Multi-source data collaborative processing method based on artificial intelligence
By integrating visual images and multi-source sensor data through collaborative processing, the technical bottlenecks of limited resources and unstable communication were overcome, enabling high-confidence data parsing and abnormal state identification, thus ensuring the accuracy and continuity of the monitoring scenario.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN XIANGYINHE SENSOR TECH CO LTD
- Filing Date
- 2026-05-25
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies, under conditions of limited computing resources and unstable communication links, struggle to build collaborative processing architectures with environmental adaptive adjustment capabilities and edge-side autonomy. This results in decreased data processing efficiency and a high risk of false alarms in anomaly identification, making it impossible to achieve high-confidence deep data analysis and processing in complex environments.
By simultaneously acquiring visual images and multi-source sensor data, spatial geometric features and physical quantity evolution trends are extracted, cross-verification deviation values are calculated, multi-source feature fusion vectors are generated, and recurrent neural networks are used to analyze historical evolution patterns. Dynamic comparison thresholds and adaptive offset compensation mechanisms are constructed, offline processing is initiated when the communication link is interrupted, and the sampling period is adjusted to achieve low-power operation.
Under conditions of limited resources and fluctuating operating conditions, it has achieved an intelligent closed loop for high-confidence data parsing and processing, ensuring that the accuracy of monitoring scene perception is no less than 98% and the early warning latency is less than 1000ms. It has environmental adaptive adjustment capabilities and edge-side autonomous capabilities.
Smart Images
Figure CN122241623A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electronic digital data processing technology, and in particular relates to a multi-source data collaborative processing method that integrates artificial intelligence. Background Technology
[0002] Currently, the collaborative processing of multi-source sensor data and visual information has become a routine method for monitoring complex systems. By integrating multi-dimensional heterogeneous signals, the system can obtain richer state representations than a single observation path, providing basic data support for decision support. However, in practical application scenarios, due to the complexity of the monitoring environment and the limited hardware resources of edge terminals, existing processing models generally rely on stable network bandwidth and cloud computing power. When the terminal is in a state of unstable communication link or limited energy supply, the data processing efficiency of the system will decrease. At the same time, the random fluctuation of environmental background noise and the physical drift of the sensor itself are coupled, resulting in a high risk of false alarms when a single discrimination logic is used to handle abnormal state identification tasks. This loss of data confidence, which is accepted by default in order to maintain the simplicity of the processing architecture, gradually transforms into a hidden cost affecting the reliability of the system during the long-term monitoring of critical facilities.
[0003] To address the aforementioned contradictions, improving computing hardware performance or increasing redundant sensor deployment inevitably leads to increased power consumption, contradicting the original intention of low-power monitoring. Furthermore, simply stacking sensors not only increases system integration complexity but also makes spatiotemporal alignment and conflict resolution between heterogeneous signals difficult due to the accumulation of errors at each observation point. Traditional linear filtering or static threshold determination methods exhibit logical lag or determination failure when facing non-stationary monitoring trend characteristics, indicating that existing edge-side processing mechanisms have reached a technical bottleneck in balancing offline autonomy and high-precision identification. For example, Chinese invention patent application CN120628304A discloses an intelligent system based on multi-sensor data fusion. The target detection system utilizes an improved DS evidence theory to resolve cross-modal data conflicts, combined with LSTM network feature weighting using an attention mechanism. However, when applied to long-term monitoring of water conservancy projects or large structural components, this technology exhibits shortcomings: the logical framework heavily relies on the statistical characteristics of the sampled data, lacks a mechanism for mutual verification between visual spatial geometric features and the physical evolution trend of the sensors, and in extreme field conditions, when environmental background noise undergoes nonlinear drift or communication links are intermittently interrupted, it lacks edge-side autonomous decision-making capabilities and dynamic closed-loop adjustment logic for risk and energy consumption, which can easily lead to false triggering or data chain breakage. Furthermore, the existing technology infrastructure pre-sets stable computing power support and continuous link connection, which leads to technical constraints and bottlenecks in offline autonomous scenarios with limited resources and fluctuating operating conditions.
[0004] Therefore, under the dual constraints of limited computing resources and unstable communication links, how to construct a collaborative processing architecture with environmental adaptive adjustment capabilities and edge-side autonomy capabilities to achieve a high-confidence data deep analysis and intelligent closed-loop process has become the technical problem to be solved by this invention. Summary of the Invention
[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A multi-source data collaborative processing method integrating artificial intelligence, comprising the following steps:
[0006] Step S1: Simultaneously acquire visual images of the monitored object and time-series sampling data from multiple sensor sources;
[0007] Step S2: Extract spatial geometric features from the visual image and identify the evolution trend of physical quantities in the time series sampling data of multi-source sensors. Calculate the mutual verification deviation value between the spatial geometric features and the evolution trend of physical quantities in the time domain. Correct the observation error of a single acquisition channel based on the mutual verification deviation value and generate a multi-source feature fusion vector.
[0008] Step S3: Map the multi-source feature fusion vector to the preset time series prediction model, analyze the historical evolution of the multi-source feature fusion vector, and output the displacement deformation trend characteristics of the monitored object;
[0009] Step S4: Obtain background bias data of the monitoring environment within the historical period, and perform adaptive offset compensation on the initial judgment benchmark based on the background bias data to construct a dynamic comparison threshold.
[0010] Step S5: Compare the displacement deformation trend characteristics with the dynamic comparison threshold. When the quantitative difference value of the displacement deformation trend characteristics deviating from the dynamic comparison threshold exceeds the safe range, it is determined that the monitored object has entered an abnormal state and an abnormal warning signal is generated.
[0011] Step S6: In the event of a communication link interruption, start the offline processing mode, complete the abnormal state identification logic locally, and store the generated offline diagnostic report.
[0012] Step S7: Identify the risk level of the abnormal state and obtain the current remaining power of the monitoring terminal. Adjust the data sampling period according to the constraint relationship between the risk level and the current remaining power to enable the monitoring terminal to enter a low-power working mode.
[0013] Preferably, step S2 specifically includes the following sub-steps: Step S21, segmenting the visual image to locate the region of interest where the monitored object is located, and calculating the coordinate displacement vector of the feature points within the region of interest; Step S22, matching the coordinate displacement vector with the physical displacement data in the time series sampling data of the multi-source sensors to determine the logical consistency between the coordinate displacement vector and the physical displacement data within the same sampling period; Step S23, when the numerical deviation between the coordinate displacement vector and the physical displacement data is greater than a preset consistency threshold, compensating the physical displacement data according to the confidence weight of the visual image to eliminate the random fluctuation component in the physical displacement data and generate a multi-source feature fusion vector.
[0014] Preferably, in step S3, the time series prediction model adopts a recurrent neural network structure with gating units to capture long-term dependencies in the multi-source feature fusion vector and output displacement deformation trend features that characterize the displacement change rate and deformation acceleration of the monitored object.
[0015] Preferably, step S4 specifically includes the following sub-steps: step S41, statistically analyze the mean and variance distribution of background parameters of the monitoring environment within the historical window; step S42, calculate the temporal drift pattern of the mean of background parameters to determine the background noise benchmark of the monitoring environment; step S43, superimpose the background noise benchmark onto the initial judgment benchmark so that the dynamic comparison threshold shifts synchronously with the background fluctuations of the controlled environment.
[0016] Preferably, in step S5, the quantitative difference value is obtained by calculating the Euclidean distance of the displacement deformation trend feature from the dynamic comparison threshold in the feature space, and the Euclidean distance is mapped to the preset risk probability distribution to determine the risk level of the monitored object.
[0017] Preferably, step S6 specifically includes the following sub-steps: step S61, performing time-series encoding on the generated offline diagnostic report according to the generation time, and storing the encoded offline diagnostic report in local non-volatile memory; step S62, after the communication link is restored, verifying the consistency of the data sequence between the monitoring terminal and the central node according to the timestamp, and retransmitting the stored offline diagnostic report in order of priority from high to low.
[0018] Preferably, step S7 specifically includes the following sub-steps: Step S71, when the risk level is lower than the preset warning threshold and the current remaining power is lower than the preset power threshold, increase the sampling interval of the multi-source sensor time series sampling data and reduce the sampling resolution of the visual image; Step S72, when the risk level exceeds the warning threshold, force a return to high-frequency sampling mode and start the real-time diagnostic task on the edge side.
[0019] Preferably, when generating an offline diagnostic report, the monitoring terminal uses a lightweight digest algorithm to extract key feature information from the offline diagnostic report and encrypts the data to ensure the integrity of the offline diagnostic report during offline storage and retransmission.
[0020] Preferably, the method is applied to water level monitoring or structural deformation monitoring scenarios in controlled physical environments. By logically coupling the spatial geometric features in visual images with the physical displacement data in the time series sampling data of multi-source sensors, an intelligent closed loop with a monitoring scenario perception accuracy of not less than 98% and an early warning delay of less than 1000ms is achieved.
[0021] Compared with existing technologies, the multi-source data collaborative processing method of this invention, which integrates artificial intelligence, has the following advantages:
[0022] 1. In the collaborative processing of multi-source data in artificial intelligence, a two-way verification mechanism between heterogeneous data streams is constructed by analyzing the correlation between visual image information and multi-source sensor data in the time dimension. The spatial distribution characteristics in the image and the evolution trend of physical parameters collected by the sensor are used for logical mutual verification to eliminate signal noise and random drift caused by a single detection path under specific environmental interference, thereby improving the confidence of the original input data under complex working conditions.
[0023] 2. By utilizing historical characteristic parameters of the monitored environment, an adaptive adjustment logic for the comparison threshold is established, enabling the system's judgment benchmark to be dynamically updated as the background fluctuation characteristics shift. This removes the negative interference of background noise on target state identification from a mechanism perspective, avoiding logical mis-triggering caused by the inability of fixed thresholds to adapt to dynamic working conditions.
[0024] 3. Adopting an edge-side autonomous analysis and local storage collaborative processing architecture, the diagnostic logic is maintained to run independently when the network communication link is interrupted. The integrity of the monitoring data is protected by the generated offline reports. Combined with the retransmission mechanism after the link is restored, the timing alignment of the data link is achieved, ensuring the continuity and reliability of the entire data processing process under extreme conditions. Attached Figure Description
[0025] Figure 1 This is the overall logical flowchart of the multi-source data collaborative processing method of the present invention;
[0026] Figure 2 This is a logical framework diagram of the edge-side autonomy and energy efficiency balance scheduling of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0028] A multi-source data collaborative processing method integrating artificial intelligence includes the following steps:
[0029] Step S1: Simultaneously acquire visual images of the monitored object and time-series sampling data from multiple sensor sources;
[0030] Step S2: Extract spatial geometric features from the visual image and identify the evolution trend of physical quantities in the time series sampling data of multi-source sensors. Calculate the mutual verification deviation value between the spatial geometric features and the evolution trend of physical quantities in the time domain. Correct the observation error of a single acquisition channel based on the mutual verification deviation value and generate a multi-source feature fusion vector.
[0031] Step S3: Map the multi-source feature fusion vector to the preset time series prediction model, analyze the historical evolution of the multi-source feature fusion vector, and output the displacement deformation trend characteristics of the monitored object;
[0032] Step S4: Obtain background bias data of the monitoring environment within the historical period, and perform adaptive offset compensation on the initial judgment benchmark based on the background bias data to construct a dynamic comparison threshold.
[0033] Step S5: Compare the displacement deformation trend characteristics with the dynamic comparison threshold. When the quantitative difference value of the displacement deformation trend characteristics deviating from the dynamic comparison threshold exceeds the safe range, it is determined that the monitored object has entered an abnormal state and an abnormal warning signal is generated.
[0034] Step S6: In the event of a communication link interruption, start the offline processing mode, complete the abnormal state identification logic locally, and store the generated offline diagnostic report.
[0035] Step S7: Identify the risk level of the abnormal state and obtain the current remaining power of the monitoring terminal. Adjust the data sampling period according to the constraint relationship between the risk level and the current remaining power to enable the monitoring terminal to enter a low-power working mode.
[0036] Preferably, step S2 specifically includes the following sub-steps: Step S21, segmenting the visual image to locate the region of interest where the monitored object is located, and calculating the coordinate displacement vector of the feature points within the region of interest; Step S22, matching the coordinate displacement vector with the physical displacement data in the time series sampling data of the multi-source sensors to determine the logical consistency between the coordinate displacement vector and the physical displacement data within the same sampling period; Step S23, when the numerical deviation between the coordinate displacement vector and the physical displacement data is greater than a preset consistency threshold, compensating the physical displacement data according to the confidence weight of the visual image to eliminate the random fluctuation component in the physical displacement data and generate a multi-source feature fusion vector.
[0037] Preferably, in step S3, the time series prediction model adopts a recurrent neural network structure with gating units to capture long-term dependencies in the multi-source feature fusion vector and output displacement deformation trend features that characterize the displacement change rate and deformation acceleration of the monitored object.
[0038] Preferably, step S4 specifically includes the following sub-steps: step S41, statistically analyze the mean and variance distribution of background parameters of the monitoring environment within the historical window; step S42, calculate the temporal drift pattern of the mean of background parameters to determine the background noise benchmark of the monitoring environment; step S43, superimpose the background noise benchmark onto the initial judgment benchmark so that the dynamic comparison threshold shifts synchronously with the background fluctuations of the controlled environment.
[0039] Preferably, in step S5, the quantitative difference value is obtained by calculating the Euclidean distance of the displacement deformation trend feature from the dynamic comparison threshold in the feature space, and the Euclidean distance is mapped to the preset risk probability distribution to determine the risk level of the monitored object.
[0040] Preferably, step S6 specifically includes the following sub-steps: step S61, performing time-series encoding on the generated offline diagnostic report according to the generation time, and storing the encoded offline diagnostic report in local non-volatile memory; step S62, after the communication link is restored, verifying the consistency of the data sequence between the monitoring terminal and the central node according to the timestamp, and retransmitting the stored offline diagnostic report in order of priority from high to low.
[0041] Preferably, step S7 specifically includes the following sub-steps: Step S71, when the risk level is lower than the preset warning threshold and the current remaining power is lower than the preset power threshold, increase the sampling interval of the multi-source sensor time series sampling data and reduce the sampling resolution of the visual image; Step S72, when the risk level exceeds the warning threshold, force a return to high-frequency sampling mode and start the real-time diagnostic task on the edge side.
[0042] Preferably, when generating an offline diagnostic report, the monitoring terminal uses a lightweight digest algorithm to extract key feature information from the offline diagnostic report and encrypts the data to ensure the integrity of the offline diagnostic report during offline storage and retransmission.
[0043] Preferably, the method is applied to water level monitoring or structural deformation monitoring scenarios in controlled physical environments. By logically coupling the spatial geometric features in visual images with the physical displacement data in the time series sampling data of multi-source sensors, an intelligent closed loop with a monitoring scenario perception accuracy of not less than 98% and an early warning delay of less than 1000ms is achieved.
[0044] Example 1: In the application of water level monitoring and dam deformation early warning in large-scale water conservancy reservoir areas, when the environment faces fluctuations in reservoir water level and communication links, the method provided by this invention collects visual images of the dam structure and time-series sampling data from multiple sources of sensors at the monitoring terminal. It uses image processing operators to extract spatial geometric features representing dam displacement from the visual images and identifies the evolution trend of physical quantities reflecting the dam's compressive state in the time-series sampling data from multiple sources of sensors. The system calculates the cross-verification deviation value between the spatial geometric features and the evolution trend of physical quantities within the same time-domain window, and performs compensation and correction on the observation error generated by a single acquisition channel based on the confidence weight determined by the cross-verification deviation value, generating a multi-source feature fusion vector reflecting the dam's deformation state.
[0045] The multi-source feature fusion vector is input into a recurrent neural network structure with gated units. By analyzing the evolution of the multi-source feature fusion vector within historical periods, the system calculates and outputs displacement deformation trend features representing the displacement change rate and deformation acceleration of the monitored object. During this process, the system collects the mean and variance distribution of background parameters of the monitoring environment within a historical window to obtain background bias data. Adaptive offset compensation is performed on the initial judgment benchmark using the background bias data, and a dynamic comparison threshold that shifts with the background noise of the environment is constructed. When the calculated displacement deformation trend feature deviates from the Euclidean distance of the dynamic comparison threshold in the feature space by more than a safe range, the system determines that the dam has entered an abnormal deformation state and generates an abnormal warning signal. By using environmental adaptive logic to remove background noise interference, the recognition accuracy is maintained above 98%, and the abnormal warning delay is controlled within 1000ms. In the process of issuing early warning signals and subsequent risk classification, the system uses a Poisson distribution model deployed in local memory to visualize the preset risk probability distribution logic. The processor inputs the calculated Euclidean distance quantization value into the Poisson distribution model, performs smooth integral calculation along its probability density function, and outputs the structural failure confidence probability in the range of 0 to 100%. The system further defines response rules based on preset thresholds: if the confidence probability is between 0 and 30%, it is output as a low-risk level; between 30% and 70%, it is output as a medium-risk level; and above 70%, it is marked as a high-risk level with the potential for immediate disconnection. Thus, through strict integral calculation and calibration mapping path, the underlying transmission logic from Euclidean distance to the visualized risk classification is established. To achieve the quantification of this risk probability, the system uses abnormal deformation samples from historical operating conditions to calibrate the Poisson distribution parameters and set the expected value of the Poisson distribution. It equals the historical average of displacement change within a unit sampling period; when the Euclidean distance calculation result is input into the model, the confidence probability of structural failure is obtained by calculating the cumulative distribution value of the probability density function within a specific interval; this parameter During the first stable operation monitoring cycle after equipment installation, displacement noise samples under no-load conditions are automatically calculated and written into non-volatile memory as a benchmark for subsequent risk assessment.
[0046] When the wireless communication link is interrupted due to lightning strikes or rainfall at the monitoring site, the monitoring terminal activates the edge autonomous mode according to the preset offline processing logic. It independently performs the identification of abnormal states locally and performs time-series encoding and storage of the generated offline diagnostic reports. At the same time, the system identifies the risk level of the abnormal state in real time and obtains the current remaining power of the terminal battery. When the risk level is found to be lower than the warning threshold and the current remaining power is insufficient, the terminal enters a low-power working mode by increasing the sampling interval and reducing the image resolution. When the risk level is detected to exceed the warning threshold, the high-frequency sampling mode is restored to ensure the continuity of the diagnostic task. After the communication link is restored, the terminal verifies the consistency of the data sequence with the central node according to the timestamp and completes the retransmission of the offline diagnostic reports according to the priority from high to low. Thus, the integrity of the dam health monitoring process is maintained under the constraints of limited resources and fluctuating operating conditions.
[0047] Example 2: In the physical verification test of simulating the monitoring of minute deformations in a dam body, the system faces technical bottlenecks such as zero-point drift of sensor signals and loss of visual features caused by environmental disturbances. The test platform is built on a simulated dam model with a multi-stage precision stepping thrust mechanism. Data sources include sequential images acquired by an industrial camera with a resolution of no less than 3840×2160 and a 12-channel pressure sensor array with an accuracy better than 0.05%FS arranged on the pressure-bearing surface of the dam foundation. The sampling frequency of the pressure sensor array is set to 100Hz to capture the transient response characteristics of the dam body under pulsed dynamic loads. The sampling period for the core parameters is specified. The system is configured to balance the real-time performance of data acquisition with the energy efficiency of edge computing power at the terminal, taking into account the background vibration spectrum bandwidth in the monitoring environment. When the signal frequency is above 20Hz, in order to satisfy the sampling theorem and suppress signal spectral aliasing, the sampling frequency... It tends towards the upper limit of the value range, i.e., the sampling period. Set to 10ms.
[0048] During the experiment, the system acquired the original stress waveform containing Gaussian white noise with a signal-to-noise ratio of 20dB and a low-contrast visual image after simulated rain occlusion. The system then activated a multi-source collaborative mode to extract spatial geometric feature values from the visual image. The value is 0.52 mm, and the physical displacement value fed back by the pressure sensor array is identified. The value is 0.45mm; as an intermediate calculation step for transparency, the system calculates the mutual verification deviation value. It fluctuates at 0.07 mm, and is based on the confidence weight of visual features. For physical displacement values Superimposed compensation logic; where This represents the mutual verification deviation value between spatial geometric characteristics and the evolution trend of physical quantities in the time domain. These are the spatial geometric feature values in a visual image. This is the physical displacement value fed back by the pressure sensor array. As the confidence weight for visual features, the standard deviation of the multi-source feature fusion vector in the temporal domain evolution is reduced from 0.12 in the original data to 0.03. Mapping this vector to a recurrent neural network structure with gating units, the output displacement deformation trend feature exhibits a smooth evolution in the feature space relative to the Euclidean distance of the dynamic comparison threshold. Before performing the aforementioned Euclidean distance quantization comparison, to bridge the dimensionality difference between the kinematic tensor feature of displacement deformation trend and the static scalar threshold, the system pre-processes the baseline scalar threshold based on the nonlinear activation function extension layer within the model. Orthogonal dimensionality reduction is achieved by extracting the preset time constants of dam stiffness and damping coefficient, smoothly mapping a single dynamic comparison threshold to the origin coordinates of a three-dimensional reference space containing displacement tolerance axis, velocity critical axis, and acceleration upper limit axis, and projecting the displacement deformation trend feature vector output by the model into this three-dimensional reference space according to the same base scale. By calculating the standard straight-line distance from the dynamic trend coordinate point to the reference origin in the aligned unified multidimensional orthogonal space, the substantial dimensionality reduction of Euclidean distance is obtained, and the final recognition accuracy is 98.7% with an early warning response time of 820ms.
[0049] To verify the regularity of the technical effect, a multi-dimensional control system was set up in the experiment. In the control group where the visual cross-validation step was removed, the false alarm rate of deformation recognition increased from 0.3% to 4.8% due to the influence of simulated temperature drift interference. In the gradient validation, as the power of the actively introduced environmental background noise increased from 10% to 50%, the displacement deformation trend characteristics output by the sample group of the present invention remained anchored within the judgment envelope, showing strong stability. When the sampling period of the key parameter was further adjusted... When the sampling frequency exceeds 100ms, the feature extraction process suffers performance degradation due to the sampling frequency being less than twice the signal frequency. The recognition accuracy drops below 85%, and the evolution details of minute displacements cannot be restored. The experimental results confirm that the parameter range defined in this invention is the preferred range for achieving high-precision monitoring. Through deep fusion of multi-dimensional heterogeneous data and prediction of time series trends, the system achieves a closed loop from raw data acquisition to intelligent decision feedback.
[0050] Example 3: This example combines Figures 1 to 2 This section explains the multi-source data collaborative processing method that integrates artificial intelligence, such as... Figure 1As shown, the method executes step S1 to acquire visual images of the object and time-series sampling data from multiple sources of sensors. Then, step S2 calculates the feature cross-verification deviation to correct the error and generate a fusion vector. Step S3 maps the fusion vector to the displacement deformation trend output by the prediction model. In step S4, a dynamic threshold is constructed based on the background bias data to compensate the initial benchmark. Then, step S5 compares the trend features with the dynamic threshold to determine the abnormal state and issue an early warning. In the event of a communication link interruption, step S6 is triggered to start offline processing to complete the abnormal identification and store the report. Finally, step S7 adjusts the sampling period based on the risk and remaining power to enter a low-power working mode.
[0051] like Figure 2 As shown, the system initially operates in a synchronous acquisition state and is associated with an initial acquisition benchmark. After extracting spatial geometric features and the evolution trend of physical quantities, it transitions to a feature mutual verification and correction state to perform heterogeneous data stream verification. Under the logical path of correcting single-channel observation errors and generating fusion vectors, it enters a trend prediction operation state to implement time series model mapping. When the quantization difference value exceeds the safe range, it transitions to an anomaly warning state to present features that deviate from the dynamic comparison threshold. In addition, if the system is in a synchronous acquisition state and meets the judgment conditions of low risk level and remaining power below the threshold, it switches to a low-power operating mode for energy efficiency balancing scheduling. If, in the low-power operating mode, the risk level exceeds the warning threshold and high-frequency sampling is forcibly restored, the system transitions from the low-power operating mode to an edge autonomous offline state to perform link interruption / offline storage. At the same time, if the system detects a communication link interruption in the synchronous acquisition state, it directly enters the edge autonomous offline state. Finally, under the feedback conditions of link recovery and completion of offline diagnostic report retransmission, the system returns from the edge autonomous offline state to the initial synchronous acquisition state.
[0052] Example 4: In a reservoir dam monitoring environment with intermittent water mist interference, the dam structure is obscured by water vapor, leading to a decrease in visual feature contrast. Furthermore, the sensor signals inside the dam body experience zero-point drift due to temperature and humidity cycling interference. The system utilizes a processing platform with a computing power of at least 4 TOPS to perform convolution processing on pixel units in the visual image. It calculates the brightness difference between the horizontal and vertical directions to determine the edge intensity distribution and uses coordinate displacement vector calculation commands to lock feature markers on the dam surface. The region of interest containing these feature markers is defined as containing… The feature matrix of pixel units; where The number of pixels on one side of the region of interest is a positive integer.
[0053] The system establishes confidence weights in the local cache. A real-time update queue for analyzing visual images in continuous... Image information entropy within each frame period, and noise variance under environmental interference are calculated. and according to the function Determine the confidence level of visual features, and then calculate the confidence level weights. As a multiplication operator acting on spatial geometric eigenvalues At the same time, the mutual verification deviation value will be... The reciprocal of the value is used as a scaling factor for the physical displacement value. The sensitivity coefficient k was determined by fitting historical experimental data and set to 0.42 in this scenario to ensure that the confidence weights remain constant even as the variance of visual noise increases. It decreases linearly; parameters The value is 0.01, used to avoid denominator overflow and weight zeroing when there is no significant noise interference in the environment; this weight coefficient is pre-written into the processing platform during the system initialization phase by performing regression analysis on the entropy values and recognition accuracy of 500 known occluded samples to ensure statistically significant reliability under complex working conditions. In the arithmetic logic unit, a weighted summation is performed to generate a multi-source feature fusion vector reflecting the deformation potential energy of the dam structure. This vector is used as an input tensor into a recurrent neural network with gated units. The hidden layer extracts the first-order derivative features reflecting deformation displacement acceleration through the state transition equation. To eliminate the influence of dimensional mismatch in heterogeneous data, normalization collaborative preprocessing is performed before entering the recurrent neural network. The pixel coordinate displacement is converted into millimeter-level physical quantities through spatial geometric feature extraction operators, and the high-frequency sensor sampling values are resampled to be synchronized with the visual frame rate using a time-aligned interpolation algorithm for sensor data, so that visual features and pressure features form a temporally aligned two-dimensional tensor matrix under a unified physical time scale. represents the confidence weight of visual features, which is a dimensionless parameter; This is the preset sensitivity coefficient; The noise variance of the image signal; To prevent compensation values where the denominator is zero; These are spatial geometric eigenvalues. This represents the mutual verification deviation value between spatial geometric characteristics and the evolution trend of physical quantities in the time domain. This is the physical displacement value. The number of frame periods is a positive integer. Based on the optimal estimation theory of multi-source sensor data fusion, the lower the signal-to-noise ratio of the measurement channel, the smaller the confidence weight assigned to a specific channel. The system establishes a real-time update queue of confidence weights in a local cache to extract visual images in continuous... Image information entropy quantization calculation within each frame period, noise variance under environmental interference. The controller calls the inverse proportional compensation logic to calculate the confidence weight of visual features. Calculate the confidence weight of physical displacement Among them, the introduction The number of frame periods is defined as a positive integer; parameter The preset sensitivity coefficient is set to be a dimensionless constant greater than zero; variable The variance of the noise extracted from the image signal in the representation operation is taken as a value greater than zero; constant. As a compensation value to prevent calculation overflow caused by a denominator of zero; variable The deviation between the corresponding spatial geometric characteristics and the evolution trend of physical quantities in the time domain is verified by the international standard unit of millimeters; constant. The representative system establishes a tolerance upper limit for deviation in millimeters through calibration experiments. The system then calculates and outputs the visual channel feature correction amount. and sensor channel feature correction amount Input baseline Represents spatial geometric eigenvalues. Represents the physical displacement value; the processor reads the feature correction amount. and A fixed-length two-dimensional structured feature matrix is constructed sequentially and used as the multi-source feature fusion vector. use In the data format, within the operational logic of the feature correction operator for specific applications, to implement the compensation benchmark positively correlated with the signal-to-noise ratio, the confidence weight in the multiplication logic is... The actual representation is a normalized penalty denominator, and the system's built-in inverse proportional regulator is used in execution. Before calculating the features, the output of the original formula is pre-transformed with its reciprocal, meaning the actual effective correction matrix elements are... This mechanism ensures that the environmental noise variance quantized from image information entropy is... During a surge, the relay base As the multiplier increases, the effective multiplier weights that ultimately map to the geometric features of that channel decrease, thus correcting the unbalanced drift of single-channel eigenvalues caused by environmental noise. This follows the physical closed-loop criterion for suppressing high-noise channels in the optimal estimation theory of multi-source sensors, resulting in a multi-source feature fusion vector. As an input tensor, it enters the hidden layer of a recurrent neural network with gated units, and the first derivative features reflecting deformation displacement acceleration are extracted through the state transition equation.
[0054] The system identifies the risk level of abnormal states and obtains the current remaining battery power of the monitoring terminal. When the risk level of abnormal states exceeds a preset threshold and the current remaining battery power of the terminal drops below 20%, the system activates an energy-saving management mode based on sampling rate attenuation, adjusting the data sampling period. Based on the relation The system is adjusted to a low-frequency state, reducing processor power consumption by lowering the data bus wake-up frequency. When the dam displacement change rate exceeds the threshold of 2.5 mm per hour, the system interrupts the low-power operation mode and resumes 100Hz data sampling. This ensures the generation of offline diagnostic reports with time-consistent output even under disconnection conditions. After the wireless communication link is restored, a timestamp alignment algorithm is used to retransmit historical data, keeping the false alarm rate caused by environmental noise below 0.4%. The adjusted data sampling period, The initial data sampling period, The compensation coefficient is determined based on the current remaining battery power. This compensation coefficient is used to ensure a deterministic scheduling balance between terminal battery life and core perception latency during the edge autonomy phase. A piecewise linear mapping mechanism is used for dynamic calculation and assignment. Specifically, in the microcontroller's execution logic, when the current remaining power is... When the value is within a continuous range of 10% to 20%, the system is based on the expression. Calculate this coefficient; when the hardware detects the remaining battery power. When the energy storage core is below the 10% critical warning line, in order to prevent over-discharge of the energy storage core and strive to maintain the safety warning link at the end, the system forcibly executes high-priority protection logic, adjusting the coefficient... The compensation coefficient is locked at a constant system experience upper limit value of 5. The piecewise linear mapping mechanism is adopted, and its specific assignment logic is as follows: when the remaining power is... When the value is in the 10% to 20% range, the coefficient ranges from 1 to 2; this mapping is derived based on the energy consumption model and aims to extend the data sampling period. This linearly extends the lifespan of the monitoring terminal while maintaining the baseline sampling power consumption until the remaining power drops to zero, ensuring that the minimum sampling density of key displacement changes is maintained per unit time even under extremely low power conditions. This parameter boundary calibration procedure ensures that the sampling frequency scheduling is not simply stuck on the intention of fuzzy positive correlation, but also obtains the guarantee of a quantitative mapping mechanism that can be directly called and executed at the software and hardware level.
[0055] Example 5: In a sensing system deployed in the pressure zone downstream of a concrete gravity dam, the system acquires the initial electrical signal of the pressure sensing array monitoring the dam body under zero hydrostatic pressure conditions during the initial operation phase, and records the initial pixel coordinates of feature markers in the visual image at this time. A regression operator is then used to adjust the voltage output of the 12 pressure sensors. With physical displacement value A linear relationship fitting function is used to determine the compensation constant reflecting the zero-point drift of the sensor. Meanwhile, for the visual acquisition channel, the actual length represented by a unit pixel is calculated using the geometric dimensions of the dam surface. This provides initial benchmark data with consistent dimensions for the cross-validation of the evolution trend of spatial geometric features and physical quantities. By removing the initial deviation of the hardware caused by installation stress or environmental factors, the physical parameters input into the multi-source feature fusion vector reflect the real displacement increment generated by the force on the dam structure.
[0056] Sensitivity coefficient for recurrent neural networks During calibration, the system uses a locally stored historical operating condition dataset as the input source. The historical operating condition dataset includes 500 labeled samples with known deformation potential energies and corresponding image noise distribution features. The noise variance is simulated in the arithmetic logic unit. The evolution process from 0.05 to 0.8 reflects the system monitoring confidence weight. The convergence state is calculated and the rate of change of recognition accuracy is determined when the recognition accuracy changes with respect to the parameter. When the derivative converges to a preset interval, the value at this point is determined as the sensitivity coefficient. The system obtains the historical displacement variance envelope of the dam structure under stable loads, and uses this as a benchmark to establish a linear coefficient for the dynamic comparison threshold shifting with environmental background noise. After parameter calibration, the weight allocation rules are stored in the non-volatile memory of the processing platform, enabling the system to automatically trigger error compensation logic and generate abnormal early warning signals based on real-time acquired physical signals during engineering applications. The sensitivity coefficient, The noise variance of the image signal; The confidence weights are for visual features.
[0057] Example 6: In the procedure for determining the operating parameters of the dam body feature marker identification operator, the system acquires the camera installation distance and the physical scale data of the monitored location to determine the number of pixels on one side of the region of interest. The coverage area of the region of interest is selected as three times the physical diameter of the feature marker points to reserve space for capturing the displacement vector. A 3×3 convolution operator is loaded into the image processing module as the image gradient extraction operator. This is achieved by using the horizontal convolution operator... The weight distribution is determined as follows The system convolves the convolution operator with the feature matrix within the region of interest to obtain the edge intensity distribution. It then searches for local maxima in the intensity response map to locate the coordinate centers of the feature markers. This parameter is used for calibration to eliminate gradient detection bias caused by uneven ambient lighting distribution. This represents the number of pixels on one side of the region of interest. This is the horizontal convolution operator.
[0058] When setting the procedure for retransmission scheduling of offline diagnostic reports in edge autonomous mode, the system loads risk-level-based data into the local storage space of the monitoring terminal. With timestamp delay The determination model, in which the transmission priority of each set of data sequences to be retransmitted is determined. Follows a linear decay function as defined below: The system will transmit the generated offline diagnostic reports according to transmission priority. They are arranged in descending order of priority in the queue, and their transmission priority is assigned after the wireless communication link is established. Higher deformation data reduces the sensing latency of high-risk signals under communication-constrained conditions and maintains the integrity of monitoring data; among which, For transmission priority, Risk level score To generate the time difference between the current time and the time point of the data generation. is the time decay constant.
[0059] Determine the adjusted data sampling period At that time, the system invokes the risk level in the energy-saving management mode. With current remaining power The coupling determination rule, when the remaining power Reduced to below 20% and risk level When the temperature falls below the warning threshold, the system performs downlink adjustment of the sampling frequency and adjusts the compensation coefficient. The incremental value was determined to be positively correlated with the degree of power deficit, and the data sampling period was increased. To reduce the frequency of monitoring; and once a risk level is identified... If the warning threshold is exceeded, the system will ignore the current remaining battery power. Constraints and forced compensation coefficients Reset to 1 to adjust the data sampling period. By restoring the baseline level to 10ms, this logic of determining the power constraint weight based on risk level scores ensures the synergy between the monitoring terminal's lifespan and the real-time capture of key deformation signals under energy-constrained conditions, thus achieving reliable monitoring of the dam's safety status; among which, The adjusted data sampling period, Risk level score This represents the current remaining battery percentage. The compensation coefficient is a dimensionless parameter. This is the initial data sampling period.
[0060] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.
Claims
1. A multi-source data collaborative processing method integrating artificial intelligence, characterized in that, Includes the following steps: Step S1: Simultaneously acquire visual images of the monitored object and time-series sampling data from multiple sensor sources; Step S2: Extract spatial geometric features from the visual image and identify the evolution trend of physical quantities in the time series sampling data of multi-source sensors. Calculate the mutual verification deviation value between the spatial geometric features and the evolution trend of physical quantities in the time domain. Correct the observation error of a single acquisition channel based on the mutual verification deviation value and generate a multi-source feature fusion vector. Step S3: Map the multi-source feature fusion vector to a preset time series prediction model, analyze the historical evolution of the multi-source feature fusion vector, and output the displacement deformation trend features of the monitored object. The time series prediction model adopts a long short-term memory network architecture, containing two hidden layers, each with 64 neurons. The model input receives a fixed-length feature matrix of dimension 2, and dynamically controls the retention and updating of historical feature information through forget gates and input gates. During training, mean squared error is used as the loss function, and the model weights are iteratively updated using the Adam optimization algorithm, with a learning rate of 0.
001. The model output is connected to a fully connected layer, using the Softmax function as the activation mechanism, and outputs the displacement deformation trend features. Step S4: Obtain background bias data of the monitoring environment within the historical period, and perform adaptive offset compensation on the initial judgment benchmark based on the background bias data to construct a dynamic comparison threshold. Step S5: Compare the displacement deformation trend characteristics with the dynamic comparison threshold. When the quantitative difference value of the displacement deformation trend characteristics deviating from the dynamic comparison threshold exceeds the safe range, it is determined that the monitored object has entered an abnormal state and an abnormal warning signal is generated. Step S6: In the event of a communication link interruption, start the offline processing mode, complete the abnormal state identification logic locally, and store the generated offline diagnostic report. Step S7: Identify the risk level of the abnormal state and obtain the current remaining power of the monitoring terminal. Adjust the data sampling period according to the constraint relationship between the risk level and the current remaining power to enable the monitoring terminal to enter a low-power working mode.
2. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: Step S21, segment the visual image to locate the region of interest where the monitored object is located, and calculate the coordinate displacement vector of the feature points within the region of interest; Step S22: Match the coordinate displacement vector with the physical displacement data in the time series sampling data of the multi-source sensor to determine the logical consistency between the coordinate displacement vector and the physical displacement data in the same sampling period; Step S23: When the numerical deviation between the coordinate displacement vector and the physical displacement data is greater than the preset consistency threshold, compensate the physical displacement data according to the confidence weight of the visual image to eliminate the random fluctuation component in the physical displacement data and generate a multi-source feature fusion vector.
3. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, In step S3, the time series prediction model adopts a recurrent neural network structure with gated units to capture long-term dependencies in the multi-source feature fusion vector and output displacement deformation trend features that characterize the displacement change rate and deformation acceleration of the monitored object.
4. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: Step S41, statistically analyze the mean and variance distribution of background parameters of the monitoring environment within the historical window; Step S42, calculate the temporal drift pattern of the mean background parameters to determine the background noise benchmark of the monitoring environment; Step S43, superimpose the background noise benchmark onto the initial judgment benchmark so that the dynamic comparison threshold shifts synchronously with the background fluctuations of the controlled environment.
5. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, In step S5, the quantitative difference value is obtained by calculating the Euclidean distance of the displacement deformation trend feature from the dynamic comparison threshold in the feature space, and the Euclidean distance is mapped to the preset risk probability distribution to determine the risk level of the monitored object.
6. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, Step S6 specifically includes the following sub-steps: Step S61, the generated offline diagnostic report is time-coded according to the generation time, and the encoded offline diagnostic report is stored in local non-volatile memory; Step S62: After the communication link is restored, the consistency of the data sequence between the monitoring terminal and the central node is verified according to the timestamp, and the stored offline diagnostic report is retransmitted in order of priority from high to low.
7. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, Step S7 specifically includes the following sub-steps: Step S71, when the risk level is lower than the preset warning threshold and the current remaining power is lower than the preset power threshold, increase the sampling interval of the multi-source sensor time series sampling data and reduce the sampling resolution of the visual image. Step S72: When the risk level exceeds the warning threshold, force a return to high-frequency sampling mode and start real-time diagnostic tasks on the edge side.
8. The multi-source data collaborative processing method integrating artificial intelligence according to claim 6, characterized in that, When generating offline diagnostic reports, the monitoring terminal uses a lightweight digest algorithm to extract key feature information from the offline diagnostic reports and encrypts the data to ensure the integrity of the offline diagnostic reports during offline storage and retransmission.
9. The multi-source data collaborative processing method integrating artificial intelligence according to claim 1, characterized in that, The method is applied to water level monitoring or structural deformation monitoring scenarios in controlled physical environments. By logically coupling the spatial geometric features in visual images with the physical displacement data in the time series sampling data of multi-source sensors, an intelligent closed loop with a monitoring scenario perception accuracy of no less than 98% and an early warning delay of less than 1000ms is achieved.