An online baffle blocking identification method and system for a thickener rake frame

Abstract

The application relates to the technical field of fault monitoring and identification of mining equipment, in particular to a thickener rake frame jamming online identification method and system, to solve the problems that existing schemes rely on fixed current threshold or single-point prediction, it is difficult to distinguish normal working condition current floating and jamming additional load, it is easy to be misreported by transient disturbance, and working condition self-adaptive ability is poor, the present application only relies on motor current and basic running state signal, adopts a one-dimensional diffusion current representation model, generates a multi-level noise current sequence through a forward noise chain, generates an interpretable current trajectory and an adaptive normal fluctuation limit parameter based on a shared backbone double-output head structure, extracts the out-of-bound part of the deviation current to form an additional load current sequence, judges through a continuous positive value counting and increment constraint continuous accumulation evidence mechanism, outputs a three-level identification result and a jamming risk index, and converts the three-level identification result and the jamming risk index into industrial signals to send to an electric control system to realize linkage protection.

Description

Technical Field

[0001] This application relates to the field of fault monitoring and identification technology for mining equipment, specifically to an online identification method and system for jamming of a thickener scraper frame. Background Technology

[0002] The thickener is a core piece of equipment in the solid-liquid separation and reclaimed water processes of a mineral processing plant. Its rake frame continuously rotates to disrupt the settling layer structure and push the underflow towards the discharge port. Its operational stability directly affects the underflow concentration, clarified water quality, and the continuity of subsequent processes. Under conditions such as high-concentration slurry, increased proportion of fine particles, changes in slurry viscosity, fluctuations in feed rate, adjustments to the dosing regime, or abnormal rise in mud level, the rake frame is prone to gradually increasing obstruction or even jamming. This manifests as increased drive load, increased energy consumption, and abnormal stress on the rake arm. In severe cases, it can lead to overload shutdown, reducer damage, rake arm bending, or even damage to the tank structure, causing shutdown for tank cleaning and interruption of the mineral processing process.

[0003] The signals that the on-site electrical control system can usually reliably acquire are the real-time current of the rake frame drive motor, as well as basic status signals such as operating enable, direction, and speed. However, the motor current is superimposed with the load drift under normal operating conditions and the additional load accumulation under abnormal operating conditions. The two sources are coupled with each other on the current curve, which makes it necessary for online identification to both fit the mineral processing mechanism and output stable conclusions that can be used for electrical control linkage without adding complex sensors.

[0004] Existing methods for identifying anomalies in thickener scraper frames often employ threshold and trend-based judgment schemes based on motor current / power / torque signals. For example, setting an upper limit for current, a threshold for the current rise slope, and a delay for exceeding the limit, combined with moving averages and low-pass filtering to suppress noise; or using linear regression and time series prediction to generate a reference current, triggering an alarm based on the difference between the real-time value and the reference value; some schemes combine process parameters such as concentration, mud layer height, feed rate, and chemical dosage to perform rule fusion or train a classification model to achieve multi-source hierarchical alarms, and connect to a PLC / DCS to perform deceleration, reversal, or shutdown protection.

[0005] However, the threshold, slope, and delay parameters mostly rely on manual experience for tuning, making it difficult to adaptively adjust according to ore properties, feeding system, and dosing strategy. Under normal operating conditions, load fluctuations can easily lead to false alarms or mask the development process of gradual jamming. Moreover, the reference current is mostly point prediction or simple fitting, lacking explicit characterization of the normal fluctuation range, and the difference discrimination lacks a dynamic and unified benchmark. In addition, it is mostly based on the original current peak or single over-limit trigger, which is easily affected by short-term impacts and transient deviations during operating condition switching, and cannot stably reflect the continuous accumulation characteristics of the additional load of jamming. Summary of the Invention

[0006] This invention aims to establish a calculable representation of the evolution of current under normal operating conditions and its reasonable fluctuation range under the constraints of relying solely on the real-time current of the drive motor and basic operating status signals in thickener beneficiation. This allows for the reliable separation and continuous identification of the additional load current caused by scraper frame jamming, thereby generating online jamming identification results that can be used for closed-loop linkage of the electrical control system. Based on this, an online identification method for thickener scraper frame jamming is proposed, including the following steps:

[0007] S1. Obtain the real-time current sampling value of the thickener rake frame drive motor and form a real-time current sequence; receive the rake frame running status signal and form a running status sequence to obtain the identification input;

[0008] S2. Based on the identified input, call the one-dimensional diffusion current characterization model. The one-dimensional diffusion current characterization model includes a forward noise chain and a reverse noise chain. The forward noise chain superimposes random disturbances on the real-time current sequence to form a multi-level noise current sequence. The reverse noise chain uses the real-time current sequence as a condition to denoise the multi-level noise current sequence from high noise level to low noise level, generating an interpretable current trajectory based on the normal operating condition assumption, and simultaneously outputs the normal fluctuation limit parameter characterizing the normal operating condition current floating range.

[0009] S3. Calculate the deviation current sequence based on the real-time current sequence and the interpretable current trajectory, and extract the part of the deviation current sequence that exceeds the normal fluctuation limit according to the normal fluctuation limit parameter to form the additional load current sequence.

[0010] S4. Perform continuous motor current rise discrimination on the extra load current sequence, initialize continuous accumulation evidence and continuous positive value count, read the extra load current value sequentially according to the sampling time and calculate the extra load current increment value, perform incremental update or decay update on the continuous accumulation evidence and set upper and lower limits according to the joint judgment result of the continuous positive value count and the extra load current increment value, and generate a jamming risk index and jamming status indicator with a value range of 0 to 1 based on the comparison result of continuous accumulation evidence and continuous threshold parameter.

[0011] S5. Generate online identification results based on the jamming status indicator and jamming risk index. The online identification results include at least three statuses: normal, suspected jamming, and confirmed jamming.

[0012] Preferably, in step S2, the execution process of the forward noise chain includes the following steps:

[0013] S201. Initialize the diffusion level identifier to the lowest level and use the real-time current sequence as the zero noise level sequence; read the pre-stored noise intensity table, which contains noise intensity values ​​corresponding one-to-one with each diffusion level.

[0014] S202. According to the diffusion level from low to high, generate a random perturbation sequence with zero mean and unit variance for each level, scale it according to the corresponding noise intensity, and then add it point by point to the current sequence of the previous level to obtain the noise current sequence corresponding to the current diffusion level, and finally form a multi-level noise current sequence covering all diffusion levels.

[0015] Preferably, the execution process of the reverse denoising chain includes the following steps:

[0016] S211. Calculate the current difference between adjacent sampling points based on the real-time current sequence, generate a current increment sequence, and set the increment of the first point of the sequence to 0; concatenate the real-time current sequence and the current increment sequence according to the channel dimension to form a conditional feature tensor; use the current level sequence of the multi-level noise current sequence as the denoising feature tensor.

[0017] S212. Map the current diffusion level identifier to a level indicator vector, and generate a diffusion level embedding vector through two fully connected layers; the diffusion level embedding vector is used to modulate the channel features of the convolutional layer inside the inverse denoising chain, so that the interpretable current trajectory output under different noise levels is semantically consistent with the normal fluctuation limit parameter.

[0018] S213. Input the denoising feature tensor and the conditional feature tensor into the shared characterization backbone network to extract and fuse features, and generate the interpretable current trajectory and normal fluctuation boundary parameters of the current level synchronously through dual output heads.

[0019] S214. Update the current level sequence of the multi-level noise current sequence step by step in order of diffusion level from high to low, until the diffusion level indicator drops to the lowest level, and output the final interpretable current trajectory and normal fluctuation limit parameters.

[0020] The reverse denoising chain adopts a shared representation backbone network plus a dual-output head structure. The shared representation backbone network is a one-dimensional residual convolutional network, which sequentially includes an input convolutional layer, a slowly varying residual stacked layer, a fast varying residual stacked layer, and a fusion convolutional layer. The input convolutional layer performs channel alignment and concatenation of the denoised feature tensor and the conditional feature tensor to generate a basic feature map. The slowly varying residual stacked layer consists of 3 residual blocks and uses increasing dilation rate convolution to extract the slow-varying current drift feature. The fast varying residual stacked layer consists of 3 residual blocks and uses small convolutional kernels to extract the short-time current fluctuation feature. The fusion convolutional layer concatenates and compresses the two feature paths to generate a fusion feature map.

[0021] The dual output head includes an interpretable current trajectory output head and a normal fluctuation limit parameter output head, which share the same fused feature map. They are respectively output through one-dimensional convolution to interpretable current trajectory and normal fluctuation limit parameter, so that interpretable current trajectory and normal fluctuation limit parameter are generated synchronously under the same operating conditions.

[0022] Preferably, in S2, the multi-level noise current sequence starts from the diffusion level. arrive The stepwise denoising process follows the diffusion transfer formula:

[0023]

[0024] in, The diffusion level is identified as The current level noise current sequence at that time, After noise reduction Level of current sequence; The diffusion level is identified as Interpretable current trajectories generated at that time; diffusion level The corresponding denoising step size coefficient, diffusion level The corresponding noise reinjection coefficients are pre-stored in the diffusion transfer table; diffusion level The corresponding noise intensity, diffusion level The corresponding zero-mean, unit-variance random perturbation sequence; The sampling time number is used. All sequences have the same length, and the point-by-point operations correspond one-to-one with the time index.

[0025] Preferably, step S3 specifically includes the following steps:

[0026] S301. Read the real-time current values ​​sequentially according to the sampling time of the real-time current sequence, and read the baseline current values ​​at the same sampling time in the interpretable current trajectory. Subtract them point by point and combine them in time order to generate the deviation current sequence.

[0027] S302. Read the normal fluctuation limit parameter at the sampling time, compare the deviation current sequence with the normal fluctuation limit parameter point by point, take the result of the deviation current minus the normal fluctuation limit parameter when the deviation current is greater than the normal fluctuation limit parameter, and take zero when the deviation current is not greater than the normal fluctuation limit parameter to generate the initial additional load current sequence.

[0028] S303. Perform minimum additional load threshold parameter filtering on the initial additional load current sequence point by point, set the values ​​of those less than the minimum additional load threshold parameter to zero, and keep the sequence length consistent with the real-time current sequence to form an additional load current sequence.

[0029] Preferably, in step S4, the calculation process of the continuously accumulating evidence and the blocking risk index includes:

[0030] Initialize continuous accumulation of evidence Continuous positive value counting Read the additional load current values ​​in the order of sampling time. Calculate the additional load current increment at adjacent sampling times. ;

[0031] like Then, consecutive positive values ​​are counted. If Then, consecutive positive values ​​are counted. ;

[0032] The continuous accumulation of evidence Update using the following formula:

[0033]

[0034] in, This is the upper limit parameter for evidence; A positive coefficient of increase in evidence; This is the continuous sampling threshold parameter; To enhance the judgment threshold parameter; The evidence attenuation coefficient is defined as a value in the interval [0,1].

[0035] Blockage Risk Index Calculate using the following formula:

[0036]

[0037] in, This is a continuous threshold parameter; the range of the jamming risk index is 0 to 1. When the jamming risk index is 1 and the jamming status is determined; when At that time, the risk index of blockage And determine the normal status indicator.

[0038] Preferably, step S1 specifically includes: acquiring drive motor current sampling values ​​according to a preset sampling period, and forming a current sampling value sequence after dimensional unification conversion; and then, according to a preset sequence length parameter... Perform a sliding truncation on the current sample value sequence to generate a length of The real-time current sequence, wherein the preset sequence length parameter The length of the input sequence is the same as that of the inverse denoising chain;

[0039] The system receives a running enable signal, a running direction signal, and a speed feedback signal to form a running state sequence. When the running enable signal is valid and the speed feedback signal is greater than the speed threshold parameter, the real-time current sequence and the running state sequence are aligned and spliced ​​according to the time reference to form the identification input.

[0040] Preferably, step S5 includes reading a suspected threshold parameter; if the current status is normal and the jamming risk index is less than the suspected threshold parameter, a normal online identification result is generated; if the jamming risk index is greater than or equal to the suspected threshold parameter, a suspected jamming online identification result is generated; if the current status is jamming, a confirmed jamming online identification result is generated.

[0041] Preferably, it also includes S6, wherein:

[0042] The online identification results are discretized and mapped to the following card identification signal values: 0 for normal state, 1 for suspected card jamming, and 2 for confirmed card jamming.

[0043] Read the full-scale value and quantization ratio, and linearly map the jamming risk index in the 0-1 range to the corresponding amplitude risk signal value;

[0044] The jamming identification signal value and risk signal value are sent to the dense machine's electrical control system through the industrial communication interface at a preset refresh cycle for electrical control linkage control of deceleration, reversal, and shutdown protection.

[0045] An online identification system for thickener rake frame jamming includes a data acquisition module, a one-dimensional diffusion characterization module, an additional load extraction module, a continuous discrimination module, and a result output module, wherein:

[0046] The data acquisition module is used to collect the current sampling value and operating status signal of the thickener rake frame drive motor, generate real-time current sequence and operating status sequence through dimension conversion and sliding truncation, and splice them together to form an identification input when the effective operating conditions are met, and output to the one-dimensional diffusion characterization module.

[0047] The one-dimensional diffusion characterization module has a built-in one-dimensional diffusion current characterization model consisting of a forward noise chain and a reverse noise chain. It is used to receive the identification input, perform step-by-step noise reduction on the multi-level noise current sequence based on the real-time current sequence, synchronously generate interpretable current trajectory and normal fluctuation boundary parameters, and output them to the additional load extraction module.

[0048] The one-dimensional diffusion current characterization model consists of a forward noise-adding chain and a reverse denoising chain. The forward noise-adding chain applies hierarchical random perturbations to the real-time current sequence to form a multi-level noise current sequence. The reverse denoising chain performs denoising on the multi-level noise current sequence to generate an interpretable current trajectory and outputs normal fluctuation limit parameters. The core neural network of this model is located in the reverse denoising chain. Instead of using the common single-head output method of one-dimensional denoising networks, the reverse denoising chain adopts a structure of shared representation backbone network plus dual output heads. It explicitly expresses the normal rise caused by dense motor conditions using normal fluctuation limit parameters, enabling the continuous determination of the motor current rise to obtain a calculable and interpretable normal range.

[0049] The neural network input data of the inverse denoising chain includes: the current level sequence of the multi-level noise current sequence and the real-time current sequence; the neural network calculates the current increment sequence based on the real-time current sequence and concatenates the real-time current sequence and the current increment sequence to form a conditional feature tensor, the dimension of which is the sequence of length L multiplied by 2 channels; the current level sequence of the multi-level noise current sequence is used as the denoising feature tensor, the dimension of which is the sequence of length L multiplied by 1 channel; the inverse denoising chain also receives the diffusion level identifier and generates a diffusion level embedding vector through two fully connected layers. The first fully connected layer has 64 neurons and the second fully connected layer has 128 neurons. The diffusion level embedding vector is used to modulate the feature calculation of subsequent convolutional layers, so that the same network outputs interpretable current trajectories and normal fluctuation boundary parameters with consistent semantics under different noise levels.

[0050] The intermediate layer of the inverse denoising chain is a one-dimensional residual convolutional backbone network, which includes an input convolutional layer, a slowly varying residual stacked layer, a rapidly varying residual stacked layer, and a fusion convolutional layer. The input convolutional layer concatenates the denoised feature tensor and the conditional feature tensor after channel alignment, and uses one-dimensional convolution to form the basic feature map, with 64 convolutional channels. The slowly varying residual stacked layer consists of three residual blocks, each containing two one-dimensional convolutional layers with 64 convolutional channels. The convolution uses an increasing dilation rate to expand the receptive field, representing the slow load drift caused by changes in slurry viscosity, feed, and reagent addition in the interpretable current trajectory. The rapidly varying residual stacked layer also consists of three residual blocks, each containing two one-dimensional convolutional layers with 64 convolutional channels. The convolution uses small kernels to represent short-term current fluctuations, allowing the normal fluctuation boundary parameters to vary with the fluctuation intensity. The fusion convolutional layer concatenates the outputs of the slowly varying residual stacked layer and the fast varying residual stacked layer, and compresses them into a 64-channel fusion feature map through one-dimensional convolution.

[0051] The output layer of the inverse denoising chain has a dual-output head structure: the interpretable current trajectory output head applies a one-dimensional convolution to the fused feature map and outputs a single-channel sequence of length L to obtain the interpretable current trajectory; the normal fluctuation limit parameter output head applies a one-dimensional convolution to the fused feature map and outputs a single-channel sequence of length L to obtain the normal fluctuation limit parameter. The dual output heads share the same fused feature map, enabling the interpretable current trajectory and the normal fluctuation limit parameter to be generated synchronously under the same conditions. The normal fluctuation limit parameter changes with the operating pattern of the real-time current sequence, supporting the chain that calculates the deviation current based on the real-time current sequence and the interpretable current trajectory and extracts the additional load current sequence based on the normal fluctuation limit parameter. Structurally, the interpretable range of normal fluctuation is solidified into a referenceable parameter object, rather than relying on a fixed current threshold or a fixed slope threshold.

[0052] The additional load extraction module is used to receive the real-time current sequence, the interpretable current trajectory and the normal fluctuation limit parameter, calculate the deviation current sequence and extract the part that exceeds the normal fluctuation limit, and generate an additional load current sequence after filtering by a minimum threshold, and output it to the continuous discrimination module.

[0053] The continuous discrimination module is used to receive the additional load current sequence, calculate the continuous accumulated evidence, and combine it with the continuous threshold parameter to generate a jamming risk index and a jamming status identifier, and output them to the result output module.

[0054] The result output module is used to receive the jamming status indicator and jamming risk index, generate a three-level online identification result including normal, suspected jamming, and confirmed jamming, and convert it into jamming identification signal and risk signal that can be read by the electronic control system, and send it to the thickener electronic control system.

[0055] The beneficial effects of this invention are as follows:

[0056] 1. Compared to fixed thresholds / slope thresholds / delay triggers based on motor current, or direct differential discrimination after forming a reference current using linear regression or single-point prediction, this invention introduces a one-dimensional diffusion model into the current characterization stage of thickener beneficiation equipment. Under real-time current sequence conditions, it simultaneously generates interpretable current trajectories and normal fluctuation boundary parameters based on normal operating conditions, making the reasonable range of normal upward movement a calculable, referable, and adaptively changing quantity according to operating conditions. This design incorporates slow-varying drifts and short-term fluctuations caused by slurry viscosity, feed fluctuations, and reagent changes into the same stochastic evolution framework. By using a shared backbone network combined with dual output heads, it synchronously provides baselines and boundaries within the same semantic space, thereby reducing the inconsistency problem of operating condition migration caused by relying on manual experience to tune fixed thresholds. At the same time, the solution only uses current and basic operating status signals that can be stably obtained on-site, without relying on additional process measurement points, making it easy to implement in existing electrical control systems.

[0057] 2. In the discrimination chain, this application first constructs a deviation current from the real-time current to the interpretable current trajectory, and then uses normal fluctuation limit parameters to extract the out-of-limit portion point by point to form an additional load current sequence. This allows subsequent discrimination to revolve around the load exceeding the normal interpretable range, rather than directly judging the original current peak or a single over-limit. Subsequently, through a continuous positive value counting and a mechanism of increasing or decreasing evidence accumulation with enhanced trend constraints, the additional load caused by jamming is gradually accumulated and mapped into a stable increase in evidence, suppressing transient deviations and false triggers caused by short-term shocks and operating condition switching, and outputting a risk index and three-state results that can be used for linkage control. This effect can be quantitatively verified by comparison and playback, such as by counting the number of false alarms per unit time, the proportion of false alarms under the same trigger threshold, and the advance amount from the risk index crossing the suspected threshold to the confirmed jamming. For example, under the conditions of 1s sampling and a window length L of 300, the number of false alarms and the early warning time during the gradual jamming process can be compared with the fixed threshold and delay of this scheme when there are frequent fluctuations in ore feeding or chemical addition, to form a verifiable engineering quantitative basis. Attached Figure Description

[0058] Figure 1 This is a flowchart of the process for obtaining and recognizing input according to the present invention;

[0059] Figure 2 This is a flowchart illustrating the calling process of the one-dimensional diffusion current characterization model of the present invention.

[0060] Figure 3 A flowchart is generated for the additional load current sequence;

[0061] Figure 4 Flowchart for determining the continuous rise of motor current;

[0062] Figure 5 Generate a flowchart for the online recognition results;

[0063] Figure 6 Flowchart for sending card identification signals and risk signals;

[0064] Figure 7 This is a diagram illustrating the additional load.

[0065] Figure 8 A schematic diagram for synchronous generation of dual output heads;

[0066] Figure 9 This is a diagram illustrating the continuous accumulation of evidence.

[0067] Figure 10 This is a schematic diagram of the network architecture. Detailed Implementation

[0068] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0069] Example 1:

[0070] like Figure 1-10 As shown, an online identification method for jamming of a thickener rake frame includes the following steps:

[0071] S1, Input Recognition Structure: (e.g., ...) Figure 1 As shown, current sampling values ​​are collected from the current detection channel of the thickener rake frame drive motor according to a preset sampling period, and the current sampling values ​​are converted into current values ​​with uniform dimensions and then arranged into a current sampling value sequence in time order.

[0072] The current sampling value sequence is processed according to the preset sequence length parameter. Perform a sliding cut, generating a length of The real-time current sequence, wherein the preset sequence length parameter The length of the input sequence is the same as that of the inverse denoising chain;

[0073] The system receives operation enable signals, operation direction signals, and speed feedback signals from the dense electromechanical control system, and forms an operation status sequence according to a time base consistent with the preset sampling period.

[0074] When the operation enable signal is valid and the speed feedback signal is greater than the speed threshold parameter, the real-time current sequence and the operation status sequence are spliced ​​together according to the time base to form the identification input.

[0075] S2, Calling the one-dimensional diffusion current characterization model: (e.g.) Figure 2 As shown, the operation enable signal is determined to be valid and the speed feedback signal is greater than the speed threshold parameter based on the operation state sequence. When the determination condition is met, the one-dimensional diffusion current characterization model is called and the real-time current sequence is used as the condition input of the reverse denoising chain.

[0076] A forward noise chain is executed on the real-time current sequence, and random perturbations are superimposed level by level according to the noise intensity corresponding to the diffusion level label to form a multi-level noise current sequence and update the diffusion level label to the highest noise level.

[0077] In the reverse denoising chain, the current level sequence of the multi-level noise current sequence is set as the denoising feature tensor, and the current increment sequence is calculated based on the real-time current sequence. The real-time current sequence and the current increment sequence are concatenated by channel to form a conditional feature tensor.

[0078] The diffusion level identifier is input into two fully connected layers to generate a diffusion level embedding vector, where the first fully connected layer has 64 neurons and the second fully connected layer has 128 neurons. The diffusion level embedding vector is then used to calculate the convolutional channel features of the modulated one-dimensional residual convolutional backbone network.

[0079] The input convolutional layer is used to align the channels of the denoised feature tensor and the conditional feature tensor and then concatenate them. A 64-channel basic feature map is generated by one-dimensional convolution.

[0080] The basic feature map is sequentially input into the slow-varying residual stacking layer and the fast-varying residual stacking layer. The slow-varying residual stacking layer consists of three residual blocks and uses increasing dilation rate convolution to extract slow load drift features. The fast-varying residual stacking layer consists of three residual blocks and uses small convolution kernel convolution to extract short-term fluctuation features. The output channels of the slow-varying residual stacking layer and the fast-varying residual stacking layer are concatenated and compressed by a fusion convolution layer to form a 64-channel fusion feature map.

[0081] The fused feature map adopts a shared representation backbone network plus a dual output head structure. One-dimensional convolution is applied to the interpretable current trajectory output head to generate an interpretable current trajectory with the same length as the real-time current sequence. One-dimensional convolution is applied to the normal fluctuation limit parameter output head to synchronously generate a normal fluctuation limit parameter with the same length as the real-time current sequence. Based on the diffusion level label, the current level sequence of the multi-level noise current sequence is updated sequentially from high noise level to low noise level until the diffusion level label is the lowest noise level.

[0082] When the one-dimensional diffusion current characterization model is invoked based on the identified input, it does not employ fixed thresholds, linear regression, or single-point prediction methods. Instead, it uses a structure of a forward noise chain and a reverse noise chain to probabilistically characterize the real-time current sequence, outputting parameters that can explain the current trajectory and the normal fluctuation boundary. The one-dimensional diffusion current characterization model decomposes the current changes during unblocked operation into an evolutionary process of progressively superimposed disturbances through a forward noise chain, thus unifying the current rise caused by changes in slurry viscosity, feed fluctuations, and reagent addition changes into a unified source of random disturbances.

[0083] The one-dimensional diffusion current characterization model then generates an interpretable current trajectory under normal operating condition assumptions through a reverse denoising chain under real-time current sequence conditions. The interpretable current trajectory expresses the evolution direction and level of the current when the equipment is not jammed. At the same time, the reverse denoising chain outputs normal fluctuation limit parameters during the generation of the interpretable current trajectory. The normal fluctuation limit parameters are used to limit the reasonable upward range of the real-time current sequence around the interpretable current trajectory.

[0084] Compared with existing technologies, the normal fluctuation limit parameter and the interpretable current trajectory are generated synchronously by the same inverse denoising chain. The normal fluctuation limit parameter adapts to the operating state of the real-time current sequence, thereby turning the interpretable range of normal float into a calculable and referable parameter object, replacing the fixed current threshold or fixed slope threshold that was previously set based on experience. Furthermore, the one-dimensional diffusion current characterization model uses the real-time current sequence as the only condition input and does not introduce additional process variables. It can still provide the corresponding normal fluctuation limit parameter for normal float under different operating conditions, so that the real-time current sequence is decomposed into deviation current and the path of the additional load current sequence in the additional load current sequence extraction step, and has a consistent judgment benchmark.

[0085] S3, Extraction of Additional Load Current Sequence: (e.g.) Figure 3 As shown, real-time current values ​​are read sequentially according to the sampling time of the real-time current sequence, and baseline current values ​​at the same sampling time in the interpretable current trajectory are read. The values ​​are subtracted point by point and combined in time order to generate the deviation current sequence.

[0086] Read the normal fluctuation limit parameter at the sampling time, compare the deviation current sequence with the normal fluctuation limit parameter point by point, take the result of the deviation current minus the normal fluctuation limit parameter when the deviation current is greater than the normal fluctuation limit parameter, and take zero when the deviation current is not greater than the normal fluctuation limit parameter to generate the initial additional load current sequence.

[0087] The initial additional load current sequence is filtered point by point using the minimum additional load threshold parameter. Values ​​less than the minimum additional load threshold parameter are set to zero, and the sequence length is kept consistent with the real-time current sequence to form the additional load current sequence.

[0088] S4. Calculation of Continuously Accumulated Evidence and Risk Index: (e.g.) Figure 3 As shown, the additional load current sequence is input into the motor current rise continuous discrimination, the continuous accumulation evidence is initialized to zero, the continuous positive value count is initialized to zero, and the continuous threshold parameter, continuous sampling threshold parameter, evidence increment coefficient, evidence attenuation coefficient, enhancement judgment threshold parameter and evidence upper limit parameter are read.

[0089] The additional load current values ​​are read sequentially according to the sampling time of the additional load current sequence, and the difference between the additional load current values ​​at adjacent sampling times is used to obtain the additional load current increment value.

[0090] The consecutive positive count is incremented when the additional load current value is greater than zero, and the consecutive positive count is set to zero when the additional load current value is less than or equal to zero.

[0091] When the continuous positive count is greater than or equal to the continuous sampling threshold parameter and the additional load current increment is greater than the enhanced judgment threshold parameter, the additional load current value is weighted and accumulated using the evidence increment coefficient to update the continuously accumulated evidence, and the continuously accumulated evidence is limited to no more than the evidence upper limit parameter.

[0092] When the number of consecutive positive values ​​is less than the continuous sampling threshold parameter or the incremental value of the additional load current is less than or equal to the enhancement judgment threshold parameter, the evidence attenuation coefficient is used to attenuate and update the continuously accumulated evidence, and the continuously accumulated evidence is limited to not less than zero.

[0093] The blocking risk index is calculated by comparing the continuous accumulated evidence with the continuous threshold parameter. When the continuous accumulated evidence is greater than or equal to the continuous threshold parameter, the blocking risk index is set to one. When the continuous accumulated evidence is less than the continuous threshold parameter, the blocking risk index is mapped to a range of zero to one according to the linear ratio of the continuous accumulated evidence to the continuous threshold parameter.

[0094] The continuously accumulated evidence is compared with the continuous threshold parameter. When the continuously accumulated evidence is greater than or equal to the continuous threshold parameter, the stuck state is identified. When the continuously accumulated evidence is less than the continuous threshold parameter, the normal state is identified.

[0095] When performing continuous motor current rise discrimination on the additional load current sequence, the continuous motor current rise discrimination does not directly judge the real-time current sequence, nor does it judge the current peak or single over-limit trigger. Instead, it calculates continuous accumulated evidence for the additional load current sequence and outputs the jamming risk index and jamming status indicator under the constraint of continuous threshold parameter.

[0096] The calculation of continuous accumulation evidence is based on the temporal continuity and cumulativeity of the additional load current sequence: when the additional load current sequence remains positive and shows an increasing trend during continuous sampling, the continuous accumulation evidence is incrementally updated; when the additional load current sequence falls back to near zero or shows a significant decline, the continuous accumulation evidence is decayed and updated. This calculation logic maps the gradual accumulation of additional load caused by jamming to a monotonic accumulation process of continuous accumulation evidence, ensuring that short-term current surges and short-term deviations caused by operating condition switching do not constitute a stable source of growth for continuous accumulation evidence. The continuous threshold parameter serves as the sole constraint on the continuous accumulation evidence to the jamming state indicator. When the continuous accumulation evidence reaches the continuous threshold parameter, a jamming risk index is generated and a jamming state indicator is determined; when the continuous accumulation evidence is below the continuous threshold parameter, a normal state indicator is determined.

[0097] Compared with existing technologies, the input object for continuously accumulating evidence is the additional load current sequence, which is obtained by the normal fluctuation limit parameter. The continuous judgment of the motor current rise is thus continuously judged around the additional load that exceeds the range that can be explained by normal operating conditions, so as to achieve the mechanism alignment of the progressive jamming, which is different from the false judgment path of operating condition floating caused by trend judgment of real-time current sequence.

[0098] S5. Generation of Level 3 Online Recognition Results: (e.g., ...) Figure 5 As shown, the jamming status indicator and jamming risk index are read, and the suspected threshold parameter is read.

[0099] When the jamming status is marked as normal, the jamming risk index is compared with the suspected threshold parameter. If the jamming risk index is less than the suspected threshold parameter, it is generated as normal; if the jamming risk index is greater than or equal to the suspected threshold parameter, it is generated as suspected jamming.

[0100] A confirmation block is generated when the block status indicator is set to "block status".

[0101] S6, Industrial signal output: such as Figure 6 As shown, a card block identification signal value is generated by state mapping based on the online identification result. When the online identification result is normal, the card block identification signal value is zero; when the online identification result is suspected card blockage, the card block identification signal value is one; and when the online identification result is confirmed card blockage, the card block identification signal value is two.

[0102] Read the full-scale value and quantization ratio, and generate a risk signal value by amplitude mapping according to the jamming risk index. When the jamming risk index is zero, the risk signal value is zero; when the jamming risk index is one, the risk signal value is the full-scale value; when the jamming risk index is between zero and one, the risk signal value is determined according to the quantization ratio.

[0103] Read the industrial communication interface and refresh cycle, and send the card identification signal value and risk signal value to the dense electromechanical control system according to the refresh cycle.

[0104] like Figure 7 As shown in the figure, the gray band represents the real-time motor current, the thin blue band is the baseline of the interpretable current trajectory generated by the model, and the light blue area is the normal fluctuation boundary that adapts to the operating conditions. The thickness widens or narrows over time to characterize the normal buoyancy caused by changes in slurry viscosity, feed, and reagent addition. Only when the gray current band locally exceeds the light blue area is it extracted as an additional load, i.e., a source of risk. This embodiment uses a dynamic reasonable range instead of a fixed threshold or a fixed slope threshold to structurally separate normal fluctuations from abnormal loads, significantly reducing false alarms under operating condition fluctuations, and providing calculable and referable boundary parameter objects for subsequent continuous discrimination.

[0105] like Figure 8As shown in the figure, the input includes a noisy current block, i.e., the current noise level sequence of the diffusion process, and a conditional current block, i.e., the real-time current. The current increment ΔI is calculated within the network and concatenated with I to form a 2-channel conditional feature. Both types of inputs enter the shared residual convolutional backbone of the same inverse denoising network, where consistent semantic fusion features are extracted under diffusion level embedding (FC64→FC128) modulation. The network then branches into two output heads: the left head outputs the baseline: the interpretable current trajectory Q, and the right head outputs the envelope: the normal fluctuation limit parameter B. The shared backbone in this embodiment ensures that the two outputs are generated synchronously under the same conditions and are semantically aligned; the dual-head design explicitly parameterizes the normal interpretable range, improving interpretability and reducing redundant modeling and computational overhead.

[0106] like Figure 9 As shown, the continuous discrimination logic is metaphorically represented by a water tank or power tank: the scattered red pulses on the left represent short-term additional load impacts. After entering the accumulation container, the liquid level hardly increases or quickly drops, indicating insensitivity to transient disturbances. The continuously increasing red blocks on the right represent gradual additional loads. Continuous injection causes the liquid level to rise steadily and exceed the threshold line, triggering a red alarm. This diagram corresponds to the continuous discrimination of S4, emphasizing the accumulation of evidence over time rather than triggering an alarm after a single exceedance. Its advantages include resistance to short-term impacts, reduced false alarms caused by random spikes, and earlier and more stable warnings for slow deterioration processes such as gradual jamming / continuous friction.

[0107] Example 2:

[0108] Example 1 is executed through an online identification system for thickener rake frame obstruction, including a data acquisition module, a one-dimensional diffusion characterization module, an additional load extraction module, a continuous discrimination module, and a result output module, wherein:

[0109] The data acquisition module is used to collect the current sampling value and operating status signal of the thickener rake frame drive motor. After dimension conversion and sliding truncation, it generates a real-time current sequence and operating status sequence. When the effective operating conditions are met, the sequence is spliced ​​together to form an identification input and output to the one-dimensional diffusion characterization module.

[0110] The one-dimensional diffusion characterization module has a built-in one-dimensional diffusion current characterization model consisting of a forward noise chain and a reverse noise chain. It is used to receive and identify the input, and to denoise the multi-level noise current sequence step by step with the real-time current sequence as the condition. Simultaneously, it generates interpretable current trajectory and normal fluctuation boundary parameters, and outputs them to the additional load extraction module.

[0111] The one-dimensional diffusion current characterization model consists of a forward noise-adding chain and a reverse denoising chain. The forward noise-adding chain applies hierarchical random perturbations to the real-time current sequence to form a multi-level noise current sequence. The reverse denoising chain performs denoising on the multi-level noise current sequence to generate an interpretable current trajectory and outputs normal fluctuation limit parameters. The core neural network of this model is located in the reverse denoising chain. Instead of using the common single-head output method of one-dimensional denoising networks, the reverse denoising chain adopts a structure of shared representation backbone network plus dual output heads. It explicitly expresses the normal rise caused by dense motor conditions using normal fluctuation limit parameters, enabling the continuous determination of the motor current rise to obtain a calculable and interpretable normal range.

[0112] The neural network input data of the inverse denoising chain includes: the current level sequence of the multi-level noise current sequence and the real-time current sequence; the neural network calculates the current increment sequence based on the real-time current sequence and concatenates the real-time current sequence and the current increment sequence to form a conditional feature tensor, the dimension of which is the sequence of length L multiplied by 2 channels; the current level sequence of the multi-level noise current sequence is used as the denoising feature tensor, the dimension of which is the sequence of length L multiplied by 1 channel; the inverse denoising chain also receives the diffusion level identifier and generates a diffusion level embedding vector through two fully connected layers. The first fully connected layer has 64 neurons and the second fully connected layer has 128 neurons. The diffusion level embedding vector is used to modulate the feature calculation of subsequent convolutional layers, so that the same network outputs interpretable current trajectories and normal fluctuation boundary parameters with consistent semantics under different noise levels.

[0113] The intermediate layer of the inverse denoising chain is a one-dimensional residual convolutional backbone network, which includes an input convolutional layer, a slowly varying residual stacked layer, a rapidly varying residual stacked layer, and a fusion convolutional layer. The input convolutional layer concatenates the denoised feature tensor and the conditional feature tensor after channel alignment, and uses one-dimensional convolution to form the basic feature map, with 64 convolutional channels. The slowly varying residual stacked layer consists of three residual blocks, each containing two one-dimensional convolutional layers with 64 convolutional channels. The convolution uses an increasing dilation rate to expand the receptive field, representing the slow load drift caused by changes in slurry viscosity, feed, and reagent addition in the interpretable current trajectory. The rapidly varying residual stacked layer also consists of three residual blocks, each containing two one-dimensional convolutional layers with 64 convolutional channels. The convolution uses small kernels to represent short-term current fluctuations, allowing the normal fluctuation boundary parameters to vary with the fluctuation intensity. The fusion convolutional layer concatenates the outputs of the slowly varying residual stacked layer and the fast varying residual stacked layer, and compresses them into a 64-channel fusion feature map through one-dimensional convolution.

[0114] The output layer of the inverse denoising chain has a dual-output head structure: the interpretable current trajectory output head applies a one-dimensional convolution to the fused feature map and outputs a single-channel sequence of length L to obtain the interpretable current trajectory; the normal fluctuation limit parameter output head applies a one-dimensional convolution to the fused feature map and outputs a single-channel sequence of length L to obtain the normal fluctuation limit parameter. The dual output heads share the same fused feature map, enabling the interpretable current trajectory and the normal fluctuation limit parameter to be generated synchronously under the same conditions. The normal fluctuation limit parameter changes with the operating pattern of the real-time current sequence, supporting the chain that calculates the deviation current based on the real-time current sequence and the interpretable current trajectory and extracts the additional load current sequence based on the normal fluctuation limit parameter. Structurally, the interpretable range of normal fluctuation is solidified into a referenceable parameter object, rather than relying on a fixed current threshold or a fixed slope threshold.

[0115] The additional load extraction module is used to receive real-time current sequences, interpretable current trajectories, and normal fluctuation limit parameters, calculate the deviation current sequence, and extract the part that exceeds the normal fluctuation limit. After filtering by the minimum threshold, the additional load current sequence is generated and output to the continuous discrimination module.

[0116] The continuous discrimination module is used to receive additional load current sequences, calculate continuous accumulated evidence, and combine it with continuous threshold parameters to generate a jamming risk index and jamming status indicator, which are then output to the result output module.

[0117] The output module receives the jamming status indicator and jamming risk index, generates a three-level online identification result including normal, suspected jamming, and confirmed jamming, and converts it into jamming identification signal and risk signal that can be read by the electronic control system, and sends it to the thickener electronic control system.

[0118] Example 3:

[0119] Based on Embodiments 1 and 2, an online identification method for thickener rake frame jamming is provided, including the following steps:

[0120] S1. The thickener rake frame drive motor is equipped with a current detection channel, and the controller operates according to a preset sampling period. Trigger acquisition. The current detection channel provides a current sample value at each sampling time, and the current sample value is recorded as... ,in This is the sampling time sequence number. The controller reads this information via the analog-to-digital converter interface. Read the scaling factor With bias ,right The unit conversion is performed by converting the units into uniform units. Multiply by the proportionality factor And superimposed bias To obtain a current value with uniform dimensions The unified dimension is the ampere dimension of the motor stator current. The controller sequentially samples the data. Write the current sample value sequence, which consists of the current values ​​of consecutive sample points.

[0121] The controller reads the preset sequence length parameter. Preset sequence length parameter The length is consistent with the input sequence of the inverse denoising chain. The controller performs a sliding truncation on the current sample value sequence, using a length of... The circular cache stores the latest A current value; when a new current value is collected. At that time, Write to the circular buffer and overwrite the earliest written current value to ensure the circular buffer always remains continuous. There are [number] sampling points. The controller reads the circular buffer in chronological order to form a buffer of length [length]. The real-time current sequence, denoted as The composition of the real-time current sequence is written as The time interval of the real-time current sequence is determined by the preset sampling period. limited.

[0122] The controller reads the operation enable signal, running direction signal, and speed feedback signal from the dense generator's electronic control system. The operation enable signal is denoted as... The direction of travel signal is denoted as The speed feedback signal is denoted as The controller operates at a preset sampling period. Consistent time base acquisition , , The operating status sequence is written according to the sampling time order. For scenarios where the speed feedback signal is an analog input, the controller reads the proportional coefficient. With bias The speed sampling values ​​are subjected to a dimensional transformation by multiplying the speed sampling values ​​by a scaling factor. And superimposed bias Obtain the speed feedback signal The controller reads the speed threshold parameter. Speed ​​threshold parameter Used to define the minimum rotational speed boundary for the rake frame to be in an effective rotational state.

[0123] The controller determines the operation enable signal and the speed feedback signal at each sampling time. When the operation enable signal is valid and the speed feedback signal is valid... Greater than the speed threshold parameter The controller constructs a recognition input. It employs the same sliding truncation strategy as the real-time current sequence for the operating state sequence, resulting in a sequence of length [length missing]. The sequence of running states, denoted as The composition of the running state sequence is written as The state vector of each sampling point .

[0124] The controller aligns and concatenates the real-time current sequence and the operating state sequence point by point according to the sampling time to form the recognition input, which is denoted as . When the input is composed of sampling points, it is written as... The identification vector for each sampling point When the input is organized in matrix form, it is written as: for The data matrix consists of four rows and four columns. The row index corresponds to the sampling time, and the four columns are, in order, the current value, the operation enable signal, the operation direction signal, and the speed feedback signal. The length of the input sequence is determined by a preset sequence length parameter. This constraint allows the recognition input to be consistent with the length of the input sequence of the inverse denoising chain and to be used for the construction of the conditional input for the one-dimensional diffusion current characterization model.

[0125] S2. The controller receives the operating status sequence and reads the operating enable signal. Speed ​​feedback signal With speed threshold parameter ,in This is the sampling time sequence number. For effective and Greater than The one-dimensional diffusion current characterization model is invoked, and the length is... The real-time current sequence is denoted as As a conditional input to the inverse denoising chain, the real-time current sequence is composed of continuous... The current values ​​at each sampling time are arranged in chronological order.

[0126] The controller performs a forward noise chain on the real-time current sequence and reads the diffusion level identifier. Level range parameters ,in This is the highest noise level. The controller will... Initialize to zero, setting the real-time current sequence to a zero-noise level sequence. The controller reads the noise intensity table. ,in The diffusion level is identified as The corresponding noise intensity. The controller increments the diffusion level indicator from 1 to... During the process, a random perturbation sequence is generated for each diffusion level identifier. each Generated by a random number generator with zero mean and unit variance; the controller sorts the random perturbation sequence according to noise intensity. Scale and stack point by point onto the previous diffusion level sequence To obtain the current diffusion level sequence The diffusion level label will be updated to the highest noise level.

[0127] In the reverse denoising chain, the controller organizes the current diffusion level sequence into a denoised feature tensor. The denoising feature tensor has a time length of A sequence tensor with one channel. The controller calculates the current increment sequence based on the real-time current sequence. The current increment sequence at the sampling time The value is determined by the sampling time. The current value is obtained by subtracting the current value at the previous sampling time, and the increment at the first point of the sequence is set to zero. The controller concatenates the real-time current sequence and the current increment sequence by channel to form a conditional feature tensor. The conditional feature tensor has a time length of Furthermore, it is a sequence tensor with two channels, with the two channels corresponding to the real-time current sequence and the current increment sequence, respectively.

[0128] The controller will identify the diffusion level. Two fully connected layers are input to generate diffusion-level embedding vectors. The controller maps diffusion level identifiers to level indicator vectors. The level indicator vector has a length of The vector is set to one only at the position corresponding to the diffusion level indicator. The first fully connected layer generates an intermediate vector of length 64 using the level indicator vector as input, and the second fully connected layer generates a diffusion level embedding vector of length 128 using the intermediate vector as input. The controller uses the diffusion level embedding vector to calculate the convolutional channel features of the modulated one-dimensional residual convolutional backbone network, and applies channel scaling modulation and channel bias modulation driven by the diffusion level embedding vector to the output channels of the convolutional layers.

[0129] The controller concatenates the denoised feature tensor and the conditional feature tensor along the channel dimension to form a three-channel input tensor. The input tensor has a time length of The input tensor has three channels: the current diffusion level sequence, the real-time current sequence, and the current increment sequence. The controller uses an input convolutional layer to apply a one-dimensional convolution to the input tensor to generate a basic feature map. The basic feature map retention time is The system has 64 channels, and each channel of the basic feature map is a response sequence calculated by sliding the corresponding convolutional kernel along the time dimension. The controller sequentially inputs the basic feature map into a slow-varying residual stack layer and a fast-varying residual stack layer. The slow-varying residual stack layer consists of three residual blocks with 64 convolutional channels within each block, and the convolutional dilation rates of the three residual blocks are set in ascending order of 1, 2, and 4. The fast-varying residual stack layer also consists of three residual blocks with 64 convolutional channels within each block, and the convolutional kernel length is set to 3 with a dilation rate of 1. The controller concatenates the output channels of the slow-varying residual stack layer and the fast-varying residual stack layer, and then compresses them through a fusion convolutional layer to form a 64-channel fused feature map. .

[0130] The controller employs a shared representation backbone network plus a dual-output head structure for the fused feature map. The interpretable current trajectory output head applies a one-dimensional convolution to the fused feature map to generate an interpretable current trajectory. This can be interpreted as the current trajectory having a time length of The single-channel sequence; the normal fluctuation limit parameter output head applies a one-dimensional convolution to the fused feature map to synchronously generate the normal fluctuation limit parameter. The normal fluctuation limit parameter is the time length. The single-channel sequence. The network parameters of the one-dimensional diffusion current characterization model were determined through offline training and stored in the controller storage medium.

[0131] The controller reads the diffusion transfer table. ,in For the noise reduction step size coefficient, The noise reinjection coefficient is used. The controller performs a denoising update once per diffusion level, using an interpretable current trajectory as the denoising target and using noise reinjection to maintain the randomness of the diffusion chain; the noise reinjection uses a random perturbation sequence. Generated according to the zero-mean unit variance rule and scaled according to noise intensity. The controller updates the current diffusion level sequence from the highest noise level to the lowest noise level according to the diffusion level identifier. The update is completed using the following formula, starting from the diffusion level identifier. The diffusion level is identified as Sequence transfer:

[0132]

[0133] in, The diffusion level is identified as The current level sequence at that time; The diffusion level is identified as The current level sequence at that time; The diffusion level is identified as Interpretable current trajectories generated at that time; The diffusion level is identified as follows in the diffusion transfer table. The corresponding denoising step size coefficient; The diffusion level is identified as follows in the diffusion transfer table. The corresponding noise reinjection coefficient; The diffusion level in the noise intensity table is labeled as The corresponding noise intensity; The diffusion level is identified as The corresponding random perturbation sequence; The sampling time sequence number; The length of the real-time current sequence; , , , All are time lengths The sequence is such that point-by-point operations correspond one-to-one with the time index.

[0134] The controller updates the diffusion level identifier in descending order until the diffusion level identifier is the lowest noise level, and completes the generation process of the one-dimensional diffusion current characterization model at the interpretable current trajectory and normal fluctuation limit parameters corresponding to the lowest noise level.

[0135] S3, The controller simultaneously obtains a length of [length] under the same time reference. The real-time current sequence, interpretable current trajectory, and normal fluctuation limit parameters, among which The preset sequence length parameter, This is the sampling time sequence number. The real-time current sequence is denoted as... The real-time current sequence consists of a vector of current values, written as... ,in Sampling time The current value. The interpretable current trajectory is denoted as... The current trajectory can be interpreted as consisting of a vector of baseline current values, written as... ,in Sampling time The baseline current value. The normal fluctuation limit parameter is denoted as... The normal fluctuation limit parameter consists of a limit value vector, written as ,in Sampling time The limit values. The current values, baseline current values, and limit values ​​all use the same current dimension.

[0136] The controller reads the current values ​​from the real-time current sequence in the order of sampling time, and reads the baseline current value at the same sampling time in the interpretable current trajectory. It then performs point-by-point subtraction at each sampling time to obtain the deviation current value. The controller combines the deviation current values ​​from each sampling time in chronological order to form a deviation current sequence, denoted as […]. The deviation current sequence is written as ,in Sampling time The deviation current value is determined. To facilitate point-by-point calculations over time, the controller concatenates the real-time current sequence, interpretable current trajectory, and normal fluctuation limit parameters column-wise to form an aligned data matrix. The aligned data matrix is The row consists of three columns: current value, baseline current value, and limit value. The row index is the sampling time number. to .

[0137] The controller reads the normal fluctuation limit parameters at each sampling time and compares the deviation current sequence with the normal fluctuation limit parameters point by point. At the sampling time... When the deviation current value is greater than the limit value, the difference between the deviation current value and the limit value is calculated as the initial additional load current value; at the sampling time... When the deviation current value is not greater than the limit value, the initial additional load current value is set to zero. The controller combines the initial additional load current values ​​at each sampling time in chronological order to form an initial additional load current sequence, which is denoted as [initial additional load current sequence]. The initial additional load current sequence is written as ,in Sampling time The initial additional load current value. Point-by-point comparison and point-by-point calculation in the aligned data matrix. The process is completed under the condition that the row index is consistent, so that the deviation current value and the limit value are paired at the same sampling time.

[0138] The controller reads the minimum additional load threshold parameter. The minimum additional load threshold parameter uses the same dimensions as the current. The controller performs point-by-point filtering on the initial additional load current sequence at the sampling time. The initial additional load current is less than The value will be set to zero at the sampling time. The initial additional load current is greater than or equal to The values ​​remain unchanged. The controller combines the filtered values ​​from each sampling time in chronological order to form an additional load current sequence, denoted as . The additional load current sequence is written as ,in Sampling time The additional load current value. The controller maintains the additional load current sequence length as... And keep the sampling time index of the additional load current sequence consistent with the real-time current sequence.

[0139] S4. The controller inputs the additional load current sequence into the motor current and continuously judges the rise. This is the sampling time sequence number. The preset sequence length parameter is used. The additional load current sequence is denoted as... Additional load current sequence is The additional load current values ​​are composed of samples in chronological order, and the additional load current values ​​are denoted as follows: Sampling time index Pick to When the controller continuously accumulates evidence during the process of determining whether the motor current is rising, it records this as... Counting consecutive positive values ​​is denoted as Sampling time corresponding Set to zero, and set the sampling time to zero. corresponding Set to zero. The controller reads the continuous threshold parameter. Continuous sampling threshold parameter Evidence increment coefficient Evidence attenuation coefficient Enhanced judgment threshold parameters With the upper limit of evidence parameter ,in For positive coefficients, For the value to fall into coefficient, and Use with additional load current value Consistent current dimensions.

[0140] The controller reads the additional load current values ​​sequentially according to the sampling time. The incremental value of the additional load current is obtained by subtracting the additional load current values ​​at adjacent sampling times. Additional load current increment From sampling time Additional load current value Downsampling time The additional load current value is obtained at the sampling time. The additional load current increment is set to zero. The controller constructs a discrimination data matrix. The discriminant data matrix is Row 2, column 2, the first column is the sampling time. to The sequence of additional load current values, with the second column representing the sampling time. to The sequence of additional load current increments, such that each row corresponds to the same sampling time. Binary data set.

[0141] The controller updates the consecutive positive value count point by point based on the discrimination data matrix. Sampling time satisfy When, the count will be continuously positive. Counting consecutive positive values ​​at the previous sampling time. Add one to the base; sampling time satisfy When, the count will be continuously positive. Set to zero. The controller performs an enhanced trend determination at each sampling time. The condition for enhanced trend determination is a continuous count of positive values. Greater than or equal to the continuous sampling threshold parameter And the additional load current increment value Greater than the enhanced judgment threshold parameter .

[0142] The controller updates the continuously accumulated evidence at each sampling time based on the enhancement trend determination result. The continuous accumulated evidence update adopts a unified calculation formula that includes incremental update, decay update and amplitude limiting:

[0143]

[0144] in, Sampling time The continuous accumulation of evidence, Sampling time The continuous accumulation of evidence, This is the upper limit parameter for evidence. The evidence increment factor, Sampling time Additional load current value, Sampling time Counting consecutive positive values, For continuous sampling threshold parameters, Sampling time The additional load current increment, To enhance the judgment threshold parameter, The evidence attenuation coefficient, For upper limit operators, To constrain the lower bound operator, For logical AND, For logical OR.

[0145] At each sampling time, the controller generates a jamming risk index and determines the jamming status indicator based on continuously accumulated evidence and a continuous threshold parameter. The jamming risk index is denoted as... The value of the jamming risk index is limited to [specific range]. Sampling time satisfy At that time, the risk index will be blocked. Set it to 1, and record the blocking status as 1. Set to 1; sampling time satisfy In time, evidence will continue to accumulate. Divide by the duration threshold parameter Obtain the ratio value and use it as the jamming risk index. And mark the stuck status. Set to zero. The controller will sample the time. Corresponding jamming risk index With the stuck status indicator Write the online discrimination result variable for real-time reading and control linkage of the dense electromechanical control system.

[0146] S5, Controller at sampling time Read the jamming status indicator and jamming risk index, and read the suspected threshold parameters. The jamming status indicator is denoted as... ,in Indicates a normal status indicator. This indicates the jamming status. The jamming risk index is denoted as... , For interval The real number within. The suspected threshold parameter is denoted as... , For interval The real number within is used to distinguish between normal and suspected jamming under normal status indication conditions.

[0147] The controller will sample the time. The decision input is organized as a decision vector. The decision vector is used to drive the generation of online recognition results. The controller records the online recognition results as... , The value is one of three: normal, suspected jamming, or confirmed jamming. To ensure stable transmission of online identification results in industrial communication, the controller synchronously generates a ternary state vector of the online identification results. ,in This indicates that the online recognition result is normal and all other components are zero. This indicates that the online identification result is suspected to be a jam and all other components are zero. This indicates that the online identification result confirms the card is blocked and all other components are zero.

[0148] Controller in When entering the normal state indicator branch, for and Perform a comparison operation. The comparison operation compares dimensionless scalars of the same dimension, and the comparison result is a Boolean value. ,in express , express The controller is in The online recognition result is generated as normal and set to normal. ; Controller in The online identification result is generated as a suspected blockage, and then... The controller is not modified in this branch. and Based solely on suspected threshold parameters Discretize and map the online recognition results.

[0149] Controller in Upon entering the jam status indicator branch, the online identification result is directly generated as "confirmed jam," and then set... The controller is not in this branch. and Performing comparison operations makes the blocking status flag a deterministic constraint on the online identification result, thereby ensuring that the blocking status flag triggered by continuously accumulated evidence can be unconditionally inherited by the online identification result and used in the online processing logic.

[0150] S6, Controller at refresh time Read the online recognition results and the jamming risk index. The online recognition results are recorded as follows: The online identification result is one of three values: normal, suspected jamming, or confirmed jamming. The jamming risk index is denoted as... , For interval Real numbers within.

[0151] The controller performs state mapping based on the online identification results to generate a card resistance identification signal value, which is denoted as [value missing]. , It is a non-negative integer. If the online recognition result is normal, then... Set as When the online identification result indicates suspected obstruction, Set as When the online identification result confirms the card is blocked, Set as The controller organizes the card identification signal value into a one-dimensional scalar and uses it for writing to industrial communication registers.

[0152] The controller reads the full-scale value and quantization ratio, and performs amplitude mapping to generate a risk signal value. The full-scale value is denoted as... , This represents the maximum amplitude of the risk signal. The quantization ratio is denoted as... , `value` is a positive integer used to define the discrete quantization level of the full-scale value. The controller records the risk signal value as... , With full scale value A consistent amplitude dimension is adopted. The jamming risk index satisfies... At that time, Set as The jamming risk index meets the requirements. At that time, Set as The jamming risk index meets the requirements. At that time, the controller first calculates the quantization level index. Quantization-level index Through the and The product is rounded to the nearest whole number, and then... Limited to Within the integer range; the controller then uses and The ratio represents the quantized proportionality coefficient, and this proportionality coefficient is compared with the full-scale value. Multiply to obtain the risk signal value The controller will Arrange the data vectors to be sent in sequence. The data vectors sent are used to construct write payloads on the industrial communication interface.

[0153] The controller reads the industrial communication interface and refresh cycle. The industrial communication interface is denoted as... The refresh cycle is denoted as , This is a positive time parameter. The controller operates at a refresh cycle. In the triggered periodic task, a data vector will be sent. Write to industrial communication interface The corresponding data area includes a card resistance identification signal value writing unit and a risk signal value writing unit; the card resistance identification signal value writing unit stores... Risk signal values ​​are stored in the writing unit. The controller maintains the element order of the transmitted data vector after each write operation, enabling the dense electromechanical control system to parse the jamming identification signal value and the risk signal value at a fixed position.

Claims

1. A method for online identification of jamming in a thickener rake frame, characterized in that: Includes the following steps: S1. Obtain the real-time current sampling value of the thickener rake frame drive motor and form a real-time current sequence; receive the rake frame running status signal and form a running status sequence to obtain the identification input; S2. Based on the identified input, call the one-dimensional diffusion current characterization model. The one-dimensional diffusion current characterization model includes a forward noise chain and a reverse noise chain. The forward noise chain superimposes random disturbances on the real-time current sequence to form a multi-level noise current sequence. The reverse noise chain uses the real-time current sequence as a condition to denoise the multi-level noise current sequence from high noise level to low noise level, generating an interpretable current trajectory based on the normal operating condition assumption, and simultaneously outputs the normal fluctuation limit parameter characterizing the normal operating condition current floating range. S3. Calculate the deviation current sequence based on the real-time current sequence and the interpretable current trajectory, and extract the part of the deviation current sequence that exceeds the normal fluctuation limit according to the normal fluctuation limit parameter to form the additional load current sequence. S4. Perform continuous motor current rise discrimination on the extra load current sequence, initialize continuous accumulation evidence and continuous positive value count, read the extra load current value sequentially according to the sampling time and calculate the extra load current increment value, perform incremental update or decay update on the continuous accumulation evidence and set upper and lower limits according to the joint judgment result of the continuous positive value count and the extra load current increment value, and generate a jamming risk index and jamming status indicator with a value range of 0 to 1 based on the comparison result of continuous accumulation evidence and continuous threshold parameter. S5. Generate online identification results based on the jamming status indicator and jamming risk index. The online identification results include at least three statuses: normal, suspected jamming, and confirmed jamming.

2. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: In step S2, the execution process of the forward noise chain includes the following steps: S201. Initialize the diffusion level identifier to the lowest level and use the real-time current sequence as the zero noise level sequence; read the pre-stored noise intensity table, which contains noise intensity values ​​corresponding one-to-one with each diffusion level. S202. According to the diffusion level from low to high, generate a random perturbation sequence with zero mean and unit variance for each level, scale it according to the corresponding noise intensity, and then add it point by point to the current sequence of the previous level to obtain the noise current sequence corresponding to the current diffusion level, and finally form a multi-level noise current sequence covering all diffusion levels.

3. The online identification method for thickener rake frame jamming according to claim 2, characterized in that: The execution process of the reverse denoising chain includes the following steps: S211. Calculate the current difference between adjacent sampling points based on the real-time current sequence, generate a current increment sequence, and set the increment of the first point of the sequence to 0; concatenate the real-time current sequence and the current increment sequence according to the channel dimension to form a conditional feature tensor; use the current level sequence of the multi-level noise current sequence as the denoising feature tensor. S212. Map the current diffusion level identifier to a level indicator vector, and generate a diffusion level embedding vector through two fully connected layers; the diffusion level embedding vector is used to modulate the channel features of the convolutional layer inside the inverse denoising chain, so that the interpretable current trajectory output under different noise levels is semantically consistent with the normal fluctuation limit parameter. S213. Input the denoising feature tensor and the conditional feature tensor into the shared characterization backbone network to extract and fuse features, and generate the interpretable current trajectory and normal fluctuation boundary parameters of the current level synchronously through dual output heads. S214. Update the current level sequence of the multi-level noise current sequence step by step in order of diffusion level from high to low, until the diffusion level indicator drops to the lowest level, and output the final interpretable current trajectory and normal fluctuation limit parameters. The reverse denoising chain adopts a shared representation backbone network plus a dual-output head structure. The shared representation backbone network is a one-dimensional residual convolutional network, which sequentially includes an input convolutional layer, a slowly varying residual stacked layer, a fast varying residual stacked layer, and a fusion convolutional layer. The input convolutional layer performs channel alignment and concatenation of the denoised feature tensor and the conditional feature tensor to generate a basic feature map. The slowly varying residual stacked layer consists of 3 residual blocks and uses increasing dilation rate convolution to extract the slow-varying current drift feature. The fast varying residual stacked layer consists of 3 residual blocks and uses small convolutional kernels to extract the short-time current fluctuation feature. The fusion convolutional layer concatenates and compresses the two feature paths to generate a fusion feature map. The dual output head includes an interpretable current trajectory output head and a normal fluctuation limit parameter output head, which share the same fused feature map. They are respectively output through one-dimensional convolution to interpretable current trajectory and normal fluctuation limit parameter, so that interpretable current trajectory and normal fluctuation limit parameter are generated synchronously under the same operating conditions.

4. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: In S2, the multi-level noise current sequence starts from the diffusion level. arrive The stepwise denoising process follows the diffusion transfer formula: in, The diffusion level is identified as The current level noise current sequence at that time, After noise reduction Level of current sequence; The diffusion level is identified as Interpretable current trajectories generated at that time; diffusion level The corresponding denoising step size coefficient, diffusion level The corresponding noise reinjection coefficients are pre-stored in the diffusion transfer table; diffusion level The corresponding noise intensity, diffusion level The corresponding zero-mean, unit-variance random perturbation sequence; The sampling time number is used. All sequences have the same length, and the point-by-point operations correspond one-to-one with the time index.

5. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: S3 specifically includes the following steps: S301. Read the real-time current values ​​sequentially according to the sampling time of the real-time current sequence, and read the baseline current values ​​at the same sampling time in the interpretable current trajectory. Subtract them point by point and combine them in time order to generate the deviation current sequence. S302. Read the normal fluctuation limit parameter at the sampling time, compare the deviation current sequence with the normal fluctuation limit parameter point by point, take the result of the deviation current minus the normal fluctuation limit parameter when the deviation current is greater than the normal fluctuation limit parameter, and take zero when the deviation current is not greater than the normal fluctuation limit parameter to generate the initial additional load current sequence. S303. Perform minimum additional load threshold parameter filtering on the initial additional load current sequence point by point, set the values ​​of those less than the minimum additional load threshold parameter to zero, and keep the sequence length consistent with the real-time current sequence to form an additional load current sequence.

6. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: In step S4, the calculation process for the continuously accumulating evidence and the blocking risk index includes: Initialize continuous accumulation of evidence Continuous positive value counting Read the additional load current values ​​in the order of sampling time. Calculate the additional load current increment at adjacent sampling times. ; like Then, consecutive positive values ​​are counted. If Then, consecutive positive values ​​are counted. ; The continuous accumulation of evidence Update using the following formula: in, This is the upper limit parameter for evidence; A positive coefficient of increase in evidence; This is the continuous sampling threshold parameter; To enhance the judgment threshold parameter; The evidence attenuation coefficient is defined as a value in the interval [0,1]. Blockage Risk Index Calculate using the following formula: in, This is a continuous threshold parameter; the range of the jamming risk index is 0 to 1. When the jamming risk index is 1 and the jamming status is determined; when At that time, the risk index of blockage And determine the normal status indicator.

7. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: S1 specifically includes: acquiring drive motor current sampling values ​​according to a preset sampling period, and forming a current sampling value sequence after dimensional conversion; and then, according to a preset sequence length parameter... Perform a sliding truncation on the current sample value sequence to generate a length of The real-time current sequence, wherein the preset sequence length parameter The length of the input sequence is the same as that of the inverse denoising chain; The system receives a running enable signal, a running direction signal, and a speed feedback signal to form a running state sequence. When the running enable signal is valid and the speed feedback signal is greater than the speed threshold parameter, the real-time current sequence and the running state sequence are aligned and spliced ​​according to the time reference to form the identification input.

8. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: S5 includes reading the suspected threshold parameter. If the current status is normal and the jamming risk index is less than the suspected threshold parameter, a normal online identification result is generated. If the jamming risk index is greater than or equal to the suspected threshold parameter, a suspected jamming online identification result is generated. If the current status is jamming, a confirmed jamming online identification result is generated.

9. The online identification method for thickener rake frame jamming according to claim 1, characterized in that: It also includes S6, in which: The online identification results are discretized and mapped to the following card identification signal values: 0 for normal state, 1 for suspected card jamming, and 2 for confirmed card jamming. Read the full-scale value and quantization ratio, and linearly map the jamming risk index in the 0-1 range to the corresponding amplitude risk signal value; The jamming identification signal value and risk signal value are sent to the dense electromechanical control system through the industrial communication interface according to the preset refresh cycle.

10. A thickener rake frame jamming online identification system, used to execute the thickener rake frame jamming online identification method according to any one of claims 1-9, characterized in that: It includes a data acquisition module, a one-dimensional diffusion characterization module, an additional load extraction module, a continuous discrimination module, and a result output module, among which: The data acquisition module is used to collect the current sampling value and operating status signal of the thickener rake frame drive motor, generate real-time current sequence and operating status sequence through dimension conversion and sliding truncation, and splice them together to form an identification input when the effective operating conditions are met, and output to the one-dimensional diffusion characterization module. The one-dimensional diffusion characterization module has a built-in one-dimensional diffusion current characterization model consisting of a forward noise chain and a reverse noise chain. It is used to receive the identification input, perform step-by-step noise reduction on the multi-level noise current sequence based on the real-time current sequence, synchronously generate interpretable current trajectory and normal fluctuation boundary parameters, and output them to the additional load extraction module. The additional load extraction module is used to receive the real-time current sequence, the interpretable current trajectory and the normal fluctuation limit parameter, calculate the deviation current sequence and extract the part that exceeds the normal fluctuation limit, and generate an additional load current sequence after filtering by the minimum threshold, and output it to the continuous discrimination module. The continuous discrimination module is used to receive the additional load current sequence, calculate the continuous accumulated evidence and combine it with the continuous threshold parameter to generate the jamming risk index and jamming status identifier, and output them to the result output module. The result output module is used to receive the jamming status indicator and jamming risk index, generate a three-level online identification result including normal, suspected jamming, and confirmed jamming, and convert it into jamming identification signal and risk signal that can be read by the electronic control system, and send it to the thickener electronic control system.

Images