A water conservancy construction hazard source identification system and method based on a brain-like spiking neural network

Abstract

The application relates to the technical field of water conservancy construction, and discloses a water conservancy construction hazard source identification system and method based on a brain-like pulse neural network, which comprises an information acquisition terminal and an edge processing device. The information acquisition terminal is used for collecting multi-source time sequence information of a water conservancy construction area, performing time correlation and pulsing treatment on the multi-source time sequence information, obtaining multi-channel pulse time sequence data, and sending the multi-channel pulse time sequence data to the edge processing device. The edge processing device comprises a pretreatment module, an anomaly identification module, a brain-like time sequence identification module and a risk decision module. The brain-like time sequence identification module comprises an input coding layer, a time sequence reasoning layer and an output identification layer. The risk decision module is used for determining a risk level according to the hazard source identification result and generating early warning control information based on the risk level.

Description

Technical Field

[0001] This invention relates to the field of water conservancy construction technology, and in particular to a water conservancy construction hazard identification system and method based on a brain-like spiking neural network. Background Technology

[0002] During the construction of reservoirs, dikes, rivers, cofferdams, dams, and other water conservancy projects, the work area is usually adjacent to water bodies, slopes, or temporary work boundaries, and is accompanied by the cross-operation of construction machinery. To reduce construction risks, it is usually necessary to monitor the wearing status of protective equipment by workers, their activities near water or edges, and the relative relationship between workers and construction equipment, so as to promptly identify potential hazards that may cause falls, equipment collisions, or violations of regulations.

[0003] Currently, safety monitoring at water conservancy construction sites mainly relies on manual inspections, fixed video surveillance, or target recognition based on single-frame images. The monitoring range and continuity of manual inspections are limited by the inspection cycle and personnel allocation. Fixed video surveillance is easily affected by the monitoring angle, obstruction, and changes in the work area. Recognition methods based on single-frame images usually focus on the spatial characteristics at a certain moment, making it difficult to continuously reflect dangerous states with time-evolution characteristics, such as workers gradually approaching the edge area, continuous imbalance in posture, continuous lack of protective equipment, or the movement trajectory of construction equipment gradually approaching the workers. Moreover, water conservancy construction sites are prone to water mist, heavy rain, and communication fluctuations, which may cause short-term changes in image information, posture information, or environmental perception information. Furthermore, existing recognition methods, lacking multi-source temporal correlation and working condition differentiation, are prone to identifying inherent environmental fluctuations as real dangerous anomalies. On the other hand, using a uniform threshold or uniform noise reduction intensity weakens the risk characteristics of water and edge proximity, leading to false alarms or missed alarms. Summary of the Invention

[0004] In order to overcome at least one of the defects mentioned above in the prior art, the present invention provides a hydraulic construction hazard source identification system and method based on a brain-like spiking neural network, so as to solve the problems of difficulty in continuously identifying dynamic hidden dangers under complex hydraulic construction conditions, and difficulty in distinguishing between environmental disturbances and real anomalies.

[0005] Firstly, this application provides a hydraulic construction hazard identification system based on a brain-like spiking neural network, comprising an information acquisition terminal and an edge processing device, wherein... The information acquisition terminal is used to collect multi-source time-series information of the water conservancy construction area, perform time correlation and pulse processing on the multi-source time-series information to obtain multi-channel pulse time-series data, and send the multi-channel pulse time-series data to the edge processing device; The edge processing device includes a preprocessing module, an anomaly identification module, a brain-like time sequence recognition module, and a risk decision module. The preprocessing module is used to process the multi-channel pulse time sequence data to obtain the time sequence data to be identified. The anomaly identification module is used to acquire reference time series data, determine the working condition weight based on the acquired working condition information, and determine the credibility weight based on the acquired data quality information. The working condition weight and the credibility weight are used to fuse the time series deviation of the time series data to be identified relative to the reference time series data. Based on the time series deviation, the module distinguishes between inherent environmental fluctuations and abnormal hazard sources, and determines the target time series data corresponding to the abnormal hazard source from the time series data to be identified. The brain-like temporal recognition module includes an input encoding layer, a temporal reasoning layer, and an output recognition layer. The input encoding layer is used to map the target temporal data to feature channels corresponding to different hazard source recognition tasks. The temporal reasoning layer is used to extract and fuse temporal features of each feature channel through a spiking neural network. The output recognition layer is used to determine the hazard source recognition result based on the extraction and fusion result of the temporal reasoning layer. The risk decision module is used to determine the risk level based on the hazard source identification results, and generate early warning and control information based on the risk level.

[0006] In one embodiment, the information acquisition terminal includes a visual perception module, an environmental perception module, a posture perception module, and a signal processing module, wherein, The visual perception module includes an image acquisition unit with overlapping fields of view, used to acquire visual temporal information of workers, protective equipment and work equipment, and to determine the spatial relationship and motion state between workers, protective equipment and work equipment based on the visual temporal information; The environmental sensing module includes an environmental sensor for acquiring environmental parameters, and the environmental sensing module is used to generate environmental time-series information of the water conservancy construction area based on the environmental parameters. The attitude perception module includes an inertial measurement unit, which is used to acquire the attitude angle and motion trajectory of the operator, and to form the attitude change information and motion state information of the operator based on the acquired attitude angle and motion trajectory. The visual timing information, the environmental timing information, the posture change information, and the motion state information together constitute the multi-source timing information. The signal processing module is connected to the visual perception module, the environmental perception module, and the posture perception module respectively, and is used to perform time correlation and pulse processing on the multi-source timing information to form the multi-channel pulse timing data.

[0007] In one embodiment, the signal processing module includes a noise reduction unit, a normalization unit, a timing synchronization unit, and a pulse coding unit, wherein, The noise reduction unit is used to filter out instantaneous interference information and invalid information in the multi-source timing information; The standardization unit is used to unify the data range and convert the data format of different data types in the multi-source time series information; The timing synchronization unit performs time window matching on the multi-source timing information with different sampling frequencies according to a unified time reference; The pulse coding unit is used to determine the pulse firing state based on the signal strength and signal change degree corresponding to the data channels of the operator, protective equipment and operating equipment, and to generate multi-channel pulse timing data that retains the sequential relationship of the multi-source timing information based on the pulse firing state.

[0008] In one embodiment, the anomaly detection module includes a reference data unit, a working condition weight determination unit, a confidence weight determination unit, and an anomaly comparison unit, wherein, The reference data unit is used to select reference time series data that matches the current water conservancy construction area from a preset reference time series database according to the construction scenario type, operation stage and environmental conditions. The operating condition weight determination unit is used to determine the operating condition weight based on the water level change state, water mist state, precipitation state, illumination state, and disturbance state. The credibility weight determination unit is used to determine the credibility weight based on the sensor signal-to-noise ratio, image quality, attitude stability, and communication status. The anomaly comparison unit is used to determine the weighted time series deviation corresponding to the data channels of workers, protective equipment and work equipment using the working condition weight and the confidence weight, and compares the weighted time series deviation with the anomaly identification conditions to distinguish between inherent environmental fluctuations and abnormal hazard sources.

[0009] In one embodiment, the anomaly comparison unit is further configured to determine the channel fusion weight based on the weighted timing deviation corresponding to each of the feature channels, and set the fusion priority of the corresponding feature channels according to the reciprocal relationship of each of the weighted timing deviations, so that feature channels with smaller weighted timing deviations and higher data reliability obtain higher fusion priorities.

[0010] In one embodiment, the input coding layer includes a personnel protection feature channel, a personnel position and posture feature channel, and an equipment operation feature channel, wherein, The personnel protection feature channel is used to characterize the presence, wearing status, and duration of the protective equipment. The personnel position and posture feature channel is used to characterize the spatial changes, posture changes, and movement trends of the workers. The equipment operation feature channel is used to characterize the positional changes, trajectory changes, and motion states of the operating equipment relative to the operator; The temporal inference layer is used to extract temporal features from each of the feature channels according to the fusion priority, and to perform feature fusion between the feature channels to identify personnel protection anomalies, water and edge risks, and equipment operation risks.

[0011] In one embodiment, the temporal reasoning layer includes a cluster of excitatory neurons, a cluster of inhibitory neurons, and a balance regulation unit, wherein, The excitatory neuron cluster is used to receive the temporal features of the corresponding feature channels and generate an excitatory response based on the temporal features; The inhibitory neuron cluster is connected to the excitatory neuron cluster to receive the excitatory response and to feed back an inhibitory signal to the excitatory neuron cluster through the inhibitory connection. The balance adjustment unit is connected to the excitatory neuron cluster, the inhibitory neuron cluster, and the anomaly detection module, respectively, and is used to obtain the group response state of the excitatory neuron cluster and the inhibitory neuron cluster. Based on the response deviation between the two group response states and the environmental interference judgment result output by the anomaly detection module, the unit adjusts the basic activation level of the inhibitory neuron cluster and the inhibitory effect strength of the inhibitory connection. When the anomaly detection module determines that environmental interference is enhanced, the balance adjustment unit increases the basic activation level of the inhibitory neuron cluster and enhances the inhibitory effect of the inhibitory connection on the excitatory neuron cluster, so as to suppress the transient temporal features associated with the environmental interference. When the temporal inference layer detects that the temporal features corresponding to the hazard source are continuously and gradually changing from the corresponding feature channels, the balance adjustment unit reduces the suppression trigger frequency for the corresponding feature channels in order to preserve the continuous temporal features associated with the evolution of the hazard source.

[0012] In one embodiment, the temporal inference layer further includes an adaptive learning unit connected to both the anomaly detection module and the balance adjustment unit, wherein... The adaptive learning unit is used to obtain the pulse firing times of the pre-level neuron and the post-level neuron. The pre-level neuron and the post-level neuron are connected through excitatory connections for transmitting excitatory responses or through inhibitory connections. Based on the pulse firing order and pulse time interval between the pre-level neuron and the post-level neuron, a connection strength adjustment amount is determined. The connection strength of the excitatory connection or the inhibitory connection is updated according to the connection strength adjustment amount, and the updated connection strength is limited to the corresponding preset connection strength range. The adaptive learning unit is also used to obtain the type and duration of environmental interference determined by the anomaly identification module, as well as the hazard source features and hazard source feature durations extracted by the temporal inference layer from each of the feature channels, determine a target temporal learning window based on the duration of environmental interference and the duration of hazard source features, and select pulse temporal data that participate in the connection strength update within the target temporal learning window.

[0013] In one embodiment, the output recognition layer is connected to the temporal inference layer and is used to receive the inference results obtained by the temporal inference layer after extracting and fusing temporal features of each feature channel, and to count the number of output pulses corresponding to different hazard source identification tasks within a continuously set preset observation window; The output recognition layer is further configured to determine the recognition confidence information corresponding to each hazard source recognition result based on the number of output pulses corresponding to each hazard source recognition task, the consistency of output results within adjacent preset observation windows, and the channel fusion weight of each feature channel participating in the corresponding hazard source recognition task, and send the hazard source recognition result and the recognition confidence information to the risk decision module. The risk decision module is connected to the output recognition layer and is used to compare the recognition confidence information with the risk judgment threshold respectively, and determine the risk level corresponding to different confidence intervals based on the comparison results. The risk decision module is also used to generate corresponding early warning and control information according to the risk level. The early warning and control information includes at least one of terminal early warning instructions, on-site linkage early warning instructions, management personnel notification information and platform early warning instructions. The module sends the terminal early warning instructions to the information collection terminal, the on-site linkage early warning instructions to the on-site linkage equipment, and the management personnel notification information and the platform early warning instructions to the management platform. The management platform is communicatively connected to the edge processing device and is used to receive and store the hazard identification results, the identification confidence information, the risk level and the corresponding early warning record, obtain the review information for the hazard identification results and the risk level, generate model update data for adjusting the brain-like temporal recognition module based on the review information, and send the model update data to the edge processing device. The edge processing device is also used to update at least one of the model parameters and recognition rules of the brain-like temporal recognition module according to the model update data, and to perform incremental calibration on the updated brain-like temporal recognition module using the newly added temporal data collected in the current water conservancy construction area.

[0014] Secondly, this application provides a method for identifying hazardous sources in water conservancy construction based on a neuromorphic spiking neural network, which is executed based on the aforementioned system for identifying hazardous sources in water conservancy construction based on a neuromorphic spiking neural network. The method for identifying hazardous sources in water conservancy construction includes: Multi-source time-series information from the water conservancy construction area is collected, and the multi-source time-series information is subjected to time correlation and pulse processing to obtain multi-channel pulse time-series data. The multi-channel pulse timing data is processed to obtain the timing data to be identified; Obtain reference time series data, and use operating condition weight and confidence weight to fuse the time series deviation of the time series data to be identified relative to the reference time series data. Based on the fused time series deviation, distinguish between inherent environmental fluctuations and abnormal hazard sources, and determine the target time series data corresponding to the abnormal hazard source from the time series data to be identified. The target time series data is mapped to personnel protection feature channel, personnel position and posture feature channel and equipment operation feature channel, and a mapping relationship between each data channel and the corresponding feature channel is established. Based on the weighted time series deviation of the data channel corresponding to each feature channel, the channel fusion weight and fusion priority of each feature channel are determined to obtain the hazard source identification result. The risk level is determined based on the hazard source identification results, and early warning and control information is generated based on the risk level.

[0015] In summary, the hydraulic construction hazard identification system and method based on a brain-like spiking neural network provided by this invention has the following advantages over existing technologies: (1) Through the information collection terminal, the multi-source time series information of the water conservancy construction area can be time-correlated and pulsed, so that data from different sources and different sampling frequencies can form unified multi-channel pulse time series data in time order, retain the order of occurrence and development of dangerous states, introduce working condition weight and confidence weight, fuse the deviation between the time series data to be identified and the reference time series data, reduce false alarms caused by inherent environmental fluctuations, and then through the brain-like time series recognition module, the target time series data can be task-oriented mapped and time series reasoned, which can identify dangerous states with continuity and gradual change. (2) The signal processing module processes the multi-source time-series information step by step. The noise reduction unit can reduce the interference of low-quality data on subsequent recognition. The standardization unit can solve the differences in the dimensions, range and format of image data, environmental parameters and posture data, so that different types of data can enter the unified processing flow. The timing synchronization unit performs time window matching on data with different sampling frequencies according to a unified time reference to avoid time misalignment between visual frames, posture data and environmental data. The pulse coding unit determines the pulse firing state according to the signal strength and degree of change. While compressing the amount of data, it retains the temporal relationship and change characteristics of multi-source information, which is conducive to reducing the computational burden of the edge processing device and improving the sensitivity of the spiking neural network to temporal changes. (3) Excitatory neuron clusters can enhance the continuous feature response related to the hazard source, and inhibitory neuron clusters can suppress random noise and environmental disturbances. When environmental disturbances are enhanced, the basic activation degree or inhibitory connection strength of inhibitory neurons can be increased, which can strengthen the suppression of instantaneous features such as rainstorm noise, water surface fluctuations, and light flicker. When the hazard source features are continuously and gradually changing, the suppression frequency of the corresponding feature channel can be reduced, which can preserve the evolutionary features such as personnel continuously approaching the edge, gradually losing posture, or equipment continuously approaching personnel. This achieves a dynamic balance between suppressing environmental noise and preserving the real hazard trend, and reduces false alarms or missed alarms caused by uniform noise reduction intensity. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the abnormal data source selection process of a water conservancy construction hazard source identification system based on a brain-like spiking neural network according to an embodiment of this application; Figure 2 This is a structural block diagram of an information acquisition terminal for a water conservancy construction hazard identification system based on a neuromorphic spiking neural network according to an embodiment of this application; Figure 3 This is a flowchart illustrating a method for identifying hazardous sources in water conservancy construction based on a neuromorphic spiking neural network, according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention; however, those skilled in the art will recognize the applicability of other processes and / or the use of other materials.

[0020] In one embodiment, see Figure 1 and Figure 2 As shown, this application provides a hydraulic construction hazard identification system based on a brain-like spiking neural network, including an information acquisition terminal and an edge processing device.

[0021] The information acquisition terminal is used to collect multi-source time-series information of the water conservancy construction area, perform time correlation and pulse processing on the multi-source time-series information to obtain multi-channel pulse time-series data, and send the multi-channel pulse time-series data to the edge processing device.

[0022] The information acquisition terminal is used to continuously collect multi-source time-series information and form multi-channel pulse time-series data within the water conservancy construction area; the edge processing device is used to perform preprocessing, anomaly identification, brain-like time-series recognition, and risk decision-making on the multi-channel pulse time-series data sequentially. The information acquisition terminal and the edge processing device are connected via a wireless communication module or a wired communication module.

[0023] The edge processing device is set in the central control box or computing node near the construction area of ​​the water conservancy construction site. The information acquisition terminal and the edge processing device can be connected by 5G, 4G, Wi-Fi or industrial Ethernet communication. When the communication link is temporarily interrupted, the information acquisition terminal can cache the multi-channel pulse timing data and continue to perform acquisition according to the most recently received control parameters. After the communication is restored, the cached data is retransmitted in the order of timestamps.

[0024] Specifically, the edge processing device includes a preprocessing module, an anomaly identification module, a neuromorphic time series recognition module, and a risk decision-making module. The preprocessing module receives multi-channel pulse time series data sent by the information acquisition terminal and performs data cleaning, standardization, time series synchronization, and validity checks on it, outputting time series data to be identified. The anomaly identification module is used to acquire reference time series data, determine the operating condition weight based on the acquired operating condition information, and determine the credibility weight based on the acquired data quality information. It uses the operating condition weight and the credibility weight to fuse the time series deviation of the time series data to be identified relative to the reference time series data, and distinguishes between inherent environmental fluctuations and other time series data based on the time series deviation. An abnormal hazard source is identified, and target time-series data corresponding to the abnormal hazard source is determined from the time-series data to be identified. The brain-like time-series recognition module includes an input encoding layer, a time-series inference layer, and an output recognition layer. The input encoding layer is used to map the target time-series data to feature channels corresponding to different hazard source identification tasks. The time-series inference layer is used to extract and fuse time-series features of each feature channel through a spiking neural network. The output recognition layer is used to determine the hazard source identification result based on the extraction and fusion result of the time-series inference layer. The risk decision module is used to determine the risk level based on the hazard source identification result and generate early warning and control information based on the risk level.

[0025] It should be noted that inherent environmental fluctuations refer to changes in perceived data caused by water mist, heavy rain, water ripples, water surface reflection, strong light backlight, wind and waves, normal equipment vibration, or short-term communication fluctuations; abnormal hazard sources refer to suspected safety anomalies that still exhibit persistence, directionality, or multi-channel correlation after correction by operating condition weights and confidence weights; the abnormal hazard sources output by the anomaly identification module are suspected anomalies to be identified, and are not equivalent to the final hazard source category. The final hazard source category is determined by the brain-like temporal recognition module. The anomaly identification module constitutes the pre-filter unit of the brain-like temporal recognition module, which can first exclude obvious inherent environmental fluctuations and only send suspected abnormal data into the spiking neural network, thereby reducing invalid inference operations and reducing misidentification caused by water mist, water surface ripples, and changes in light and shadow.

[0026] The information collection terminal can be installed on the smart safety helmet worn by the worker, or it can be placed on a wearable vest, the exterior of construction machinery, a fixed monitoring bracket, or a mobile inspection device. In this embodiment, a smart safety helmet is preferably used as the carrier of the information collection terminal, so that image, posture, and environmental data can be continuously collected as the worker moves, reducing the monitoring blind spots caused by changes in the work surface or obstruction of fixed cameras.

[0027] Specifically, the information acquisition terminal includes a visual perception module, an environmental perception module, a posture perception module, and a signal processing module. The visual perception module includes an image acquisition unit with overlapping fields of view, used to acquire visual temporal information of workers, protective equipment, and work equipment, and to determine the spatial relationship and motion state between workers, protective equipment, and work equipment based on the visual temporal information. The environmental perception module includes an environmental sensor for acquiring environmental parameters, and is used to form environmental temporal information of the water conservancy construction area based on the environmental parameters. The posture perception module includes an inertial measurement unit for acquiring the posture angles and motion trajectories of workers, and forming posture change information and motion state information of workers based on the acquired posture angles and motion trajectories. The visual temporal information, the environmental temporal information, the posture change information, and the motion state information together constitute the multi-source temporal information. The signal processing module is connected to the visual perception module, the environmental perception module, and the posture perception module respectively, and is used to perform time correlation and pulse processing on the multi-source temporal information to form the multi-channel pulse temporal data.

[0028] The helmet shell houses a main control circuit board, an inertial measurement unit, an environmental sensor, a communication module, and a rechargeable battery. An audio-visual and vibration warning unit is located on the rear of the shell. At least two image acquisition units are provided, with the field of view edges of each image acquisition unit overlapping at least partially. The main control circuit board registers the overlapping images according to the factory-calibrated intrinsic, extrinsic, and distortion parameters, thereby reducing the visual blind spots caused by a single camera turning its head or being obstructed by equipment. This design can simultaneously meet the needs of convenient wearing and continuous acquisition from multiple perspectives.

[0029] It should also be noted that the binocular vision unit includes two high-definition cameras, an image acquisition chip, and an image preprocessing circuit. The two high-definition cameras are positioned at the front of the helmet brim or on both sides of the helmet shell, forming a partially overlapping field of view. At least one high-definition camera can be equipped with an infrared night vision lens or an infrared supplementary lighting component to improve image acquisition quality at night, in tunnels, in shadow areas, or under low-light conditions. The binocular parallax principle is used to determine the three-dimensional position of personnel, construction equipment, water boundaries, slope edges, or edge protection facilities. The specific processing includes: defogging, exposure compensation, distortion correction, and stereo correction of the synchronously acquired left and right images; identifying water boundaries, bank edges, or guardrail boundaries; calculating the parallax value of the target pixels; determining the target depth based on the parallax value and the binocular baseline; further calculating the spatial distance between the worker and the work boundary or construction equipment. The vision perception module can also identify the presence, position, and wearing status of safety helmets, safety ropes, life jackets, reflective clothing, and other protective equipment. The equipment operation data is mainly identified by cameras to recognize the appearance, location and trajectory of the equipment. There is no need to connect to the internal controller of the construction equipment, nor to read internal data such as engine speed and hydraulic pressure. Therefore, it can be adapted to different types of construction equipment such as excavators, cranes, loaders, compactors and transport vehicles.

[0030] The safety helmet is made of ABS, HDPE or other insulating engineering plastics that meet impact protection requirements. The electronic compartment is equipped with an elastic sealing ring around its perimeter, and the exposed interfaces use waterproof connectors.

[0031] The environmental sensing module includes temperature and humidity sensors, air pressure sensors, water level sensors, water mist sensors, vibration sensors, and light sensors. The temperature and humidity sensors are used to identify high humidity, condensation tendency, and temperature drift compensation conditions; the water mist sensors are used to characterize the concentration of diffused water mist; the air pressure sensors, water level sensors, and vibration signals can help characterize rainstorms, floods, wind waves, or mechanical disturbances. Each sensor is connected to the terminal controller via a digital bus or analog conditioning circuit and shares a unified clock.

[0032] The attitude perception module is a six-axis inertial sensor containing a three-axis accelerometer and a three-axis gyroscope. It is fixed inside the safety helmet and establishes a stable attitude relationship with the helmet. Its output is calibrated with zero bias and transformed to obtain pitch angle, roll angle, yaw angle, angular velocity, linear acceleration and center of gravity change characteristics, which are used to distinguish standing, walking, standing still, bending over, tilting, slipping and body deviation. If necessary, a magnetometer can be added to form a nine-axis attitude unit, but this is not a limitation.

[0033] The signal processing unit module is implemented by a low-power microprocessor, neuromorphic processing chip, or programmable logic device. Its input is connected to each sensing unit, and its output is connected to the communication unit. The processing flow includes outlier removal, bandpass or median noise reduction, channel normalization, unified timestamp alignment, sliding window segmentation, and Poisson pulse coding. The audible and visual warning unit includes one or more of a buzzer, a voice device, and a warning light. In high-noise scenarios, a vibration reminder device can also be added. The power module uses a rechargeable battery and provides power detection, overcurrent, and undervoltage protection.

[0034] It should also be noted that the information acquisition terminal may include a communication module, an early warning output module, and a power supply module. The communication module can use 5G, 4G, Wi-Fi, LoRa, UWB, or industrial Ethernet communication methods; the early warning output module may include one or more of a buzzer, warning light, vibrator, and voice broadcaster; the power supply module may use a rechargeable battery and be configured with power detection, overcurrent protection, overvoltage protection, and undervoltage protection circuitry.

[0035] Specifically, the signal processing module includes a noise reduction unit, a normalization unit, a timing synchronization unit, and a pulse coding unit. The noise reduction unit is used to filter out instantaneous interference and invalid information in the multi-source timing information. The normalization unit is used to unify the data range and convert the data format of different data types in the multi-source timing information. The timing synchronization unit performs time window matching on the multi-source timing information with different sampling frequencies according to a unified time reference. The pulse coding unit is used to determine the pulse firing state according to the signal strength and signal change degree corresponding to the data channels of the workers, protective equipment, and working equipment, and generate multi-channel pulse timing data that retains the sequential relationship of the multi-source timing information according to the pulse firing state.

[0036] The noise reduction unit can employ median filtering, Gaussian filtering, Kalman filtering, bandpass filtering, outlier removal, or moving average methods to process different types of data. For example, it can perform dehazing, brightness equalization, and isolated noise removal on image data; zero-bias correction and bandpass filtering on inertial data; and amplitude limiting filtering and outlier removal on environmental data.

[0037] The standardization unit transforms image features, distance data, environmental parameters, and attitude parameters into a unified data range. For example, different physical quantities can be normalized to the range of 0 to 1, and each data point can be assigned a data channel identifier, terminal identifier, timestamp, and validity identifier.

[0038] The image acquisition unit can acquire images at a rate of 20 to 60 frames per second, the inertial measurement unit can acquire attitude information at a rate of 50 to 200 Hz, and the environmental sensor can acquire environmental information at a rate of 1 to 10 Hz. The timing synchronization unit uses a unified millisecond clock or image frame time as a reference to perform nearest-neighbor matching, linear interpolation, or time window statistics on data from different sampling frequencies. For data that cannot be matched within a preset tolerance range, no forgery padding is performed; instead, an invalid data flag is written at the corresponding location.

[0039] Specifically, the pulse coding unit can employ Poisson pulse coding, where the normalized continuous signal strength and rate of change are mapped to an adaptive pulse firing rate, and the determination of whether to generate a pulse at the current time step is based on the following formula:

[0040] in, Indicates the first Each data channel at time The probability of generating a pulse within the corresponding time step; No. An adaptive pulse firing rate determined by the signal strength and signal variation of each data channel; The discrete time step; It is a natural exponential function.

[0041] When the signal strength is high or the signal changes rapidly, the pulse firing rate increases accordingly; when the signal is in a stable state, the pulse firing rate decreases accordingly. Through the above pulse coding method, the time sequence of dangerous states such as personnel gradually approaching the edge area, continuous lack of protective equipment, and gradual deviation of equipment trajectory can be preserved while reducing the amount of data and the burden of edge computing.

[0042] Specifically, the anomaly identification module includes a reference data unit, a working condition weight determination unit, a credibility weight determination unit, and an anomaly comparison unit. The reference data unit is used to select reference time series data that matches the current water conservancy construction area from a preset reference time series database based on the construction scenario type, operation stage, and environmental conditions. The working condition weight determination unit is used to determine the working condition weight based on water level change status, water mist status, precipitation status, illumination status, and disturbance status. The credibility weight determination unit is used to determine the credibility weight based on sensor signal-to-noise status, image quality status, attitude stability status, and communication status. The anomaly comparison unit is used to determine the weighted time series deviation corresponding to the data channels of workers, protective equipment, and operating equipment using the working condition weight and the credibility weight, and compare the weighted time series deviation with the anomaly identification conditions to distinguish between inherent environmental fluctuations and abnormal hazard sources.

[0043] The reference time-series database is built from historical data without hazardous sources under different water conservancy construction scenarios. This database can include data on normal sunny days, heavy rain, water mist, strong sunlight / backlight, water level fluctuations, surface waves, and normal equipment vibration. The reference data unit selects the reference time-series data closest to the current working conditions based on the current construction area type, construction stage, weather conditions, water level, and equipment operating status.

[0044] It should also be noted that the operating condition weight is used to characterize the degree of influence of environmental conditions on the current data channel. For visual channels with significant water mist, their effective operating condition weight can be reduced; for visual channels with stable water levels and normal lighting, their effective operating condition weight can be increased. For attitude channels, if the inertial measurement unit operates stably and no abnormal vibrations occur, a higher effective operating condition weight can be maintained. The operating condition weight determination unit first performs dimensionless normalization on water mist concentration, water level fluctuation amplitude, and backlight reflection intensity, and then couples them according to the degree of influence of various environmental factors on the perception results of water conservancy construction. Specifically, the operating condition weight can be determined by the following formula:

[0045]

[0046] in, For the first The working condition weights of each sample; and These are the real-time water mist concentration and the critical threshold for water mist saturation, respectively. and These are the real-time water level fluctuation amplitude and the maximum water level fluctuation threshold, respectively. and These are backlight reflection intensity and strong light saturation threshold, respectively. + + , , and For fixed coupling coefficients under operating conditions, in this application The operating condition weights in this application The value range is 0.3-0.9.

[0047] It is understandable that multi-source time-series information is not simply a splicing of multiple sensor data, but rather that each pulse data is associated with a time identifier, data channel identifier, terminal identifier, and data validity identifier. Therefore, the edge processing device can trace the original sensing source corresponding to a certain identification result, and when a certain channel is affected by water mist, condensation, or packet loss, it can call other time synchronization channels for cross-validation.

[0048] In some embodiments, the operating condition weight determination unit divides the water mist concentration, water level fluctuation amplitude, and backlight reflection intensity by their respective saturation thresholds, thereby compressing environmental parameters of different dimensions into a comparable numerical range. (Coefficient) This is used to reflect the different impacts of various operating conditions on perceived data. For example, water mist may simultaneously affect the light transmittance of the camera and some environmental sensors, and its coefficient can be higher than that of short-term light reflection that only affects the visual channel. This keeps the operating condition weights within the effective boundary, thereby preventing extreme rainstorms from excessively suppressing all real dangerous features and preventing the environmental filtering mechanism from failing under normal operating conditions.

[0049] It should also be noted that, unlike the operating condition weight which reflects the degree of environmental interference, the confidence weight is used to characterize whether the current sample itself is credible. The confidence weight can integrate sensor signal-to-noise ratio, image clarity, attitude stability, and communication integrity, specifically expressed as follows:

[0050] in, For the first The credibility weight of each sample; The signal-to-noise ratio is the sensing signal-to-noise ratio; Image sharpness coefficient; This is the attitude stability coefficient; The communication integrity quality coefficient has a value between 0 and 1. When the device is stable, the picture is clear and there is no packet loss, it can be 1.0. When there is slight fogging, it can be 0.6. Samples below the effective threshold are not included in the calculation.

[0051] Among them, the credibility weight corresponds to whether a sample itself is credible; the sensor signal-to-noise quality coefficient can be obtained by mapping the ratio of effective signal power to noise power; the image clarity quality coefficient can be obtained by combining gradient, contrast, occlusion ratio, and overexposure ratio; the attitude stability quality coefficient can be obtained by acceleration modulus deviation, gyroscope zero bias, and sensor self-test status; and the communication integrity quality coefficient can be obtained by packet loss rate, latency, and retransmission count. Each sub-coefficient is between 0 and 1. If any key quality item fails severely, the overall credibility of the sample can be rapidly reduced.

[0052] After obtaining the operating condition weight and the confidence weight, the anomaly comparison unit applies both weights to the residual between the real-time time series data and the reference time series data, and obtains the double-weighted root mean square error using the following formula:

[0053] in, The score is a double-weighted anomaly detection score; For the first Real-time valid data for each sample; To match the reference data for no hidden dangers under the working conditions; This represents the number of valid samples.

[0054] The anomaly comparison unit not only compares the time-series deviations at a single time point, but also analyzes the directionality, persistence, and periodicity of the deviations within a continuous observation window. When the deviation in the visual channel changes periodically and the period of change is consistent with the period of water surface ripples, while the attitude channel does not show a corresponding change, it can be identified as an inherent environmental fluctuation. When the distance between the personnel and the work boundary decreases continuously, and the pitch angle, roll angle, or center of gravity shift increases continuously, even if water surface fluctuations exist at the same time, the corresponding data should be retained as target time-series data.

[0055] Specifically, the anomaly comparison unit is also used to determine the channel fusion weight based on the weighted time series deviation corresponding to each of the feature channels, and to set the fusion priority of the corresponding feature channels according to the reciprocal relationship of each of the weighted time series deviations, so that feature channels with smaller weighted time series deviations and higher data reliability can obtain higher fusion priority.

[0056] In this process, after the anomaly detection module has identified suspected anomalies, the uncertainty deviations for the personnel protection feature channel, the personnel position and posture feature channel, and the equipment operation feature channel are calculated separately. Here, the weighted time-series deviation reflects the degree of uncertainty of the corresponding feature channel after being affected by noise, operating condition mismatch, and data quality degradation, rather than directly indicating the degree of danger. Therefore, the smaller the weighted time-series deviation, the more stable and reliable the data in that feature channel is, and the higher its fusion priority can be obtained.

[0057] Among them, after obtaining weighted time series biases for the three types of time series feature channels—personnel protection, personnel position and attitude, and equipment operation—the channel fusion weights can be calculated according to the inverse relationship of error to give higher fusion priority to channels with smaller biases and higher data reliability.

[0058] in, For the first Fusion weights for time-series feature channels; This represents the weighted root mean square error for the corresponding feature channel; In this embodiment, the total number of core feature channels is [number]. Channels with smaller errors and higher reliability are given higher priority; the priority of attitude channels is increased when dew forms, and the priority of associated visual channels is decreased when there is periodic water surface swaying.

[0059] When the reliability of visual temporal information is reduced due to water mist, condensation, or lens contamination, the anomaly comparison unit reduces the fusion priority of personnel protective visual features or equipment visual features and increases the fusion priority of personnel position and posture feature channels. When obvious periodic water surface swaying, water surface ripples, or light and shadow flickering are detected, the fusion priority of visual feature channels associated with the corresponding periodic changes is reduced. When personnel position, posture angle, and center of gravity changes show a consistent edge offset trend within a continuous time window, even if the visual channel is partially interfered with by water mist, the fusion priority of personnel position and posture feature channels is maintained at a high level. Through the above dynamic weight allocation method, it is not necessary to use fixed weights for all data channels. Instead, the information contribution level can be dynamically adjusted among visual, posture, and equipment operation features according to the real-time working conditions and data quality of the water conservancy construction site.

[0060] This invention presets a fixed selection threshold. ,when If the current data fluctuation is determined to be a normal environmental fluctuation caused by water mist, water surface swaying, or changes in light and shadow, and there is no safety hazard, then the current round of reasoning is terminated directly; When a suspected hazard source is identified as having a temporal anomaly, it is automatically fed into a brain-like spiking neural network for in-depth temporal feature verification. If false anomalies or missed detections occur in three consecutive inferences, the assignment range of the working condition weight and the confidence weight can be automatically adjusted to dynamically adapt to real-time changes in the working conditions at the construction site.

[0061] Specifically, the anomaly comparison unit first calculates separately for each data channel. Then, its directionality, persistence, and periodicity within a continuous observation window are statistically analyzed. If the deviation, even briefly, exceeds a fixed selection threshold... However, if the change is the same as the water surface ripple period and no corresponding change appears in the attitude channel, it can be marked as an inherent environmental fluctuation. If the distance between the person and the boundary decreases monotonically in the continuous window, while the pitch angle and center of gravity shift continue to increase, then even if there is water surface periodic interference, the target's time series data will be retained for subsequent identification.

[0062] For example, when If a system is identified as a false anomaly three times consecutively by manual review, it does not directly modify the anomaly threshold to evade detection. Instead, it first checks whether the reference operating conditions match, whether the sensors are drifting, and whether the normalized threshold for the operating condition weights is outdated. Only when the data quality meets the requirements and the manual review is consistent will the weight range be fine-tuned within a limited range, thereby avoiding compromising the system's security sensitivity due to individual erroneous reviews.

[0063] Specifically, the input encoding layer includes a personnel protection feature channel, a personnel position and posture feature channel, and an equipment operation feature channel. The personnel protection feature channel is used to characterize the presence, wearing status, and duration of protective equipment. The personnel position and posture feature channel is used to characterize the spatial changes, posture changes, and movement trends of the operator. The equipment operation feature channel is used to characterize the position changes, trajectory changes, and action status of the operating equipment relative to the operator. The temporal inference layer is used to extract temporal features from each of the feature channels according to the fusion priority, and to perform feature fusion between the feature channels to identify personnel protection anomalies, water and edge risks, and equipment operation risks. The personnel protection characteristic channel can input information such as the presence of a safety helmet, the position of the safety helmet, the safety rope attachment mark, the life jacket wearing mark, the reflective clothing wearing mark, and the duration of the protective status. Through continuous time-series judgment, it can distinguish between workers briefly raising their hands to adjust their safety helmets and workers not wearing their safety helmets correctly for a long time, as well as between workers temporarily obstructing their safety ropes and workers not wearing their safety ropes for an extended period. The personnel position and posture characteristic channel can input information such as the distance between the personnel and the water boundary, the edge of the bank, or the edge protection boundary, the rate of distance change, the direction of movement, the pitch angle, the roll angle, the center of gravity shift, and the dwell time. When the distance between the personnel and the edge boundary continuously decreases, while the pitch angle and the center of gravity shift continuously increase, it can form a continuous gradual change characteristic of the water and edge risk. The equipment operation characteristic channel can input information such as the position of the construction equipment relative to the workers, the equipment's external trajectory, the direction of movement, the speed of movement, the appearance of the equipment during start-up and shutdown, and the dwell time in the dangerous area. The system can determine whether equipment is gradually entering a water-adjacent restricted area, continuously approaching workers, or lingering in a weak slope area based on equipment outline, cab location, working device location, and continuous image trajectory. The input encoding layer employs a multi-scale spatiotemporal pulse fusion encoding method, encoding instantaneous changes within a short time scale, continuous changes in motion within a medium time scale, and trend changes within a long time scale. The short time scale is used to identify slips, sudden tilts, and sudden equipment approach; the medium time scale is used to identify persistent lack of protective equipment and personnel lingering in dangerous areas; and the long time scale is used to identify personnel slowly approaching the water or construction equipment gradually shifting towards a dangerous area. The three feature channels can maintain relatively independent neuron clusters, and limited cross-channel connections are set in the temporal inference layer to enable mutual verification of personnel protection, personnel posture, and equipment operation characteristics, while preventing strong noise from one visual channel from spreading to other feature channels through a large number of cross-channel connections.

[0064] Specifically, the temporal inference layer includes an excitatory neuron cluster, an inhibitory neuron cluster, and a balance regulation unit. The excitatory neuron cluster receives temporal features from corresponding feature channels and generates an excitatory response based on these features. The inhibitory neuron cluster is connected to the excitatory neuron cluster to receive the excitatory response and feeds back an inhibitory signal to the excitatory neuron cluster through inhibitory connections. The balance regulation unit is connected to the excitatory neuron cluster, the inhibitory neuron cluster, and the anomaly detection module, respectively, and is used to acquire the group response states of the excitatory and inhibitory neuron clusters, and to determine the environmental interference based on the response deviation between the two group response states and the environmental interference result output by the anomaly detection module. As a result, the basal activation level of the inhibitory neuron cluster and the inhibitory effect of the inhibitory connections are adjusted. When the anomaly detection module determines that environmental interference is enhanced, the balance adjustment unit increases the basal activation level of the inhibitory neuron cluster and enhances the inhibitory effect of the inhibitory connections on the excitatory neuron cluster to suppress the instantaneous temporal features associated with the environmental interference. When the temporal inference layer detects that the temporal features corresponding to the hazard source are continuously and gradually changing from the corresponding feature channels, the balance adjustment unit reduces the inhibition trigger frequency for the corresponding feature channels to preserve the continuous temporal features associated with the evolution of the hazard source.

[0065] In this process, inhibitory neuron clusters can receive excitatory responses generated by excitatory neuron clusters through excitatory connections, thereby clearly forming an excitatory driving link from excitatory neuron clusters to inhibitory neuron clusters; inhibitory neuron clusters can feed back inhibitory signals to excitatory neuron clusters through inhibitory connections, forming a feedback inhibition link between inhibitory neuron clusters and excitatory neuron clusters.

[0066] The connections between excitatory and inhibitory neurons are categorized into four types: E to E, E to I, I to E, and I to I. Initially, the positive weights for E to E can be between 0.6 and 0.8, the positive weights for E to I can be between 0.4 and 0.6, the negative weights for I to E can be between -0.7 and -0.5, and the weakly positive weights for I to I can be between 0.2 and 0.3. Positive weights are limited to between 0 and 1, and negative weights are limited to between -1 and 0. These ranges are implementation examples and can be changed in practice based on training results. In high-humidity water mist conditions, the lower limit of the I to E inhibition weights is locked to prevent over-inhibition; in water surface reflection scenarios, the E to E weights corresponding to light and shadow clutter are reduced; and in slow offset scenarios near edges, the propagation priority of the attitude channel is preserved.

[0067] During temporal reasoning, each neuron iterates over the current pulse input and historical membrane potential. When the membrane potential exceeds the firing threshold, an output pulse is generated and the neuron is reset. Specifically, this can be updated using the following formula:

[0068] in, Indicates the neuron at time t. The membrane potential of the next timing step; The membrane potential attenuation coefficient is set to 0.85 in this application; For a moment Input timing pulse current; The reset potential is set to 0.2 in this application; It is a pulse-triggered function; the membrane potential accumulates with continuous input and decays over time, preserving the gradual change process of the hazard source and avoiding the long-term accumulation of isolated noise.

[0069] It also needs to be explained that, If the size is too small, it will cause people to slowly approach the edge or the cumulative effect of a long-term lack of protective equipment to disappear too quickly; An excessively large value may cause short-term disturbances that have ended to persist for an extended period. An optimal value of 0.85 is suitable for balancing gradual hazards and instantaneous disturbances in hydraulic engineering scenarios, but it can be adjusted using a validation set at different sampling frequencies. The reset potential is used to clear some of the accumulated response after neuron firing, allowing the network to continue responding to subsequent changes in hazard.

[0070] The balancing unit can calculate the average response of the excitatory and inhibitory neuron clusters every 10 time steps. When the excitatory response is too high and the anomaly detection module determines that there is interference from heavy rain, water mist, or strong light, the absolute value of the I-to-E inhibitory connection is increased; when the excitatory response is too low and the personnel position and posture feature channel has continuous gradual changes, the absolute value of the I-to-E inhibitory connection or the inhibition trigger frequency is appropriately reduced; for persistent water mist, a water mist-specific inhibitory sub-cluster can be set in the inhibitory neuron cluster of the temporal inference layer; for periodic pulses of water surface light and strong backlight, the forward transmission weight of the corresponding E-to-E connection can be reduced; for low-frequency temporal disturbances caused by slow rise and fall of water level, periodic decoupling or frequency band separation can be performed; for dangerous features of personnel near the edge, the excitatory transmission priority of the personnel position and posture feature channel is maintained.

[0071] Based on this, the temporal inference layer also includes an adaptive learning unit connected to the anomaly detection module and the balance adjustment unit respectively. The adaptive learning unit is used to obtain the pulse firing time of each of the pre-level and post-level neurons. The pre-level and post-level neurons are connected through excitatory connections for transmitting excitatory responses or through inhibitory connections. Based on the pulse firing order and pulse time interval between the pre-level and post-level neurons, the unit determines the connection strength adjustment amount, updates the connection strength of the excitatory or inhibitory connections according to the connection strength adjustment amount, and limits the updated connection strength within the corresponding preset connection strength range. The adaptive learning unit is also used to obtain the type and duration of environmental interference determined by the anomaly detection module, as well as the hazard source features and hazard source feature durations extracted by the temporal inference layer from each of the feature channels. Based on the duration of environmental interference and the duration of hazard source features, the unit determines a target temporal learning window and selects pulse temporal data to participate in the connection strength update within the target temporal learning window.

[0072] The adaptive learning unit updates the synaptic connection strength based on the firing order and time interval of the preceding and following neurons, which can be specifically expressed as:

[0073] in, This represents the change in synaptic weight; For adaptive learning rate, the base value is set to 0.001, and it can be 0.005 for incremental edge learning; This represents the pulse timing of the preceding neuron; This represents the pulse moment of the subsequent neuron; It is a timing window constant, with a base value of 20ms; it strengthens dangerous features with sequential causal relationships and weakens random noise, and after the update, the weights are limited to the preset boundary.

[0074] For data with ordered pulse timing, consistent cross-channel responses, and verified by manual verification as genuine hazard sources, the adaptive learning unit can enhance the corresponding E-to-E connections to strengthen the real hazard characteristics. For data with disordered pulse timing, obvious periodicity, and inconsistency with the attitude channel, the unit can reduce the connection strength of the corresponding excitatory connections and enhance the inhibitory effect of the corresponding I-to-E inhibitory connections. When the anomaly identification module determines that the current construction area is under water mist conditions, the adaptive learning unit expands the target temporal learning window to avoid updating network connections based solely on short-term fogging images. For data associated with periodic water surface swaying, water surface ripples, or light and shadow flickering, its participation in connection strength updates is reduced or suspended. When the personnel position and attitude feature channel detects a continuous offset of the worker relative to the work boundary, and this offset is accompanied by changes in pitch angle, roll angle, or center of gravity, the corresponding pulse timing data is retained for connection strength updates, and the update priority of the corresponding connections in the personnel position and attitude feature channel is increased. After each connection strength update, the adaptive learning unit compares the update result with the corresponding preset connection strength range. When the connection strength exceeds the upper limit, the upper limit value is used; when the connection strength is below the lower limit, the lower limit value is used. Boundary constraints can prevent the network from overactivating due to the continuous increase of excitatory weights, and can also prevent the real danger features from being completely suppressed due to the excessive absolute value of inhibitory weights.

[0075] Specifically, the output identification layer is connected to the temporal inference layer and is used to receive the inference results obtained by the temporal inference layer after extracting and fusing temporal features from each feature channel. Within continuously set preset observation windows, it counts the number of output pulses corresponding to different hazard source identification tasks. The output identification layer is also used to determine the identification confidence information corresponding to each hazard source identification result based on the number of output pulses corresponding to each hazard source identification task, the consistency of output results within adjacent preset observation windows, and the channel fusion weights of each feature channel participating in the corresponding hazard source identification task. The output identification layer then sends the hazard source identification results and the identification confidence information to the risk decision module. The risk decision module is connected to the output identification layer and is used to compare the identification confidence information with risk judgment thresholds and determine the risk level corresponding to different confidence intervals based on the comparison results. The risk decision module is also used to generate corresponding early warning control information based on the risk level. The early warning control information includes terminal early warning indicators. The system acquires at least one of the following: a terminal warning command, a field linkage warning command, a management personnel notification message, and a platform warning command. The terminal warning command is sent to the information acquisition terminal, the field linkage warning command is sent to the field linkage equipment, and the management personnel notification message and the platform warning command are sent to the management platform. The management platform is communicatively connected to the edge processing device and is used to receive and store the hazard identification result, the identification confidence information, the risk level, and the corresponding warning record. It also acquires verification information for the hazard identification result and the risk level, generates model update data for adjusting the brain-like temporal recognition module based on the verification information, and sends the model update data to the edge processing device. The edge processing device is further used to update at least one of the model parameters and recognition rules of the brain-like temporal recognition module based on the model update data, and incrementally calibrates the updated brain-like temporal recognition module using newly collected temporal data from the current water conservancy construction area.

[0076] The output recognition layer counts the number of output pulses corresponding to different hazard source recognition tasks within a preset observation window, and combines the output consistency of continuous observation windows with the feature channel fusion weights involved in the recognition to obtain the recognition confidence level using the following formula:

[0077] in, Hazard source category The confidence level of identification; To observe the categories within the window The number of pulses; The total number of all output pulses within the observation window; To ensure consistent output across continuous windows; For the first Channel to Category The normalized response; , , represent the combination coefficients respectively.

[0078] In the identification of abnormal personnel protection, the system continuously acquires head images, helmet positions, safety rope attachment status, and duration. It can distinguish between short-term adjustment actions and continuous violations such as not wearing a helmet at all, prolonged deviation, not attaching a safety rope near water, or not reattaching it after removal. In the identification of risks near water and edges, the system integrates the distance between personnel and the boundary, pitch angle, center of gravity shift, direction of movement, and duration of dwelling. When personnel are 1m to 1.5m from the edge and leaning forward, it can be judged as a warning risk. When they continue to approach within 1m without protection, it can be judged as a high-risk risk. Specific thresholds can be calibrated according to the construction organization design. In the identification of risks in equipment operation, the system continuously extracts the position, working range, start and stop, operating posture, and duration of dwelling near water of excavators, loaders, or compactors. It identifies deviations beyond the range, illegal start and stop, unattended idling, prolonged dwelling in dangerous areas, or operation near weak slope areas. It also avoids misjudging normal turning or short-term pauses as violations by using trajectory continuity and changes in the relative distance between personnel and machines.

[0079] The risk assessment threshold preferably includes a first risk assessment threshold and a second risk assessment threshold, with the second risk assessment threshold being greater than the first risk assessment threshold. In this embodiment, the first risk assessment threshold can be 0.6, and the second risk assessment threshold can be 0.8. When the identification confidence level is less than 0.6, it is determined to be a level 0 general risk, mainly corresponding to instantaneous, minor, or non-persistent anomalies, such as personnel briefly adjusting their safety helmets, equipment making normal minor turns, or personnel briefly approaching the edge area and then immediately retreating. Level 0 risks can simply be stored as records or displayed as prompts on the terminal. When the identification confidence level is greater than or equal to 0.6 and less than 0.8, it is determined to be a level 1 warning risk, mainly corresponding to persistent routine hazards, such as safety helmets being worn crookedly for extended periods, safety ropes not being properly secured, personnel continuously lingering in dangerous areas near the edge, or construction equipment continuously engaging in minor violations near the water's edge. Level 1 risks can trigger terminal-based audible, visual, vibration, or voice alerts and send notifications to management personnel. When the identification confidence level is greater than or equal to 0.8, it is classified as a Level 2 high-risk risk, primarily corresponding to urgent hazards that could directly cause personnel falls, equipment overturning, or personnel collisions. Examples include personnel continuously approaching water and losing their balance, working near water without wearing protective equipment, or construction equipment entering a high-risk water-adjacent area and remaining there for an extended period. In addition to triggering terminal alerts and management notifications, Level 2 risks can also activate equipment shutdown devices, area warning lights, on-site broadcasts, or access control devices.

[0080] The management platform stores hazard identification results, risk levels, early warnings, and handling records. It also sends equipment configuration parameters, model update data, and remote control commands to edge processing devices. The platform can save information such as hazard occurrence time, location, hazard type, risk level, handling personnel, handling results, and manual review results. Data that is manually reviewed as false alarms, omissions, or inaccurate risk levels is marked as model update samples by the management platform.

[0081] In one possible implementation, the management platform compiles on-site verification data from the past 30 days. When the system's recognition level differs from the manual review level by at least one level, the corresponding data is marked as an abnormal sample. When the same type of error occurs three times consecutively, the STDP learning rate, channel fusion weight, operating condition weight, confidence interval, or network node parameters are adjusted within a preset range. The management platform can also generate a lightweight model update package through knowledge distillation and then send the lightweight model update package to the edge processing device. After receiving the model update data, the edge processing device first performs incremental calibration and validity verification on the newly added time series data locally. After the verification result meets the preset conditions, the online neuromorphic time series recognition model is replaced to avoid unverified model updates directly affecting on-site safety recognition.

[0082] Based on the same inventive concept, see [link / reference] Figure 3 As shown, this application also provides a method for identifying water conservancy construction hazards based on a brain-like spiking neural network, which is implemented based on the aforementioned water conservancy construction hazard identification system based on a brain-like spiking neural network. The water conservancy construction hazard identification method includes: S1: Collect multi-source time-series information from the water conservancy construction area, perform time correlation and pulse processing on the multi-source time-series information, and obtain multi-channel pulse time-series data; S2: Process the multi-channel pulse timing data to obtain timing data to be identified; S3: Obtain reference time series data, use operating condition weight and confidence weight to fuse the time series deviation of the time series data to be identified relative to the reference time series data, distinguish between inherent environmental fluctuations and abnormal hazard sources based on the fused time series deviation, and determine the target time series data corresponding to the abnormal hazard source from the time series data to be identified; S4: Map the target time series data to the personnel protection feature channel, the personnel position and posture feature channel and the equipment operation feature channel, and establish the mapping relationship between each data channel and the corresponding feature channel. Based on the weighted time series deviation of the data channel corresponding to each feature channel, determine the channel fusion weight and fusion priority of each feature channel to obtain the hazard source identification result. S5: Determine the risk level based on the hazard source identification results, and generate early warning and control information based on the risk level.

[0083] The system utilizes a visual perception module, an environmental perception module, and a posture perception module to collect visual time-series information, environmental time-series information, posture change information, and motion state information in parallel. Each data point is stamped with a timestamp, data channel identifier, and data validity identifier. Noise reduction, standardization, time window matching, and pulse coding are then applied to the multi-source time-series information to form multi-channel pulse time-series data. For data that cannot be matched within a preset time tolerance, an invalid data identifier is set, and no forgery padding is performed. Reference time-series data is selected from a reference time-series database based on the current construction scenario, work stage, and environmental conditions. Working condition weights are determined based on water level, water mist, precipitation, illumination, and disturbance status. Credibility weights are determined based on sensor signal-to-noise ratio, image quality, posture stability, and communication status. A dual-weighted time-series deviation is calculated, and the persistence, directionality, and periodicity of the weighted time-series deviation within the continuous observation window are analyzed to distinguish between inherent environmental fluctuations and suspected dangerous anomalies. For inherent environmental fluctuations, the current round of deep inference is terminated; for suspected dangerous anomalies, target time-series data is extracted and mapped to personnel protection feature channels, personnel position and posture feature channels, and equipment operation feature channels, respectively. Channel fusion weights are calculated, and EI balanced time-series inference is performed according to channel fusion weights and fusion priorities. The output identification layer counts the number of output pulses corresponding to different hazard source tasks. The identification confidence information is determined by combining the consistency of the continuous observation window and the channel fusion weights. The identification confidence information is compared with the risk judgment threshold to determine the risk level, and at least one of the following is generated: terminal warning instruction, on-site linkage warning instruction, management personnel notification information, and platform warning instruction. The hazard source identification results, identification confidence information, risk level, warning records, and handling results are uploaded to the management platform. The management platform generates model update data based on the manual review information, and the edge processing device completes incremental calibration based on the model update data and the newly added time-series data.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for identifying hazardous sources in hydraulic construction based on a brain-like spiking neural network, characterized in that, This includes information collection terminals and edge processing devices, among which, The information acquisition terminal is used to collect multi-source time-series information of the water conservancy construction area, perform time correlation and pulse processing on the multi-source time-series information to obtain multi-channel pulse time-series data, and send the multi-channel pulse time-series data to the edge processing device; The edge processing device includes a preprocessing module, an anomaly identification module, a brain-like time sequence recognition module, and a risk decision module. The preprocessing module is used to process the multi-channel pulse time sequence data to obtain the time sequence data to be identified. The anomaly identification module is used to acquire reference time series data, determine the working condition weight based on the acquired working condition information, and determine the credibility weight based on the acquired data quality information. The working condition weight and the credibility weight are used to fuse the time series deviation of the time series data to be identified relative to the reference time series data. Based on the time series deviation, the module distinguishes between inherent environmental fluctuations and abnormal hazard sources, and determines the target time series data corresponding to the abnormal hazard source from the time series data to be identified. The brain-like temporal recognition module includes an input encoding layer, a temporal reasoning layer, and an output recognition layer. The input encoding layer is used to map the target temporal data to feature channels corresponding to different hazard source recognition tasks. The temporal reasoning layer is used to extract and fuse temporal features of each feature channel through a spiking neural network. The output recognition layer is used to determine the hazard source recognition result based on the extraction and fusion result of the temporal reasoning layer. The risk decision module is used to determine the risk level based on the hazard source identification results, and generate early warning and control information based on the risk level.

2. The hydraulic construction hazard identification system based on a brain-like spiking neural network as described in claim 1, characterized in that, The information acquisition terminal includes a visual perception module, an environmental perception module, an attitude perception module, and a signal processing module, wherein... The visual perception module includes an image acquisition unit with overlapping fields of view, used to acquire visual temporal information of workers, protective equipment and work equipment, and to determine the spatial relationship and motion state between workers, protective equipment and work equipment based on the visual temporal information; The environmental sensing module includes an environmental sensor for acquiring environmental parameters, and the environmental sensing module is used to generate environmental time-series information of the water conservancy construction area based on the environmental parameters. The attitude perception module includes an inertial measurement unit, which is used to acquire the attitude angle and motion trajectory of the operator, and to form the attitude change information and motion state information of the operator based on the acquired attitude angle and motion trajectory. The visual timing information, the environmental timing information, the posture change information, and the motion state information together constitute the multi-source timing information. The signal processing module is connected to the visual perception module, the environmental perception module, and the posture perception module respectively, and is used to perform time correlation and pulse processing on the multi-source timing information to form the multi-channel pulse timing data.

3. The water conservancy construction hazard identification system based on a brain-like spiking neural network as described in claim 2, characterized in that, The signal processing module includes a noise reduction unit, a normalization unit, a timing synchronization unit, and a pulse coding unit, wherein... The noise reduction unit is used to filter out instantaneous interference information and invalid information in the multi-source timing information; The standardization unit is used to unify the data range and convert the data format of different data types in the multi-source time series information; The timing synchronization unit performs time window matching on the multi-source timing information with different sampling frequencies according to a unified time reference; The pulse coding unit is used to determine the pulse firing state based on the signal strength and signal change degree corresponding to the data channels of the operator, protective equipment and operating equipment, and to generate multi-channel pulse timing data that retains the sequential relationship of the multi-source timing information based on the pulse firing state.

4. The hydraulic construction hazard identification system based on a brain-like spiking neural network as described in claim 1, characterized in that, The anomaly detection module includes a reference data unit, a working condition weight determination unit, a confidence weight determination unit, and an anomaly comparison unit, wherein... The reference data unit is used to select reference time series data that matches the current water conservancy construction area from a preset reference time series database according to the construction scenario type, operation stage and environmental conditions. The operating condition weight determination unit is used to determine the operating condition weight based on the water level change state, water mist state, precipitation state, illumination state, and disturbance state. The credibility weight determination unit is used to determine the credibility weight based on the sensor signal-to-noise ratio, image quality, attitude stability, and communication status. The anomaly comparison unit is used to determine the weighted time series deviation corresponding to the data channels of workers, protective equipment and work equipment using the working condition weight and the confidence weight, and compares the weighted time series deviation with the anomaly identification conditions to distinguish between inherent environmental fluctuations and abnormal hazard sources.

5. The water conservancy construction hazard identification system based on a brain-like spiking neural network as described in claim 4, characterized in that, The anomaly comparison unit is also used to determine the channel fusion weight based on the weighted time series deviation corresponding to each of the feature channels, and to set the fusion priority of the corresponding feature channels according to the reciprocal relationship of each of the weighted time series deviations, so that feature channels with smaller weighted time series deviations and higher data reliability can obtain higher fusion priority.

6. The hydraulic construction hazard identification system based on a neuromorphic spiking neural network as described in claim 5, characterized in that, The input coding layer includes a personnel protection feature channel, a personnel position and posture feature channel, and an equipment operation feature channel, wherein... The personnel protection feature channel is used to characterize the presence, wearing status, and duration of the protective equipment. The personnel position and posture feature channel is used to characterize the spatial changes, posture changes, and movement trends of the workers. The equipment operation feature channel is used to characterize the positional changes, trajectory changes, and motion states of the operating equipment relative to the operator; The temporal inference layer is used to extract temporal features from each of the feature channels according to the fusion priority, and to perform feature fusion between the feature channels to identify personnel protection anomalies, water and edge risks, and equipment operation risks.

7. The hydraulic construction hazard identification system based on a brain-like spiking neural network as described in claim 6, characterized in that, The temporal reasoning layer includes clusters of excitatory neurons, clusters of inhibitory neurons, and a balance regulation unit, wherein... The excitatory neuron cluster is used to receive the temporal features of the corresponding feature channels and generate an excitatory response based on the temporal features; The inhibitory neuron cluster is connected to the excitatory neuron cluster to receive the excitatory response and to feed back an inhibitory signal to the excitatory neuron cluster through the inhibitory connection. The balance adjustment unit is connected to the excitatory neuron cluster, the inhibitory neuron cluster, and the anomaly detection module, respectively, and is used to obtain the group response state of the excitatory neuron cluster and the inhibitory neuron cluster. Based on the response deviation between the two group response states and the environmental interference judgment result output by the anomaly detection module, the unit adjusts the basic activation level of the inhibitory neuron cluster and the inhibitory effect strength of the inhibitory connection. When the anomaly detection module determines that environmental interference is enhanced, the balance adjustment unit increases the basic activation level of the inhibitory neuron cluster and enhances the inhibitory effect of the inhibitory connection on the excitatory neuron cluster, so as to suppress the transient temporal features associated with the environmental interference. When the temporal inference layer detects that the temporal features corresponding to the hazard source are continuously and gradually changing from the corresponding feature channels, the balance adjustment unit reduces the suppression trigger frequency for the corresponding feature channels in order to preserve the continuous temporal features associated with the evolution of the hazard source.

8. The hydraulic construction hazard identification system based on a brain-like spiking neural network as described in claim 7, characterized in that, The temporal inference layer further includes an adaptive learning unit connected to both the anomaly detection module and the balance adjustment unit, wherein... The adaptive learning unit is used to obtain the pulse firing times of the pre-level neuron and the post-level neuron. The pre-level neuron and the post-level neuron are connected through excitatory connections for transmitting excitatory responses or through inhibitory connections. Based on the pulse firing order and pulse time interval between the pre-level neuron and the post-level neuron, a connection strength adjustment amount is determined. The connection strength of the excitatory connection or the inhibitory connection is updated according to the connection strength adjustment amount, and the updated connection strength is limited to the corresponding preset connection strength range. The adaptive learning unit is also used to obtain the type and duration of environmental interference determined by the anomaly identification module, as well as the hazard source features and hazard source feature durations extracted by the temporal inference layer from each of the feature channels, determine a target temporal learning window based on the duration of environmental interference and the duration of hazard source features, and select pulse temporal data that participate in the connection strength update within the target temporal learning window.

9. A hydraulic construction hazard identification system based on a brain-like spiking neural network as described in claim 8, characterized in that, The output recognition layer is connected to the temporal inference layer and is used to receive the inference results obtained by the temporal inference layer after extracting and fusing temporal features of each feature channel, and to count the number of output pulses corresponding to different hazard source identification tasks within a continuously set preset observation window. The output recognition layer is further configured to determine the recognition confidence information corresponding to each hazard source recognition result based on the number of output pulses corresponding to each hazard source recognition task, the consistency of output results within adjacent preset observation windows, and the channel fusion weight of each feature channel participating in the corresponding hazard source recognition task, and send the hazard source recognition result and the recognition confidence information to the risk decision module. The risk decision module is connected to the output recognition layer and is used to compare the recognition confidence information with the risk judgment threshold respectively, and determine the risk level corresponding to different confidence intervals based on the comparison results. The risk decision module is also used to generate corresponding early warning and control information according to the risk level. The early warning and control information includes at least one of terminal early warning instructions, on-site linkage early warning instructions, management personnel notification information and platform early warning instructions. The module sends the terminal early warning instructions to the information collection terminal, the on-site linkage early warning instructions to the on-site linkage equipment, and the management personnel notification information and the platform early warning instructions to the management platform. The management platform is communicatively connected to the edge processing device and is used to receive and store the hazard identification results, the identification confidence information, the risk level and the corresponding early warning record, obtain the review information for the hazard identification results and the risk level, generate model update data for adjusting the brain-like temporal recognition module based on the review information, and send the model update data to the edge processing device. The edge processing device is also used to update at least one of the model parameters and recognition rules of the brain-like temporal recognition module according to the model update data, and to perform incremental calibration on the updated brain-like temporal recognition module using the newly added temporal data collected in the current water conservancy construction area.

10. A method for identifying hazardous sources in hydraulic construction based on a brain-like spiking neural network, characterized in that, Performed based on the water conservancy construction hazard identification system based on any one of claims 1 to 9, the water conservancy construction hazard identification method includes: Multi-source time-series information from the water conservancy construction area is collected, and the multi-source time-series information is subjected to time correlation and pulse processing to obtain multi-channel pulse time-series data. The multi-channel pulse timing data is processed to obtain the timing data to be identified; Obtain reference time series data, and use operating condition weight and confidence weight to fuse the time series deviation of the time series data to be identified relative to the reference time series data. Based on the fused time series deviation, distinguish between inherent environmental fluctuations and abnormal hazard sources, and determine the target time series data corresponding to the abnormal hazard source from the time series data to be identified. The target time series data is mapped to personnel protection feature channel, personnel position and posture feature channel and equipment operation feature channel, and a mapping relationship between each data channel and the corresponding feature channel is established. Based on the weighted time series deviation of the data channel corresponding to each feature channel, the channel fusion weight and fusion priority of each feature channel are determined to obtain the hazard source identification result. The risk level is determined based on the hazard source identification results, and early warning and control information is generated based on the risk level.

Images