Fault diagnosis method and system for lifting motor of crane hook

Abstract

The invention discloses a fault diagnosis method and system for a crane hook lifting motor, and belongs to the field of motor monitoring. The method comprises the following steps: acquiring multi-modal operation data of a crane hook lifting motor in real time through a multi-sensor synchronous acquisition system, constructing a physical information neural network (PINN) diagnosis model, respectively performing fast Fourier transform (FFT) on a vibration signal, a current signal and an audio signal, extracting spectrum characteristics of each signal, and the like. According to the method, physical laws such as an electromagnetic induction law and a heat conduction equation are embedded into a neural network loss function, so that a diagnosis result has a clear physical explanation, an engineer is helped to quickly position root causes (such as winding short circuit and bearing pitting) of faults, and scientificity and pertinence of maintenance decisions are remarkably improved. The innovative frequency domain fusion mechanism dynamically distributes the weights of vibration, current and audio features based on mutual information entropy, effectively strengthens the cooperative expression of fault sensitive information, and overcomes the limitation of traditional single-mode diagnosis.

Description

Technical Field

[0001] This invention relates to motor monitoring, and more particularly to a fault diagnosis method and system for a crane hook lifting motor. Background Technology

[0002] As a critical drive component, the crane hook lifting motor's operational reliability directly impacts the overall safety of the machine. Traditional fault diagnosis relies primarily on manual experience and periodic maintenance, resulting in issues such as delayed response and high misjudgment rates. In recent years, deep learning-based intelligent diagnostic technologies have been gradually applied. For example, patent CN118964903B uses multi-sensor signal fusion to generate RGB feature maps for CNN classification, which improves accuracy to some extent, but the model lacks physical interpretability, making it difficult for engineers to understand the diagnostic basis. Patent CN117909710A introduces a Transformer model to optimize feature extraction, but it still suffers from strong reliance on cloud computing and insufficient real-time edge computing capabilities. Current mainstream technologies suffer from three common problems: First, multimodal data fusion mechanisms are crude, processing vibration, current, and audio signals only through simple splicing or channel mapping, failing to effectively uncover intermodal coupling relationships. Second, physical laws are disconnected from data-driven approaches, rendering deep learning models "black boxes" and unable to provide causal explanations based on principles of motor electromagnetics and thermodynamics. Third, diagnostic and predictive functions are separated, with existing systems focusing primarily on post-event diagnosis, making it difficult to provide early warnings for progressive faults such as bearing wear and winding aging. With the increasing demand for continuous crane operations in ports, shipyards, and other scenarios, there is an urgent need for an integrated solution that possesses physical interpretability, real-time edge response, and predictive maintenance support. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a fault diagnosis method and system for crane hook lifting motors, in order to solve the three common problems existing in current mainstream motor fault diagnosis systems.

[0004] Technical solution: A fault diagnosis method for a crane hook lifting motor, comprising the following steps: S1. Real-time acquisition of multi-modal operating data of the crane hook lifting motor through a multi-sensor synchronous acquisition system, including three-axis vibration signals, three-phase current signals, audio signals, temperature distribution data, and motor speed / position signals; S2. Construct a Physical Information Neural Network (PINN) diagnostic model. This model embeds the electromagnetic induction law of motors, thermodynamic equations, and mechanical kinematics as physical constraints into the neural network loss function, forming a hybrid loss function that includes data fitting terms and physical law constraint terms. S3. Perform Fast Fourier Transform (FFT) on the vibration signal, current signal and audio signal respectively, extract the spectral features of each signal, analyze the coupling relationship between multimodal signals by calculating the cross-correlation function, and establish a weighted fusion matrix to generate a composite fault feature vector; S4. Deploy a lightweight PINN model on edge computing nodes to perform real-time analysis of composite fault feature vectors and trigger emergency control commands when an abnormal state is detected; at the same time, compress key feature data and upload it to the cloud. S5, the cloud-based diagnostic engine receives data uploaded from edge nodes, runs the full version of the PINN model for in-depth analysis, combines the historical fault database with the equipment health index (HI) assessment model, and outputs a diagnostic report that includes fault location information, explanation of physical causes, and prediction of remaining service life. S6. Establish an edge-cloud model collaborative update mechanism: The cloud periodically sends the trained and optimized model parameters to the edge nodes, and the edge nodes integrate new fault samples through incremental learning to keep the diagnostic model dynamically evolving; S7. Generate predictive maintenance strategies based on diagnostic results, including maintenance time window suggestions, spare parts replacement lists and operation guidance plans, and output them visually through a human-machine interface; S8. When a high-risk fault is diagnosed, the crane safety protection protocol is automatically triggered to restrict the lifting and lowering of the hook and send alarm information to the monitoring terminal.

[0005] Furthermore, the multi-sensor synchronous acquisition system described in step S1 uses a hardware triggering circuit to achieve microsecond-level time synchronization, with a vibration signal sampling frequency of no less than 10kHz, a current signal sampling frequency of no less than 2kHz, and an audio signal sampling frequency of no less than 16kHz.

[0006] Furthermore, the frequency domain fusion in step S3 specifically includes: extracting 1-3kHz high-frequency resonance features from the vibration signal, extracting 0-500Hz fundamental harmonic features from the current signal, and extracting 5-8kHz mechanical friction features from the audio signal; determining the weight coefficients of each mode through feature correlation analysis, and constructing the mathematical expression for the composite feature vector as follows: in, F fused The multimodal composite fault feature vector is a comprehensive feature expression generated by dynamically weighted fusion of vibration features, current features, and audio features. Its data dimension is N×1 (N≥256), which is used to input the physical information neural network for fault classification. F vibThe vibration signal feature subvector specifically refers to the resonance features extracted from the 1-3kHz high-frequency band through Fast Fourier Transform (FFT), which includes mechanical fault-sensitive information such as bearing raceway defect frequency and gear meshing sideband, and the feature dimension is no less than 64 dimensions. F current It is a feature subvector of the current signal, specifically referring to the harmonic features extracted in the 0-500Hz fundamental frequency band, including electrical fault feature components such as motor winding imbalance and rotor bar breakage, with a feature dimension of no less than 32 dimensions. F audio The audio signal feature subvector specifically refers to the mechanical triboacoustic features extracted in the 5-8kHz frequency band through a bandpass filter, including acoustic emission features of faults such as abnormal wear of gearboxes and failure of bearing lubrication, with a feature dimension of no less than 48 dimensions. α , β , γ The dynamic weighting coefficients are calculated based on mutual information entropy. α Vibration characteristic weighting coefficient (value range 0.35-0.65), dynamically adjusted based on the correlation between the vibration signal and the current fault mode. β The current characteristic weighting coefficient (range 0.25-0.55) increases with the increase of current harmonic distortion. γ Audio feature weight coefficients (range 0.15-0.45) adaptively decrease based on background noise intensity. They satisfy the normalization constraint α+β+γ=1. The weight coefficients are updated in real-time via the mutual information entropy calculation module. Update logic: If mechanical fault characteristics are detected, the weight of α is increased (Δ≥0.2); If the current THD distortion rate is >5%, then increase the β weight (Δ≥0.15); If the ambient noise is >80dB, reduce the weight of γ (Δ≤-0.1).

[0007] Furthermore, the physical constraints include the motor stator and rotor air gap magnetic flux density equation, the winding temperature rise differential equation, and the gearbox transmission ratio relationship. During the neural network training process, the data consistency is verified in real time through the physical residual calculation module. When the deviation of the physical law exceeds the threshold, the network weight is automatically adjusted.

[0008] A fault diagnosis system for a crane hook lifting motor, comprising: Multimodal data acquisition module: integrates a triaxial MEMS accelerometer, a closed-loop Hall current sensor, a wideband microphone array, an infrared thermal imager, and a photoelectric encoder, and is connected to the edge computing unit via an industrial fieldbus; Edge diagnostic terminal: It adopts an embedded GPU computing platform, with built-in lightweight PINN diagnostic model, real-time signal processing algorithm library and device control interface, supporting millisecond-level fault response; Cloud-based analytics platform: Deployed on a high-performance computing cluster, it includes a physical information neural network training engine, a fault knowledge graph database, a remaining life prediction module, and a maintenance decision generator; Collaborative communication gateway: It adopts 5G industrial modules to realize encrypted data transmission from the edge to the cloud, and supports differential compression of model parameters and breakpoint resume function; Human-machine interaction terminal: Equipped with an industrial touch screen and mobile application terminal, it realizes three-dimensional fault location visualization, dynamic demonstration of physical mechanism and maintenance work order management functions; Safety linkage control unit: Deeply integrated with the crane PLC system, it automatically activates the mechanical braking device and cuts off the motor drive power when a high-risk fault such as broken shaft, short circuit in winding, or broken gear is diagnosed.

[0009] Furthermore, the edge diagnostic terminal adopts a modular hardware architecture, including a signal conditioning circuit board (supporting IEPE power supply and anti-aliasing filtering), a Jetson Xavier NX core computing module, an isolated digital output board (DO), and four PoE gigabit Ethernet interfaces.

[0010] Furthermore, the cloud-based analysis platform's fault knowledge graph database contains a library of typical crane motor fault cases, which are associated with fault feature patterns, root cause chains, and maintenance solutions. Intelligent matching and reasoning of fault cases are achieved through graph neural networks.

[0011] Furthermore, the safety linkage control unit has built-in multi-level protection logic, including: Level 1 response (issuing an early warning when fault characteristics first appear), Level 2 response (reducing load operation when the fault probability exceeds the threshold), and Level 3 response (immediately shutting down when a high-risk fault is identified). The response actions at each level are transmitted to the actuator through the Profisafe protocol, which is certified for safety.

[0012] Beneficial Effects: This invention embeds physical laws such as the law of electromagnetic induction and the equation of heat conduction into the loss function of a neural network, giving the diagnostic results a clear physical interpretation. This helps engineers quickly locate the root cause of faults (such as winding short circuits and bearing pitting), significantly improving the scientific rigor and relevance of maintenance decisions. The innovative frequency domain fusion mechanism dynamically allocates the weights of vibration, current, and audio features based on mutual information entropy, effectively strengthening the collaborative expression of fault-sensitive information and overcoming the limitations of traditional single-mode diagnosis, achieving a qualitative leap in the accuracy of identifying complex faults. The lightweight PINN model achieves millisecond-level anomaly response at the edge, and combined with cloud-based deep analysis to form a hierarchical diagnostic architecture, meeting both the real-time emergency response needs of high-risk faults and ensuring the refined analysis capabilities of complex faults. Integrating fault diagnosis and remaining life prediction functions, it dynamically generates predictive maintenance strategies based on equipment health indicators, promoting the transformation of operation and maintenance models from reactive maintenance to proactive prevention, significantly reducing the risk of unplanned downtime and the probability of cascading accidents. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0014] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] Example This embodiment focuses on fault diagnosis of the hook lifting motor (model: YZR355M2-8, rated power: 90kW, rated speed: 740r / min) of a 20-foot container gantry crane used in ports. This crane handles an average of 300 TEUs per day in an environment with salt spray, dust, and intermittent impact loads. Traditional scheduled maintenance often results in unplanned downtime due to delayed diagnosis (8-10 downtimes per year, with losses exceeding 50,000 RMB per downtime). This embodiment, by deploying the fault diagnosis method and system described in claims 1-8, achieves real-time monitoring, precise location, and predictive maintenance of motor faults, solving the problems of "black box diagnosis," "delayed response," and "blind maintenance" inherent in traditional technologies.

[0016] Implementation of the fault diagnosis system: Building a multimodal data acquisition module: The following sensors are integrated: A triaxial MEMS accelerometer: Analog Devices ADXL355, measuring range ±10g, resolution 0.001g, deployed at the front and rear bearing housings and gearbox housing of the motor (3 measuring points in total), supports IEPE power supply, sampling frequency set to 12kHz (meeting ≥10kHz requirements), used to acquire 1-3kHz high-frequency resonance characteristics (e.g., bearing outer ring defect frequency 1850Hz, gear meshing frequency 2100Hz); Closed-loop Hall current sensor: LEM LA28-NP, measuring range 0-500A, accuracy 0.5%, connected in series with the motor's three-phase power supply line, sampling frequency set to 2.5kHz (meeting ≥2kHz requirements), used to extract 0-500Hz fundamental harmonic characteristics (e.g., 50Hz fundamental wave, 150Hz third harmonic); Wideband microphone array: Knowles SPK06. The 41HT4H-B array, with a frequency response of 20Hz-20kHz, is deployed on the outside of the motor ventilation shroud (3 array units, 15cm spacing), with a sampling frequency set to 18kHz (meeting the ≥16kHz requirement), to capture 5-8kHz mechanical friction characteristics (such as the friction sound of bearing lubrication failure at 6.2kHz, and the impact sound of gear tooth breakage at 7.5kHz); the infrared thermal imager: FLIRA655sc, resolution 640×512, temperature measurement range -20℃-120℃, frame rate 30fps, focused on the end of the motor stator winding and the outer ring of the bearing to acquire temperature distribution data (sampling interval 1s); the photoelectric encoder: Pepperl+Fuchs RVI58N-011K1R61N-1024, resolution 1024 lines, installed on the motor output shaft end, sampling frequency 1kHz, to acquire motor speed (accuracy ±0.1r / min) and position signals in real time.

[0017] Using a Xilinx Artix-7 FPGA as the core, a 100MHz high-precision clock signal is generated and connected to the sampling trigger pins of all sensors via synchronous trigger lines (length error ≤5cm), achieving a timestamp deviation of ≤5μs for each mode of data. Simultaneously, an 8-channel synchronous sampling ADC (ADIAD7606, sampling rate 200kSPS) is integrated into the signal conditioning circuit board to synchronously quantize the analog signals output by the sensors, avoiding characteristic deviations caused by asynchronous sampling.

[0018] The signal conditioning circuit board includes: ① IEPE power supply module (output 24VDC, maximum current 500mA, meeting the power supply requirements of the accelerometer); ② anti-aliasing filter (Butterworth 4th order, vibration signal cutoff frequency 5kHz, current signal cutoff frequency 1kHz, audio signal cutoff frequency 8kHz); ③ signal amplification module (adjustable gain, range 1-100 times, adaptable to the output amplitude of different sensors); ④ isolation module (using optocoupler isolation, withstand voltage 2500VAC, suppressing port power grid interference).

[0019] Design and deployment of edge diagnostic terminals: The edge diagnostic terminal adopts a modular design, with core hardware including: a computing core: NVIDIA Jetson Xavier NX (6-core ARM Cortex-A78, 384-core Volta GPU, 8GB LPDDR4, 16GB eMMC), supporting CUDA 11.4 to meet the real-time inference requirements of lightweight PINN models; an interface module: 4 PoE Gigabit Ethernet interfaces (supporting IEEE 802.3af, for power supply and data transmission to microphone arrays and infrared thermal imagers); 2 RS485 interfaces (for communication with the crane PLC, baud rate 115200bps, no parity); 1 isolated digital output board (8-channel DO, output voltage 24VDC, maximum current 2A, used to trigger emergency control commands); a power supply module: wide input voltage (12-36VDC), output 5V / 12V / 24V, with overvoltage and overcurrent protection to adapt to power supply fluctuations in port vehicles; and a cooling module: passive cooling + fan-assisted cooling (fan start / stop threshold 55℃) to ensure stable operation of the equipment in a 35℃ high-temperature environment.

[0020] The edge terminal software is developed based on the Linux Ubuntu 20.04 system. Its core functional modules include: Real-time data receiving module: It uses the UDP protocol to receive multimodal data (data packet size 1024 bytes, receive buffer 1MB), and achieves data alignment by timestamp sorting, with a processing latency of ≤1ms; Lightweight PINN model: Based on TensorRT quantization (INT8 precision), the input is a 256-dimensional composite fault feature vector (Ffused), and the output is the probability of 5 types of faults (winding short circuit, bearing pitting, gear tooth breakage, rotor bar breakage, normal), with an inference latency of ≤5ms. Model simplification strategy: Reduce the number of hidden layers (from 10 layers to 5 layers), reduce the number of neurons per layer (from 256 to 128), and retain physical constraint terms (stator and rotor air gap magnetic flux density equation, winding temperature rise differential equation); Anomaly detection and emergency control module: Preset fault probability thresholds (Level 1 warning ≥ 30%, Level 2 emergency ≥ 60%). When an anomaly is detected, a control signal is output through the DO board (such as triggering the motor to reduce speed to 50% of the rated speed), and an anomaly log (including timestamp, feature vector, and fault probability) is generated simultaneously; Data compression and upload module: Use the LZ4 compression algorithm (compression ratio of approximately 3:1) to compress key feature data (feature vectors 10 seconds before and after the anomaly moment), and upload it to the cloud through the 5G module with an upload rate ≥ 1Mbps and a data loss rate ≤ 0.1%.

[0021] Building a cloud-based analytics platform: The cloud platform is deployed in a high-performance computing cluster in the port data center. The cluster contains 10 computing nodes, each configured with: an Intel Xeon Gold 6330 CPU (28 cores), an NVIDIA A100 GPU (40GB VRAM), a 1TB SSD, and 100GB DDR4 memory. The nodes are interconnected through an InfiniBand network (100Gbps bandwidth) to support distributed training and inference.

[0022] The software functional modules include: PINN training engine: based on the PyTorch 2.0 framework, it implements the training of the full PINN model (10 hidden layers, 256 neurons per layer, ReLU activation function). The training data includes historical fault samples (5000 groups, covering 5 types of faults) and real-time uploaded edge data. Physical constraints are implemented through a custom loss function, as shown in the following formula: Loss total =0.6×Loss data +0.4×Loss phy Among them, Loss data For MSE loss (fitting multimodal data), Loss phyThe L2 norm of the physical residuals is used to verify the consistency of physical quantities such as air gap magnetic flux density and temperature rise, with a residual threshold of ≤5%. Training is conducted once a week, with 1000 iterations per cycle, achieving a model accuracy of ≥98%. The fault knowledge graph database is built based on the Neo4j graph database and includes three core entities: ① fault type (e.g., "winding short circuit"); ② fault characteristics (e.g., "current third harmonic amplitude ≥20%"); ③ repair plan (e.g., "replace winding insulation paper, bake at 120℃, and keep warm for 4 hours"). The relationships between entities include "feature-fault" (weights based on mutual information entropy) and "fault-repair" (association based on historical repair success rate). Case matching is achieved through graph neural network (GAT), with a matching accuracy of ≥95% and a response time of ≤100ms; Remaining life (RUL) prediction module: based on equipment health index (HI) assessment model, HI is calculated as: HI=w1×(vibration peak / rated value)+w2×(current THD)+w3×(winding temperature / rated temperature) (w1=0.4, w2=0.3, w3=0.3). RUL is predicted using an LSTM network (input is the HI sequence of the past 30 days, output is the HI trend of the next 90 days), with a prediction error ≤10%; Maintenance decision generator: Based on the diagnostic results (fault type, RUL), predictive maintenance strategies are generated: ① Maintenance time window: determined based on RUL (e.g., if RUL≤30 days, maintenance is recommended within 7 days); ② Spare parts replacement list: associated with BOM (e.g., "bearing model 6319, quantity 1; insulation paper model DMD, quantity 5m"); ③ Operation guidance plan: includes step breakdown (e.g., "1. Power off, 2. Remove motor end cover, 3. Check winding insulation resistance, 4. Replace spare parts, 5. No-load test run") and precautions (e.g., "apply anti-rust grease in salt spray environment").

[0023] Collaborative communication gateway and security linkage control unit: The collaborative communication gateway uses Huawei's 5G industrial module (MH5000-31), supporting SA standalone networking (frequency band n48), with an uplink speed of ≥100Mbps and a downlink speed of ≥500Mbps. Core functions include: Encrypted transmission: TLS 1.3 protocol is used to encrypt edge-to-cloud data, with the key updated every 24 hours to prevent data leakage; Differential compression: PINN model parameters (approximately 100MB) are differentially compressed using the BSDiff algorithm, resulting in a compressed data size of ≤10MB, reducing bandwidth usage; Resume interrupted transmission: Model parameters are transmitted in blocks (1MB each), with edge nodes recording the IDs of received blocks. If transmission is interrupted, only unreceived blocks are downloaded upon reconnection, ensuring the integrity of parameter transmission.

[0024] The safety linkage control unit is integrated with the crane PLC (Siemens S7-1500F) via the Profisafe protocol (safety level SIL2), and has built-in three-level protection logic: Level 1 response: When the fault probability is ≥30%, the PLC outputs a warning signal (controls the audible and visual alarm in the cab, frequency 1Hz, red light flashing), and simultaneously displays warning information on the human-machine interface (such as "Abnormal bearing vibration, please pay attention"); Level 2 response: When the fault probability is ≥60%, the PLC controls the motor to operate at reduced load (load from 100% of rated value). Reduce to 50%), limit the lifting speed of the hook (from 1.5m / s to 0.5m / s), and record the operating parameters at the time of unloading; Level 3 response: when a high-risk fault is detected (such as "broken shaft" or "winding short circuit", characterized by a vibration characteristic frequency of 2× rotational frequency and a current THD≥15%), the PLC immediately outputs a braking signal (triggering the hydraulic brake, braking time ≤0.5s), and at the same time cuts off the motor drive power (disconnects contactor KM1), and sends an alarm message to the port monitoring center (including fault type, location, and time).

[0025] Human-computer interaction terminal: The Weintek MT8102iE (10.1-inch, 1280×800 resolution, capacitive touch) was selected and installed in the crane cab. The display includes: real-time operation monitoring: multimodal data curves (vibration time / frequency domain, current harmonics, temperature distribution), refresh rate 1 second / time; fault diagnosis results: fault type, probability, location (3D model annotation, such as "pitting of the motor rear bearing"), explanation of physical causes (such as "bearing raceway wear leads to increased vibration peak, consistent with mechanical kinematics"); maintenance work orders: displaying pending work orders (maintenance time, spare parts list, responsible person), supporting work order confirmation and feedback.

[0026] The mobile application is developed based on Android 13 and supports mobile phones (Android 10 and above) and tablets. Core functions include: Remote monitoring: real-time viewing of motor operation data and diagnostic results, and receiving alarm push notifications (push notification latency ≤ 3s); 3D fault visualization: rendering a 3D model of the motor using the Unity engine, and displaying detailed information by clicking on the fault location (e.g., "Short circuit in winding A phase, location: 3rd slot on the left side of the motor stator"); Maintenance management: viewing work order progress, uploading repair photos (e.g., the replaced bearing), and generating repair reports.

[0027] Fault diagnosis methods include: (1) Synchronous acquisition of multimodal data During the operation of the crane lifting a 20-foot container (load 25 tons), the multimodal data acquisition module operates according to the following parameters: Sampling frequency: vibration signal 12kHz, current signal 2.5kHz, audio signal 18kHz, temperature data 1Hz, speed / position signal 1kHz; Acquisition duration: continuous acquisition for 24 hours, generating one data file (approximately 500MB) per hour, stored on the edge terminal's SSD (capacity 1TB); Synchronization verification: the timestamp deviation of each sensor data is checked by the synchronization pulse (period 1ms) generated by the FPGA clock module. The measured deviation is ≤3μs, meeting the microsecond-level synchronization requirements.

[0028] (2) Construction of PINN diagnostic model Three core physical laws are selected as constraints, and a neural network loss function is embedded: the stator and rotor air gap magnetic flux density equation: B(θ,t)=Bm×sin(ωt-pθ+φ), where Bm is the magnetic flux density amplitude (0.8T under normal operating conditions), ω is the electric angular velocity (2π×50rad / s), p is the number of pole pairs (4 pairs), θ is the mechanical angle, and φ is the initial phase; the winding temperature rise differential equation: dT / dt=(Pk×(T-T0)) / C, where P is the winding loss (1.2kW under normal operating conditions), k is the heat dissipation coefficient (0.05kW / ℃), T0 is the ambient temperature (measured at 30℃), and C is the winding heat capacity (0.5kJ / ℃); the gearbox transmission ratio relationship: n1 / n2=i, where n1 is the motor speed, n2 is the hook drum speed, and i is the transmission ratio (20:1).

[0029] The loss function consists of the data fitting term (Loss) data ) and physical law constraint terms (Loss) phy Composed of ( ), the weighting coefficients were determined through cross-validation (0.6:0.4): Loss data =(1 / N)×Σ||y pred -y true || 2 Where N is the number of samples, y pred y represents the failure probability output by the model. true Labels (e.g., "winding short circuit" is labeled [1,0,0,0,0]); Loss phy =(1 / 3)×(||B pred -B true || 2 +||T pred -T true || 2 +||n2 pred -n2 true || 2 ), where B pred T pred n2pred B is the physical quantity predicted by the model. true T true n2 true These are measured values.

[0030] (3) Multimodal feature fusion The collected vibration, current, and audio signals were processed using FFT (16384 FFT points, 0.73Hz frequency resolution): Vibration signal: Features in the 1-3kHz frequency band were extracted, including the bearing outer ring defect frequency (1850Hz, calculated as f0=0.3×n×z, where n is the motor speed 740r / min and z is the number of bearing rolling elements 8) and the gear meshing frequency (2100Hz, calculated as f0). m =n×z_g / 60, z g (Gear with 28 teeth), forming a 64-dimensional F vib Current signal: Extract features from the 0-500Hz frequency band, including the 50Hz fundamental amplitude, the 150Hz third harmonic amplitude, and the 250Hz fifth harmonic amplitude; calculate the current THD (Total Harmonic Distortion); and form a 32-dimensional F-wave signal. current Audio signal: Extract features from the 5-8kHz frequency band, including the amplitude and spectral entropy of 6.2kHz (bearing lubrication failure friction sound) and 7.5kHz (gear tooth breakage impact sound), forming a 48-dimensional F... audio .

[0031] Calculate the initial values ​​of α, β, and γ using mutual information entropy (MI): Calculate F vib With MI=0.85 and F for the fault label, current With MI=0.72 and F fault tag, audio The MI of the fault label is 0.68; the initial weights are: α = 0.85 / (0.85+0.72+0.68) = 0.38, β = 0.72 / 2.25 = 0.32, γ = 0.68 / 2.25 = 0.30 (satisfying α+β+γ=1).

[0032] The weights are adjusted based on real-time operating conditions: if the detected vibration peak exceeds 10% of the normal operating condition (1.2g) (i.e., 1.32g), it is determined to be a mechanical fault characteristic, so α is increased to 0.38+0.2=0.58, β is adjusted to 0.32-0.1=0.22, and γ is adjusted to 0.30-0.1=0.20; if the current THD is 6%>5%, β is increased to 0.22+0.15=0.37, α is adjusted to 0.58-0.1=0.48, and γ is adjusted to 0.20-0.05=0.15; if the measured ambient noise is 75dB<80dB, γ remains at 0.15.

[0033] Final composite feature vector: F fused =0.48×Fvib +0.37×F current +0.15×F audio The dimensions are 64+32+48=144, which are expanded to 256 dimensions through zero padding for input to the PINN model.

[0034] (4) Real-time diagnosis and emergency response at the edge The edge diagnostic terminal receives a Ffused signal every 10ms and inputs it into the lightweight PINN model for inference: when the probability of detecting a "bearing pitting" fault is 45% (≥30%), a first-level response is triggered: a warning signal is output to the driver's cab audible and visual alarm, and at the same time, "risk of bearing pitting at the rear of the motor, current probability 45%" is displayed on the touch screen, and the Ffused signal at that moment is recorded (vibration characteristics: 1850Hz amplitude is 1.8 times the normal value, current THD=4%, audio 6.2kHz amplitude is 1.5 times the normal value); after 1 hour, the fault probability rises to 65% (≥60%), triggering a second-level response: a speed reduction command is sent to the PLC via the RS485 interface, the motor speed is reduced from 740r / min to 370r / min, the hook lifting speed is reduced from 1.5m / s to 0.5m / s, and the Ffused signal for 10 seconds before and after that moment is compressed (compressed data size 800KB) and uploaded to the cloud via the 5G module.

[0035] (5) Cloud-based in-depth analysis and diagnostic report generation The cloud receives data uploaded from the edge and inputs it into the complete PINN model for in-depth analysis: combined with the fault knowledge graph, it matches the historical case "Port crane bearing pitting case in May 2023" (features: abnormal amplitude at 1850Hz, normal THD, abnormal at 6.2kHz, repair solution: replace bearing 6319), with a matching similarity of 98%; the HI evaluation model calculates the current HI=0.65 (normal HI≤0.3, fault threshold HI≥0.5), and the LSTM network predicts RUL=25 days (that is, HI will reach 0.9 after 25 days, requiring shutdown for maintenance).

[0036] A diagnostic report is generated, including: Fault location: Pitting corrosion in the rear bearing (model 6319) of the motor, with 3D marking of the fault location; Physical cause: Wear of the bearing raceway leads to enhanced high-frequency resonance characteristics (1850Hz) of vibration, which conforms to the law of "defects causing impact vibration" in mechanical kinematics; Winding temperature rise is normal (45℃), ruling out the influence of electrical faults; Remaining life: 25 days (confidence level 90%); Maintenance recommendation: Schedule maintenance within 15 days, replace bearing 6319, and check the gearbox grease at the same time.

[0037] (6) Edge-cloud model collaborative update New fault samples uploaded from the edge every week are collected in the cloud (such as the "bearing pitting" sample in this case), merged with historical samples (a total of 5001 groups), and the full version of the PINN model is retrained. After training, the model accuracy is improved from 98% to 98.5%. The cloud performs differential compression on the new model parameters (100MB) (8MB after compression) and distributes them to the edge terminal through the 5G module. After receiving the parameters, the edge terminal integrates the new samples using an incremental learning method: freezes the first 3 layers of the model (feature extraction layer) and only updates the last 2 layers (classification layer). The training iteration is 100 rounds, taking ≤5 minutes. The updated model improves the recognition accuracy of "bearing pitting" from 95% to 97%.

[0038] (7) Predictive maintenance decision generation and visualization output The cloud-based maintenance decision generator generates a maintenance strategy based on the diagnostic report: Maintenance time window: It is recommended to complete maintenance within 10-15 days (avoiding peak port operation hours: 8:00-18:00 daily); Spare parts replacement list: Bearing 6319 (quantity 1, model matching motor speed 740r / min), grease (model lithium-based grease 2#, quantity 1kg), seals (quantity 2); Operation instructions: ① Power off and tag, ② Remove the motor rear end cover, ③ Remove the old bearing and clean the journal, ④ Install the new bearing (heat to 80℃ for easier installation), ⑤ Add grease (fill to 1 / 3 of the bearing's internal space), ⑥ Run under no-load for 30 minutes to check vibration and temperature, ⑦ Run under load for 1 hour to confirm normal operation.

[0039] The maintenance strategy is visualized through the human-machine interface terminal: the industrial touch screen displays the maintenance work order and supports the operator to confirm "received"; the mobile application pushes the work order to the maintenance personnel's mobile phone, including the spare parts procurement link (linked to the port spare parts warehouse) and maintenance step animation (3D demonstration of the bearing replacement process).

[0040] (8) High-risk fault safety protection (simulation verification) To verify the response to high-risk faults, a "broken shaft" fault was artificially set up (an eccentric block was installed on the motor output shaft to simulate the characteristics of a broken shaft): the edge terminal detected that the amplitude of 2 × frequency (24.7Hz, calculation formula: 2 × 740 / 60≈24.7Hz) in the vibration characteristics was 5 times the normal value, the fault probability "broken shaft" = 95%, triggering a level 3 response; the safety linkage control unit sent a braking command to the PLC through the Profisafe protocol, activated the hydraulic brake within 0.3 seconds, and the motor speed dropped to 0 within 0.5 seconds; at the same time, the motor drive power was cut off (contactor KM1 was disconnected) to prevent the hook from falling due to shaft breakage; an alarm message was immediately sent to the port monitoring center (including the fault type "broken shaft", the time "April 10, 2024, 14:30", and the location "motor output shaft"), and an emergency alarm was pushed to the mobile application terminal (red pop-up window, accompanied by a prompt sound).

[0041] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A fault diagnosis method for a crane hook lifting motor, characterized in that, Includes the following steps: S1. Real-time acquisition of multi-modal operating data of the crane hook lifting motor through a multi-sensor synchronous acquisition system, including three-axis vibration signals, three-phase current signals, audio signals, temperature distribution data, and motor speed / position signals; S2. Construct a Physical Information Neural Network (PINN) diagnostic model. This model embeds the electromagnetic induction law of motors, thermodynamic equations, and mechanical kinematics as physical constraints into the neural network loss function, forming a hybrid loss function that includes data fitting terms and physical law constraint terms. S3. Perform Fast Fourier Transform (FFT) on the vibration signal, current signal and audio signal respectively, extract the spectral features of each signal, analyze the coupling relationship between multimodal signals by calculating the cross-correlation function, and establish a weighted fusion matrix to generate a composite fault feature vector; S4. Deploy a lightweight PINN model on edge computing nodes to perform real-time analysis of composite fault feature vectors and trigger emergency control commands when an abnormal state is detected; at the same time, compress key feature data and upload it to the cloud. S5, the cloud-based diagnostic engine receives data uploaded from edge nodes, runs the full version of the PINN model for in-depth analysis, combines the historical fault database with the equipment health index (HI) assessment model, and outputs a diagnostic report that includes fault location information, explanation of physical causes, and prediction of remaining service life. S6. Establish an edge-cloud model collaborative update mechanism: The cloud periodically sends the trained and optimized model parameters to the edge nodes, and the edge nodes integrate new fault samples through incremental learning to keep the diagnostic model dynamically evolving; S7. Generate predictive maintenance strategies based on diagnostic results, including maintenance time window suggestions, spare parts replacement lists and operation guidance plans, and output them visually through a human-machine interface; S8. When a high-risk fault is diagnosed, the crane safety protection protocol is automatically triggered to restrict the lifting and lowering of the hook and send alarm information to the monitoring terminal.

2. The fault diagnosis method for the crane hook lifting motor according to claim 1, characterized in that, The multi-sensor synchronous acquisition system described in step S1 uses a hardware triggering circuit to achieve microsecond-level time synchronization. The sampling frequency of the vibration signal is not less than 10kHz, the sampling frequency of the current signal is not less than 2kHz, and the sampling frequency of the audio signal is not less than 16kHz.

3. The fault diagnosis method for the crane hook lifting motor according to claim 1, characterized in that, Step S3, frequency domain fusion, specifically includes: extracting 1-3kHz high-frequency resonance features from the vibration signal, extracting 0-500Hz fundamental harmonic features from the current signal, and extracting 5-8kHz mechanical friction features from the audio signal; determining the weight coefficients of each mode through feature correlation analysis, and constructing the mathematical expression for the composite feature vector as follows: in, F fused This is a feature vector of a multimodal composite fault. F vib For vibration signal feature vectors, F current For current signal feature vectors, F audio For audio signal feature vectors, α , β , γ These are dynamic weighting coefficients calculated based on mutual information entropy.

4. The fault diagnosis method for the crane hook lifting motor according to claim 1, characterized in that, The physical constraints include the motor stator and rotor air gap magnetic flux density equation, the winding temperature rise differential equation, and the gearbox transmission ratio relationship. During the neural network training process, the data consistency is verified in real time through the physical residual calculation module. When the deviation of the physical law exceeds the threshold, the network weight is automatically adjusted.

5. A fault diagnosis system for a crane hook lifting motor, characterized in that, include: Multimodal data acquisition module: integrates a triaxial MEMS accelerometer, a closed-loop Hall current sensor, a wideband microphone array, an infrared thermal imager, and a photoelectric encoder, and is connected to the edge computing unit via an industrial fieldbus; Edge diagnostic terminal: It adopts an embedded GPU computing platform, with built-in lightweight PINN diagnostic model, real-time signal processing algorithm library and device control interface, supporting millisecond-level fault response; Cloud-based analytics platform: Deployed on a high-performance computing cluster, it includes a physical information neural network training engine, a fault knowledge graph database, a remaining life prediction module, and a maintenance decision generator; Collaborative communication gateway: It adopts 5G industrial modules to realize encrypted data transmission from the edge to the cloud, and supports differential compression of model parameters and breakpoint resume function; Human-machine interaction terminal: Equipped with an industrial touch screen and mobile application terminal, it realizes three-dimensional fault location visualization, dynamic demonstration of physical mechanism and maintenance work order management functions; Safety linkage control unit: Deeply integrated with the crane PLC system, it automatically activates the mechanical braking device and cuts off the motor drive power when a high-risk fault such as broken shaft, short circuit in winding, or broken gear is diagnosed.

6. The fault diagnosis system for the crane hook lifting motor according to claim 5, characterized in that, The edge diagnostic terminal adopts a modular hardware architecture, including a signal conditioning circuit board, a Jetson Xavier NX core computing module, an isolated digital output board, and four PoE gigabit Ethernet interfaces.

7. The fault diagnosis system for the crane hook lifting motor according to claim 5, characterized in that, The cloud-based analytics platform's fault knowledge graph database contains a library of typical crane motor fault cases, which are associated with fault feature patterns, root cause chains, and maintenance solutions. It uses graph neural networks to achieve intelligent matching and reasoning of fault cases.

8. The fault diagnosis system for the crane hook lifting motor according to claim 5, characterized in that, The safety linkage control unit has built-in multi-level protection logic, including: Level 1 response: issuing an early warning when fault characteristics first appear; Level 2 response: reducing load when the fault probability exceeds the threshold; Level 3 response: immediately shutting down when a high-risk fault is identified; the response actions at each level are transmitted to the actuator through the Profisafe protocol with safety certification.

Images