination, calculates the cosine distance between the vectors to obtain the first deviation degree, such as a distance value of 0.15, reflecting the feature difference of the instances in the combination. Then, the server traverses other combinations in the database except the reference combination, calculates the posture angle difference between each instance in the reference combination and the instances in other combinations, and selects target combinations with an angle difference within a set threshold of 2 degrees, such as a combination containing "31-degree normal work" and "33-degree imbalance", and adds the "33-degree imbalance" instance in the target combination to the reference combination to form a pending safety state containing three instances. The server extracts the feature vectors of the three instances in the pending safety state, calculates the average cosine distance between the vectors to obtain the second deviation degree, such as an average value of 0.22, reflecting the feature difference of the expanded combination.Finally, the server calculates the mean square error of each instance first state feature and second state feature to obtain the attitude feature deviation degree, and simultaneously calculates the absolute difference between the first deviation degree and the second deviation degree to obtain the attitude vector deviation degree, sums the two deviation degrees according to a weight of 6:4 as a loss function, and updates the convolution kernel weight, the LSTM hidden layer parameter and the full connection layer bias of the CNN-LSTM network through the Adam optimizer reverse propagation, after 50 rounds of repeated iteration training, the loss function converges to below 0.03, the server saves the final network parameter, and obtains the trained attitude signal coding model.

[0076] In the embodiment of the present application, the safety state identifier of the posture sensing data instance is one or more, and the second state feature of each posture sensing data instance in the plurality of posture sensing data instance combinations is determined based on the safety state identifier of each posture sensing data instance in the plurality of posture sensing data instance combinations. The implementation can be performed by the following examples.

[0077] For any one safety state identifier of each posture sensing data instance in the plurality of posture sensing data instance combinations, a preset safety state level corresponding to the any one safety state identifier is determined from a plurality of preset safety state levels, and the posture feature of the preset safety state level corresponding to the any one safety state identifier is taken as the posture feature of the any one safety state identifier.

[0078] For any one posture sensing data instance in the plurality of posture sensing data instance combinations, the second state feature of the any one posture sensing data instance is determined based on the posture features of one or more safety state identifiers of the any one posture sensing data instance.

[0079] In the embodiment of the present application, for example, when determining the second state feature of the posture sensing data instance, the server first calls the pre-constructed preset safety state level system, which includes three preset levels of "safe", "low risk" and "high risk", and each level corresponds to the posture feature benchmark value verified by clustering of 1000 typical posture instances in the early stage, i.e., 0.1 for the "safe" level, 0.2 for the "low risk" level, and 0.8 for the "high risk" level. Then, the server iterates through the safety state identifiers of each posture sensing data instance, which are independently labeled by three experts with more than 10 years of experience in high-altitude operation safety management. Some boundary posture instances have multiple identifiers due to slight differences in expert judgment. For example, for the posture instance of "body and wall angle of 32 degrees, one foot suspended 10 cm and lasting for 2 seconds" collected in the simulation laboratory, expert A labels it as "low risk", expert B labels it as "slight imbalance (low risk)", and expert C labels it as "safe-low risk transition". The server performs level matching for each identifier of this instance: for the "low risk" identifier, it is directly mapped to the preset "low risk" level, and the posture feature benchmark value 0.2 is extracted; for the "slight imbalance (low risk)" identifier, the core risk level in the label text is analyzed as "low risk", and the benchmark value 0.2 is also matched and extracted; for the "safe-low risk transition" identifier, according to the preset boundary identifier judgment rule (the level with more obvious risk tendency is used for transition type identifier), it is determined that the core corresponds to the preset "low risk" level, and the benchmark value 0.2 is extracted. Then, the server aggregates the posture features corresponding to the three identifiers of this instance, as the values are the same, 0.2 is directly determined as the second state feature of the instance. For another example, for the posture instance of "body and wall angle of 45 degrees, both hands are separated from the safety rope and last for 1 second" collected in the real project, 3 experts all label it as "high risk", the server directly matches the preset "high risk" level, and extracts the benchmark value 0.8 as the second state feature. For the instance with only one identifier (such as the normal posture instance of "angle of 28 degrees, both feet are stably stepped on the work platform", only labeled as "safe"), the server directly determines the benchmark value 0.1 of the preset "safe" level as the second state feature of the instance. Through the above identifier-level matching and multi-identifier aggregation logic, the server ensures that the second state feature of each posture sensing data instance is highly consistent with the safety state labeled by experts, and provides accurate supervision basis for parameter updating of the subsequent multi-dimensional posture feature extraction network.

[0080] In the embodiment of the present application, before the preset safety state level corresponding to the any one safety state identifier is taken as the posture feature of the any one safety state identifier, the following implementation is provided.

[0081] According to the safety state identifier of each attitude sensing data in the plurality of attitude sensing data combinations, a state correlation mapping tensor is determined, wherein any data in the state correlation mapping tensor represents a cooperative intensity value of the any data in the first state dimension corresponding to a preset safety state level and the any data in the second state dimension corresponding to a preset safety state level being observed at the same time, and one safety state identifier is one preset safety state level;

[0082] Based on the attitude disturbance tensor template and the state correlation mapping tensor, an attitude feature of each preset safety state level is determined, wherein the dimension of the attitude disturbance tensor template is greater than the number of preset safety state levels.

[0083] In this embodiment of the invention, for example, the server first retrieves attitude sensing data instances from the company's high-altitude operation attitude database and their safety status identifiers annotated by senior safety experts. Some boundary attitude instances have multiple levels of annotation, such as "safe-low-risk transition" and "low-risk-high-risk transition," corresponding to three preset safety status levels: "safe," "low-risk," and "high-risk." The server constructs a 3×3 initial state association tensor using these three levels as the first and second state dimensions, and iterates through the annotation information of all instances: 60,000 instances are annotated only as "safe," 20,000 instances are annotated only as "low-risk," 10,000 instances are annotated only as "high-risk," 5,000 instances are annotated as "safe-low-risk," 4,000 instances are annotated as "low-risk-high-risk," and 1,000 instances are annotated as "safe-high-risk." The server calculates the effective association value of each state dimension intersection unit as "simultaneous annotation count / total number of instances", obtaining the state association mapping tensor: the collaboration strength between the first dimension "safe" and the second dimension "safe" is 0.6 (60,000 / 100,000), "safe-low risk" is 0.05 (5,000 / 100,000), and "safe-high risk" is 0.01 (1,000 / 100,000); the collaboration strength between the first dimension "low risk" and the second dimension "low risk" is 0.2, "low risk-safe" is 0.05, and "low risk-high risk" is 0.04; the collaboration strength between the first dimension "high risk" and the second dimension "high risk" is 0.1, "high risk-low risk" is 0.04, and "high risk-safe" is 0.01. Subsequently, the server loads a pre-built 5-dimensional attitude perturbation tensor template (dimensions greater than 3 preset levels), the template containing 5 128-dimensional unassigned data slices. The server randomly selects a "safe" level from the undetermined pose feature levels, and chooses the first unassigned slice from the template as its initial pose feature. Next, it queries the maximum cooperative strength value of 0.05 (corresponding to the "low-risk" level) in the "safe" dimension of the state association mapping tensor, and selects the second unassigned slice with the highest cosine similarity to the "safe" feature from the template as the pose feature for the "low-risk" level. Then, it queries the maximum cooperative strength value of 0.04 (corresponding to the "high-risk" level) in the "low-risk" dimension, and selects the third unassigned slice with the highest similarity to the "low-risk" feature as the pose feature for the "high-risk" level. This completes the assignment of pose features for all preset safe state levels, ensuring that the differences between features match the cooperative strength of the association mapping tensor.

[0084] In this embodiment of the invention, the security status identifiers of the attitude sensing data instances are multiple. The step of determining the state association mapping tensor based on the security status identifier of each attitude sensing data instance in the combination of multiple attitude sensing data instances can be implemented through the following example.

[0085] Obtain an initial state-associated tensor, wherein each state unit of the first state dimension of the initial state-associated tensor corresponds to a preset security state level, and each state unit of the second state dimension of the initial state-associated tensor corresponds to a preset security state level.

[0086] For any one of the multiple attitude sensing data combinations, based on the multiple safety state identifiers of the attitude sensing data, a valid association value is written at the intersection of the state dimensions of the state unit corresponding to the first state dimension and the state unit corresponding to the second state dimension. The state unit corresponding to the first state dimension is the first feature dimension position in the initial state association tensor corresponding to the preset safety state level of one safety state identifier of the attitude sensing data. The state unit corresponding to the second state dimension is the second feature dimension position in the initial state association tensor corresponding to the preset safety state level of other safety state identifiers of the attitude sensing data besides the one safety state identifier.

[0087] The effective association values ​​at the intersection of the state dimensions of each state unit in the first state dimension and each state unit in the second state dimension in the initial state association tensor are superimposed to calculate the state association mapping tensor.

[0088] In this embodiment of the invention, for example, the server first retrieves an initial state-related tensor with preset safety status levels of "Safe (S)", "Low Risk (L)", and "High Risk (H)" from the company's high-altitude operation posture database. This tensor is a 3×3 zero matrix, where the row cells of the first state dimension correspond to S, L, and H respectively, and the column cells of the second state dimension correspond to S, L, and H respectively. All initial cell values ​​in the matrix are 0. Next, the server processes a posture instance collected from a simulation laboratory, which involves "the body forming a 32-degree angle with the wall, one foot dangling 10cm in the air for 2 seconds". This instance is labeled as "Safe", "Low Risk", and "Low Risk" by three senior safety experts, meaning that multiple safety statuses are identified as S and L. The server first selects an identifier S of the instance as the first state dimension reference, locates the row (row 1) corresponding to the first dimension S of the initial tensor, and then selects another identifier L of the instance other than S as the second state dimension reference, locates the column (column 2) corresponding to the second dimension L, and writes a valid association value 1 at the intersection of the two rows and two columns of state dimensions; then the server selects another identifier L of the instance as the first state dimension reference, locates the row (row 2) corresponding to the first dimension L, selects an identifier S other than L as the second state dimension reference, locates the column (column 1) corresponding to the second dimension S, and writes a valid association value 1 at the intersection of the two rows and two columns of state dimensions. The server iterates through all instances in the database with multiple security status identifiers using the same logic: for 5000 instances labeled S and L, each instance writes 1 to the intersection of SL (row 1, column 2) and LS (row 2, column 1); for 4000 instances labeled L and H, each instance writes 1 to the intersection of LH (row 2, column 3) and HL (row 3, column 2); for 1000 instances labeled S and H, each instance writes 1 to the intersection of SH (row 1, column 3) and HS (row 3, column 1). Finally, the server performs superposition calculations on the intersection units of each state dimension of the initial state association tensor: the superposition value of the intersection unit is 5000 for SL, 5000 for LS, 4000 for LH, 4000 for HL, 1000 for SH, and 1000 for HS. The diagonal units SS, LL, and HH correspond to the superposition values ​​of 60,000, 20,000, and 10,000 instances with only a single identifier, respectively (each instance with a single identifier is written with 1 in its own diagonal unit). Finally, a complete state association mapping tensor is obtained, and the value of each unit in the tensor accurately reflects the cooperative strength of two preset security state levels being observed simultaneously.

[0089] In this embodiment of the invention, the determination of the attitude features for each preset safety state level based on the attitude perturbation tensor template and the state association mapping tensor can be implemented through the following example.

[0090] Randomly select the first preset safety state level from the preset safety state levels where the attitude characteristics have not been determined;

[0091] A data slice at the first feature dimension position in the unassigned state is randomly selected from the attitude perturbation tensor template and used as the attitude feature of the first preset safety state level.

[0092] Based on the posture features of the first preset safety state level and the posture perturbation tensor template, determine the posture features of the preset safety state level corresponding to each effective association value in the first feature dimension position corresponding to the first preset safety state level in the state association mapping tensor in the second state dimension.

[0093] If there is a preset safety state level with undetermined posture features, then a first preset safety state level is randomly selected from the preset safety state levels with undetermined posture features to start the loop until there is no preset safety state level with undetermined posture features.

[0094] In this embodiment of the invention, for example, the server first loads a preset set of safety state levels {S (safe), L (low risk), H (high risk)} and a 5×128-dimensional attitude perturbation tensor template (5 unassigned data slices, each slice being a 128-dimensional feature vector). At this point, the attitude features for all three levels are undetermined. The server randomly selects "Safe (S)" from the set of undetermined levels as the first preset safety state level. Then, it randomly selects the second slice from the 5 unassigned slices of the attitude perturbation tensor template (the 128-dimensional values ​​include a trunk tilt correlation dimension of 0.3, a limb coordination dimension of 0.1, etc.), determines it as the attitude feature of S, and marks the slice as "assigned". Next, the server retrieves the previously constructed state correlation mapping tensor and locates the first feature dimension row (row 1) corresponding to S. The effective correlation values ​​in this row are, in order: coordination strength with L 5000 and coordination strength with H 1000. The server uses the pose feature of S as a benchmark and calculates the cosine similarity between the remaining unassigned slices (slices 1, 3, 4, and 5) in the template and the feature of S. The result shows that slice 3 has the highest similarity to the feature of S (0.82), which conforms to the logic of the highest cooperative strength between S and L in the state association mapping tensor. Therefore, slice 3 is identified as the pose feature of L and marked as "assigned". At this time, only H remains as the level of unassigned pose features. The server selects H again from the unassigned set as the new first preset security state level and retrieves the first feature dimension row (row 3) corresponding to H in the state association mapping tensor. The highest effective association value in this row is the cooperative strength of 4000 with L. The server uses the pose feature of L (slice 3) as a benchmark and calculates the cosine similarity between the remaining unassigned slices (slices 1, 4, and 5) in the template and the feature of L. It finds that slice 4 has the highest similarity to the feature of L (0.76). Therefore, slice 4 is identified as the pose feature of H and marked as "assigned". At this point, all preset safety state levels have had their pose characteristics determined, the server stops looping, and the pose feature allocation process is complete. Throughout the process, the server strictly adheres to the cooperative strength logic of the state association mapping tensor and the unassigned slices of the pose perturbation tensor template to ensure that the pose features at each level not only meet the perturbation randomness requirements but also maintain high similarity with the features at the highest association strength level, providing an accurate state feature benchmark for the subsequent training of the pose signal encoding model.

[0095] In this embodiment of the invention, the step of determining the pose feature of the preset safety state level corresponding to each valid association value in the first feature dimension position of the first preset safety state level in the state association mapping tensor according to the pose feature of the first preset safety state level and the pose perturbation tensor template can be implemented through the following example.

[0096] Based on each valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, determine the second preset security state level corresponding to the second state dimension where the maximum valid association value is not determined and the pose feature is not determined.

[0097] The data slice at the first feature dimension position that has the highest similarity to the first preset label and is in an unassigned state is determined from the attitude perturbation tensor template and is used as the attitude feature of the second preset security state level.

[0098] If, in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, there exists a preset security state level corresponding to the second state dimension with an undetermined attitude feature and a valid association value, then starting from each valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, the system determines the second preset security state level with the largest valid association value in the second state dimension and an undetermined attitude feature, and loops until there is no preset security state level corresponding to the second state dimension with an undetermined attitude feature and a valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor.

[0099] In this embodiment of the invention, for example, the server first retrieves the first feature dimension row corresponding to the first preset security state level S (safe) in the state association mapping tensor. The effective association values ​​in this row are, in order, a coordination strength of 5000 with L (low risk) and a coordination strength of 1000 with H (high risk). The server sorts these effective association values ​​and determines that the maximum effective association value is 5000, which corresponds to the preset security state level L in the second state dimension. Since L has not yet determined its attitude features, L is marked as the second preset security state level. Next, the server loads a 5×128-dimensional attitude perturbation tensor template. In this template, the attitude features of S have occupied the second slice, and the remaining unallocated slices are the first, third, fourth, and fifth slices. The server uses the attitude features of S (the second slice) as a reference to calculate the cosine similarity between the remaining unallocated slices and the features of S. The result shows that the third slice has the highest similarity to the features of S (0.82), which conforms to the logic of the highest coordination strength between S and L in the state association mapping tensor. Therefore, the third slice is determined as the attitude features of L and marked as "allocated". At this point, the server re-checks the valid association values ​​within the first feature dimension row corresponding to S. It finds that H, corresponding to a cooperation strength of 1000 with H, still has no determined pose feature, and thus enters the loop process. The server re-determines the current maximum valid association value as 1000 based on the valid association values ​​within the first feature dimension row corresponding to S, with its corresponding second state dimension level being H, and H still having no determined pose feature. The server continues to use S's pose feature as a benchmark, calculating the cosine similarity between the remaining unassigned slices (slices 1, 4, and 5) in the template and S's feature. The result shows that slice 4 has the second highest similarity to S's feature (0.65), therefore slice 4 is identified as H's pose feature and marked as "assigned." Finally, the server re-checks the valid association values ​​within the first feature dimension row corresponding to S, confirming that all corresponding levels (L and H) have determined pose features, and thus stops the loop process. Throughout the process, the server strictly follows the priority of the collaborative strength of the state association mapping tensor, and matches the unassigned slices in the pose perturbation tensor template with cosine similarity to ensure that the pose features of each preset safe state level maintain a high similarity with the level with the highest association strength, while also meeting the perturbation randomness requirements of the template.

[0100] In this embodiment of the invention, the determination of the first deviation of the reference attitude sensing data set based on the attitude feature vector of at least one attitude sensing data set in the reference attitude sensing data set can be performed through the following example.

[0101] The attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination is binarized to obtain the binary attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination.

[0102] The first deviation is determined based on the attitude feature vector and binary attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination.

[0103] In this embodiment of the invention, for example, the server first retrieves a paired set of reference posture sensing data instances. This set includes "normal working posture with a 30-degree angle between the body and the wall (Instance A)" and "slightly unbalanced posture with a 32-degree angle (Instance B)" collected in a simulated laboratory. The angle difference between the two postures is 2 degrees, which conforms to the combination rule of similar posture amplitudes. The server loads the 128-dimensional posture feature vectors of the two instances in this set: In the feature vector of Instance A, the "trunk tilt correlation dimension" is 0.3, the "limb coordination dimension" is 0.2, and the "safety rope tension correlation dimension" is 0.1, with an overall average dimension of 0.28; In the feature vector of Instance B, the "trunk tilt correlation dimension" is 0.5, the "limb coordination dimension" is 0.4, and the "safety rope tension correlation dimension" is 0.2, with an overall average dimension of 0.35.

[0104] Next, the server binarizes the feature vectors of the two instances respectively: using the mean dimension of instance A (0.28) as the threshold, each dimension value is compared with the threshold, and dimensions greater than the threshold are set to 1, and those less than the threshold are set to 0, thus obtaining the binary posture feature vector of instance A, where "trunk tilt correlation dimension" is 1, "limb coordination dimension" is 0, and "safety rope tension correlation dimension" is 0; using the mean dimension of instance B (0.35) as the threshold, the binary vector of instance B is obtained in the same way, where "trunk tilt correlation dimension" is 1, "limb coordination dimension" is 1, and "safety rope tension correlation dimension" is 0.

[0105] Subsequently, the server determines the first deviation based on the original feature vector and the binary vector: first, the cosine distance between the original vector and the binary vector of each instance is calculated. The cosine distance between the original vector and the binary vector of instance A is 0.12 (reflecting the degree of difference between the original feature and the binary feature), and the cosine distance between the original vector and the binary vector of instance B is 0.15. Then, the two distance values ​​are arithmetically averaged to obtain the first deviation of the reference combination, which is 0.135. This value accurately reflects the overall degree of difference between the feature vectors of the two pose instances within the combination after binarization, providing a deviation benchmark for the parameter update of the subsequent multidimensional pose feature extraction network.

[0106] In this embodiment of the invention, the determination of the pending safety state of the reference attitude sensing data combination can be performed through the following examples.

[0107] Determine a first attitude difference value between an attitude sensor data instance in the reference attitude sensor data instance combination and an attitude sensor data instance in each other attitude sensor data instance combination, wherein each other attitude sensor data instance combination is each attitude sensor data instance combination other than the reference attitude sensor data instance combination in the plurality of attitude sensor data instance combinations;

[0108] From each of the other attitude sensing data combinations, a target attitude sensing data combination is determined whose first attitude difference value is less than the first reference attitude difference value.

[0109] Based on one attitude sensor data instance from the target attitude sensor data instance combination and the reference attitude sensor data instance combination, the undetermined safety state of the reference attitude sensor data instance combination is determined.

[0110] In an embodiment of the invention, for example, the server first retrieves combination X, which has been marked as a reference posture sensing data instance. This combination consists of "normal working posture with a body angle of 30 degrees to the wall (Instance A)" and "slightly unbalanced posture with an angle of 32 degrees (Instance B)" collected in a simulated laboratory. The posture amplitudes of the two are similar (angle difference of 2 degrees). Next, the server selects Instance B (32 degrees of slight imbalance) from combination X, which has a more risky posture, as the baseline instance. It then traverses all posture sensing data instance combinations in the database except for combination X, including combination Y (Instance C: 31 degrees normal, Instance D: 33 degrees unbalanced), combination Z (Instance E: 29 degrees normal, Instance F: 34 degrees high risk), combination M (Instance G: 28 degrees normal, Instance H: 35 degrees high risk), etc.

[0111] For each other combination, the server selects the instance in that combination whose attitude angle is closest to that of instance B as a comparison instance: in combination Y, the angle difference between instance D (33 degrees) and instance B is 1 degree; in combination Z, the angle difference between instance F (34 degrees) and instance B is 2 degrees; and in combination M, the angle difference between instance H (35 degrees) and instance B is 3 degrees. Subsequently, the server retrieves a preset first reference attitude difference value of 3 degrees and filters out the combinations corresponding to the comparison instances with an angle difference of less than 3 degrees—combination Y (angle difference of 1 degree) and combination Z (angle difference of 2 degrees). These two combinations are marked as target attitude sensing data instance combinations.

[0112] Finally, the server extracts instance D from target combination Y and instance F from target combination Z, and integrates these two instances with instances A and B from reference combination X to form a pending safety state of reference combination X containing four instances. This state covers a continuous attitude change range from 30 degrees for normal operation to 34 degrees for high-risk imbalance, preserving subtle attitude differences within the reference combination while expanding the risk gradient through the target combinations, providing more comprehensive attitude sample support for subsequent deviation calculations.

[0113] In this embodiment of the invention, the step of determining the target attitude sensing data combination from each of the other attitude sensing data combinations where the first attitude difference value is less than the first reference attitude difference value can be performed through the following example.

[0114] From a preset set of attitude sensing data sets, a target attitude sensing data set with a first attitude difference value less than a first reference attitude difference value is determined. The preset attitude sensing data set is another attitude sensing data set whose first attitude difference value is not less than a second reference attitude difference value among each other attitude sensing data set, where the first reference attitude difference value is greater than the second reference attitude difference value.

[0115] In this embodiment of the invention, for example, the server first retrieves a preset attitude difference threshold parameter, wherein the second reference attitude difference degree value is 1 degree and the first reference attitude difference degree value is 3 degrees (the first reference value is greater than the second reference value). Next, the server uses instance B with a slight imbalance of 32 degrees in reference combination X as a benchmark, and iterates through all other attitude sensing data instance combinations in the database except for combination X: combination Y contains instance D with a 33-degree imbalance (angle difference of 1 degree from instance B), combination Z contains instance F with a 34-degree imbalance (angle difference of 2 degrees), combination M contains instance H with a 35-degree imbalance (angle difference of 3 degrees), and combination N contains instance I with a 31.5-degree normal operation (angle difference of 0.5 degrees). The server first filters preset attitude sensing data instance combinations where "the first attitude difference degree value is not less than the second reference value": it calculates the angle difference between each other combination and instance B; the angle differences of combinations Y (1 degree), Z (2 degrees), and M (3 degrees) are all ≥1 degree and are included in the preset combinations; the angle difference of combination N is 0.5 degrees <1 degree and is excluded. Subsequently, the server further filters target combinations from the preset combinations that have a "first attitude difference value less than the first reference value": combination Y's 1 degree and combination Z's 2 degrees are both less than 3 degrees, meeting the requirements; combination M's 3 degrees are not less than the first reference value of 3 degrees, and are therefore excluded. Finally, the server marks combination Y and combination Z as target attitude sensing data instance combinations, completing the filtering process. Throughout the process, the server uses dual threshold filtering to both exclude redundant combinations with too small a difference from the baseline instance and ensure that the difference between the target combination and the baseline instance is within a reasonable range, providing an accurate sample range for the subsequent construction of the pending safety state.

[0116] In this embodiment of the invention, the determination of the second deviation of the reference attitude sensing data set based on the attitude feature vector of each attitude sensing data set in the undetermined safe state of the reference attitude sensing data set can be performed through the following example.

[0117] For the undetermined safety state of the reference attitude sensor data set, based on the attitude feature vectors of one attitude sensor data set and another attitude sensor data set in the reference attitude sensor data set, a second attitude difference value is determined between the two attitude sensor data sets. Based on the attitude feature vectors of one attitude sensor data set and another attitude sensor data set in the reference attitude sensor data set, a third attitude difference value is determined between the one attitude sensor data set in the reference attitude sensor data set and another attitude sensor data set in the target attitude sensor data set.

[0118] Based on the second attitude difference degree value and the third attitude difference degree value, a second deviation degree of the reference attitude sensing data combination is determined.

[0119] In an embodiment of the present invention, for example, the server first retrieves the pending safety state instances of the reference combination X. These states include a 30-degree normal operation instance A and a 32-degree slight imbalance instance B within the reference combination, as well as a 33-degree imbalance instance D of the target combination Y and a 34-degree high-risk instance F of the target combination Z. The 128-dimensional pose feature vectors of the four instances have been extracted by the previously trained CNN-LSTM network: the feature vector of instance A is [0.3, 0.1, 0.2, ...] (torso tilt correlation dimension 0.3), that of instance B is [0.5, 0.2, 0.3, ...] (torso tilt dimension 0.5), that of instance D is [0.6, 0.3, 0.4, ...] (torso tilt dimension 0.6), and that of instance F is [0.8, 0.5, 0.6, ...] (torso tilt dimension 0.8). Next, the server calculates the second attitude difference value within the reference combination: using cosine distance as a metric, the cosine of the angle between the feature vectors of instance A and instance B is calculated, yielding a difference value of 0.12, which reflects subtle attitude feature differences within the reference combination. Subsequently, the server calculates the third attitude difference value between the reference combination instances and the target combination instances: selecting instance B, which has a more significant risk tendency in the reference combination, as the benchmark, the cosine distance between its feature vectors and those of instances D and F is calculated, yielding a difference value of 0.08 between instance B and D (reflecting the feature change from slight imbalance to imbalance), and a difference value of 0.15 between instance B and F (reflecting the feature change from slight imbalance to high risk). The arithmetic mean of the two third difference values ​​is then taken, resulting in an average third difference value of 0.115. Finally, the server calculates the second deviation based on the preset weight allocation rules (30% for differences within the reference combination and 70% for differences between the reference and target combinations): multiplying the second difference value of 0.12 by the weight of 0.3 yields 0.036; multiplying the average third difference value of 0.115 by the weight of 0.7 yields 0.0805; adding the two together gives the second deviation of 0.1165. This value accurately reflects the overall difference in attitude features under the undetermined safety state, providing a key basis for subsequent parameter updates of the multidimensional attitude feature extraction network.

[0120] In this embodiment of the invention, the network parameters of the multidimensional attitude feature extraction network are updated based on the first state feature and second state feature of each attitude sensor data instance in the plurality of attitude sensor data instances and the first deviation and second deviation of the plurality of attitude sensor data instances to obtain the attitude signal coding model. This can be implemented through the following example.

[0121] Based on the first and second state features of each attitude sensor data instance in the combination of multiple attitude sensor data instances, the attitude feature deviation is determined.

[0122] Based on the first and second deviations of the combination of the multiple attitude sensing data instances, the attitude feature vector deviation is determined;

[0123] Based on the attitude feature deviation and the attitude feature vector deviation, the network parameters of the multidimensional attitude feature extraction network are updated to obtain the attitude signal coding model.

[0124] In an embodiment of the present invention, for example, the server first traverses each instance in all attitude sensing data instance combinations, extracts the first state feature of each instance calculated by the multi-dimensional attitude feature extraction network, and the second state feature mapped by the safety expert annotation. By calculating the mean square error of the first state feature and the second state feature of each instance, the mean square error of all instances is globally aggregated to obtain the overall attitude feature deviation. This deviation reflects the overall fit between the model-calculated state and the expert-annotated state.

[0125] Next, the server iterates through all attitude sensing data instance combinations, extracts the first deviation (the degree of feature difference between instances within the combination) and the second deviation (the degree of feature difference after the pending safety state is extended) obtained from feature vector analysis for each combination, and performs weighted fusion of the two deviations for each group according to the preset weight allocation rules. Then, the fusion results of all combinations are globally aggregated to obtain the attitude feature vector deviation, which reflects the feature discrimination between different attitude combinations.

[0126] Subsequently, the server constructs a loss function that integrates dual deviations, weights and sums the attitude feature deviation and attitude feature vector deviation according to preset weights, loads the initial multidimensional attitude feature extraction network, and uses an adaptive optimizer to backpropagate the loss function to each layer of the network: iteratively adjusts the convolution kernel weights of the convolutional layers to optimize the extraction accuracy of multi-sensor spatial features, corrects the bias parameters of the hidden units in the LSTM layer to improve the ability to capture temporal features, and updates the weight matrix of the fully connected layers to enhance the expression effect of the feature vectors, until the loss function converges to below the preset threshold, ensuring that the model's estimation of attitude states and differentiation of feature differences both meet the expected requirements.

[0127] Finally, the server saves the trained network parameters to obtain an attitude signal encoding model that can be used for real-time identification of dangerous postures in high-altitude operations. This model can accurately extract key features from attitude sensing data, providing reliable feature support for subsequent dangerous posture early warning.

[0128] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned high-altitude work safety harness dangerous posture warning method based on AI posture analysis. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0129] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for early warning of dangerous postures of a safety harness for high-altitude operations based on AI posture analysis, characterized in that, include: Real-time attitude sensing data of the safety harness for high-altitude operations is collected. The real-time attitude sensing data includes acceleration data, angular velocity data, and attitude angle data collected by multi-axis sensors deployed on the safety harness. The real-time attitude sensing data is preprocessed to obtain standardized attitude sensing data. The preprocessing includes data denoising, time synchronization and standardization. The standardized attitude sensing data is input into a pre-trained attitude signal encoding model to obtain the real-time attitude feature vector corresponding to the real-time attitude sensing data. Based on a preset dangerous posture feature library, the similarity between the real-time posture feature vector and each dangerous posture feature vector in the preset dangerous posture feature library is calculated. If the similarity is greater than a preset danger threshold, the current posture is determined to be a dangerous posture, and a safety warning is triggered. The safety warning includes issuing an audible and visual alarm to the high-altitude workers and sending a warning message to the monitoring terminal.

2. The method according to claim 1, characterized in that, The attitude signal coding model is trained in the following ways: Acquire multiple attitude sensor data sets and a safety status identifier for each attitude sensor data set in the multiple attitude sensor data sets, wherein the attitude sensor data set includes two attitude sensor data sets with similar attitude amplitudes; The attitude feature vector of each attitude sensing data instance in the combination of multiple attitude sensing data instances is obtained by the multi-dimensional attitude feature extraction network. Based on the attitude feature vector of each attitude sensor data instance in the combination of multiple attitude sensor data instances, the first state feature of each attitude sensor data instance in the combination of multiple attitude sensor data instances is determined, and the first state feature of the attitude sensor data instance corresponds to the safety state level of the attitude sensor data instance in the calculation result. Based on the safety status identifier of each attitude sensor data instance in the multiple attitude sensor data instance combination, a second state feature of each attitude sensor data instance in the multiple attitude sensor data instance combination is determined, and the second state feature of the attitude sensor data instance corresponds to the pre-set safety status level of the attitude sensor data instance. For a reference attitude sensing data set, a first deviation of the reference attitude sensing data set is determined based on the attitude feature vector of at least one attitude sensing data set in the reference attitude sensing data set; the reference attitude sensing data set can be any attitude sensing data set. The undetermined safe state of the reference attitude sensor data combination is determined. The undetermined safe state of the reference attitude sensor data combination includes one attitude sensor data combination from the reference attitude sensor data combination and the target attitude sensor data combination. The target attitude sensor data combination is an attitude sensor data combination other than the reference attitude sensor data combination and whose attitude difference value is within a set threshold range from the plurality of attitude sensor data combinations. Based on the attitude feature vector of each attitude sensor data instance in the undetermined safe state of the reference attitude sensor data instance combination, the second deviation degree of the reference attitude sensor data instance combination is determined. Based on the first state feature and second state feature of each attitude sensor data instance in the multiple attitude sensor data instance combination, and the first deviation and second deviation of the multiple attitude sensor data instance combination, the network parameters of the multidimensional attitude feature extraction network are updated to obtain the attitude signal coding model.

3. The method according to claim 2, characterized in that, The security status identifier of the attitude sensing data instance is one or more. The step of determining the second state feature of each attitude sensing data instance in the combination of multiple attitude sensing data instances based on the security status identifier of each instance includes: For any security status identifier of each attitude sensing data instance in the combination of multiple attitude sensing data instances, determine the preset security status level corresponding to the any security status identifier from multiple preset security status levels, and take the attitude feature of the preset security status level corresponding to the any security status identifier as the attitude feature of the any security status identifier. For any one of the multiple attitude sensing data combinations, a second state feature of the attitude sensing data is determined based on the attitude features of one or more safety state identifiers of the attitude sensing data.

4. The method according to claim 3, characterized in that, The attitude sensing data instance has multiple security status identifiers. Before using the attitude feature corresponding to a preset security status level of any one of the security status identifiers as the attitude feature of any one of the security status identifiers, the method further includes: Obtain an initial state-associated tensor, wherein each state unit of the first state dimension of the initial state-associated tensor corresponds to a preset security state level, and each state unit of the second state dimension of the initial state-associated tensor corresponds to a preset security state level. For any one of the multiple attitude sensing data combinations, based on the multiple safety state identifiers of the attitude sensing data, a valid association value is written at the intersection of the state dimensions of the state unit corresponding to the first state dimension and the state unit corresponding to the second state dimension. The state unit corresponding to the first state dimension is the first feature dimension position in the initial state association tensor corresponding to the preset safety state level of one safety state identifier of the attitude sensing data. The state unit corresponding to the second state dimension is the second feature dimension position in the initial state association tensor corresponding to the preset safety state level of other safety state identifiers of the attitude sensing data besides the one safety state identifier. The effective association values ​​at the intersection of the state dimensions of each state unit in the first state dimension and each state unit in the second state dimension in the initial state association tensor are superimposed to calculate the state association mapping tensor. Any data in the state association mapping tensor represents the cooperative strength value when the preset security state level corresponding to the data in the first state dimension and the preset security state level corresponding to the data in the second state dimension are observed simultaneously. A security state identifier is a preset security state level. Randomly select the first preset safety state level from the preset safety state levels where the attitude characteristics have not been determined; Randomly select a data slice at the first feature dimension position in the unassigned state from the attitude perturbation tensor template as the attitude feature of the first preset safety state level; Based on each valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, determine the second preset security state level corresponding to the second state dimension where the maximum valid association value is not determined and the pose feature is not determined. The data slice at the first feature dimension position that has the highest similarity to the first preset label and is in an unassigned state is determined from the attitude perturbation tensor template and is used as the attitude feature of the second preset security state level. If there exists an undetermined attitude feature and a valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, then starting from each valid association value in the first feature dimension position corresponding to the first preset security state level in the state association mapping tensor, the second preset security state level with the largest valid association value in the second state dimension and an undetermined attitude feature is determined and looped until there is no undetermined attitude feature and a valid association value in the second state dimension corresponding to the first preset security state level in the state association mapping tensor. If there is a preset safety state level with undetermined attitude features, then a first preset safety state level is randomly selected from the preset safety state levels with undetermined attitude features to start the loop until there are no preset safety state levels with undetermined attitude features. The dimension of the attitude perturbation tensor template is greater than the number of preset safety state levels.

5. The method according to claim 2, characterized in that, The determination of the first deviation of the reference attitude sensing data set based on the attitude feature vector of at least one attitude sensing data set in the reference attitude sensing data set includes: The attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination is binarized to obtain the binary attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination. The first deviation is determined based on the attitude feature vector and binary attitude feature vector of at least one attitude sensor data in the reference attitude sensor data combination.

6. The method according to claim 2, characterized in that, Determining the undetermined safety state of the reference attitude sensing data combination includes: Determine a first attitude difference value between an attitude sensor data instance in the reference attitude sensor data instance combination and an attitude sensor data instance in each other attitude sensor data instance combination, wherein each other attitude sensor data instance combination is each attitude sensor data instance combination other than the reference attitude sensor data instance combination in the plurality of attitude sensor data instance combinations; From each of the other attitude sensing data combinations, a target attitude sensing data combination is determined whose first attitude difference value is less than the first reference attitude difference value. Based on one attitude sensor data instance from the target attitude sensor data instance combination and the reference attitude sensor data instance combination, the undetermined safety state of the reference attitude sensor data instance combination is determined.

7. The method according to claim 6, characterized in that, Determining the target attitude sensing data combination from each of the other attitude sensing data combinations where the first attitude difference value is less than the first reference attitude difference value includes: From a preset set of attitude sensing data sets, a target attitude sensing data set with a first attitude difference value less than a first reference attitude difference value is determined. The preset attitude sensing data set is another attitude sensing data set whose first attitude difference value is not less than a second reference attitude difference value among each other attitude sensing data set, where the first reference attitude difference value is greater than the second reference attitude difference value.

8. The method according to claim 2, characterized in that, The determination of the second deviation of the reference attitude sensor data set based on the attitude feature vector of each attitude sensor data set in the undetermined safe state of the reference attitude sensor data set includes: For the undetermined safety state of the reference attitude sensor data set, based on the attitude feature vectors of one attitude sensor data set and another attitude sensor data set in the reference attitude sensor data set, a second attitude difference value is determined between the two attitude sensor data sets. Based on the attitude feature vectors of one attitude sensor data set and another attitude sensor data set in the reference attitude sensor data set, a third attitude difference value is determined between the one attitude sensor data set in the reference attitude sensor data set and another attitude sensor data set in the target attitude sensor data set. Based on the second attitude difference degree value and the third attitude difference degree value, a second deviation degree of the reference attitude sensing data combination is determined.

9. The method according to claim 2, characterized in that, The multidimensional attitude feature extraction network is updated based on the first state feature and second state feature of each attitude sensor data instance in the combination of multiple attitude sensor data instances, and the first deviation and second deviation of the combination of multiple attitude sensor data instances, to obtain an attitude signal coding model, including: Based on the first and second state features of each attitude sensor data instance in the combination of multiple attitude sensor data instances, the attitude feature deviation is determined. Based on the first and second deviations of the combination of the multiple attitude sensing data instances, the attitude feature vector deviation is determined; Based on the attitude feature deviation and the attitude feature vector deviation, the network parameters of the multidimensional attitude feature extraction network are updated to obtain the attitude signal coding model.

10. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-9.

Images