Aluminum electrolysis cell aluminum tapping working condition and abnormality recognition system based on machine vision
By using fractal dimension filtering and spatial weight anchoring technology, the drift and occlusion problems of image recognition systems in high temperature, high magnetic field and high dust environment were solved, and stable identification and control of aluminum electrolysis cell output conditions and anomalies were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN LIDER INTELLIGENT TECH CO LTD
- Filing Date
- 2026-06-17
- Publication Date
- 2026-07-14
AI Technical Summary
In high-temperature, high-magnetic, and high-dust environments, existing technologies cannot effectively overcome the nonlinear grayscale space drift and geometric topological displacement of non-stationary features in image recognition systems, leading to the overall collapse of the control link and failing to meet the long-term anti-disturbance control requirements of unmanned operation and maintenance.
A fractal dimension filtering operator is used to separate the foreground target feature map flow and background noise components. The image is compensated by affine transformation in combination with a spatial weight anchoring unit. The effective pixel area is monitored by the working condition determination unit. The temporal linear prediction module is used to achieve smooth updates and output stable working condition decisions.
High-purity image components were extracted in a strong magnetic and dusty environment, avoiding control chain breakage, ensuring smooth updates of operating conditions and stable output of control commands, and possessing anti-interference capabilities under extreme blind zones.
Smart Images
Figure CN122391247A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial computer vision image recognition technology, and in particular relates to a machine vision-based aluminum electrolysis cell aluminum output condition and anomaly recognition system. Background Technology
[0002] Currently, in high-temperature, high-magnetic, and high-dust industrial video feature monitoring, the common image processing method uses grayscale threshold matrix segmentation and static coordinate cropping to extract the morphology of the liquid surface edge. This approach implicitly relies on the ideal premise that the video frame stream has a stable noise distribution and the camera is in an absolutely rigid spatial position. By setting a fixed empirical gate value to remove ripple interference, the geometric topology of the target edge can be obtained. However, when the above method is applied to an extreme and non-ideal field environment, a mismatch occurs between its established static design and the reality of high dynamic evolution. On the one hand, the nonlinear spatial frequency domain noise excited by the transient strong magnetic field jump in the workshop will cause random unstable drift of the image grayscale distribution, causing the conventional fixed gate value to frequently produce over-segmentation or semantic blocking defects of morphological omission. On the other hand, long-period thermal stress and physical impact potential will cause irregular spatial geometric micro-offsets in the physical position of the camera, causing the fixed absolute pixel coordinate cropping rule to produce irreversible linear drift and semantic distortion of the anchoring center of the fluid core area. In addition, the high concentration of dynamic solid dust routinely blocks the optical channel between consecutive frames, causing the independent visual recognition operator of a single frame to collapse as the data source features are instantly lost, leading to the overall collapse of the underlying control link.
[0003] The intuitive linear improvement solutions for the aforementioned interference in the industry typically reduce noise sensitivity by significantly increasing the order of the filtering smoothing operator or increasing the mechanical rigidity and vibration resistance of the camera frame. However, such solutions not only fail to cope with the dynamic drift of time-varying signal sources, but also cause feature edge blunting distortion and signal response delay due to excessively high filtering operator depth. These improvements mostly focus on enhancing the rigidity and vibration resistance of external mechanical structures. In complex field environments, relying solely on hardware improvements is costly. Correspondingly, the backend image processing methods also have shortcomings at the software control level. For example, the authorization announcement number CN107204004B... A Chinese invention patent discloses a method and system for dynamic feature recognition in aluminum electrolysis cell video. It detects and uses optical flow to track the position sequence of image feature corner points, and subtracts the background corner point displacement to offset camera shake. This approach implicitly relies on the presence of clear and long-term stable local feature points in the industrial site. In the extreme environment of aluminum extraction station, which is accompanied by strong magnetic nonlinear jumps and high concentrations of smoke and dust, local image features are instantly lost under the sudden obscuring of smoke and dust. This causes the optical flow method to fail due to feature matching breakage and the topological principle of feature flow to fail, which cannot meet the requirements of long-term unmanned operation and maintenance anti-disturbance control.
[0004] Therefore, the technical problem to be solved by this invention is how to overcome the nonlinear gray-scale space drift of non-stationary features and calibrate in situ the geometric topological displacement induced by long-period temperature rise stress, and maintain the self-consistent and smooth update of control flow semantics under the condition that the spatiotemporal evolution features of continuous image sequences are subjected to short-term blind zone conditions of sudden dust occlusion. Summary of the Invention
[0005] This invention proposes a machine vision-based system for identifying aluminum electrolysis cell output conditions and anomalies. The system includes:
[0006] The feature decoupling unit is used to acquire the aluminum field video frame stream and use the fractal dimension filtering operator to separate the foreground target feature map stream and background noise components from the video frame stream;
[0007] The spatial weight anchoring unit, connected to the feature decoupling unit, is used to deduce the inter-frame vibration displacement of the image based on the Euclidean distance of the feature point coordinates, and to perform translation and rotation transformation compensation on the image affine transformation matrix to construct a weighted mask matrix. The weighted mask matrix is then multiplied pixel by pixel with the foreground target feature map flow to truncate background interference and output fluid edge morphology features. The background noise component is used as the input reference for spatial dynamic compensation.
[0008] The working condition determination unit, connected to the spatial weight anchoring unit, is used to monitor the effective pixel area of the region enclosed by the fluid edge morphological features. When the effective pixel area is lower than the dynamic area threshold determined based on the average effective pixel area of the historical normal frame stream for three consecutive video frame periods, a complete occlusion determination is triggered. The virtual fluid morphological features are dynamically synthesized using the temporal linear prediction module in combination with the historical centroid displacement vector, thereby completing the smooth update of the working condition decision data.
[0009] Preferably, during operation, the feature decoupling unit uses a fractal dimension filtering operator to process the video frame stream into blocks, calculates the pixel grayscale information entropy and pixel-level fractal dimension of each block's image domain, and determines blocks whose pixel grayscale information entropy is lower than the dynamic entropy threshold determined based on the baseline grayscale information entropy obtained from the statistical analysis of smokeless background frame images, and whose pixel-level fractal dimension is in the range of 1.2 to 1.8, as foreground target feature map streams, and determines the remaining image components that do not belong to the foreground target feature map streams as background noise components.
[0010] Preferably, when the working condition determination unit is running, the temporal linear prediction module extracts the historical centroid coordinates and historical area of the five consecutive normal working cycles before the complete occlusion determination in each video frame cycle after the complete occlusion determination is triggered, calculates the centroid displacement change rate and area decay rate of adjacent normal working cycles, dynamically synthesizes the virtual centroid position and virtual effective area in the current fault cycle using a linear extrapolation model, and reassembles and generates virtual fluid morphology features.
[0011] Preferably, the working condition determination unit includes a decision data smoothing module; the decision data smoothing module is connected to the temporal linear prediction module and is used to input the virtual fluid morphology features into the state machine decision model during the complete occlusion determination period, so that the change amplitude of the working condition decision data output by the state machine decision model is lower than the dynamic fluctuation threshold determined based on the transient signal-to-noise ratio of the current video frame stream.
[0012] Preferably, the spatial weight anchoring unit includes a dynamic compensation module; the dynamic compensation module is connected to the feature decoupling unit and is used to obtain the global high-frequency fluctuation frequency and average amplitude of the background noise component, and dynamically adjust the center distance attenuation coefficient of the weighted mask matrix according to the global high-frequency fluctuation frequency and average amplitude.
[0013] Preferably, the operating condition determination unit, during operation, specifically includes the following steps: calculating the effective pixel area of the closed region enclosed by the fluid edge morphology features in real time; comparing the effective pixel area with the minimum area threshold of normal operation determined based on the statistical lower limit of the pixel area of historical normal operation cycles in real time; when the effective pixel area is lower than the minimum area threshold of normal operation for three consecutive video frame cycles, it is confirmed as a complete occlusion condition and a complete occlusion fault determination signal is output.
[0014] Preferably, the spatial weight anchoring unit includes the following steps during operation: based on the fixed reference coordinates of the aluminum outlet center and the real-time extracted fluid flow direction center coordinates, the inter-frame vibration displacement is deduced by the Euclidean distance of the feature point coordinates; the azimuth translation vector and rotation angle are calculated using the inter-frame vibration displacement, and the translation and rotation matrix transformation compensation is completed on the image affine transformation matrix.
[0015] Preferably, the feature decoupling unit includes a video access interface for acquiring video streams in real time from explosion-proof industrial cameras arranged at the aluminum electrolysis cell workshop's aluminum output workstations, and decoding the video streams into single-frame image sequences to be input as video frame streams into the fractal dimension filtering operator.
[0016] Preferably, the system includes a safety alarm module connected to the working condition determination unit, which outputs visual abnormality alarm status data when the duration of completely obscuring the fault determination signal exceeds the explicit safety time threshold of 20 seconds, so that the unmanned operation and maintenance system can complete the arbitration of the workshop safety status.
[0017] Compared with existing technologies, the machine vision-based aluminum electrolysis cell aluminum output condition and anomaly identification system of the present invention has the following advantages:
[0018] 1. In the aluminum electrolysis cell aluminum output condition and anomaly identification, the local gray-level histogram distribution in the original video frame data is extracted by the feature decoupling unit to calculate the local spatial information entropy of the target pixel neighborhood window. Combined with the dynamic gating rule that monotonically increases with noise energy, the fractal dimension boundary threshold is determined. In feature comparison, the fractal dimension boundary threshold and the dynamic fractal dimension of the target pixel are used to complete the rigid condition judgment, so that the fluid geometry data below the fractal dimension boundary threshold is oriented to the main component flow, and the random noise components above the fractal dimension boundary threshold are completely included in the residual flow, realizing the extraction of high-purity image components in a strong magnetic and high dust environment.
[0019] 2. By using spatial weighted anchoring units to retrieve components in the residual flow in real time, the static boundary contour of the inherent rigid mechanical components is extracted. The feature vector of the structural invariant descriptor is calculated by combining the zero-order invariant moment and the first-order central moment. The global logic rack bounce is derived from the Euclidean distance between the feature vector and the historical calibration reference vector to complete the translation and rotation matrix transformation compensation of the orientation matrix. The reconstructed center distance weighted mask matrix is then multiplied pixel by pixel with the principal component flow. Under the conditions of long-cycle equipment thermal deformation and mechanical misalignment, the pseudo-target background interference outside the core outflow area is truncated, and finally, the fluid edge morphology features without semantic deviation are output.
[0020] 3. The working condition is judged by real-time monitoring of the effective pixel area of the closed area enclosed by the fluid edge morphology features through the temporal flow compensation module in the state arbitration unit. When the effective pixel area is lower than the preset minimum area threshold for normal operation for three consecutive video frame cycles, the complete occlusion fault judgment is triggered. Combining the historical center displacement vector and historical area change rate accumulated in the historical normal operation cycle, the virtual fluid morphology features in the synthetic fault cycle are dynamically predicted by the forward temporal linear extrapolation operator. This enables the state arbitration unit to maintain the smooth update of control command data output based on the virtual fluid morphology features, and avoids the control chain breakage caused by dense smoke occlusion. Attached Figure Description
[0021] Figure 1 This is a data flow diagram of the machine vision-based aluminum electrolysis cell aluminum output condition and anomaly identification system of the present invention.
[0022] Figure 2 This is a module architecture diagram of the machine vision-based aluminum electrolysis cell aluminum output condition and anomaly recognition system of the present invention. Detailed Implementation
[0023] 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.
[0024] A machine vision-based system for identifying aluminum electrolysis cell output conditions and anomalies, the system comprising:
[0025] The feature decoupling unit is used to acquire the aluminum field video frame stream and use the fractal dimension filtering operator to separate the foreground target feature map stream and background noise components from the video frame stream;
[0026] The spatial weight anchoring unit, connected to the feature decoupling unit, is used to deduce the inter-frame vibration displacement of the image based on the Euclidean distance of the feature point coordinates, and to perform translation and rotation transformation compensation on the image affine transformation matrix to construct a weighted mask matrix. The weighted mask matrix is then multiplied pixel by pixel with the foreground target feature map flow to truncate background interference and output fluid edge morphology features. The background noise component is used as the input reference for spatial dynamic compensation.
[0027] The working condition determination unit, connected to the spatial weight anchoring unit, is used to monitor the effective pixel area of the region enclosed by the fluid edge morphological features. When the effective pixel area is lower than the dynamic area threshold determined based on the average effective pixel area of the historical normal frame stream for three consecutive video frame periods, a complete occlusion determination is triggered. The virtual fluid morphological features are dynamically synthesized using the temporal linear prediction module in combination with the historical centroid displacement vector, thereby completing the smooth update of the working condition decision data.
[0028] Preferably, during operation, the feature decoupling unit uses a fractal dimension filtering operator to process the video frame stream into blocks, calculates the pixel grayscale information entropy and pixel-level fractal dimension of each block's image domain, and determines blocks whose pixel grayscale information entropy is lower than the dynamic entropy threshold determined based on the baseline grayscale information entropy obtained from the statistical analysis of smokeless background frame images, and whose pixel-level fractal dimension is in the range of 1.2 to 1.8, as foreground target feature map streams, and determines the remaining image components that do not belong to the foreground target feature map streams as background noise components.
[0029] Preferably, when the working condition determination unit is running, the temporal linear prediction module extracts the historical centroid coordinates and historical area of the five consecutive normal working cycles before the complete occlusion determination in each video frame cycle after the complete occlusion determination is triggered, calculates the centroid displacement change rate and area decay rate of adjacent normal working cycles, dynamically synthesizes the virtual centroid position and virtual effective area in the current fault cycle using a linear extrapolation model, and reassembles and generates virtual fluid morphology features.
[0030] Preferably, the working condition determination unit includes a decision data smoothing module; the decision data smoothing module is connected to the temporal linear prediction module and is used to input the virtual fluid morphology features into the state machine decision model during the complete occlusion determination period, so that the change amplitude of the working condition decision data output by the state machine decision model is lower than the dynamic fluctuation threshold determined based on the transient signal-to-noise ratio of the current video frame stream.
[0031] Preferably, the spatial weight anchoring unit includes a dynamic compensation module; the dynamic compensation module is connected to the feature decoupling unit and is used to obtain the global high-frequency fluctuation frequency and average amplitude of the background noise component, and dynamically adjust the center distance attenuation coefficient of the weighted mask matrix according to the global high-frequency fluctuation frequency and average amplitude.
[0032] Preferably, the operating condition determination unit, during operation, specifically includes the following steps: calculating the effective pixel area of the closed region enclosed by the fluid edge morphology features in real time; comparing the effective pixel area with the minimum area threshold of normal operation determined based on the statistical lower limit of the pixel area of historical normal operation cycles in real time; when the effective pixel area is lower than the minimum area threshold of normal operation for three consecutive video frame cycles, it is confirmed as a complete occlusion condition and a complete occlusion fault determination signal is output.
[0033] Preferably, the spatial weight anchoring unit includes the following steps during operation: based on the fixed reference coordinates of the aluminum outlet center and the real-time extracted fluid flow direction center coordinates, the inter-frame vibration displacement is deduced by the Euclidean distance of the feature point coordinates; the azimuth translation vector and rotation angle are calculated using the inter-frame vibration displacement, and the translation and rotation matrix transformation compensation is completed on the image affine transformation matrix.
[0034] Preferably, the feature decoupling unit includes a video access interface for acquiring video streams in real time from explosion-proof industrial cameras arranged at the aluminum electrolysis cell workshop's aluminum output workstations, and decoding the video streams into single-frame image sequences to be input as video frame streams into the fractal dimension filtering operator.
[0035] Preferably, the system includes a safety alarm module connected to the working condition determination unit, which outputs visual abnormality alarm status data when the duration of completely obscuring the fault determination signal exceeds the explicit safety time threshold of 20 seconds, so that the unmanned operation and maintenance system can complete the arbitration of the workshop safety status.
[0036] Example 1: In the scenario of monitoring the aluminum electrolysis cell's aluminum output operation, the system faces challenges such as high-frequency jitter in the image signal caused by strong magnetic fields and visual obstruction caused by dynamic smoke and dust. (Original video frame data...) Data is collected by an industrial camera deployed 4 meters from the aluminum outlet and transmitted to the feature decoupling unit at a sampling frequency of 25Hz. The system sets the local spatial information entropy. Computational neighborhood window For 5×5 pixels, the feature decoupling unit performs real-time statistics. The local spatial information entropy is calculated from the local gray-level histogram within the image. Based on preset dynamic gating rules, the system calculates the fractal dimension boundary threshold. The calculation formula is as follows: ,in, This is the grayscale adaptive calibration constant, with a value of 0.5. This is the grayscale shift correction factor, with a value of 0.2. The scaling factor is 0.8, and the fractal dimension threshold is... The procedure is dynamically determined in relation to the ambient background noise level as follows: Define an initial baseline state. Under steady-state conditions with no smoke or dust in the aluminum electrolysis cell and no aluminum extraction operations, 100 consecutive video frames are collected by an explosion-proof industrial camera as the baseline sequence; calculate the global grayscale information entropy of each video frame in the baseline sequence to obtain the mean baseline entropy. Standard deviation of baseline entropy In real-time monitoring, the operating condition determination unit uses... As the threshold for determining the dynamic entropy of background fluctuations, where... The background fluctuation calibration coefficient is set to 2.5, and the local spatial information entropy is calculated in real time. Exceeding the dynamic entropy threshold At that time, the system automatically triggers the scaling factor. Increase the compensation and set the step size to 0.05 to suppress non-stationary strong magnetic field interference.
[0037] The system processes raw video frame data. For each target pixel in the data, the coverage grid scale is used. and the required number of grids Calculate the fractal dimension of pixels The calculation formula is as follows: ,when At that time, the pixel data determines the geometric shape and semantics of the molten aluminum, and outputs it directionally to the main component stream. ;when At that time, the pixel is determined to be random noise and is output to the margin stream in a directional manner. It should be clarified that the output here is to the margin stream. Although the components contain random noise in the frequency domain, they completely preserve the global displacement vector caused by camera rig vibration in the spatial topology. In addition to random high-frequency components, the margin flow also includes the inherent rigid mechanical boundary around the aluminum outlet. Since the background components do not deform in a short time, its... The relative coordinate drift is equivalent to the physical vibration of the frame. The spatial weight anchoring unit extracts the residual edge features of these non-target areas and uses them as the reference input for spatial dynamic compensation. This transforms the noise that originally interfered with identification into a quantifiable displacement reference, achieving inverse dynamic compensation of the affine transformation matrix of the original image. The spatial weight anchoring unit retrieves the residual flow. The static boundary contour of the aluminum alloy mechanical component is extracted. To maintain long-term stable extraction of feature points in a strong magnetic and dusty environment, the system does not rely on unstable local point features, but instead utilizes residual flow. The edge of the aluminum outlet flange, which has the highest contrast and most consistent geometry, is used as the global anchoring reference. Specifically, the system... The closed loop curve with the most dramatic gray-level gradient change is found in the component. The geometric center of the curve is extracted as the global feature point. Since the flange is a rigid welded part, its gray-level entropy distribution in the infrared and visible light bands has extremely high spatiotemporal stability. Even under the interference of smoke and dust, the edge semantics after multi-frame mean filtering can still be accurately captured, thus providing a stable spatial coordinate origin for the derivation of Euclidean distance.
[0038] The system calculates the zeroth-order invariant moment and the first-order central moment of the boundary profile, and then maps them to generate a structure-invariant descriptor feature vector. System calculation The Euclidean distance between the target vector and the preset historical calibration reference vector is linearly translated into pixel spatial displacement, which is used as the logical rack bounce. The system utilizes Perform a translation matrix transformation on the original geometric logical anchor point coordinates, dynamically update the anchor point coordinates, and reconstruct the center distance weighted mask matrix. mask matrix With main component flow Perform pixel-by-pixel multiplication to extract the edge morphology features of the molten aluminum. The status arbitration unit monitors in real time the characteristics Effective aluminum liquid pixel area of the enclosed area When the effective area is lower than the minimum area threshold for normal operation for three consecutive video frame cycles, the system triggers a complete occlusion fault determination and retrieves the historical centroid displacement vector calibrated in the last five normal cycles before the fault. Compared with historical area change rate The state arbitration unit uses a first-order forward temporal linear extrapolation operator to dynamically synthesize the virtual aluminum liquid morphology characteristics within the current fault cycle, maintaining the control commands. Stable output, virtual fluid morphology characteristics The reconstruction utilizes image affine transformation matching matrix data and spatial displacement vectors, and the procedure is as follows: Extract fluid edge morphology features from the last normal operation image before complete occlusion determination. , as the seed point set matrix; and the historical centroid displacement vector Mapped to a two-dimensional translation matrix ,in and These correspond to the scalar increments of the barycenter coordinates on the horizontal and vertical axes of the image coordinate system, respectively; based on the historical area change rate. Determine the morphological scaling factor Construct a scaling matrix ; through matrix multiplication Performing spatial coordinate transformation, the seed point set matrix is translated to the predicted position and the contour scale is adjusted to generate a virtual fluid mask with physical motion logic. This drives the subsequent state machine decision model, ensuring the semantic logic of control flow within the visual blind spot is self-consistent, smoke and dust dissipate, and the effective pixel area is maximized. Once the system recovers to above the threshold, it automatically switches back to the normal arbitration track based on real-time feature decoupling.
[0039] Example 2: This example quantitatively verifies the robustness and engineering feasibility of the present invention in extracting edge features of aluminum effluent from an aluminum electrolysis cell under a strong magnetic field environment. An aluminum electrolysis cell aluminum effluent monitoring platform equipped with a high-definition CCD industrial camera (2048×2048 pixels resolution, 5ms exposure time, 30fps frame rate) was selected. In the experiment, the preset aluminum effluent flow rate calibration range was 5.0 kg / s to 15.0 kg / s. In the image preprocessing pipeline integrated within the feature decoupling unit, grayscale histogram statistics were performed on the video stream data using a 5×5 pixel local neighborhood window, and local spatial information entropy was introduced. As a measure of background noise intensity, it is used to determine the fractal dimension boundary threshold. The experiment compared the system recognition accuracy under different levels of Gaussian noise interference. A calibration parameter combination of a=0.5, b=0.2, and c=0.8 was set. During the test, Gaussian white noise with a signal-to-noise ratio of 18dB was actively injected into the video stream to simulate the strong electromagnetic interference environment of on-site operations. In the verification experiment, experimental group A (complete scheme) and control group B (using only the conventional edge detection operator, without feature decoupling and rigid condition adjudication) were set. When the aluminum fluid flow rate was at the median of the calibration range of 10.0 kg / s, experimental group A measured the edge morphology characteristics of the aluminum liquid. The standard deviation of pixel displacement fluctuation in the experimental group A was 0.42 pixels, while the standard deviation of edge morphology feature fluctuation in the control group B was 3.15 pixels. When the intensity gradient of the core problem variable (i.e., smoke concentration) changed, the observed data were as follows: When the smoke concentration was 50 g / m³, the effective area recognition rate of the experimental group A was 98.4%, and that of the control group B was 85.2%; when the smoke concentration was 150 g / m³, the effective area recognition rate of the experimental group A was 97.8%, and that of the control group B was 62.1%; when the smoke concentration was 300 g / m³, the effective area recognition rate of the experimental group A was 96.5%, and that of the control group B was 31.4%.
[0040] Data shows that as the dust concentration increases, the recognition accuracy of control group B decreases sharply and non-linearly, while experimental group A maintains high stability. Experimental group A efficiently removes random noise components through rigid conditional decision logic, thereby achieving steady-state output of features in a high-dust environment. To further verify the rationality of the above numerical range, an out-of-range control group C (where a=0.2, b=0.1, c=1.2) was established. Under the same working conditions, the edge features of the molten aluminum in control group C... The jump deviation in consecutive video frames reached 12.0 pixels, causing the arbitration unit of the operating condition to fail to converge. This indicates that... The dynamic gating rule parameters must meet the preset calibration range; deviation from this range will cause the main component stream to... The presence of excessive noise demonstrates that the parameter range defined in this invention is a necessary boundary to ensure the long-term stable operation of the system.
[0041] When the system performs fault arbitration, it monitors the effective aluminum liquid pixel area. ,when When the resolution is below 2500 pixels for three consecutive cycles, the system enters the virtual morphological feature synthesis mode. A first-order forward temporal linear extrapolation operator is used to dynamically synthesize the virtual aluminum liquid morphology. The calculation formula is as follows: ,in, This describes the morphological characteristics of molten aluminum in the previous cycle. This represents the historical area change rate (pixels / frame). The linear extrapolation here, using the center-of-gravity displacement vector (pixels), aims to address short-term visual blindness caused by the instantaneous surge of dust. Its effectiveness is based on the physical inertia of the aluminum outlet flow rate over a time interval of seconds. To prevent errors from diverging over time, the system employs a dual constraint: firstly, the weight of the linear extrapolation monotonically decreases with increasing obscuring period; secondly, if the duration of complete obscuring reaches the 20-second safety threshold, the system determines that the extrapolation model can no longer represent the actual operating condition. In this case, the safety alarm module will forcibly take over the system and output an abnormal command. This combination of temporal linear prediction and explicit safety thresholds ensures short-term smoothness of the control chain while guaranteeing industrial safety under extremely long-term blind zones. Experiments show that when dust completely obscures the aluminum outlet, causing visual signal interruption, this virtual morphology synthesis logic ensures the control commands are executed. The fluctuation deviation remained within 5%, confirming its anti-interference capability and semantic continuity under extreme blind zone conditions.
[0042] Example 3: This example combines Figures 1 to 2 This section describes a machine vision-based system for identifying aluminum electrolysis cell output conditions and anomalies. Figure 1 As shown, the system's processing flow begins with receiving the video frame stream from the aluminum extraction site. This video frame stream is then input to the feature decoupling unit, which uses a fractal dimension filtering operator to separate the target from the noise in the video frame stream. This results in the output of a foreground target feature map stream and a background noise component, with the background noise component serving as a spatial dynamic compensation reference. The foreground target feature map stream and the background noise component are then combined and input to the spatial weighting anchoring unit. This unit performs operations such as inferring the vibration displacement compensation affine transformation matrix and constructing a weighted mask matrix to truncate background interference. It then outputs fluid edge morphology features downstream. These features are input to the operating condition determination unit, which performs operations such as triggering a complete occlusion determination based on the effective area threshold and maintaining smooth updates of the operating condition by combining historical displacement vectors. Finally, it outputs operating condition decision data. Simultaneously, the operating condition determination unit, in conjunction with historical data, calls the synthesis logic and the temporal linear prediction module to establish a dotted-line connection feedback mechanism. This temporal linear prediction module extracts historical centroid coordinates and historical areas and dynamically synthesizes virtual fluid morphology features, which are then fed back to the operating condition determination unit.
[0043] like Figure 2 As shown, the overall module layout includes a feature decoupling unit, a spatial weight anchoring unit, a working condition determination unit, and a safety alarm module. The feature decoupling unit points downward and connects to the spatial weight anchoring unit, the spatial weight anchoring unit points downward and connects to the working condition determination unit, and the working condition determination unit points downward and connects to the safety alarm module.
[0044] Example 4: Under the long-cycle production conditions of aluminum electrolysis cells, the extremely uneven distribution of cathode current can lead to asymmetric erosion of the side furnace walls. If this trend is not identified in time, it will induce the risk of perforation of the cell body. This example constructs a defensive anomaly early warning procedure based on multidimensional image entropy. When a small deformation occurs in the side furnace walls, an algorithm will issue an alarm. The test platform uses a panoramic infrared thermal imaging system deployed on the side of the cell, with a sampling frequency of 10Hz. The system uses a 64×64 pixel grid as the initial feature extraction unit and calculates the feature extraction value of each grid. Gray covariance matrix at time step To eliminate local pixel oversaturation noise induced by electromagnetic fields, the system sets the information gain weighting factor λ=0.65, and calculates the feature vector space projection magnitude of this cell. The calculation formula is as follows: ,in, For matrix trace operation operator, The system uses a preset feature projection matrix to map high-dimensional image features to a one-dimensional morphological evolution space. After running under normal conditions for 24 hours, the system records the feature magnitude values. mean with standard deviation The system is set to trigger an early warning as follows: when the average modulus value of five consecutive monitoring periods reaches a certain threshold... satisfy When this is considered a trigger for lateral erosion risk, in the simulated operating conditions, when a 5% grayscale step change in the pixel area is artificially introduced to simulate early weak points in the furnace wall, the monitoring unit captures... The value jumped to 16.2 within 3 sampling periods, immediately triggering a defensive early warning mechanism.
[0045] To verify the stability of the algorithm in complex dynamic environments, the experimental group of this invention compared the risk identification capabilities under different working conditions. The experimental data showed that in the control group without feature projection mapping, the false negative rate for a 5% grayscale step change was as high as 45% due to the interference of environmental thermal radiation background. However, the system of this invention, which adopts the above-mentioned feature projection mapping mechanism, reduced the false negative rate to 3% under the same interference intensity. This data verification shows that feature projection through matrix trace operation enhances the local subtle deformation, effectively overcoming the problem of insufficient feature contrast in thermal background of traditional image recognition methods. After identifying the erosion risk, the system retrieves the corresponding thermal conductivity coefficient matrix, calculates the furnace wall thinning rate through numerical simulation, and feeds this parameter back to the anode movement control module to adjust the anode position to balance side heat dissipation, thereby achieving closed-loop defense against the erosion trend.
[0046] Example 5: During the continuous operation phase of the aluminum electrolysis cell, the system addresses operational logic drift caused by electrode wear by constructing an abnormal state baseline early warning procedure based on spatiotemporal correlation. Upon initial deployment, the system reads the unbiased operating condition image sequence of the aluminum outlet over 72 hours to construct a spatial feature baseline library. The system extracts the texture contrast parameters of each 64×64 pixel grid. And calculate the feature evolution gradient. The calculation formula is as follows: ,in, for Time of the first Texture contrast of each grid For the previous sampling time, For the sampling time interval, when The eigenvalues fluctuated beyond the dynamic distribution standard deviation set by the baseline library within 10 sampling periods. At that time, the recognition unit executes parameter recalibration logic to automatically correct the input weight allocation of the feature extraction unit in order to offset the signal attenuation caused by the structural loss of the electrode. When the system detects a sudden change in visual contrast caused by the peeling off of the oxide film on the electrode surface, it determines the disturbance as an abnormal electrode state rather than an abnormal process operation based on the updated feature evolution gradient threshold.
[0047] In another dynamic calibration experiment for adapting to multi-task operating conditions, by changing the electrolyzer's operating voltage to trigger changes in the electrolyte flow characteristics, the system executes a pre-parameter calibration procedure. When a voltage jump exceeding 0.5V is detected, the sliding window update mechanism of the operating condition feature matching library is triggered, adjusting the semantic weights of the images within the current operating window. By redistributing the data and using a weighted average method to suppress high-frequency false features in the image caused by surface ripples in the fluid, the stability of the abnormal operation condition determination is maintained. According to actual tests, this procedure still maintains an accuracy rate of over 90% in identifying abnormal operation conditions even under extreme conditions where the physical displacement of the electrodes causes a 2.5-degree deflection of the monitoring angle. This confirms the necessity of feature space adaptive baseline calibration for improving the system's ability to cope with fluctuations in complex physical environments.
[0048] Example 6: During the on-site deployment phase of the aluminum electrolysis cell aluminum output condition and anomaly identification system, to address the individual differences in edge recognition accuracy caused by different cell types, magnetic field distributions, and visual detector installation angles, a standardized pre-calibration procedure was performed before the system went online. This involved reading the background image from the aluminum outlet camera when the cell wall temperature was at an initial steady state of 450°C, and extracting the static background grayscale reference for each pixel. The system calculates the features of the calibrated image after environmental fluctuation compensation by executing an environmental fluctuation suppression algorithm. The calculation formula is as follows: ,in, This is the original video frame image captured at the current sampling time. As a static background grayscale reference, For environmental noise intensity, The system's preset stability adjustment offset is set to 0.05. The calibration unit simulates the gradual change process of the aluminum fluid's geometry within the camera's field of view, injecting grayscale gradient change signals at a step size of 0.2 units / second per frame, and simultaneously recording the feature evolution gradient output by the recognition unit. Based on this, the steady-state judgment threshold for abnormal flow deviation state is determined. ,when The real-time value exceeds When the condition is determined, the identification unit outputs a logic instruction to trigger the state transition to the working condition determination unit.
[0049] The system monitors the deflection angle of the aluminum outlet fluid. When the angle changes, the instantaneous rate of change of that angle is calculated. The state criterion is calculated using the following formula: ,in, This represents the change in the deflection angle within adjacent sampling periods. The sampling time interval, To determine the anomaly threshold based on historical operating condition statistics, the system sets the following judgment criteria in the identification logic: If 8 consecutive sampling periods in to Within the value range of , the recognition unit determines that the fluid is in a dynamic fluctuation state and locks the recognition output; if Values exceeding The identification unit triggers an abnormal flow deviation alarm and sends a flow adjustment command to the anode control module. Field test data confirms that under the condition that the monitoring angle deflects by 2.5 degrees due to the physical displacement of the electrode, the calibration logic reduces the false judgment rate of abnormal flow deviation from 15% in the original scheme to 0.5%, achieving stable identification of aluminum output conditions under complex physical environments.
[0050] 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 machine vision-based system for identifying aluminum electrolysis cell output conditions and anomalies, characterized in that, The system includes: The feature decoupling unit is used to acquire the aluminum field video frame stream and use the fractal dimension filtering operator to separate the foreground target feature map stream and background noise components from the video frame stream; The spatial weight anchoring unit, connected to the feature decoupling unit, is used to deduce the inter-frame vibration displacement of the image based on the Euclidean distance of the feature point coordinates, and to perform translation and rotation transformation compensation on the image affine transformation matrix to construct a weighted mask matrix. The weighted mask matrix is then multiplied pixel by pixel with the foreground target feature map flow to truncate background interference and output fluid edge morphology features. The background noise component is used as the input reference for spatial dynamic compensation. The working condition determination unit, connected to the spatial weight anchoring unit, is used to monitor the effective pixel area of the region enclosed by the fluid edge morphological features. When the effective pixel area is lower than the dynamic area threshold determined based on the average effective pixel area of the historical normal frame stream for three consecutive video frame periods, a complete occlusion determination is triggered. The virtual fluid morphological features are dynamically synthesized using the temporal linear prediction module in combination with the historical centroid displacement vector, thereby completing the smooth update of the working condition decision data.
2. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, During operation, the feature decoupling unit uses a fractal dimension filtering operator to process the video frame stream into blocks, calculates the pixel grayscale entropy and pixel-level fractal dimension of each block's image domain, and determines blocks whose pixel grayscale entropy is lower than the dynamic entropy threshold determined by the baseline grayscale entropy obtained from the statistical analysis of smokeless background frame images, and whose pixel-level fractal dimension is in the range of 1.2 to 1.8, as foreground target feature map streams. The remaining image components that do not belong to the foreground target feature map streams are determined as background noise components.
3. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, When the working condition determination unit is running, the temporal linear prediction module extracts the historical centroid coordinates and historical area of the five consecutive normal working cycles before the complete occlusion determination in each video frame cycle after the complete occlusion determination is triggered. It also calculates the centroid displacement change rate and area decay rate of adjacent normal working cycles, uses a linear extrapolation model to dynamically synthesize the virtual centroid position and virtual effective area in the current fault cycle, and reassembles and generates virtual fluid morphology features.
4. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, The working condition determination unit includes a decision data smoothing module; the decision data smoothing module is connected to the time-series linear prediction module and is used to input the virtual fluid morphology features into the state machine decision model during the complete occlusion determination period, so that the change amplitude of the working condition decision data output by the state machine decision model is lower than the dynamic fluctuation threshold determined based on the transient signal-to-noise ratio of the current video frame stream.
5. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, The spatial weight anchoring unit includes a dynamic compensation module; the dynamic compensation module is connected to the feature decoupling unit and is used to obtain the global high-frequency fluctuation frequency and average amplitude of the background noise component, and dynamically adjust the center distance attenuation coefficient of the weighted mask matrix according to the global high-frequency fluctuation frequency and average amplitude.
6. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, When the working condition determination unit is running, it specifically includes the following steps: real-time calculation of the effective pixel area of the closed area enclosed by the fluid edge morphology features; real-time comparison of the effective pixel area with the minimum area threshold of normal operation determined based on the statistical lower limit of the pixel area of the historical normal operation cycle; when the effective pixel area is lower than the minimum area threshold of normal operation for three consecutive video frame cycles, it is confirmed as a complete occlusion working condition and a complete occlusion fault determination signal is output.
7. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, When the spatial weighted anchoring unit is running, it specifically includes the following steps: based on the fixed reference coordinates of the aluminum outlet center and the real-time extracted fluid flow direction center coordinates, the inter-frame vibration displacement is deduced by the Euclidean distance of the feature point coordinates; the azimuth translation vector and rotation angle are calculated using the inter-frame vibration displacement, and the translation and rotation matrix transformation compensation is completed on the image affine transformation matrix.
8. The aluminum electrolysis cell aluminum output condition and anomaly identification system based on machine vision according to claim 1, characterized in that, The feature decoupling unit includes a video access interface, which is used to acquire video streams in real time from explosion-proof industrial cameras arranged at the aluminum electrolysis cell workshop's aluminum output workstations, and decode the video streams into single-frame image sequences to be input as video frame streams into the fractal dimension filtering operator.
9. A machine vision-based aluminum electrolysis cell aluminum output condition and anomaly identification system according to claim 1, characterized in that, The system includes a safety alarm module connected to the working condition determination unit. When the duration of the fault determination signal being completely obscured exceeds the explicit safety time threshold of 20 seconds, it outputs visual abnormality alarm status data so that the unmanned operation and maintenance system can complete the arbitration of the workshop safety status.