Intelligent monitoring and control method and device for fault of overflow type ball mill

Abstract

The application discloses a kind of intelligent monitoring and control method and device for overflow type ball mill, belong to intelligent monitoring technical field of mine machinery.The device is through integration signal conditioning and adaptive acquisition hardware subsystem, edge inference calculation core of physical priori guidance and industrial control execution interface unit, real-time acquisition overflow type ball mill Multi-source perception signal, and through hardware closed loop control loop generation isometric vibration signal flow;Subsequently call fault mechanism mapping knowledge base, initialize the operation parameter of feature extraction circuit in each processing channel, obtain the decoupled component-specific feature vector;Finally, real-time calculation of physical state index, adaptive weighted fusion is carried out to component-specific feature vector, and the health state evaluation instruction and control instruction of key component are output in parallel in fault discriminator.The application can effectively overcome the problem that existing general monitoring equipment cannot adapt to speed drift, lack of edge intelligence and control response lag from hardware level.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for mining machinery, specifically relating to a method and device for intelligent fault monitoring and control of overflow ball mills. Background Technology

[0002] In modern large-scale mining grinding processes, the stable operation of overflow ball mills is directly related to the safety of the mineral processing line. Due to the lack of forced discharge and the harsh alternating load environment, the spindle speed of this equipment exhibits a non-linear change with the hardness of the ore and the liquid level, leading to frequent occurrences of liner loosening, gear uneven loading, and bearing temperature rise.

[0003] Currently, on-site monitoring of this type of heavy equipment mainly includes three technical solutions. The first is to use a general industrial data acquisition card or vibration transmitter to acquire vibration signals through fixed-frequency sampling and transmit the data to a host computer for offline analysis. The second is to use an external encoder in conjunction with a high-end acquisition instrument, which uses encoder pulses to trigger A / D conversion to achieve equal-interval sampling in the angle domain. The third is to use a general edge gateway or data acquisition and monitoring control to transmit multi-source data such as vibration, temperature, and current to the cloud or central server, and use back-end algorithms for fault diagnosis and condition assessment.

[0004] However, these existing methods all have certain limitations. First, the fixed-frequency sampling method of general industrial data acquisition cards or vibration transmitters cannot adapt to the speed changes under non-steady-state operating conditions, and frequency domain distortion is prone to occur at the source end. Although the method of using external hardware encoders with high-end acquisition instruments can achieve synchronous sampling, its installation accuracy is low, and the encoder is easily damaged in dusty environments, resulting in high maintenance costs. Although general edge gateways have certain transmission capabilities, they cannot understand the firmware logic of the specific faults in overflow ball mills. They can only transmit data and cannot actively adjust the monitoring strategy based on the physical characteristics of the liners or gears. When the equipment has a weak early fault, the device lacks the physical prior frequency focusing capability, causing the key features to be submerged in background noise and failing to trigger protection actions.

[0005] Therefore, this invention proposes a fault intelligent monitoring and control method and device for overflow ball mills, aiming to accurately align signals in the time and angular domains at the hardware level, thereby avoiding the inherent flaw of asynchronous sampling in traditional equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a method and device for intelligent fault monitoring and control of overflow ball mills, which effectively overcomes the problems of existing general monitoring equipment being unable to adapt to speed drift, lacking edge intelligence, and having lag in control response from the hardware level.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a fault intelligent monitoring and control method for overflow ball mills, comprising the following steps:

[0008] S1. Data Acquisition and Closed-Loop Control: Real-time acquisition of multi-source sensing signals from the overflow ball mill, including at least the spindle speed signal, slurry level signal, and raw vibration signals collected by vibration sensors deployed on key components; Based on the spindle speed signal and the slurry level signal, the trigger cycle of the analog-to-digital converter is dynamically adjusted through a hardware closed-loop control loop to synchronously acquire the raw vibration signals with equal-angle step alignment, generating a hardware-aligned equal-angle vibration signal stream;

[0009] S2. Physically Prior-Guided Feature Extraction: The fault mechanism mapping knowledge base pre-installed in the edge computing hardware is invoked to divert the equal-angle vibration signal stream to different physical meaning processing channels according to its fault correlation; using the prior information in the fault mechanism mapping knowledge base, the operation parameters of the feature extraction circuit in each processing channel are initialized, so that the feature extraction circuit directionally enhances the physical frequency band related to the fault of a specific component, and extracts the decoupled component-specific feature vectors in parallel from each processing channel;

[0010] S3. State Awareness and Dynamic Decision-Making: Real-time calculation of physical state indicators of the original vibration signal or equal-angle vibration signal stream; Based on the physical state indicators, dynamic gating signals are generated, and the component-specific feature vectors output by each processing channel are adaptively weighted and fused to generate a global state feature representation; The global state feature representation is input into the component-specific fault discriminator, and health status assessment instructions for key components of the overflow ball mill are output in parallel; According to the health status assessment instructions, when a fault risk is determined, control signals for controlling the operating state of the overflow ball mill are automatically generated and output.

[0011] Furthermore, in step S1, dynamically adjusting the trigger cycle of the analog-to-digital converter through a hardware closed-loop control loop specifically includes:

[0012] The effective spindle speed is obtained by performing a moving average filter on the real-time acquired spindle speed signal. Its expression is:

[0013]

[0014] In the formula, For buffer depth, The instantaneous value of the spindle speed in the i-th sampling period;

[0015] Based on the slurry level signal, the level weighting coefficient is obtained by looking up a table. ;

[0016] Based on the reference sampling period, the effective rotational speed and the liquid level weighting coefficient The theoretical triggering period is calculated based on the closed-loop control law. Its expression is:

[0017]

[0018] In the formula, N is the preset number of virtual sampling points per cycle;

[0019] The theoretical triggering cycle is limited to a preset safety threshold by a hardware saturation clamping circuit, thus generating the final execution cycle.

[0020] The hardware timer is updated using the final execution cycle, and the analog-to-digital converter is directly triggered to perform multi-channel synchronous sampling.

[0021] Furthermore, in step S2, the fault mechanism mapping knowledge base is pre-set with the association between key components and fault characteristics. The association includes at least component type, fault type, dominant frequency range, location of sensitive measuring point, and dominant vibration direction.

[0022] Specifically, the equal-angle vibration signal stream is diverted to different physical meaning processing channels according to the fault mechanism mapping knowledge base, and the equal-angle vibration signal stream is routed to the low-frequency channel corresponding to the liner monitoring, the medium-frequency channel corresponding to the gear monitoring, and the high-frequency channel corresponding to the bearing monitoring.

[0023] Further, in step S2, initializing the operational parameters of the feature extraction circuits in each processing channel using the prior information in the fault mechanism mapping knowledge base specifically includes:

[0024] For each processing channel, obtain its target center frequency. ;

[0025] Using the center frequency of the target Matching wavelet basis functions The convolution kernel register of the convolutional neural network within this channel is initialized, enabling the feature extraction circuit to possess hardware sensitivity to the target physical frequency band from the initial power-on stage. and Preset parameters for firmware. Where n is the sampling frequency and n is the discrete time index.

[0026] Furthermore, for the high-frequency channel, a hardware-level three-stage noise suppression sub-step is included before the feature extraction step:

[0027] Impulsive rarefaction sub-step: Calculate the spectral kurtosis of the high-frequency channel signal. And using a spectral kurtosis calculation circuit, only spectral kurtosis is allowed. Higher than the preset threshold The impact component passes through, resulting in a sparsified signal;

[0028] Envelope enhancement sub-step: Perform Hilbert transform on the sparsified signal to extract its envelope to enhance the impact energy, and obtain the envelope-enhanced signal;

[0029] Amplitude gating sub-step: Using a hardware comparator, filter out signals in the envelope enhancement signal that are below the background noise threshold. The components are used to obtain the denoised signal used for feature extraction.

[0030] Furthermore, in step S3, the physical state index includes at least the kurtosis index of the original signal. and the energy centroid used for quantifying energy distribution under chemical conditions ;

[0031] The step of generating a dynamic gating signal based on the physical state indicators specifically includes:

[0032] When the kurtosis index of the original signal is detected When the gain is increased, a gate signal is generated to improve the gain of the high-frequency channel. The gating signal The expression is:

[0033]

[0034] In the formula, For the Sigmoid function, This represents the statistical mean of the high-frequency channel characteristics. and These are the learnable parameters and bias terms for the high-frequency channel, respectively.

[0035] When the energy center of gravity is detected When shifting to lower frequencies with significant oscillations, a gating signal is generated to improve the gain of the low-frequency channel. Its expression is:

[0036]

[0037] In the formula, For the deep features of low frequency channels Norm, This represents the effective value of the Y-axis vibration at the FE end. and These are the learnable parameters and bias terms for the low-frequency channel, respectively.

[0038] When the energy center of gravity is detected When the frequency band is in the mid-frequency range and sideband asymmetry occurs, a gate signal is generated to improve the gain of the mid-frequency channel. Its expression is:

[0039]

[0040] In the formula, This represents the statistical mean of the mid-frequency channel characteristics. The energy difference in the gear meshing sideband detected by the hardware. and These are the learnable parameters and bias terms for the intermediate frequency channel, respectively.

[0041] The component-specific feature vectors output from each processing channel are adaptively weighted and fused to generate a global state feature representation. The formula for the global state feature representation is as follows:

[0042]

[0043] in, This represents the global state features. , , These are component-specific feature vectors output from the low-frequency, mid-frequency, and high-frequency channels, respectively. This represents the element-wise multiplication operator.

[0044] Furthermore, in step S3, after the step of outputting health status assessment instructions for key components of the overflow ball mill in parallel, a multi-objective collaborative calibration step is also included:

[0045] Introduce a smoothing damping coefficient with hysteresis to the health status assessment command. Smooth logic through instructions The transient decision signal is converted into a steady-state control signal containing a confidence interval; the instruction smoothing logic... The expression is:

[0046]

[0047] In the formula, To control the number of sampling frames in the period, No. Frame data corresponds to the first Output gain of class state, As the target reference state, This represents the total number of state categories.

[0048] By setting up dedicated gain compensation loops for each key component and introducing an automatic gain control mechanism based on energy contribution, the sensitivity of the component's dedicated eigenvector is dynamically calibrated, generating the component detection deviation. The detection deviation The expression is:

[0049]

[0050] In the formula, These are the detection deviations for the liner, gear, and bearing channels, respectively. This is the gain balance coefficient;

[0051] Calculate the theoretical gating state vector based on the physical state indices. And by calculating the current gate output value The mean square error of the theoretical gated state vector is used to generate a physical consistency check item. The physical consistency check item The expression is:

[0052]

[0053] in, For the first The current gated output value of the frame data;

[0054] Based on the instruction smoothing logic Component inspection deviation Physical consistency check item Through overall control strategy Generate the final control signal, where This is the strategy correction coefficient.

[0055] A fault intelligent monitoring and control device for overflow ball mills, comprising:

[0056] The signal conditioning and adaptive acquisition hardware subsystem includes:

[0057] A multi-channel analog front end is used to connect to speed sensors, level transmitters and vibration sensors to acquire multi-source sensing signals in real time.

[0058] The speed pulse shaping circuit is used to shape the raw signal from the speed probe into a standard pulse.

[0059] An FPGA-based sampling timing controller, with built-in hardware counters, phase latches, and saturation clamping circuits, is used to dynamically adjust the trigger period of the analog-to-digital converter based on the multi-source sensing signals through a hardware closed-loop control loop, thereby generating an equally aligned vibration signal stream.

[0060] The core of edge reasoning computation guided by physical priors includes:

[0061] Heterogeneous processing system-on-a-chip, including general-purpose processing cores and hardware acceleration cores;

[0062] Non-volatile memory is used to store a fault mechanism mapping knowledge base, which is pre-set with the association between key components and fault characteristics;

[0063] The hardware acceleration core is used to initialize the operation parameters of the feature extraction circuit according to the prior information in the fault mechanism mapping knowledge base, extract the decoupled component-specific feature vectors in parallel, generate dynamic gating signals based on physical state indicators for adaptive weighted fusion, and finally output health status assessment instructions.

[0064] An industrial control execution interface unit, comprising:

[0065] The digital I / O interface is used to directly output an emergency stop signal to the safety interlock circuit of the overflow ball mill according to the health status assessment command.

[0066] An industrial fieldbus interface is used to send closed-loop control commands to the actuator to adjust feeding or operating parameters based on the health status assessment commands.

[0067] Furthermore, the FPGA-based sampling timing controller is specifically used for:

[0068] The theoretical triggering period is calculated based on the real-time effective rotational speed and the closed-loop control law corrected by the liquid level weighting coefficient.

[0069] The built-in saturation clamping circuit limits the theoretical triggering cycle to a preset safety threshold, generating the final execution cycle.

[0070] The final execution cycle is written to the reload register of the programmable timer to update the interrupt frequency;

[0071] When the timer interrupt overflows, the analog-to-digital converter is directly triggered to perform multi-channel synchronous conversion, and the converted digital signal frame is labeled with angle-time dual tags.

[0072] Furthermore, the hardware acceleration core is a DSP core, which contains wavelet convolution kernel registers corresponding to the component types in the fault mechanism mapping knowledge base; the DSP core is used to call the dominant frequency range in the fault mechanism mapping knowledge base, generate the corresponding wavelet basis function, and load the parameters of the wavelet basis function into the wavelet convolution kernel register to complete the initialization of the feature extraction circuit.

[0073] The edge inference computing core is also used to calculate the kurtosis index and energy centroid of the original signal in real time, and dynamically generate gain control signals for each processing channel through a state-aware gating circuit based on the calculation results.

[0074] The device also includes a dynamic gain adjustment circuit, which, according to the gain control signal, attenuates the signal of the channel where the high-energy component is located and amplifies the signal of the channel where the weak fault component is located through automatic gain control logic, so as to achieve equalization state capture under multi-source coupling conditions.

[0075] The beneficial effects of the above technical solution are as follows:

[0076] 1. This invention achieves precise angular sampling under unsteady conditions through hardware closed-loop control with dual-source feedback of rotational speed and liquid level. Specifically, a hardware closed-loop control loop based on dual-source feedback of rotational speed and liquid level is constructed within the monitoring terminal. The rotational speed pulses from the magnetoelectric probe and the slurry level data from the level gauge interface are synchronously latched via a high-speed I / O interface. The theoretical trigger interval is dynamically calculated using load characteristics, and the final execution cycle is generated through boundary saturation protection by a saturation clamping circuit. This directly drives a programmable timer to control an analog-to-digital converter to complete synchronous acquisition with equiangular step size alignment, generating a hardware-aligned equiangular vibration signal stream. This method eliminates the need for an external encoder, relying on hardware-level closed-loop control to achieve adaptive tracking of the sampling frequency to rotational speed fluctuations. This fundamentally solves the problems of spectral ambiguity and order sideband interference caused by fixed-frequency sampling under unsteady conditions, ensuring dual alignment of the vibration signal in both the time and angular domains, and significantly improving the sensing accuracy of the device under unstable conditions.

[0077] 2. This invention embeds physical prior logic into the hardware processing flow by pre-setting a fault mechanism mapping knowledge base, thereby achieving targeted enhancement of component-specific features. Specifically, a fault mechanism mapping knowledge base is pre-set within the device chip, deeply embedding the physical mechanism of the equipment into the hardware processing flow. Multi-channel vibration signals are distributed to dedicated monitoring channels for liners, gears, and bearings through logic splitting. In each channel, wavelet basis functions matching the target physical frequency band are used to initialize the convolution kernel register, enabling the device to possess hardware sensitivity to specific fault features upon power-up. This method, by using physical prior logic as a guide for feature extraction, avoids the poor physical interpretability of purely data-driven models, significantly reducing false alarms and false negatives.

[0078] 3. This invention dynamically adjusts channel gain through a state-aware gating circuit to achieve a reverse balance of energy contribution from multiple sources and early detection of minor faults. Specifically, this invention constructs a state-aware gating circuit by real-time calculation of two physical state indices: the kurtosis index and the energy centroid frequency of the original signal. This dynamically generates gain control signals for each monitoring channel and adaptively weights and fuses the component-specific feature vectors output from each processing channel to generate a global state feature representation. When the energy of the liner impact signal is too high, the system automatically reduces the gain weight of the low-frequency channel while increasing the sensitivity of the mid-frequency and high-frequency channels, achieving a reverse balance of energy contribution from multiple sources. This mechanism solves the energy competition problem of signal coupling between multiple components from the hardware level, preventing high-energy liner impact components from masking low-energy fault characteristics such as minor damage to gears and bearings. This allows the device to maintain relatively stable monitoring sensitivity of each component even under drastic fluctuations in operating conditions, significantly improving the ball mill's safety protection capability against sudden faults. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the device structure of the present invention;

[0080] Figure 2 This is a schematic diagram of the device principle of the present invention;

[0081] Figure 3 This is a flowchart of the method of the present invention;

[0082] Figure 4 A schematic diagram of the structure and operating logic of the sampling cycle closed-loop control module for an overflow ball mill.

[0083] Figure 5 This is a schematic diagram of the operating logic of the multi-task parallel inference processing center. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0085] It should be noted that, unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0086] This invention designs an intelligent fault monitoring and control device for overflow ball mills, the structure of which is as follows: Figure 1 As shown, the working principle is as follows Figure 2 As shown, the system includes a signal conditioning and adaptive acquisition hardware subsystem, a physical prior-guided edge inference computing core, and an industrial control execution interface unit. All of these modules are installed in a dustproof and shockproof industrial-grade chassis, and the modules are connected to each other through an internal high-speed data bus to achieve real-time interaction between data and instructions.

[0087] The signal conditioning and adaptive acquisition hardware subsystem includes:

[0088] A multi-channel analog front end is used to connect to vibration sensors installed in the bearing housing, gearbox, and cylinder of the overflow ball mill, speed sensors installed near the main shaft, and slurry level transmitters. In this embodiment, the vibration sensor is an IEPE / ICP type sensor, and the slurry level transmitter is a 4-20mA type transmitter.

[0089] The speed pulse shaping circuit is used to shape the original signal from the speed probe into a standard pulse; specifically, it receives the original signal from the magnetoelectric speed probe, shapes it into a standard transistor-to-transistor logic (TTL) pulse through a Schmitt trigger, and then sends it to the capture unit of the controller.

[0090] The FPGA-based sampling timing controller uses a Field-Programmable Gate Array (FPGA) as its underlying logic core and incorporates a hardware counter, phase latch, and saturation clamping circuit. It dynamically adjusts the trigger period of the analog-to-digital converter based on the multi-source sensing signals through a hardware closed-loop control circuit to generate an equiangularly aligned vibration signal stream. Specifically:

[0091] The physical prior-guided edge inference computing core, used for fault diagnosis and state decision-making, includes a heterogeneous processing system-on-a-chip (SoC) and non-volatile memory. The heterogeneous SoC uses a dual-core heterogeneous architecture, comprising a general-purpose processing core (ARM) and a hardware acceleration core (DSP). The general-purpose processing core runs the embedded operating system and handles task scheduling and communication protocol stacks; the hardware acceleration core is specifically designed for parallel execution of convolution operations. To achieve millisecond-level real-time inference, the wavelet basis function-initialized convolution kernels and gating weight parameters are pre-loaded into the DSP's cache.

[0092] Non-volatile memory is used to store a fault mechanism mapping knowledge base. This knowledge base pre-defines the association between key components and fault characteristics, and records the characteristic frequency band addresses corresponding to faults of different components. During the inference core initialization phase, the system calls the prior data in the knowledge base to guide the feature extraction circuit to lock onto and focus on specific physical frequency bands, thereby improving the targeting and efficiency of fault feature extraction.

[0093] The edge inference computing core also includes a state-aware gating circuit. The input of this circuit is connected to the physical state index calculation module, and the output is connected to the gain control terminal of each processing channel. This circuit is used to calculate the kurtosis index of the original signal in real time. and energy center of gravity The gain control signals for each processing channel are dynamically generated.

[0094] The industrial control execution interface unit includes:

[0095] The digital I / O interface includes optocoupler-isolated relay output nodes. Once the device diagnoses a serious fault, it directly closes the emergency stop relay and activates the safety interlock circuit of the overflow ball mill main motor.

[0096] The industrial fieldbus interface integrates an RS485 or CAN bus transceiver and supports Modbus RTU or Profibus protocols. It is used to send closed-loop control commands to the frequency converter or feeding system according to health status assessment instructions, so as to realize real-time adjustment of the cylinder speed or feed rate, thereby completing the closed-loop control of the ball mill's operating status.

[0097] Meanwhile, this device is physically connected to the electrical control system of the overflow ball mill through an industrial control execution interface unit.

[0098] Based on the above hardware architecture, this invention proposes a fault intelligent monitoring and control method for overflow ball mills, the process of which is as follows: Figure 3 As shown, it includes data acquisition and closed-loop control, physical prior-guided feature extraction, and state perception and dynamic decision-making. The structure and operating logic of the sampling period closed-loop control module are as follows: Figure 4 As shown, the core processing flow of physical prior-guided feature extraction, state perception, and dynamic decision-making is as follows: Figure 5 As shown, the specific steps include:

[0099] S1. Data Acquisition and Closed-Loop Control: Real-time acquisition of multi-source sensing signals from the overflow ball mill, including at least the spindle speed signal, slurry level signal, and raw vibration signals collected by vibration sensors deployed on key components; Based on the spindle speed signal and the slurry level signal, the trigger cycle of the analog-to-digital converter is dynamically adjusted through a hardware closed-loop control loop to synchronously acquire the raw vibration signals with equal-angle step alignment, generating a hardware-aligned equal-angle vibration signal stream.

[0100] The specific steps are as follows:

[0101] The controller parameters are set and a sampling reference is established. Specifically, after the monitoring terminal is powered on, the rated speed of the overflow ball mill is read from the non-volatile memory. And the number of virtual sampling points per loop Calculate the baseline interruption period and will Write the initial register of the hardware timer as the steady-state operating reference.

[0102] The reference interruption period The calculation formula is:

[0103]

[0104] Real-time acquisition of multi-source sensing signals from the overflow ball mill: Specifically, the multi-source sensing module synchronously latches the spindle speed signal from the magnetoelectric probe via a high-speed I / O interface at a refresh rate higher than 1kHz. Analog quantities of the original vibration signals of each vibration channel The spindle speed signal is filtered using an internal moving average filter circuit. Preprocessing is performed to calculate the first... Effective speed per control cycle Its expression is:

[0105]

[0106] in, This is the buffer depth, used to filter out pulse jitter caused by electromagnetic interference in the field. For the first Instantaneous spindle speed value for each sampling period.

[0107] Because overflow ball mills experience sudden surges in liquid level under unsteady conditions, such as sudden load increases causing "bloating" or idling, this directly leads to a shift in the material's center of gravity and a sudden change in resistance torque. This mechanical inertia causes phase lag in synchronous sampling based solely on rotational speed measurement. To address this issue, this invention introduces a feedforward load compensation mechanism based on liquid level sensing. The specific steps are as follows: First, the adaptive sampling controller reads data from the slurry level gauge interface to obtain the slurry level signal and converts it into a liquid level percentage. Then, using the table lookup method, as shown in Table 1, based on the current liquid level percentage... Within the specified range, the corresponding liquid level weighting coefficient is retrieved from the firmware. When a heavy load and high liquid level are detected, the positive coefficient is used as the sampling compensation; when an underload and low liquid level are detected, the negative coefficient is used as the sampling compensation. This enables feedforward load compensation for speed fluctuations based on physical mechanisms at the hardware level, ensuring that equiangular alignment is not distorted under extreme working conditions.

[0108] Finally, the controller is based on the reference sampling period and effective speed. and the liquid level weighting coefficient The theoretical trigger interval for the current cycle is calculated based on the closed-loop control law. The expression for the solution of the closed-loop control law is:

[0109]

[0110] Meanwhile, during the sampling process, the sampling frequency is corrected in real time by a hardware arithmetic logic unit (ALU) to ensure that the mechanical angle increment corresponding to each acquisition command remains constant.

[0111] Table 1 Mapping Table of Slurry Level and Load Compensation Weighting Coefficients

[0112]

[0113] It should be noted that the liquid level percentage The current liquid level is the percentage of the effective inner diameter of the cylinder or the full scale of the sensor. Specific threshold parameters can be written into the non-volatile memory based on the initial calibration of the field equipment.

[0114] Furthermore, to prevent the sampling system from crashing due to a sudden drop in speed or overspeed, the safety logic unit uses a hardware saturation clamping circuit to limit the theoretical trigger cycle. Within a preset safety threshold, the final execution cycle is generated. Its expression is:

[0115]

[0116] in, and These are the minimum and maximum values ​​of the hardware security threshold locked in the register, respectively.

[0117] The controller will execute the final cycle. Write to the reload register of the programmable timer to update the interrupt frequency in real time and realize the dynamic update of the hardware timer; when the timer interrupt overflows, the hardware directly triggers the analog-to-digital conversion trigger signal, instructing the analog-to-digital conversion module to perform multi-channel synchronous conversion; at the same time, the data link layer marks all digital signal frames with angle-time dual tags and stores them in the first-in-first-out buffer (FIFO) for subsequent diagnostic units to view, realizing the dual correspondence of sampling points in the time domain and angle domain.

[0118] S2. Physically Prior-Guided Feature Extraction: The fault mechanism mapping knowledge base, i.e., the fault mechanism mapping table, pre-installed in the edge computing hardware is invoked to divert the equal-angle vibration signal stream to different physical meaning processing channels according to its fault correlation; using the prior information in the fault mechanism mapping knowledge base, the operation parameters of the feature extraction circuit in each processing channel are initialized, so that the feature extraction circuit directionally enhances the physical frequency band related to the fault of a specific component, and extracts the decoupled component-specific feature vectors in parallel from each processing channel.

[0119] The specific steps are as follows:

[0120] Based on the dynamic characteristics of key components of the overflow ball mill, a fault mechanism mapping table is pre-set in the non-volatile memory of the device, as shown in Table 2, to guide the processor to perform physical-level feature routing of the input signal stream.

[0121] Table 2 Fault Mechanism Mapping Table

[0122]

[0123] Table 2 This indicates the characteristic frequency of liner loosening. Indicates the spindle rotation frequency. Indicates the gear meshing frequency. Indicates the resonant frequency of the bearing. The resonant bandwidth represents the characteristic frequency of bearing failure.

[0124] The central processing unit, by calling the pre-defined mapping firmware, distributes the acquired multi-channel vibration data to three logical processing channels based on fault correlation:

[0125] (1) Liner monitoring channel: to monitor changes in cylinder stiffness and low-frequency oscillation components caused by liner loosening;

[0126] (2) Gear monitoring channel: monitors gear misalignment, meshing asymmetry caused by wear, and intermediate frequency modulation components;

[0127] (3) Bearing monitoring channel: monitors the high-frequency impact components caused by the temperature rise of the main bearing, oil film instability and rolling element damage.

[0128] To address the high-frequency broadband noise generated by steel ball collisions in the bearing monitoring channel, a hardware-level three-stage noise suppression sub-step is used for preprocessing before feature extraction. This preprocessing includes the following sub-steps:

[0129] Impact rarefaction step

[0130] Calculate the spectral kurtosis of high-frequency channel signals It then invokes the spectral kurtosis calculation circuit, compares the real-time calculated value with the fault-free threshold in the register, and only allows spectral kurtosis to be calculated. Higher than the preset threshold The impact component is passed through a filter to obtain a sparsified signal. .

[0131] Among them, the spectral kurtosis of the high-frequency channel signal The calculation formula is:

[0132]

[0133] In the formula, The first high-frequency channel signal The instantaneous amplitude of each sampling point.

[0134] Sparse signal The expression is:

[0135]

[0136] In the formula, A kurtosis threshold given by the hardware. This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0137] Envelope enhancer step

[0138] For sparsified signals Perform a Hilbert transform to extract its envelope to enhance the impact energy, thus obtaining the envelope-enhanced signal. The Hilbert transform is expressed as follows:

[0139]

[0140] Envelope enhancement signal The calculation formula is:

[0141]

[0142] In the formula, This represents the output signal after the Hilbert transform. Let Hilbert transform function, and These represent the real and imaginary parts of the analytic signal, respectively.

[0143] Amplitude gating sub-step

[0144] The envelope enhancement signal is filtered out using a hardware comparator. Below the background noise threshold The invalid components are removed to obtain the denoised signal used for feature extraction. Its expression is:

[0145]

[0146] In a specific hardware implementation, the above three sub-steps are completed by the digital signal processor (DSP) sequentially calling the corresponding instruction modules. Specifically, the DSP first calls the impulse sparsity instruction, which triggers the spectral kurtosis calculation circuit to execute the impulse sparsity sub-step; then it calls the envelope enhancement instruction, which triggers the Hilbert transform to execute the envelope enhancement sub-step; and finally it calls the amplitude gating instruction, which triggers the hardware comparator to execute the amplitude gating sub-step.

[0147] In the feature extraction layer of each processing channel, the target center frequency is obtained. The target center frequency Matching the physical frequency band with a specific fault characteristic to be detected; utilizing the center frequency of the target. The matching wavelet basis function parameters are used to initialize the convolution kernel register of the convolutional neural network within the channel, enabling the feature extraction circuit to have hardware sensitivity to the target physical frequency band from the initial power-on stage.

[0148] The expression for the initialization function is:

[0149]

[0150] In the formula, and Preset parameters for firmware. Sampling frequency, This represents the discrete-time sequence number.

[0151] Meanwhile, to prioritize the processing of data from high-risk channels, a basic bias vector is injected into the attention mechanism register based on equipment maintenance statistics. And through dynamic coefficients Adjusting the final attention weight .in, This is a quantified value of the historical failure probability. This is a quantified value for the severity level of the fault. This serves as the baseline value for equipment health.

[0152] Final attention weight The expression is:

[0153]

[0154] In the formula, The weight vectors are automatically learned by the attention mechanism module through training.

[0155] in, Specifically, it is an interpolation coefficient or gating weight used in the attention features learned based on data-driven learning. and the underlying bias vector injected based on equipment operation and maintenance statistics The proportions are allocated among them. The range of values ​​is Furthermore, it is a dynamic variable that changes in real time with the input signal, used to determine whether the system trusts the features extracted in real time or the fixed physical experience at the current moment.

[0156] S3. State Awareness and Dynamic Decision-Making: Real-time calculation of physical state indicators of the original vibration signal or equal-angle vibration signal stream; Based on the physical state indicators, dynamic gating signals are generated, and the component-specific feature vectors output by each processing channel are adaptively weighted and fused to generate a global state feature representation; The global state feature representation is input into the component-specific fault discriminator, and health status assessment instructions for key components of the overflow ball mill are output in parallel; According to the health status assessment instructions, when a fault risk is determined, control signals for controlling the operating state of the overflow ball mill are automatically generated and output.

[0157] The physical state indices include at least the kurtosis indices used for the original signal. and the energy centroid used for quantifying energy distribution under chemical conditions Its expression is:

[0158]

[0159] In the formula, The power spectral density is obtained from the short-time Fourier transform. and These are the signal mean and standard deviation, respectively. It represents the time-domain amplitude of the original vibration signal or the constant-angle vibration signal stream.

[0160] The step of generating a dynamic gating signal based on the physical state indicators, wherein the dynamic gating signal is implemented through a state-aware gating circuit, specifically includes:

[0161] When the kurtosis index of the original signal is detected When the gain is increased, a gate signal is generated to improve the gain of the high-frequency channel. The gating signal The expression is:

[0162]

[0163] in, For the Sigmoid function, The statistical mean of the high-frequency channel characteristics, together with the other two, characterize the impact activity within the system. and These are the learnable parameters and bias terms for the high-frequency channel, respectively.

[0164] When the energy center of gravity is detected When shifting to lower frequencies with significant oscillations, a gating signal is generated to improve the gain of the low-frequency channel. Its expression is:

[0165]

[0166] in, For the deep features of low frequency channels Norms are used to characterize the overall energy intensity of the feature space; The effective value of the Y-axis vibration at the FE end is used to reflect the swing amplitude of the cylinder of the physical entity; and These are the learnable parameters and bias terms for the low-frequency channel, respectively. .

[0167] When the energy center of gravity is detected When the frequency band is in the mid-frequency range and sideband asymmetry occurs, a gate signal is generated to improve the gain of the mid-frequency channel. Its expression is:

[0168]

[0169]

[0170] in, This represents the statistical mean of the mid-frequency channel characteristics. The energy difference in the gear meshing sidebands detected by the hardware is used to measure the degree of degradation of gear meshing symmetry; and These are the learnable parameters and bias terms for the intermediate frequency channel, respectively. This represents the power spectral density.

[0171] An element-wise multiplier is used to fuse the features of each channel with the gate weights to obtain a state-adaptive global feature representation. Its expression is:

[0172]

[0173] in, This represents the element-wise multiplication operator. These are component-specific feature vectors output from the low-frequency, mid-frequency, and high-frequency channels, respectively.

[0174] The reasoning process of the aforementioned logic enforcement device follows the laws of physical evolution, enhancing the robustness of the system under unsteady conditions.

[0175] The fused features after energy dynamic balance processing, i.e., global feature representation The data is input in parallel to the fault discriminators of each component in the multi-task inference model, and the health status assessment results of the liner, gear and bearing are decoupled and output synchronously. When the health status assessment value of any component is lower than the preset safety threshold, the device immediately generates a corresponding differentiated control command to perform protective operations including emergency stop or speed adaptive adjustment.

[0176] To facilitate real-time output of the health status of each component, the device has three built-in independent decision output registers, corresponding to the liner, gear, and bearing, respectively. The status codes of each register are defined as follows: the liner status register includes the status codes of: normal, loose, and detached; the gear status register includes the status codes of: normal, worn, and unbalanced load; and the bearing status register includes the status codes of: normal, damaged, and temperature rise.

[0177] Meanwhile, to further improve the accuracy and reliability of the diagnosis, a multi-task collaborative arbitration mechanism is adopted to jointly analyze the composite vibration signals and generate the final control command under joint constraints. Specifically, the main arbitration unit in the joint arbitration assesses the health status of the entire system, while multiple decoupled monitoring channels separately confirm the working conditions of the liner, gears, and bearings. Based on this, a decoupled monitoring strategy is executed, with each monitoring channel independently locking and analyzing its corresponding characteristic frequency band based on a fixed physical mechanism model. This prevents the device from ignoring weak bearing fault characteristics when processing high-energy liner impact signals. Simultaneously, the device incorporates an energy dynamic balance algorithm to perceive the energy distribution of the input signal across the entire frequency band in real time. When a strong energy impact, such as a liner impact, is detected, the algorithm automatically reduces the weight of that channel and amplifies the gain of weak signal channels such as bearings, thereby preventing high-energy signals from dominating the judgment process and causing missed detections.

[0178] To suppress output jitter under strong interference conditions, solve the signal shielding problem caused by differences in signal energy levels among multiple components, and ensure that device decisions conform to physical evolution laws, a closed-loop calibration algorithm is introduced to perform multi-target collaborative calibration of the commands after generating the initial control commands. Specifically, this includes smoothing the main monitoring commands, sensitivity calibration of dedicated channels for components, and physical consistency verification, enabling the device to accurately lock the state of each component and physically constrain the control logic under complex operating conditions.

[0179] (1) Smooth processing of master monitoring commands

[0180] To address the issue of jittering in the device's output signal and its tendency to trigger false noise under strong interference conditions, a hysteresis-based instruction smoothing logic is introduced into the main monitoring output circuit.

[0181] When obtaining health assessment instructions, a smoothing damping coefficient with hysteresis is introduced. Smooth logic through instructions The transient decision signal is transformed into a steady-state control signal containing a confidence interval, where the smoothing logic... The expression is:

[0182]

[0183] In the formula, To control the number of sampling frames in the period, No. Frame data corresponds to the first Output gain of class state As the target reference state, This represents the total number of state categories.

[0184] (2) Sensitivity calibration of the component's dedicated channel

[0185] Design dedicated gain compensation circuits for each key component of the overflow ball mill to solve the hardware detection problem of insufficient response to weak features in mixed signals from multiple components. The key components include liners, gears, and bearings.

[0186] Specifically, deviation calibration terms are set in the signal conditioning circuits of the liner monitoring channel, gear monitoring channel, and bearing monitoring channel, respectively, and the detection deviation is obtained by linear weighting using a gain balance coefficient. Its expression is:

[0187]

[0188] In the formula, These are the detection deviations for the liner, gear, and bearing channels, respectively. This is the gain balance coefficient.

[0189] Meanwhile, to address the signal shielding issue caused by the large differences in vibration energy levels among key components in the overflow ball mill, the device incorporates a dynamic gain adjustment circuit and introduces an automatic gain control mechanism based on energy contribution to adjust the balance coefficient. Real-time optimization is performed. Specifically, the liner plate experiences high energy from direct impact with the steel ball, while the early damage signal of the bearing is weak. Therefore, a corresponding mapping relationship is established using the negative feedback mapping between gain and physical energy. The root mean square value of the corresponding physical frequency band for each channel is calculated in real time to generate a dynamic gain factor.

[0190]

[0191] In the formula, For dynamic gain factor, The preset sensitivity value of the component is described. This is the circuit noise floor smoothing term. Let be the root mean square value of the i-th channel signal.

[0192] For the main bearing channel, a high sensitivity preset value is set in the hardware. This artificially assigns higher priority to weak signal channels. Through this automatic peak-shaving and valley-filling mechanism, the system dynamically attenuates the signal dominance of high-energy components during monitoring, while amplifying the signal amplitude of weak faulty components. This ensures that the device does not ignore early abnormal characteristics of the bearing due to the dominance of liner impact noise, achieving balanced state capture under multi-source coupling conditions.

[0193] (3) Physical consistency verification

[0194] To prevent logic drift in the gating aggregation module during multi-channel fusion and to ensure that device decisions conform to physical evolution laws, a physical consistency check is introduced. Specifically, the theoretical gating state vector is first calculated based on physical state indices. Then, the mean squared error is used to calculate the difference between the current execution state and the target state, thereby correcting the controller's decisions in reverse and obtaining the physical consistency check item. Its expression is:

[0195]

[0196] in, For the first The current gated output value of the frame data.

[0197] In summary, the final overall control strategy objective function is... The expression is:

[0198]

[0199] In the formula, This is a strategy correction coefficient used to balance instruction smoothing logic. and detection deviation Physical consistency check item The relationship between these factors determines the degree to which the system is punished by the constraints of physical laws.

[0200] Through strategy correction coefficient This not only ensures the accuracy of fault identification, but also guarantees the rationality of the physical logic during the multi-task processing stage of the device, thus achieving stable operation of the closed-loop monitoring system.

[0201] This invention adapts the structural characteristics and fault mechanisms of overflow ball mills to the monitoring device, realizing physical-level decoupling and dynamic adaptation of the operating states of multiple components. This is beneficial for the device to detect early hidden dangers under strong noise conditions and output reliable control signals, and has important engineering value.

[0202] Finally, it should be noted that any parts of this invention not described in detail are prior art. Those skilled in the art will understand that the above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A fault intelligent monitoring and control method for overflow ball mills, characterized in that, Includes the following steps: S1. Data Acquisition and Closed-Loop Control: Real-time acquisition of multi-source sensing signals from the overflow ball mill, including at least the spindle speed signal, slurry level signal, and raw vibration signals collected by vibration sensors deployed on key components; Based on the spindle speed signal and the slurry level signal, the trigger cycle of the analog-to-digital converter is dynamically adjusted through a hardware closed-loop control loop to synchronously acquire the raw vibration signals with equal-angle step alignment, generating a hardware-aligned equal-angle vibration signal stream; S2. Physically Prior Guided Feature Extraction: The fault mechanism mapping knowledge base pre-installed in the edge computing hardware is invoked to divert the equal-angle vibration signal stream to different physical meaning processing channels according to its fault correlation. Using the prior information in the fault mechanism mapping knowledge base, the operation parameters of the feature extraction circuit in each processing channel are initialized, so that the feature extraction circuit can directionally enhance the physical frequency band related to the fault of a specific component, and extract the decoupled component-specific feature vectors in parallel from each processing channel. S3. State Awareness and Dynamic Decision-Making: Real-time calculation of physical state indicators of the original vibration signal or equal-angle vibration signal stream; Based on the physical state indicators, dynamic gating signals are generated, and the component-specific feature vectors output by each processing channel are adaptively weighted and fused to generate a global state feature representation; The global state feature representation is input into the component-specific fault discriminator, and health status assessment instructions for key components of the overflow ball mill are output in parallel; According to the health status assessment instructions, when a fault risk is determined, control signals for controlling the operating state of the overflow ball mill are automatically generated and output.

2. The intelligent fault monitoring and control method for an overflow ball mill according to claim 1, characterized in that, In step S1, dynamically adjusting the trigger cycle of the analog-to-digital converter through a hardware closed-loop control circuit specifically includes: The effective spindle speed is obtained by performing a moving average filter on the real-time acquired spindle speed signal. Its expression is: In the formula, For buffer depth, The instantaneous value of the spindle speed in the i-th sampling period; Based on the slurry level signal, the level weighting coefficient is obtained by looking up a table. ; Based on the reference sampling period, the effective rotational speed and the liquid level weighting coefficient The theoretical triggering period is calculated based on the closed-loop control law. Its expression is: In the formula, N is the preset number of virtual sampling points per cycle; The theoretical triggering cycle is limited to a preset safety threshold by a hardware saturation clamping circuit, thus generating the final execution cycle. The hardware timer is updated using the final execution cycle, and the analog-to-digital converter is directly triggered to perform multi-channel synchronous sampling.

3. The intelligent fault monitoring and control method for an overflow ball mill according to claim 1, characterized in that, In step S2, the fault mechanism mapping knowledge base is pre-set with the association between key components and fault characteristics. The association includes at least component type, fault type, dominant frequency range, location of sensitive measuring point, and dominant vibration direction. Specifically, the equal-angle vibration signal stream is diverted to different physical meaning processing channels according to the fault mechanism mapping knowledge base, and the equal-angle vibration signal stream is routed to the low-frequency channel corresponding to the liner monitoring, the medium-frequency channel corresponding to the gear monitoring, and the high-frequency channel corresponding to the bearing monitoring.

4. The intelligent fault monitoring and control method for an overflow ball mill according to claim 3, characterized in that, In step S2, initializing the operational parameters of the feature extraction circuits in each processing channel using the prior information in the fault mechanism mapping knowledge base specifically includes: For each processing channel, obtain its target center frequency. ; Using the center frequency of the target Matching wavelet basis functions The convolution kernel register of the convolutional neural network within this channel is initialized, enabling the feature extraction circuit to possess hardware sensitivity to the target physical frequency band from the initial power-on stage. and Preset parameters for firmware. Where n is the sampling frequency and n is the discrete time index.

5. The intelligent fault monitoring and control method for an overflow ball mill according to claim 3, characterized in that, For the high-frequency channel, a hardware-level three-stage noise suppression sub-step is included before the feature extraction step: Impulsive rarefaction sub-step: Calculate the spectral kurtosis of the high-frequency channel signal. And using a spectral kurtosis calculation circuit, only spectral kurtosis is allowed. Higher than the preset threshold The impact component passes through, resulting in a sparsified signal; Envelope enhancement sub-step: Perform Hilbert transform on the sparsified signal to extract its envelope to enhance the impact energy, and obtain the envelope-enhanced signal; Amplitude gating sub-step: Using a hardware comparator, filter out signals in the envelope enhancement signal that are below the background noise threshold. The components are used to obtain the denoised signal used for feature extraction.

6. The intelligent fault monitoring and control method for an overflow ball mill according to claim 1, characterized in that, In step S3, the physical state index includes at least the kurtosis index of the original signal. and the energy centroid used for quantifying energy distribution under chemical conditions ; The step of generating a dynamic gating signal based on the physical state indicators specifically includes: When the kurtosis index of the original signal is detected When the gain is increased, a gate signal is generated to improve the gain of the high-frequency channel. The gating signal The expression is: In the formula, For the Sigmoid function, This represents the statistical mean of the high-frequency channel characteristics. and These are the learnable parameters and bias terms for the high-frequency channel, respectively. When the energy center of gravity is detected When shifting to lower frequencies with significant oscillations, a gating signal is generated to improve the gain of the low-frequency channel. Its expression is: In the formula, For the deep features of low frequency channels Norm, This represents the effective value of the Y-axis vibration at the FE end. and These are the learnable parameters and bias terms for the low-frequency channel, respectively. When the energy center of gravity is detected When the frequency band is in the mid-frequency range and sideband asymmetry occurs, a gate signal is generated to improve the gain of the mid-frequency channel. Its expression is: In the formula, This represents the statistical mean of the mid-frequency channel characteristics. The energy difference in the gear meshing sideband detected by the hardware. and These are the learnable parameters and bias terms for the intermediate frequency channel, respectively. The component-specific feature vectors output from each processing channel are adaptively weighted and fused to generate a global state feature representation. The formula for the global state feature representation is as follows: in, This represents the global state features. , , These are component-specific feature vectors output from the low-frequency, mid-frequency, and high-frequency channels, respectively. This represents the element-wise multiplication operator.

7. The intelligent fault monitoring and control method for an overflow ball mill according to claim 1, characterized in that, In step S3, after the step of outputting health status assessment instructions for key components of the overflow ball mill in parallel, a multi-objective collaborative calibration step is also included: Introduce a smoothing damping coefficient with hysteresis to the health status assessment command. Smooth logic through instructions The transient decision signal is converted into a steady-state control signal containing a confidence interval; the instruction smoothing logic... The expression is: In the formula, To control the number of sampling frames in the period, No. Frame data corresponds to the first Output gain of class state, As the target reference state, This represents the total number of state categories. By setting up dedicated gain compensation loops for each key component and introducing an automatic gain control mechanism based on energy contribution, the sensitivity of the component's dedicated eigenvector is dynamically calibrated, generating the component detection deviation. The detection deviation The expression is: In the formula, These are the detection deviations for the liner, gear, and bearing channels, respectively. This is the gain balance coefficient; Calculate the theoretical gating state vector based on the physical state indices. And by calculating the current gate output value The mean square error of the theoretical gated state vector is used to generate a physical consistency check item. Its expression is: in, For the first The current gated output value of the frame data; Based on the instruction smoothing logic Component inspection deviation Physical consistency check item Through overall control strategy Generate the final control signal, where This is the strategy correction coefficient.

8. A fault intelligent monitoring and control device for an overflow ball mill for implementing the method of any one of claims 1-7, characterized in that, include: The signal conditioning and adaptive acquisition hardware subsystem includes: A multi-channel analog front end is used to connect to speed sensors, level transmitters and vibration sensors to acquire multi-source sensing signals in real time. The speed pulse shaping circuit is used to shape the raw signal from the speed probe into a standard pulse. An FPGA-based sampling timing controller, with built-in hardware counters, phase latches, and saturation clamping circuits, is used to dynamically adjust the trigger period of the analog-to-digital converter based on the multi-source sensing signals through a hardware closed-loop control loop, thereby generating an equally aligned vibration signal stream. The core of edge reasoning computation guided by physical priors includes: Heterogeneous processing system-on-a-chip, including general-purpose processing cores and hardware acceleration cores; Non-volatile memory is used to store a fault mechanism mapping knowledge base, which is pre-set with the association between key components and fault characteristics; The hardware acceleration core is used to initialize the operation parameters of the feature extraction circuit according to the prior information in the fault mechanism mapping knowledge base, extract the decoupled component-specific feature vectors in parallel, generate dynamic gating signals based on physical state indicators for adaptive weighted fusion, and finally output health status assessment instructions. An industrial control execution interface unit, comprising: The digital I / O interface is used to directly output an emergency stop signal to the safety interlock circuit of the overflow ball mill according to the health status assessment command. An industrial fieldbus interface is used to send closed-loop control commands to the actuator to adjust feeding or operating parameters based on the health status assessment commands.

9. The intelligent fault monitoring and control device for an overflow ball mill according to claim 8, characterized in that, The FPGA-based sampling timing controller is specifically used for: The theoretical triggering period is calculated based on the real-time effective rotational speed and the closed-loop control law corrected by the liquid level weighting coefficient. The built-in saturation clamping circuit limits the theoretical triggering cycle to a preset safety threshold, generating the final execution cycle. The final execution cycle is written to the reload register of the programmable timer to update the interrupt frequency; When the timer interrupt overflows, the analog-to-digital converter is directly triggered to perform multi-channel synchronous conversion, and the converted digital signal frame is labeled with angle-time dual tags.

10. The intelligent fault monitoring and control device for an overflow ball mill according to claim 8, characterized in that, The hardware acceleration core is a DSP core, which contains wavelet convolution kernel registers corresponding to the component types in the fault mechanism mapping knowledge base. The DSP core is used to call the dominant frequency range in the fault mechanism mapping knowledge base, generate the corresponding wavelet basis function, and load the parameters of the wavelet basis function into the wavelet convolution kernel register to complete the initialization of the feature extraction circuit. The edge inference computing core is also used to calculate the kurtosis index of the original signal in real time. and energy center of gravity Based on the calculation results, the gain control signal for each processing channel is dynamically generated through the state-aware gating circuit. The device also includes a dynamic gain adjustment circuit, which, according to the gain control signal, attenuates the signal of the channel where the high-energy component is located and amplifies the signal of the channel where the weak fault component is located through automatic gain control logic, so as to achieve equalization state capture under multi-source coupling conditions.

Images