Brain-like chip hardware acceleration method and system

Abstract

The application discloses a brain-like chip hardware acceleration method and system, relates to the related field of brain-like computing, and comprises the following steps: collecting heterogeneous time sequence data streams in real time through a multi-modal sensor array on an industrial device; performing event detection to generate a sparse pulse event stream; deploying a brain-like chip, which internally integrates a pulse neural network and an on-chip network of a memory-computing integrated architecture; inputting the sparse pulse event stream into the brain-like chip, distributing the pulse events through the on-chip network, performing asynchronous event-driven parallel processing on the sparse pulse event stream by using a dynamic calculation distribution mechanism of the pulse neural network, and obtaining a real-time state feature vector of the industrial device; and fusing and reasoning with a pre-stored device health knowledge graph to generate a control instruction or a fault early warning signal. The application solves the technical problems of slow calculation speed, low efficiency and poor real-time performance of existing data processing systems, achieves real-time and efficient processing of industrial data, and further improves the technical effects of device operation stability and reliability.

Description

Technical Field

[0001] This application relates to the field of neuromorphic computing, and in particular to neuromorphic chip hardware acceleration methods and systems. Background Technology

[0002] The increasing real-time requirements for industrial equipment condition monitoring and fault early warning pose a severe challenge to the computational efficiency of data processing systems. Currently, the industry generally adopts computing architectures based on general-purpose processors, accelerating data processing by increasing clock speed, adding cores, or introducing GPU coprocessing. However, these traditional acceleration solutions are limited by the memory wall problem of the von Neumann architecture, and their fixed-period synchronous computing mode is incompatible with the sparse event characteristics of sensor data, resulting in a significant waste of computing resources on invalid data and encountering bottlenecks in acceleration efficiency.

[0003] Currently, data processing systems suffer from technical problems such as slow computing speed, low efficiency, and poor real-time performance. Summary of the Invention

[0004] This application provides a neuromorphic chip hardware acceleration method and system. It employs a multimodal sensor array from industrial equipment to collect heterogeneous time-series data streams in real time, generates sparse pulse event streams through event detection, and deploys a neuromorphic chip integrating a memory-computing architecture with a spiking neural network and an on-chip network. The sparse pulse event streams are input into the neuromorphic chip, distributed by the on-chip network, and processed in parallel using the dynamic computation and allocation mechanism of the spiking neural network to obtain real-time state feature vectors. These feature vectors are then fused with a local equipment health knowledge graph for reasoning to generate control commands or fault warning signals. This approach solves the technical problems of slow computation speed, low efficiency, and poor real-time performance in existing data processing systems, achieving real-time and efficient processing of industrial data and thus improving the stability and reliability of equipment operation.

[0005] This application provides a neuromorphic chip hardware acceleration method, comprising: acquiring heterogeneous time-series data streams in real time through a multimodal sensor array on an industrial device; performing event detection on the heterogeneous time-series data streams to generate sparse pulse event streams; deploying a neuromorphic chip, wherein the neuromorphic chip integrates a spiking neural network and an on-chip network with an in-memory computing architecture; inputting the sparse pulse event streams into the neuromorphic chip, distributing the pulse events through the on-chip network, and using the dynamic computation allocation mechanism of the spiking neural network to perform asynchronous event-driven parallel processing on the sparse pulse event streams to obtain a real-time state feature vector of the industrial device; fusing and reasoning the real-time state feature vectors with a device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial device; wherein the dynamic computation allocation mechanism is configured to activate the corresponding neurons for computation when a pulse event arrives.

[0006] In a possible implementation, event detection is performed on the heterogeneous time-series data stream to generate a sparse pulse event stream, and the following processing is performed: extracting a first time-series data stream from the heterogeneous time-series data stream; calculating a first data change rate of the first time-series data stream in real time; when the first data change rate is greater than a preset first noise tolerance threshold, it is determined to be a valid event, and a first pulse signal is generated, and multiple first pulse signals constitute the sparse pulse event stream; wherein, the first pulse signal includes the sensor ID, intensity, and timestamp of the valid event.

[0007] In a possible implementation, the sparse pulse event stream is input to the neuromorphic chip, the pulse events are distributed through the on-chip network, and the dynamic computation allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. The following processing is performed: the neuromorphic chip integrates an on-chip network router; the sparse pulse event stream is routed to the computing core corresponding to the spiking neural network through the on-chip network router; each computing core performs asynchronous event-driven parallel processing based on the received pulse events by adjusting the membrane potential and synaptic weights of neurons to obtain the real-time state feature vector of the industrial equipment.

[0008] In a possible implementation, the following process is performed: each computing core is loaded with pre-trained initial synaptic weights, which are obtained by: collecting historical sensor data; training and converging the neural network using the historical sensor data to obtain initial weight values; and loading the initial weight values ​​onto the physical synapses of the neuromorphic chip to obtain initial synaptic weights.

[0009] In a possible implementation, the synaptic weights are adjusted by performing the following processing: based on the pulse timing-dependent plasticity rule, the values ​​of the synaptic weights are dynamically adjusted according to the firing sequence of the preceding and following neurons; wherein, when the pulse firing time of the first neuron is always earlier than the pulse firing time of the second neuron within a preset time period, the first synaptic weight connecting the first neuron and the second neuron is strengthened, and conversely, the first synaptic weight is weakened; wherein, the first neuron is a neuron in the previous layer, and the second neuron is a neuron in the next layer.

[0010] In a possible implementation, the real-time state feature vector is fused and reasoned with the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment. The following processing is performed: the relevance weights between the real-time state feature vector and entity nodes in the device health knowledge graph are calculated; historical relevant cases and rules are retrieved according to a preset relevant weight threshold; the historical relevant cases and relevant rules are used together as the reasoning basis to reason about the real-time state feature vector to generate control commands or fault warning signals for the industrial equipment; wherein, the device health knowledge graph stores the component relationships, historical fault modes, and expert operation and maintenance procedures of the industrial equipment in a graph structure.

[0011] In a possible implementation, a heterogeneous time-series data stream is acquired in real time using a multimodal sensor array on an industrial device, and the following processing is performed: a broadband acoustic signature signal generated by the operation of the industrial device is acquired using a contact piezoelectric microphone to obtain an acoustic signature time-series data stream; vibration acceleration signals of the mechanical structure of the device in three orthogonal axes are acquired using a MEMS vibration sensor to obtain a vibration time-series data stream; dynamic visual information of the device's operating status is acquired using an event camera to obtain a video time-series data stream; and tension signals of the transmission chain are acquired using a Wheatstone bridge composed of resistance strain gauges to obtain a tension time-series data stream; wherein, the acoustic signature time-series data stream, the vibration time-series data stream, the video time-series data stream, and the tension time-series data stream constitute the heterogeneous time-series data stream.

[0012] This application also provides a neuromorphic chip hardware acceleration system, comprising: a heterogeneous time-series data stream acquisition module for real-time acquisition of heterogeneous time-series data streams through a multimodal sensor array on an industrial device; an event detection module for event detection of the heterogeneous time-series data streams and generation of sparse pulse event streams; a neuromorphic chip deployment module for deploying a neuromorphic chip, wherein the neuromorphic chip integrates a spiking neural network with an in-memory computing architecture and an on-chip network; a neuromorphic computing module for inputting the sparse pulse event streams into the neuromorphic chip, distributing the pulse events through the on-chip network, and using the dynamic computing allocation mechanism of the spiking neural network to perform asynchronous event-driven parallel processing on the sparse pulse event streams to obtain a real-time state feature vector of the industrial device, wherein the dynamic computing allocation mechanism is configured to activate the corresponding neuron for computation when a pulse event arrives; and a fusion reasoning module for fusion reasoning of the real-time state feature vectors with a device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial device.

[0013] The proposed neuromorphic chip hardware acceleration method and system first acquires heterogeneous time-series data streams in real time using a multimodal sensor array on industrial equipment. Then, it performs event detection on the heterogeneous time-series data streams to generate a sparse pulse event stream. Next, it deploys a neuromorphic chip, which integrates a memory-based spiking neural network and an on-chip network. The sparse pulse event stream is then input into the neuromorphic chip, where the on-chip network distributes the pulse events. Utilizing the dynamic computation allocation mechanism of the spiking neural network, the sparse pulse event stream undergoes asynchronous event-driven parallel processing to obtain a real-time state feature vector of the industrial equipment. The dynamic computation allocation mechanism is configured to activate corresponding neurons for computation when a pulse event arrives. Finally, the real-time state feature vector is fused and reasoned with a pre-stored equipment health knowledge graph on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment. This achieves the technical effect of real-time and efficient processing of industrial data, thereby improving the stability and reliability of equipment operation. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0015] Figure 1 This is a flowchart illustrating the neuromorphic chip hardware acceleration method provided in an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the structure of a neuromorphic chip hardware acceleration system provided in an embodiment of this application.

[0017] Figure labeling: Heterogeneous time-series data stream acquisition module 10, event detection module 20, neuromorphic chip deployment module 30, neuromorphic computing module 40, fusion reasoning module 50. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This application provides a method for hardware acceleration of neuromorphic chips, such as... Figure 1 As shown, the method includes:

[0020] Step S100: Collect heterogeneous time-series data streams in real time through a multimodal sensor array on an industrial device.

[0021] Specifically, a multimodal sensor array refers to a collection of various types of sensors that digitize analog signals through synchronous sampling circuits and analog-to-digital converters, enabling the acquisition of multi-dimensional information about the equipment. Heterogeneous time-series data streams refer to collections of time-series data from different sources, in different formats, and with different characteristics, such as simultaneously containing different types of data streams like sound, vibration, and images. By utilizing sensors installed in different parts of industrial equipment with different sensing functions, various types of time-series data generated during equipment operation are simultaneously collected. These data are timestamped using a unified time reference. Because these data differ in format and characteristics, they form heterogeneous time-series data streams.

[0022] In one possible implementation, a heterogeneous time-series data stream is acquired in real time using a multimodal sensor array on an industrial device. Step S100 further includes step S110, which involves acquiring a wideband acoustic signature signal generated by the operation of the industrial device using a contact piezoelectric microphone to obtain an acoustic signature time-series data stream. Specifically, a contact piezoelectric microphone is a sensor that uses the piezoelectric effect to convert sound vibrations into electrical signals. By contacting the surface of the device with the contact piezoelectric microphone, the sound vibrations generated by the device's operation are sensed, converted into electrical signals, and acoustic signature signals with a wide frequency range are acquired, forming an acoustic signature time-series data stream that varies over time.

[0023] For example, when monitoring wind turbine blades, a contact piezoelectric microphone is fixed at the blade root near the hub. When the wind turbine is running, the blades rotate under wind force. Friction between the blades and the air, the blades' own vibrations, and any slight deformations that may occur generate sounds of different frequencies. The piezoelectric microphone converts these sound vibrations into electrical signals, collecting signals with frequencies ranging from tens of hertz to thousands of hertz, forming a broadband acoustic signature signal. These signals are recorded chronologically to obtain a temporal data stream reflecting the blade's operating status.

[0024] Step S120: The vibration acceleration signals of the equipment's mechanical structure in three orthogonal axes are acquired using a MEMS vibration sensor to obtain a vibration time-series data stream. Specifically, the MEMS vibration sensor is a vibration-sensing sensor manufactured based on microelectromechanical systems technology. Using the MEMS vibration sensor, the vibration acceleration of the equipment's mechanical structure in three mutually perpendicular directions, namely the X, Y, and Z directions in space, is measured simultaneously. The vibration acceleration reflects the intensity of the equipment's vibration. The acceleration data is recorded over time to form a vibration time-series data stream.

[0025] For example, when monitoring the main shaft of a wind turbine, a MEMS vibration sensor is installed near the main shaft bearing housing. When the wind turbine is operating, the main shaft rotates at high speed driven by wind power, while simultaneously bearing various loads transmitted from the blades. This causes the main shaft to vibrate to varying degrees in the X, Y, and Z directions. The MEMS vibration sensor simultaneously measures the vibration acceleration in these three directions and records these measurements in chronological order, forming a vibration time-series data stream that reflects the vibration status of the main shaft during operation.

[0026] Step S130: Dynamic visual information of the device's operating status is acquired through an event camera to obtain a video temporal data stream. Specifically, the event camera is an event-driven visual sensor. Instead of acquiring images at a fixed frame rate, this sensor triggers event recording based on changes in pixel brightness in the scene, acquiring visual changes during device operation to form a video temporal data stream.

[0027] For example, when monitoring the movement of a wind turbine nacelle, an event camera is used to point at the nacelle. When the external wind speed or direction changes, or when the wind turbine itself performs operations such as yaw or pitch control, the position and attitude of the nacelle will change, causing changes in pixel brightness in the scene. When the event camera detects a change in pixel brightness, it will trigger event recording, forming a video time-series data stream reflecting the dynamic movement of the nacelle.

[0028] Step S140: Tension signals from the transmission chain are acquired using a Wheatstone bridge constructed from resistance strain gauges to obtain a tension time-series data stream. The acoustic signature time-series data stream, the vibration time-series data stream, the video time-series data stream, and the tension time-series data stream constitute the heterogeneous time-series data stream. Specifically, a resistance strain gauge is a sensor that converts mechanical deformation into resistance changes. The strain gauges are arranged into a Wheatstone bridge circuit. A Wheatstone bridge is a circuit used to accurately measure minute changes in resistance. The Wheatstone bridge is installed in the transmission chain of the equipment. When the transmission chain is subjected to tension, the resistance strain gauges deform, causing a change in resistance. This change in resistance is measured by the Wheatstone bridge and converted into a tension signal, which is recorded over time to form the tension time-series data stream. The tension time-series data stream, together with the acquired acoustic signature, vibration, and video time-series data streams, constitutes a heterogeneous time-series data stream reflecting the overall operating status of the industrial equipment.

[0029] For example, when monitoring the tower of a wind turbine, resistance strain gauges are attached to key stress-bearing areas of the tower, with multiple strain gauges forming a Wheatstone bridge. When the wind turbine is running, the tower is subjected to enormous loads transmitted from the blades and the force of the wind, resulting in a certain degree of deformation. This deformation causes changes in the resistance of the strain gauges, and the Wheatstone bridge outputs an electrical signal proportional to the tension experienced by the tower. These electrical signals are recorded sequentially over time, for example, recording the tension value once per second, forming a tension time-series data stream that reflects the stress on the tower.

[0030] Step S200: Perform event detection on the heterogeneous time-series data stream to generate a sparse pulse event stream.

[0031] Specifically, event detection is the identification of specific changes or patterns in a data stream. It refers to identifying significant change points from a continuous data stream, identifying meaningful events from a heterogeneous time-series data stream, and converting them into sparse pulse signal forms. The sparse pulse event stream is a set of events represented in the form of pulse signals. The pulse signals are only generated when a valid event is detected, and there is no data output at non-event times. It is sparser and has less data volume than the original data stream.

[0032] In one possible implementation, event detection is performed on the heterogeneous time-series data stream to generate a sparse pulse event stream. Step S200 further includes step S210, extracting a first time-series data stream by traversing the heterogeneous time-series data stream. Specifically, traversal refers to accessing each element in the data set in a certain order, and the first time-series data stream refers to any specific type of data stream separated from the heterogeneous time-series data stream. A data selector or multiplexer selects one type of time-series data stream from the parallel data streams for processing based on a preset sensor ID. For example, a configuration register specifies the sensor channel to be extracted, and the hardware logic intercepts the sampling points of that channel in real time to form a continuous sub-data stream.

[0033] Step S220: Calculate the first data change rate of the first time-series data stream in real time. Specifically, the first data change rate is an indicator that measures how quickly the data value in the first time-series data stream changes over time, characterizing the slope of the signal change over time and reflecting the degree of state abrupt change. A hardware differential or a subtractor module in a field-programmable gate array is used to perform differential operations on consecutive sampling points. For example, for a signal with a sampling period of T, the formula for calculating the change rate at time t is: the current sampled value minus the previous sampled value, divided by the sampling period. The result is stored in a register for threshold comparison.

[0034] Step S230: When the first data change rate is greater than a preset first noise tolerance threshold, it is determined to be a valid event, and a first pulse signal is generated. Multiple first pulse signals constitute the sparse pulse event stream. Each first pulse signal includes the sensor ID, intensity, and timestamp of the valid event. Specifically, the first data change rate corresponds one-to-one with the first noise tolerance threshold. The first noise tolerance threshold is used to distinguish between normal data fluctuations and valid events. It is a threshold set based on the statistical characteristics of sensor noise and is used to filter out minor fluctuations caused by environmental noise. The comparator circuit compares the calculated first data change rate with the first noise tolerance threshold stored in the configuration register. If it is greater than the threshold, a valid event is detected, triggering pulse encoding to generate a fixed-format digital pulse. This pulse carries event attribute information, including three parts: sensor ID, intensity, and timestamp. The sensor ID is encoded by a hardware address, the intensity is the quantized value of the first data change rate, and the timestamp is read from a global clock counter. Multiple such pulse signals form the sparse pulse event stream.

[0035] Step S300: Deploy a neuromorphic chip, which integrates a spiking neural network and an on-chip network with an in-memory computing architecture.

[0036] Specifically, in-memory computing architecture is a hardware design that integrates data storage and computation, eliminating data transfer overhead. Spiking neural networks are networks modeled after brain neurons, used for data processing; computing units communicate via pulse sequences, exhibiting event-driven characteristics. On-chip networks are multi-core interconnect architectures within a chip, used for internal data transmission and communication.

[0037] A neuromorphic chip is installed in the system. This chip employs an application-specific integrated circuit (ASIC) design, where the computing core and memory are integrated using three-dimensional stacking technology. The in-memory computing architecture utilizes resistive random access memory (RRAM) or phase-change memory (PCM) to form a cross-switch array, with synaptic weights directly stored in memristor cells, achieving physical unification of multiply-accumulate operations and storage. The on-chip network adopts a mesh topology, consisting of routing nodes and links, supporting multicast communication.

[0038] Step S400: The sparse pulse event stream is input into the neuromorphic chip. The pulse events are distributed through the on-chip network. The dynamic calculation allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. The dynamic calculation allocation mechanism is configured to activate the corresponding neuron for calculation when a pulse event arrives.

[0039] Specifically, the sparse pulse event stream generated by the detection is input to the on-chip network interface controller via chip pins. The neuromorphic chip, through the on-chip network, allocates the pulse events to the corresponding processing units based on the target core ID in the pulse packet header. A spiking neural network dynamically allocates computing resources based on event arrival; when a pulse arrives, only the neurons connected to that pulse are activated, while other neurons remain dormant to save power. Pulse events are processed in parallel, ultimately extracting the device's real-time state feature vector. This vector is a multi-dimensional feature extracted from the firing patterns of neurons in the network's output layer, representing the device's current operating state. The dynamic computation allocation mechanism is an energy-saving strategy that triggers the computation of relevant neurons only when an input event occurs. Asynchronous event-driven processing means that the processing is triggered by events, rather than at fixed time intervals. The real-time state feature vector is a set of features that reflects the device's current operating state.

[0040] In one possible implementation, the sparse pulse event stream is input to the neuromorphic chip, the pulse events are distributed through the on-chip network, and the dynamic computation allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. Step S400 further includes step S410, where the neuromorphic chip integrates an on-chip network router, and the sparse pulse event stream is routed to the corresponding computing core of the spiking neural network through the on-chip network router. Specifically, the neuromorphic chip has an on-chip network router, which is a component in the on-chip network used for data routing, responsible for transmitting data from the source address to the destination address. The router parses the destination address in the input pulse packet, competes for the output port through an arbiter, and forwards the data slice to the next-hop node using a wormhole switching method until it is transmitted to the computing core in the spiking neural network responsible for processing the event. The computing core is the unit in the spiking neural network that performs specific computation tasks, and each computing core is equipped with an input buffer for temporarily storing arriving pulse events.

[0041] In step S420, each computing core, based on the received pulse events, performs asynchronous event-driven parallel processing by adjusting the membrane potential and synaptic weights of neurons to obtain the real-time state feature vector of the industrial equipment. Specifically, each computing core changes the membrane potential of its internal neurons and the synaptic weights of the connections between neurons according to the received pulse event information, processing the pulse events in parallel asynchronously. That is, multiple computing cores process independently and simultaneously according to the arrival of events, ultimately extracting the real-time state feature vector of the equipment. Specifically, when a pulse arrives, the synaptic current is injected into the membrane capacitance through a transconductance amplifier, and the membrane voltage decays exponentially. When the voltage exceeds a threshold, a comparator triggers pulse delivery and resets the membrane voltage. The synaptic weight values ​​are stored in a non-volatile memory unit, and the injected current magnitude is controlled by a digital-to-analog converter. The membrane potential is an electrical quantity that simulates the membrane voltage of biological neurons and is used to integrate the input pulse. The synaptic weight is a parameter representing the strength of connections between neurons and affects the amplitude of pulse transmission.

[0042] In one possible implementation, step S420 further includes step S421, wherein each of the computing cores is loaded with pre-trained initial synaptic weights, the initial synaptic weights being obtained by: collecting historical sensor data; using the historical sensor data to train and converge the neural network to obtain initial weight values; and loading the initial weight values ​​onto the physical synapses of the neuromorphic chip to obtain initial synaptic weights.

[0043] Specifically, the initial synaptic weights are the initial weight values ​​determined through offline training before network deployment. Physical synapses are the actual circuit structures in a neuromorphic chip that realize neuronal connections. Trained synaptic weights are pre-set in the computational core of the neuromorphic chip. These weights are obtained by collecting sensor data from previous device operation, using the backpropagation algorithm on a general-purpose computer to train the neural network to a stable state, and then quantizing the trained weight matrix and programming it into the weight memory of the neuromorphic chip via a JTAG or PCIe interface. For example, weight values ​​are mapped to the conductance values ​​of resistive switching memory, and the resistance value is set by writing circuitry.

[0044] In one possible implementation, adjusting the synaptic weights, step S420 further includes step S422, which dynamically adjusts the value of the synaptic weights based on the pulse timing-dependent plasticity rule and the firing sequence of the preceding and following neurons. Specifically, when the pulse firing time of the first neuron is always earlier than the pulse firing time of the second neuron within a preset time period, the first synaptic weight connecting the first neuron and the second neuron is enhanced; conversely, the first synaptic weight is weakened. Here, the first neuron is a neuron in the previous layer, and the second neuron is a neuron in the next layer.

[0045] Specifically, pulse timing-dependent plasticity is a biological plasticity rule based on adjusting weights according to the order of pulses. Each synaptic circuit contains timing detection logic that records the pulse firing times of preceding and following neurons. These neurons are neurons in the neural network that have a sequential information transmission relationship. If, within a preset time period, the pulse firing time of the preceding neuron is always earlier than that of the following neuron, the weight of the corresponding synapse is increased; otherwise, it is decreased. The adjustment is achieved through a charge pump circuit, changing the resistance value of the memristor.

[0046] Step S500: The real-time state feature vector is fused and reasoned with the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment.

[0047] Specifically, the equipment health knowledge graph is a structured knowledge base describing equipment components, failure modes, and maintenance knowledge. The equipment health knowledge graph is stored in the embedded memory of the neuromorphic chip as a graph database. The similarity between the real-time state feature vector and nodes in the equipment health knowledge graph is calculated using cosine distance. When the similarity exceeds a preset threshold, the corresponding entity node is activated. The edges connecting the activated nodes are traversed, and the currently activated node pattern is matched against rules in a pre-existing rule base. Successfully matched rules trigger corresponding actions, outputting digital control commands or fault warning signals through the chip's general-purpose input / output interface or communication interface. The control commands are used to adjust the operating parameters of the industrial equipment, and the fault warning signals are used to indicate potential equipment failure risks.

[0048] In one possible implementation, the real-time state feature vector is fused and reasoned with a device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment. Step S500 further includes step S510, calculating the correlation weight between the real-time state feature vector and entity nodes in the device health knowledge graph, and retrieving historical relevant cases and rules based on a preset correlation weight threshold. The device health knowledge graph stores component relationships, historical fault modes, and expert maintenance procedures for the industrial equipment in a graph structure. Specifically, the correlation weight is a numerical value used to quantify the degree of association between the real-time state feature vector and a specific entity node in the device health knowledge graph. The preset correlation weight threshold is a threshold value used to filter out weakly correlated information. Historical relevant cases are records of past faults or operating conditions stored in the retrieved device health knowledge graph that have been diagnosed and whose correlation with the real-time state feature vector is greater than the preset correlation weight threshold. The relevant rules are logical statements summarized by experts, which are more correlated with the real-time state feature vector than a preset relevant weight threshold, and indicate what kind of operation should be triggered under specific conditions, such as "IF condition A AND condition B THEN execute action C".

[0049] Specifically, during the construction of the device health knowledge graph, each entity node, such as a specific fault mode, a specific device component, or a specific maintenance action, has a corresponding feature vector representation. During inference, the hardware circuitry calculates the cosine similarity between the real-time state feature vector and the feature vectors of these nodes in parallel; this calculation result is the relevance weight. An approximate nearest neighbor search hardware accelerator is used to quickly identify all entity nodes with relevance weights greater than a threshold. These activated nodes are then used as keywords, and an index search is performed in the case library and rule library to find all historical relevant cases and rules containing these keywords. These are then loaded into a cache for use in the next step of inference.

[0050] Step S520: Using the historical relevant cases and the relevant rules as the basis for reasoning, the real-time state feature vector is inferred to generate control commands or fault warning signals for the industrial equipment. Specifically, the real-time state feature vector is compared with the retrieved historical relevant cases. For example, the Euclidean distance with each historical case is calculated, and the top K most similar cases are identified. The conclusions of these cases are used to provide data support and confidence for decision-making. The conditional parts (IF parts) of all relevant rules are checked in parallel to see if they are satisfied by the currently active set of nodes. The final decision is determined by a vote or weighted average of the case matching results and the rule execution results.

[0051] For example, the real-time feature vector is matched with the historical case library, and the three most similar cases all point to "spindle imbalance". Simultaneously, the rule engine checks that the condition of the rule "IF feature M appears AND feature N appears THEN output 'spindle imbalance warning'" is met. Both conditions reinforce this conclusion, thus generating a high-confidence "spindle imbalance" fault warning signal. If the inference result indicates that operating parameters need adjustment, for example, if the rule "IF bearing temperature continues to rise THEN increase coolant flow" is triggered, an analog voltage signal is output via a digital-to-analog converter or a digital command is sent via the communication bus to directly control the opening of the coolant pump valve.

[0052] This application embodiment employs a multimodal sensor array of industrial equipment to collect heterogeneous time-series data streams in real time, performs event detection to generate sparse pulse event streams, and deploys a neuromorphic chip integrating a memory-computing architecture, a spiking neural network, and an on-chip network. The sparse pulse event streams are input into the neuromorphic chip, distributed by the on-chip network, and processed in parallel using the dynamic calculation and allocation mechanism of the spiking neural network to obtain real-time state feature vectors. These feature vectors are then fused with a local equipment health knowledge graph for reasoning to generate control commands or fault warning signals. This approach solves the technical problems of slow calculation speed, low efficiency, and poor real-time performance in existing data processing systems, achieving real-time and efficient processing of industrial data, thereby improving the stability and reliability of equipment operation.

[0053] In the above text, refer to Figure 1 A neuromorphic chip hardware acceleration method according to embodiments of the present invention has been described in detail. Next, reference will be made to... Figure 2 A neuromorphic chip hardware acceleration system according to an embodiment of the present invention is described.

[0054] The neuromorphic chip hardware acceleration system according to embodiments of the present invention addresses the technical problems of slow computing speed, low efficiency, and poor real-time performance in existing data processing systems, achieving real-time and efficient processing of industrial data, thereby improving the stability and reliability of equipment operation. The neuromorphic chip hardware acceleration system includes: a heterogeneous time-series data stream acquisition module 10, an event detection module 20, a neuromorphic chip deployment module 30, a neuromorphic computing module 40, and a fusion inference module 50.

[0055] The heterogeneous time-series data stream acquisition module 10 is used to acquire heterogeneous time-series data streams in real time through a multimodal sensor array on an industrial device; the event detection module 20 is used to perform event detection on the heterogeneous time-series data streams and generate sparse pulse event streams; the neuromorphic chip deployment module 30 is used to deploy a neuromorphic chip, which integrates a spiking neural network and an on-chip network with an in-memory computing architecture; the neuromorphic computing module 40 is used to input the sparse pulse event streams into the neuromorphic chip, distribute the pulse events through the on-chip network, and use the dynamic computing allocation mechanism of the spiking neural network to perform asynchronous event-driven parallel processing on the sparse pulse event streams to obtain the real-time state feature vector of the industrial device, wherein the dynamic computing allocation mechanism is configured to activate the corresponding neurons for calculation when a pulse event arrives; the fusion reasoning module 50 is used to perform fusion reasoning with the real-time state feature vectors and the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial device.

[0056] The specific configuration of the event detection module 20 is described in detail below: As mentioned above, the event detection module 20 performs event detection on the heterogeneous time-series data stream to generate a sparse pulse event stream. The event detection module 20 may further include: a first time-series data stream extraction unit for traversing and extracting a first time-series data stream from the heterogeneous time-series data stream; a first data change rate calculation unit for calculating the first data change rate of the first time-series data stream in real time; and a first pulse signal generation unit for determining a valid event when the first data change rate is greater than a preset first noise tolerance threshold and generating a first pulse signal. Multiple first pulse signals constitute the sparse pulse event stream, wherein the first pulse signal includes the sensor ID, intensity, and timestamp of the valid event.

[0057] The detailed description of the specific configuration of the neuromorphic computing module 40 is explained as follows: As mentioned above, the sparse pulse event stream is input to the neuromorphic chip, the pulse events are distributed through the on-chip network, and the dynamic computing allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. The neuromorphic computing module 40 may further include: a routing unit for the on-chip network router integrated inside the neuromorphic chip, through which the sparse pulse event stream is routed to the computing core corresponding to the spiking neural network; and a parallel processing unit for each computing core to perform asynchronous event-driven parallel processing based on the received pulse events by adjusting the membrane potential and synaptic weights of neurons to obtain the real-time state feature vector of the industrial equipment.

[0058] The parallel processing unit may further include: an initial synaptic weight loading subunit for loading pre-trained initial synaptic weights onto each of the computing cores, wherein the initial synaptic weights are obtained by: collecting historical sensor data; training and converging the neural network using the historical sensor data to obtain initial weight values; and loading the initial weight values ​​onto the physical synapses of the neuromorphic chip to obtain initial synaptic weights.

[0059] The parallel processing unit for adjusting synaptic weights may further include: a synaptic weight value adjustment subunit for dynamically adjusting the value of the synaptic weights based on the pulse timing-dependent plasticity rule and according to the firing timing of the preceding and following neurons. Specifically, when the pulse firing time of the first neuron is always earlier than the pulse firing time of the second neuron within a preset time period, the first synaptic weight connecting the first neuron and the second neuron is enhanced; conversely, the first synaptic weight is weakened. Here, the first neuron is a neuron in the previous layer, and the second neuron is a neuron in the next layer.

[0060] The detailed description of the specific configuration of the fusion reasoning module 50 is explained as follows: As mentioned above, the real-time state feature vector is fused and reasoned with the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment. The fusion reasoning module 50 may further include: a correlation weight calculation unit for calculating the correlation weight between the real-time state feature vector and the entity nodes in the device health knowledge graph, and retrieving historical related cases and related rules according to a preset correlation weight threshold. The device health knowledge graph stores the component association relationships, historical fault modes, and expert operation and maintenance procedures of the industrial equipment in a graph structure. The reasoning unit is used to use the historical related cases and the related rules as the reasoning basis to reason about the real-time state feature vector and generate control commands or fault warning signals for the industrial equipment.

[0061] The specific configuration of the heterogeneous time-series data stream acquisition module 10 is described in detail below: As mentioned above, the heterogeneous time-series data stream is acquired in real time through a multimodal sensor array on the industrial equipment. The heterogeneous time-series data stream acquisition module 10 may further include: a broadband acoustic signature signal acquisition unit for acquiring broadband acoustic signature signals generated by the operation of the industrial equipment through a contact piezoelectric microphone to obtain an acoustic signature time-series data stream; a vibration acceleration signal acquisition unit for acquiring vibration acceleration signals of the mechanical structure of the equipment in three orthogonal axes through a MEMS vibration sensor to obtain a vibration time-series data stream; a dynamic visual information acquisition unit for acquiring dynamic visual information of the operating status of the equipment through an event camera to obtain a video time-series data stream; and a tension signal acquisition unit for acquiring tension signals of the transmission chain through a Wheatstone bridge composed of resistance strain gauges to obtain a tension time-series data stream. The acoustic signature time-series data stream, the vibration time-series data stream, the video time-series data stream, and the tension time-series data stream constitute the heterogeneous time-series data stream.

[0062] The neuromorphic chip hardware acceleration system provided in this embodiment of the invention can execute the neuromorphic chip hardware acceleration method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0063] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A neuromorphic chip hardware acceleration method, characterized in that, The method includes: Heterogeneous time-series data streams are acquired in real time using multimodal sensor arrays on industrial equipment; Event detection is performed on the heterogeneous time-series data stream to generate a sparse pulse event stream; Deploy a neuromorphic chip, which integrates a spiking neural network and an on-chip network with an in-memory computing architecture; The sparse pulse event stream is input into the neuromorphic chip, the pulse events are distributed through the on-chip network, and the dynamic calculation and allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. The real-time state feature vector is fused and reasoned with the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment. The dynamic computation allocation mechanism is configured to activate the corresponding neuron for computation when a pulse event arrives.

2. The neuromorphic chip hardware acceleration method as described in claim 1, characterized in that, Event detection is performed on the heterogeneous time-series data stream to generate a sparse pulse event stream, including: Extract the first time-series data stream from the heterogeneous time-series data stream; Calculate the first data change rate of the first time-series data stream in real time; When the rate of change of the first data is greater than the preset first noise tolerance threshold, it is determined to be a valid event, and a first pulse signal is generated. Multiple first pulse signals constitute the sparse pulse event stream. The first pulse signal contains the sensor ID, intensity, and timestamp of the valid event.

3. The neuromorphic chip hardware acceleration method as described in claim 1, characterized in that, The sparse pulse event stream is input into the neuromorphic chip, the pulse events are distributed through the on-chip network, and the dynamic computation and allocation mechanism of the spiking neural network is used to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment, including: The neuromorphic chip integrates an on-chip network router. The sparse pulse event stream is routed to the computing core corresponding to the spiking neural network through the on-chip network router; Each of the aforementioned computing cores performs asynchronous event-driven parallel processing by adjusting the membrane potential and synaptic weights of neurons based on the received pulse events, thereby obtaining the real-time state feature vector of the industrial equipment.

4. The neuromorphic chip hardware acceleration method as described in claim 3, characterized in that, Each of the aforementioned computing cores is loaded with pre-trained initial synaptic weights, which are obtained in the following manner: Collect historical sensor data; The neural network is trained and converged using the historical sensor data to obtain initial weight values; The initial weight values ​​are loaded onto the physical synapses of the neuromorphic chip to obtain the initial synaptic weights.

5. The neuromorphic chip hardware acceleration method as described in claim 3, characterized in that, Adjusting synaptic weights includes: Based on the pulse timing-dependent plasticity rule, the values ​​of the synaptic weights are dynamically adjusted according to the firing sequence of related neurons. Specifically, when the pulse firing time of the first neuron is always earlier than the pulse firing time of the second neuron within a preset time period, the weight of the first synapse connecting the first neuron and the second neuron is enhanced; conversely, the weight of the first synapse is weakened. In this context, the first neuron is the neuron in the previous layer, and the second neuron is the neuron in the next layer.

6. The neuromorphic chip hardware acceleration method as described in claim 1, characterized in that, The real-time state feature vector is fused and reasoned with the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment, including: Calculate the correlation weight between the real-time status feature vector and the entity nodes in the device health knowledge graph, and retrieve historical related cases and rules based on the preset correlation weight threshold; The historical relevant cases and the relevant rules are used together as the reasoning basis to reason about the real-time state feature vector and generate control commands or fault warning signals for the industrial equipment. The equipment health knowledge graph stores the component relationships, historical failure modes, and expert operation and maintenance procedures of industrial equipment in a graph structure.

7. The neuromorphic chip hardware acceleration method as described in claim 1, characterized in that, Through multimodal sensor arrays on industrial equipment, heterogeneous time-series data streams are acquired in real time, including: Broadband acoustic signature signals generated by industrial equipment operation are collected using a contact piezoelectric microphone to obtain an acoustic signature time-series data stream. Vibration acceleration signals of the equipment's mechanical structure along three orthogonal axes are collected using MEMS vibration sensors to obtain a vibration time-series data stream. Dynamic visual information about the device's operating status is acquired by using an event camera to obtain a video time-series data stream; Tension signals from the transmission chain are acquired by using a Wheatstone bridge composed of resistance strain gauges to obtain a tension time-series data stream. The heterogeneous timing data stream comprises the voiceprint timing data stream, the vibration timing data stream, the video timing data stream, and the tension timing data stream.

8. A neuromorphic chip hardware acceleration system, characterized in that, The system is used to implement the neuromorphic chip hardware acceleration method according to any one of claims 1-7, and the system comprises: The heterogeneous time-series data stream acquisition module is used to acquire heterogeneous time-series data streams in real time through multimodal sensor arrays on industrial equipment; The event detection module is used to perform event detection on the heterogeneous time-series data stream and generate a sparse pulse event stream; A neuromorphic chip deployment module is used to deploy a neuromorphic chip, which integrates a spiking neural network and an on-chip network with an in-memory computing architecture. A neuromorphic computing module is used to input the sparse pulse event stream into the neuromorphic chip, distribute the pulse events through the on-chip network, and use the dynamic computing allocation mechanism of the spiking neural network to perform asynchronous event-driven parallel processing on the sparse pulse event stream to obtain the real-time state feature vector of the industrial equipment. The dynamic computing allocation mechanism is configured to activate the corresponding neuron for calculation when a pulse event arrives. The fusion reasoning module is used to perform fusion reasoning with the real-time state feature vector and the device health knowledge graph pre-stored locally on the neuromorphic chip to generate control commands or fault warning signals for the industrial equipment.

Images