A Smart Linkage Method and System for Energy Storage Firefighting Based on Hybrid Control

Abstract

This invention provides a method and system for intelligent linkage of energy storage and fire protection based on hybrid control. The method includes: acquiring and standardizing input multi-source data signals using a preset sampling frequency; converting the processed data signals into pulse sequences; constructing a three-layer network using an improved LIF neuron model; extracting features and generating decisions from the input pulse sequences using the three-layer network; decoding the pulse sequences output by the three-layer network; mapping the decoding results to specific linkage control commands; and executing linkage control operations according to the linkage control commands. The advantages of this invention are: it can effectively achieve multi-source data fusion and adaptive control, meeting the needs of various complex and changing application scenarios; it can ensure the real-time performance and reliability of the entire system; and it supports dynamic expansion and decision optimization, thus effectively overcoming many shortcomings of traditional predefined rule schemes.

Description

Technical Field

[0001] This invention relates to the field of fire safety technology, and in particular to a method and system for intelligent linkage of energy storage fire protection based on hybrid control. Background Technology

[0002] With the widespread application of energy storage systems in the power sector, their fire safety issues are becoming increasingly prominent. Although artificial intelligence technology has made significant progress in other fields, its application in energy storage fire control is still in its early stages. Currently, mainstream energy storage fire control systems still adopt traditional predefined rule schemes, which have the following shortcomings: 1. The control logic relies entirely on manually preset fixed rules, making it unable to cope with complex and ever-changing fire scenarios; 2. The linkage relationships between various sensors and control devices require manual configuration, resulting in poor scalability; 3. The system lacks self-learning capabilities and cannot optimize decision-making strategies from historical data; 4. Its ability to process multi-source heterogeneous data is limited, making it difficult to achieve deep information fusion. Therefore, there is an urgent need to provide an intelligent linkage method and system for energy storage fire protection that combines traditional logic control and pulse neural networks to address the aforementioned shortcomings of traditional predefined rule schemes. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method and system for intelligent linkage of energy storage fire protection based on hybrid control. By adopting a pulse neural network combined with an adaptive control strategy, it realizes intelligent mapping of multi-source input signals to linkage control commands, breaks through the limitations of traditional predefined rules, and can achieve adaptive control well while ensuring real-time performance and reliability.

[0004] This invention is implemented as follows:

[0005] Firstly, a smart linkage method for energy storage fire protection based on hybrid control, the method comprising:

[0006] Step S1: Acquire and standardize the input multi-source data signal using a preset sampling frequency, and convert the processed data signal into a pulse sequence;

[0007] Step S2: Construct a three-layer network using an improved LIF neuron model, and use the three-layer network to extract features and generate decisions from the input pulse sequence;

[0008] Step S3: Decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control commands, and execute linkage control operations according to the linkage control commands.

[0009] Furthermore, step S1 specifically includes:

[0010] Multi-source data input: 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 zone controllers, and analog signals from 200 smart detectors. Each signal is equipped with an independent opto-isolation circuit and EMC protection circuit.

[0011] Data standardization processing: For digital signals, a Schmitt trigger is used for signal shaping, and the shaped standard signal is stored in a 32-bit status word; for analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning, with a sampling frequency of 10kHz, and all analog signals are uniformly normalized to the [0, 1] interval; in terms of timing processing, a uniform 100MHz main frequency clock source is used, and all input signals are synchronously sampled at a period of 50ms.

[0012] Pulse sequence coding: Frequency coding adopts an adaptive probability distribution method to map the signal strength to a pulse frequency range of 1Hz~200Hz; Timing coding adopts a sliding mechanism with a 100ms time window to reflect the order of events through the phase difference of the pulses; Data buffering adopts a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, and is divided into real-time data area and historical data area.

[0013] Furthermore, the multi-source data input also includes: a preset parameter matrix, which adopts a hierarchical classification storage architecture, including a system global configuration layer, a device type configuration layer, an alarm threshold layer, and a linkage rule layer, and the parameter matrix supports online modification and power failure saving;

[0014] The data standardization process also includes a three-level anomaly handling mechanism: a hardware-level protection mechanism, a signal-level detection mechanism, and a data-level processing mechanism. Specifically, the hardware-level protection mechanism uses opto-isolators and EMC protection circuits to achieve overvoltage and overcurrent protection. The signal-level detection mechanism uses real-time monitoring of current and voltage thresholds to detect open circuits and short circuits. The data-level processing mechanism uses wavelet transform and digital filtering to eliminate signal interference.

[0015] Furthermore, step S2 specifically includes:

[0016] Input layer: Configured with at least 512 neurons to receive input signals, with each input signal mapped to 8 neurons by default to form redundant encoding; the neurons employ an improved LIF neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. Indicates the input current;

[0017] Hidden Layer: The network employs a hierarchical functional subnetwork, containing 64 feature extraction neurons, 128 pattern recognition neurons, and 32 decision neurons. The feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, the pattern recognition neurons are responsible for identifying typical alarm patterns, and the decision neurons are responsible for information integration and preliminary decision-making. The network connections utilize diverse mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections ranging from ±3 neurons, and high-level-to-low-level feedback connections with an intensity coefficient of 0.2. Synaptic plasticity is also achieved based on the STDP mechanism.

[0018] Output layer: Configured with multiple types of output neurons, including main control neuron group, auxiliary control neuron group and state feedback neuron group. The main control neuron group is used to generate key control instructions, the auxiliary control neuron group is used to generate secondary control instructions, and the state feedback neuron group is used to monitor the execution results and provide feedback adjustment. At the same time, multiple guarantee mechanisms are adopted in the decision-making process, including 2 / 3 voting mechanism, adaptive threshold control method and pulse sequence analysis with 100ms time window.

[0019] Furthermore, the input layer also includes: before the input signal enters the neuron, the input signal is processed by a presynaptic processing circuit, including weight modulation and delay control, with a weight range of [0,1] and a delay range of 0.5ms; at the same time, a dynamic threshold adjustment mechanism is used for the neuron.

[0020] Furthermore, step S3 specifically includes:

[0021] Pulse decoding: For pulse frequency decoding, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window, and the instantaneous frequency is calculated by an exponential moving average algorithm, with an adaptive lag interval of ±5Hz set; for pattern recognition, a multi-dimensional pulse time series analysis method is used for analysis and recognition; for decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in time and space dimensions within a 100ms sliding window.

[0022] Control mapping: It adopts a multi-channel output structure, including 8 relay outputs, 40 partition controllers and 200 smart detectors for CAN bus communication; the decoding results are mapped into specific linkage control commands through a scheduling strategy based on four-level priority. At the same time, it supports four mapping modes, including direct mapping of a single input to a single output, complex logic mapping of multiple input conditions, sequential mapping based on a predetermined timing, and linkage mapping of multiple outputs in coordination.

[0023] Execution control: When executing linkage control operations according to linkage control commands, a bistable magnetic latching relay and multiple protection mechanisms are adopted to achieve timing control with an accuracy of 0.1ms.

[0024] Secondly, a hybrid control-based intelligent linkage system for energy storage fire protection includes an input data processing and encoding module, a pulse neural network module, and an output decoding and control module.

[0025] The input data processing and encoding module is used to collect and standardize the input multi-source data signals using a preset sampling frequency, and convert the processed data signals into pulse sequences.

[0026] The spiking neural network module is used to construct a three-layer network using an improved LIF neuron model, and to extract features and generate decisions from the input spiking sequence using the three-layer network.

[0027] The output decoding and control module is used to decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control instructions, and execute linkage control operations according to the linkage control instructions.

[0028] Furthermore, the input data processing and encoding module specifically includes a multi-source input interface, a standardization processing unit, and a pulse sequence encoder;

[0029] The multi-source input interface is used for multi-source data input: 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 partition controllers, and analog signals from 200 smart detectors. Each signal is equipped with an independent opto-isolation circuit and EMC protection circuit.

[0030] The standardization processing unit is used for data standardization processing: for digital signals, a Schmitt trigger is used for signal shaping, and the shaped standard signal is stored in a 32-bit status word; for analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning, with a sampling frequency of 10kHz, and all analog signals are uniformly normalized to the [0, 1] interval; in terms of timing processing, a uniform 100MHz main frequency clock source is used, and all input signals are synchronously sampled at a period of 50ms.

[0031] The pulse sequence encoder is used for pulse sequence encoding: frequency encoding adopts an adaptive probability distribution method to map the signal strength to a pulse frequency range of 1Hz~200Hz; timing encoding adopts a sliding mechanism with a 100ms time window to reflect the order of events through the phase difference of the pulses; data buffering adopts a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, and is divided into a real-time data area and a historical data area.

[0032] Furthermore, the spiking neural network module is specifically used for:

[0033] Input layer: Configured with at least 512 neurons to receive input signals, with each input signal mapped to 8 neurons by default to form redundant encoding; the neurons employ an improved LIF neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. Indicates the input current;

[0034] Hidden Layer: The network employs a hierarchical functional subnetwork, containing 64 feature extraction neurons, 128 pattern recognition neurons, and 32 decision neurons. The feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, the pattern recognition neurons are responsible for identifying typical alarm patterns, and the decision neurons are responsible for information integration and preliminary decision-making. The network connections utilize diverse mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections ranging from ±3 neurons, and high-level-to-low-level feedback connections with an intensity coefficient of 0.2. Synaptic plasticity is also achieved based on the STDP mechanism.

[0035] Output layer: Configured with multiple types of output neurons, including main control neuron group, auxiliary control neuron group and state feedback neuron group. The main control neuron group is used to generate key control instructions, the auxiliary control neuron group is used to generate secondary control instructions, and the state feedback neuron group is used to monitor the execution results and provide feedback adjustment. At the same time, multiple guarantee mechanisms are adopted in the decision-making process, including 2 / 3 voting mechanism, adaptive threshold control method and pulse sequence analysis with 100ms time window.

[0036] Furthermore, the output decoding and control module specifically includes a pulse decoder, a control mapping unit, and an execution control unit;

[0037] The pulse decoder is used for pulse decoding: for pulse frequency decoding, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window, the instantaneous frequency is calculated by an exponential moving average algorithm, and an adaptive lag interval of ±5Hz is set; for pattern recognition, a multi-dimensional pulse time series analysis method is used for analysis and recognition; for decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in time and space dimensions within a 100ms sliding window.

[0038] The control mapping unit is used for control mapping: it adopts a multi-channel output structure, including 8 relay outputs, 40 partition controllers and 200 smart detectors for CAN bus communication; it maps the decoding results into specific linkage control commands through a scheduling strategy based on four-level priority, and supports four mapping modes during mapping, including direct mapping of a single input to a single output, complex logic mapping of multiple input conditions, sequential mapping based on a predetermined timing, and linkage mapping of multiple outputs in coordination.

[0039] The execution control unit is used to perform control: when performing linkage control operations according to linkage control instructions, it adopts a bistable magnetic latching relay and multiple protection mechanisms to achieve timing control with an accuracy of 0.1ms.

[0040] By adopting the technical solution of the present invention, at least the following beneficial effects are achieved: by using a spiking neural network combined with an adaptive control strategy, intelligent mapping of multi-source input signals to linkage control commands is realized, and linkage control operations are executed according to the mapped linkage control commands. This approach breaks through the limitations of traditional predefined rules, can effectively realize multi-source data fusion and adaptive control, meet the usage requirements of various complex and ever-changing application scenarios, and at the same time ensure the real-time performance and reliability of the entire system. Furthermore, it supports dynamic expansion and decision optimization, thereby effectively overcoming many defects of traditional predefined rule schemes. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a flowchart illustrating the execution process of an intelligent linkage method for energy storage fire protection based on hybrid control, according to the present invention.

[0043] Figure 2 This is a block diagram illustrating the principle of input data processing and encoding in this invention;

[0044] Figure 3 This is a topological diagram of the spiking neural network in this invention;

[0045] Figure 4 This is the output decoding and control mapping diagram in this invention;

[0046] Figure 5 This is a schematic diagram of the principle structure of an intelligent linkage system for energy storage fire protection based on hybrid control according to the present invention. Detailed Implementation

[0047] To better understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] It should be noted that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used solely for the convenience of describing these embodiments and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. Example 1

[0049] Please see Figures 1 to 4 As shown, the present invention provides an intelligent linkage method for energy storage fire protection based on hybrid control, the method comprising:

[0050] Step S1: Acquire and standardize the input multi-source data signal using a preset sampling frequency, and convert the processed data signal into a pulse sequence;

[0051] Step S2: Construct a three-layer network using an improved LIF neuron model, and use the three-layer network to extract features and generate decisions from the input pulse sequence;

[0052] Step S3: Decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control commands, and execute linkage control operations according to the linkage control commands.

[0053] This invention employs a spiking neural network combined with an adaptive control strategy to achieve intelligent mapping of multi-source input signals to linkage control commands, and executes linkage control operations according to the mapped linkage control commands. This approach breaks through the limitations of traditional predefined rules, effectively realizing multi-source data fusion and adaptive control, meeting the needs of various complex and ever-changing application scenarios, while ensuring the real-time performance and reliability of the entire system, and supporting dynamic expansion and decision optimization. Thus, it effectively overcomes many shortcomings of traditional predefined rule schemes.

[0054] In some embodiments of the present invention, such as Figure 2 As shown, step S1 specifically includes:

[0055] (1) Multi-source data input: This invention supports the access of multiple types of input signals: 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 partition controllers, and analog signals from 200 smart detectors. Each signal is equipped with an independent opto-isolation circuit and an EMC protection circuit. The opto-isolation technology achieves electrical isolation of the signal, which can effectively prevent signal crosstalk and electromagnetic interference, and significantly improve the anti-interference capability and signal transmission reliability of the system. The opto-isolation circuit can specifically use high-speed optical coupling devices, which have the characteristics of fast response and low loss. The EMC protection circuit includes multi-level lightning protection, surge suppression and signal filtering design, which can ensure stable signal transmission and reliable operation of the system in complex electromagnetic environments.

[0056] In practical implementation, this invention features multiple interfaces for different data sources: The in-warehouse host is equipped with 8 digital input interfaces (DI1-8), capable of collecting signals from smoke alarms, temperature alarms, gas alarms, and manual alarms. The sampling frequency can be configured within the range of 1Hz to 100Hz. The partition controller interface supports parallel access for up to 40 controllers, using CAN bus communication. It also supports hot-swapping and automatic address allocation, enabling plug-and-play intelligent access. Hot-swapping technology allows for dynamic addition and removal of devices during system operation without downtime or restart, significantly improving system flexibility and scalability. The automatic address allocation mechanism uses a protocol-based intelligent addressing algorithm, automatically and uniquely obtaining a communication address upon device access, thus avoiding address conflicts, simplifying system integration and device management processes, and significantly reducing the complexity of system deployment and maintenance. Each controller can collect 3 signals, including temperature, smoke concentration, and gas concentration. The detector interface supports up to 200 intelligent detectors connected via a CAN bus. Each intelligent detector integrates multiple sensors, including those for temperature, gas, and smoke, and uses a unified CAN communication protocol for data transmission. The sensor fault detection function employs multi-dimensional and intelligent fault diagnosis technology, specifically including: establishing multiple fault identification mechanisms such as data mutation detection, long-term drift analysis, and cross-validation by comparing the consistency, rationality, and stability of sensor output data in real time; when an anomaly is detected in a sensor, it can quickly locate the faulty sensor, automatically switch to a backup sensor or degrade the operating mode, and record the anomaly information through detailed fault logs to ensure that the system can still maintain reliable operation when some sensors fail; at the same time, the system adopts a multi-level linkage fault diagnosis mechanism, which establishes a collaborative detection model among sensor groups and combines it with a database of equipment operating status characteristics to achieve accurate fault location and graded handling, significantly improving system reliability and maintenance efficiency.

[0057] (2) Data standardization processing: For digital signals, Schmitt triggers are used to shape the signals to eliminate jitter. In the specific implementation of this invention, the input level standard can be set to high level > 20V and low level < 5V. The signal de-jittering time can be configured in the range of 10ms to 1000ms according to actual needs. The shaped standard signal is stored in a 32-bit status word.

[0058] For analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning. The sampling frequency is 10kHz, and all analog signals are normalized to the [0, 1] interval. In specific implementation, the signal conditioning process of this invention includes zero-point calibration, range calibration, linearization processing and digital filtering.

[0059] In terms of timing processing, a unified 100MHz main frequency clock source is used to synchronously sample all input signals at a period of 50ms. At the same time, a double buffering mechanism can be used to ensure the real-time performance and integrity of the data, and the timing alignment accuracy is better than 1ms.

[0060] (3) Pulse sequence encoding can be implemented by using a variety of encoding strategies to convert signals into pulse sequences: frequency encoding adopts an adaptive probability distribution method to map the signal strength to the pulse frequency range of 1Hz~200Hz, so that the signal strength and the pulse firing frequency are linearly related; timing encoding adopts a sliding mechanism of 100ms time window, and the phase difference of the pulse reflects the order of events, with a phase resolution of 0.1ms. For multiple events that occur at the same time, the encoding order can be determined according to the preset priority; data buffer adopts a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, and is divided into real-time data area and historical data area, which can support fast access to real-time data and asynchronous processing of historical data. When the buffer is full, it can automatically overwrite the oldest data.

[0061] In some embodiments of the present invention, the multi-source data input further includes: a preset parameter matrix, which adopts a hierarchical classification storage architecture, including a system global configuration layer, a device type configuration layer, an alarm threshold layer, and a linkage rule layer. Each layer has strict parameter definitions and access control. The system global configuration layer stores system configuration parameters, the device type configuration layer stores device type parameters, the alarm threshold layer stores alarm threshold parameters, and the linkage rule layer stores linkage rule parameters. Simultaneously, the parameter matrix supports online modification and power-off saving. Modifications take effect in real time and have a complete change traceability mechanism. The power-off saving function uses non-volatile storage technology to ensure that parameters are not lost in the event of a sudden power outage and can accurately restore the system configuration state. Furthermore, the parameter matrix also has a built-in security verification mechanism to check the legality of parameter modifications, preventing system anomalies due to incorrect configuration, and providing strong protection for the system's flexibility and reliability.

[0062] The data standardization process also includes a three-level anomaly handling mechanism: a hardware-level protection mechanism, a signal-level detection mechanism, and a data-level processing mechanism. Specifically, the hardware-level protection mechanism uses opto-isolators and EMC protection circuits to achieve overvoltage and overcurrent protection. The signal-level detection mechanism uses real-time monitoring of current and voltage thresholds to detect open circuits and short circuits. The data-level processing mechanism uses wavelet transform and digital filtering to eliminate signal interference.

[0063] In this invention, to achieve end-to-end mapping from pulse signals to control decisions, the spiking neural network adopts a three-layer architecture design including an input layer, a hidden layer, and an output layer, such as... Figure 3 As shown, step S2 specifically includes:

[0064] The input layer employs a dynamically configured neuron array structure: at least 512 neurons are configured to receive input signals, and each input signal is mapped to 8 neurons by default to form redundant encoding, thereby improving the system's fault tolerance. The neurons utilize a modified LIF (Leaky Integrate and Fire) neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. This represents the input current. In specific implementations of this invention, the specific number of neurons can be expanded according to the actual number of input channels connected to meet actual usage requirements.

[0065] Hidden Layer: The network adopts a hierarchical functional sub-network, containing 64 feature extraction neurons, 128 pattern recognition neurons, and 32 decision neurons. Among them, the feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, the pattern recognition neurons are responsible for identifying typical alarm patterns, and the decision neurons are responsible for completing information integration and preliminary decision-making. The network connection adopts a variety of mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections with a range of ±3 neurons, and high-level to low-level feedback connections with an intensity coefficient of 0.2. At the same time, synaptic plasticity is realized based on the STDP (SpikeTimingDependent Plasticity) mechanism. The initial learning rate is set to 0.01 and dynamically adjusted with the training progress. The stability of the network is ensured by setting upper and lower limits of weights [-1, 1].

[0066] Output Layer: This layer is configured with multiple types of output neurons, including a main control neuron group, an auxiliary control neuron group, and a state feedback neuron group. The main control neuron group generates critical control commands, the auxiliary control neuron group generates secondary control commands, and the state feedback neuron group monitors execution results and provides feedback adjustment. Multiple safeguard mechanisms are employed during the decision-making process, including a 2 / 3 voting mechanism, an adaptive threshold control method, and pulse sequence analysis within a 100ms time window. In practical implementation, the main control neuron group is primarily responsible for generating critical control commands such as fire extinguishing, smoke extraction, and power outage, and employs a 2 / 3 voting mechanism to ensure decision reliability. The auxiliary control neuron group is mainly responsible for processing secondary control commands such as alarm prompts and ventilation adjustments, and has a dynamic priority adjustment function. The state feedback neuron group is primarily responsible for real-time monitoring of execution results and providing feedback adjustment within a 100ms time window, ensuring control accuracy better than 0.1ms. The 2 / 3 voting mechanism specifically means that the signal can only pass if more than half of the neurons output the same signal; the adaptive threshold control method means that the threshold can be adjusted in real time according to the input signal, with a baseline value of 50mV; the pulse sequence analysis with a 100ms time window is to extract the spatiotemporal features of the pulse sequence for analysis, so as to improve the accuracy and reliability of decision-making.

[0067] In addition, in step S2 of the present invention, the reliability of the system can also be achieved through a multi-level and multi-dimensional guarantee mechanism, which mainly includes: hardware redundancy design, dynamic data consistency verification and security protection mechanism; specifically, the hardware redundancy design adopts a "master-slave" architecture, and each key module is equipped with at least one backup unit; the data consistency check uses multiple verification algorithms to promptly detect and isolate abnormal data; the security protection mechanism establishes a full-link abnormal handling response system to realize real-time monitoring and intelligent scheduling of the system's operating status, ensuring that the system can still maintain stable operation under abnormal conditions.

[0068] In some embodiments of the present invention, the input layer further includes: processing the input signal using a presynaptic processing circuit before the input signal enters the neuron, including weight modulation and delay control, with a weight range of [0,1], a delay range of 0.5ms, and a time resolution of 0.1ms; simultaneously, to prevent overactivation of the neuron, a dynamic threshold adjustment mechanism is also adopted for the neuron, specifically including: adopting a real-time threshold update strategy based on a sliding window with a window length of 100ms; adaptively adjusting the activation threshold of the neuron according to the intensity of the input signal, with an adjustment range of ±10mV; calculating the threshold offset using an exponential weighted average algorithm with a time constant of 20ms; and rapidly increasing the threshold when a burst signal is detected to prevent overactivation, with a baseline threshold set at 55mV, which can be adaptively adjusted according to the input frequency.

[0069] In some embodiments of the present invention, in order to realize the conversion from the output of the spiking neural network to actual control commands, such as... Figure 4 As shown, step S3 specifically includes:

[0070] Pulse Decoding: For pulse frequency decoding, advanced temporal coding conversion technology is employed. Specifically, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window, ensuring accurate pulse counting. The instantaneous frequency is calculated using an exponential moving average algorithm, guaranteeing a frequency estimate update every 1ms. An adaptive hysteresis range of ±5Hz is set to effectively filter high-frequency noise interference and suppress signal jitter.

[0071] For pattern recognition, a multi-dimensional pulse timing analysis method is employed to accurately identify various complex pulse patterns, such as burst and gradual patterns. By deeply analyzing the coordinated firing characteristics of neuronal populations, the system achieves precise capture of the network's dynamic state. Specifically, the system can track minute changes in the network state in real time, including the precise location of abrupt changes and gradual trends. For decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in both time and space dimensions within a 100ms sliding window. The time dimension captures dynamic evolution characteristics through the sliding window, while the spatial dimension performs correlation analysis on the outputs of relevant neurons. Finally, a reliability-based weighted decision mechanism comprehensively considers the confidence levels of each neuron to generate more stable and accurate linkage control commands.

[0072] Control Mapping: A multi-channel output structure is adopted to improve output flexibility. Specifically, it includes 8 relay outputs (DC24V / 5A) for critical control, 40-channel zone controller, and 200-channel intelligent detector CAN bus communication. In practical implementation, a dual CAN bus redundancy design can be used to ensure communication reliability. A four-level priority scheduling strategy maps the decoding results to specific linkage control commands. The four priority levels are emergency, high, medium, and low. This invention employs a four-level priority dynamic scheduling management mechanism, which has two key features: first, high-priority tasks can interrupt low-priority operations in real time, ensuring immediate response of critical tasks; second, through a priority inversion protection mechanism, it effectively prevents deadlock risks that may occur in traditional real-time systems, significantly improving the system's real-time performance and reliability. Meanwhile, four mapping modes are supported during mapping. The mapping strategy of this invention adopts multimodal intelligent mapping technology: (1) direct mapping of a single input to a single output to ensure one-to-one correspondence of signal transmission; (2) complex logic mapping of multiple input condition combinations to handle complex condition associations; (3) sequential mapping based on predetermined timing to achieve precise control of time-dependent tasks; and multi-output collaborative linkage mapping to support system-level linkage control.

[0073] Execution Control: When executing linkage control operations according to linkage control commands, a bistable magnetic latching relay and multiple protection mechanisms are employed to achieve timing control with an accuracy of 0.1ms. In specific implementations, the drive control uses low-power bistable magnetic latching relay technology, which supports the CAN communication protocol and achieves ultra-high precision timing control of 0.1ms at the sub-millisecond level. This not only significantly reduces system energy consumption but also possesses unique power-off state retention and automatic power-on recovery functions, ensuring stable operation of the system under complex working conditions. The multiple protection mechanisms include a safety mechanism, a monitoring feedback mechanism, and data management. Among them, the safety mechanism adopts a multi-layered, three-dimensional protection architecture. By constructing a system-level linkage protection system, it achieves proactive risk identification and intelligent isolation, significantly improving the reliability and security of the system. The monitoring feedback mechanism adopts an intelligent design, establishing a full-link state perception and predictive early warning model to achieve precise optimization of the control process. In terms of data management, it provides real-time recording of the control process, detailed logs of key events, statistical analysis of control effects, and provides strong support for system maintenance and continuous optimization through a secure remote data access interface. Example 2

[0074] Please see Figures 2 to 5 As shown, the present invention discloses an intelligent linkage system for energy storage fire protection based on hybrid control. The system includes an input data processing and encoding module, a pulse neural network module, and an output decoding and control module.

[0075] The input data processing and encoding module is used to collect and standardize the input multi-source data signals using a preset sampling frequency, and convert the processed data signals into pulse sequences.

[0076] The spiking neural network module is used to construct a three-layer network using an improved LIF neuron model, and to extract features and generate decisions from the input spiking sequence using the three-layer network.

[0077] The output decoding and control module is used to decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control instructions, and execute linkage control operations according to the linkage control instructions.

[0078] This invention employs a spiking neural network combined with an adaptive control strategy to achieve intelligent mapping of multi-source input signals to linkage control commands, and executes linkage control operations according to the mapped linkage control commands. This approach breaks through the limitations of traditional predefined rules, can effectively achieve adaptive control, meet the needs of various complex and ever-changing application scenarios, ensure the real-time performance and reliability of the entire system, and support dynamic expansion and optimization, thus effectively overcoming many shortcomings of traditional predefined rule schemes.

[0079] In some embodiments of the present invention, such as Figure 2 As shown, the input data processing and encoding module specifically includes a multi-source input interface, a standardization processing unit, and a pulse sequence encoder;

[0080] The multi-source input interface is used for multi-source data input, meaning that this invention supports the access of multiple types of input signals: 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 zone controllers, and analog signals from 200 intelligent detectors. Each signal is equipped with an independent opto-isolation circuit and EMC protection circuit. The opto-isolation technology achieves electrical isolation of the signals, effectively preventing signal crosstalk and electromagnetic interference, significantly improving the system's anti-interference capability and signal transmission reliability. The opto-isolation circuit can specifically use high-speed optical coupling devices, which have the characteristics of fast response and low loss. The EMC protection circuit includes multi-level lightning protection, surge suppression, and signal filtering designs, which can ensure stable signal transmission and reliable system operation in complex electromagnetic environments.

[0081] In practical implementation, this invention features multiple interfaces for different data sources: The in-warehouse host is equipped with 8 digital input interfaces (DI1-8), capable of collecting signals from smoke alarms, temperature alarms, gas alarms, and manual alarms. The sampling frequency can be configured within the range of 1Hz to 100Hz. The partition controller interface supports parallel access for up to 40 controllers, using CAN bus communication. It also supports hot-swapping and automatic address allocation, enabling plug-and-play intelligent access. Hot-swapping technology allows for dynamic addition and removal of devices during system operation without downtime or restart, significantly improving system flexibility and scalability. The automatic address allocation mechanism uses a protocol-based intelligent addressing algorithm, automatically and uniquely obtaining a communication address upon device access, thus avoiding address conflicts, simplifying system integration and device management processes, and significantly reducing the complexity of system deployment and maintenance. Each controller can collect 3 signals, including temperature, smoke concentration, and gas concentration. The detector interface supports up to 200 smart detectors connected via a CAN bus. Each smart detector integrates multiple sensors, including those for temperature, gas, and smoke, and uses a unified CAN communication protocol for data transmission. The sensor fault detection function employs multi-dimensional and intelligent fault diagnosis technology, specifically including: establishing multiple fault identification mechanisms such as data mutation detection, long-term drift analysis, and cross-validation by comparing the consistency, rationality, and stability of sensor output data in real time; when an anomaly is detected in a sensor, it can quickly locate the faulty sensor, automatically switch to a backup sensor or degrade the operating mode, and record the abnormal information through detailed fault logs to ensure that the system can still maintain reliable operation when some sensors fail; specific fault detection strategies include data threshold exceeding judgment, statistical variance analysis, multi-sensor cross-validation, and machine learning-based anomaly pattern recognition.

[0082] The standardization processing unit is used for data standardization processing: for digital signals, a Schmitt trigger is used to shape the signal to eliminate jitter. In specific implementation of this invention, the input level standard can be set to high level > 20V and low level < 5V. The signal de-jittering time can be configured within the range of 10ms to 1000ms according to actual needs. The shaped standard signal is stored using a 32-bit status word.

[0083] For analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning. The sampling frequency is 10kHz, and all analog signals are normalized to the [0, 1] interval. In specific implementation, the signal conditioning process of this invention includes zero-point calibration, range calibration, linearization processing and digital filtering.

[0084] In terms of timing processing, a unified 100MHz main frequency clock source is used to synchronously sample all input signals at a period of 50ms. At the same time, a double buffering mechanism can be used to ensure the real-time performance and integrity of the data, and the timing alignment accuracy is better than 1ms.

[0085] The pulse sequence encoder is used for pulse sequence encoding and can employ various encoding strategies to convert signals into pulse sequences: frequency encoding uses an adaptive probability distribution method to map signal strength to a pulse frequency range of 1Hz~200Hz, making the signal strength linearly related to the pulse firing frequency; timing encoding uses a sliding mechanism with a 100ms time window, reflecting the order of events through the phase difference of the pulses, with a phase resolution of 0.1ms, and for multiple events occurring simultaneously, the encoding order can be determined according to a preset priority; the data buffer uses a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, divided into a real-time data area and a historical data area, supporting fast access to real-time data and asynchronous processing of historical data, and automatically overwriting the oldest data when the buffer is full.

[0086] In some embodiments of the present invention, the multi-source data input further includes: a preset parameter matrix, which adopts a hierarchical classification storage architecture, including a system global configuration layer, a device type configuration layer, an alarm threshold layer, and a linkage rule layer. Each layer has strict parameter definitions and access control. The system global configuration layer stores system configuration parameters, the device type configuration layer stores device type parameters, the alarm threshold layer stores alarm threshold parameters, and the linkage rule layer stores linkage rule parameters. Simultaneously, the parameter matrix supports online modification and power-off saving. Modifications take effect in real time and have a complete change traceability mechanism. The power-off saving function uses non-volatile storage technology to ensure that parameters are not lost in the event of a sudden power outage and can accurately restore the system configuration state. Furthermore, the parameter matrix also has a built-in security verification mechanism to check the legality of parameter modifications, preventing system anomalies due to incorrect configuration, and providing strong protection for the system's flexibility and reliability.

[0087] The data standardization process also includes a three-level anomaly handling mechanism: a hardware-level protection mechanism, a signal-level detection mechanism, and a data-level processing mechanism. Specifically, the hardware-level protection mechanism uses opto-isolators and EMC protection circuits to achieve overvoltage and overcurrent protection. The signal-level detection mechanism uses real-time monitoring of current and voltage thresholds to detect open circuits and short circuits. The data-level processing mechanism uses wavelet transform and digital filtering to eliminate signal interference.

[0088] In this invention, to achieve end-to-end mapping from pulse signals to control decisions, the spiking neural network adopts a three-layer architecture design including an input layer, a hidden layer, and an output layer, such as... Figure 3As shown, the spiking neural network module is specifically used for:

[0089] The input layer employs a dynamically configured neuron array structure: at least 512 neurons are configured to receive input signals, and each input signal is mapped to 8 neurons by default to form redundant encoding, thereby improving the system's fault tolerance. The neurons utilize a modified LIF (Leaky Integrate and Fire) neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. This represents the input current. In specific implementations of this invention, the specific number of neurons can be expanded according to the actual number of input channels connected to meet actual usage requirements.

[0090] Hidden Layer: The network adopts a hierarchical functional sub-network, containing 64 feature extraction neurons, 128 pattern recognition neurons, and 32 decision neurons. Among them, the feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, the pattern recognition neurons are responsible for identifying typical alarm patterns, and the decision neurons are responsible for completing information integration and preliminary decision-making. The network connection adopts a variety of mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections with a range of ±3 neurons, and high-level to low-level feedback connections with an intensity coefficient of 0.2. At the same time, synaptic plasticity is realized based on the STDP (SpikeTimingDependent Plasticity) mechanism. The initial learning rate is set to 0.01 and dynamically adjusted with the training progress. The stability of the network is ensured by setting upper and lower limits of weights [-1, 1].

[0091] Output Layer: This layer is configured with multiple types of output neurons, including a main control neuron group, an auxiliary control neuron group, and a state feedback neuron group. The main control neuron group generates critical control commands, the auxiliary control neuron group generates secondary control commands, and the state feedback neuron group monitors execution results and provides feedback adjustment. Multiple safeguard mechanisms are employed during the decision-making process, including a 2 / 3 voting mechanism, an adaptive threshold control method, and pulse sequence analysis within a 100ms time window. In practical implementation, the main control neuron group is primarily responsible for generating critical control commands such as fire extinguishing, smoke extraction, and power outage, and employs a 2 / 3 voting mechanism to ensure decision reliability. The auxiliary control neuron group is mainly responsible for processing secondary control commands such as alarm prompts and ventilation adjustments, and has a dynamic priority adjustment function. The state feedback neuron group is primarily responsible for real-time monitoring of execution results and providing feedback adjustment within a 100ms time window, ensuring control accuracy better than 0.1ms. The 2 / 3 voting mechanism specifically means that the signal can only pass if more than half of the neurons output the same signal; the adaptive threshold control method means that the threshold can be adjusted in real time according to the input signal, with a baseline value of 50mV; the pulse sequence analysis with a 100ms time window is to extract the spatiotemporal features of the pulse sequence for analysis, so as to improve the accuracy and reliability of decision-making.

[0092] Furthermore, the spiking neural network module of this invention can also achieve system reliability through multi-level and multi-dimensional protection mechanisms, mainly including: hardware redundancy design, dynamic data consistency verification, and security protection mechanisms. Specifically, the hardware redundancy design adopts a "master-slave" architecture, with each key module configured with at least one backup unit; data consistency checks use multiple verification algorithms to promptly detect and isolate abnormal data; and the security protection mechanism establishes a full-link anomaly handling and response system to achieve real-time monitoring and intelligent scheduling of the system's operating status, ensuring that the system can still maintain stable operation under abnormal conditions.

[0093] In some embodiments of the present invention, the input layer further includes: processing the input signal using a presynaptic processing circuit before the input signal enters the neuron, including weight modulation and delay control, with a weight range of [0,1], a delay range of 0.5ms, and a time resolution of 0.1ms; simultaneously, to prevent overactivation of the neuron, a dynamic threshold adjustment mechanism is also adopted for the neuron, specifically including: adopting a real-time threshold update strategy based on a sliding window with a window length of 100ms; adaptively adjusting the activation threshold of the neuron according to the intensity of the input signal, with an adjustment range of ±10mV; calculating the threshold offset using an exponential weighted average algorithm with a time constant of 20ms; and rapidly increasing the threshold when a burst signal is detected to prevent overactivation, with a baseline threshold set at 55mV, which can be adaptively adjusted according to the input frequency.

[0094] In some embodiments of the present invention, in order to realize the conversion from the output of the spiking neural network to actual control commands, such as... Figure 4 As shown, the output decoding and control module specifically includes a pulse decoder, a control mapping unit, and an execution control unit;

[0095] The pulse decoder is used for pulse decoding. For pulse frequency decoding, advanced temporal coding conversion technology is employed. Specifically, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window to achieve accurate pulse counting. The instantaneous frequency is calculated using an exponential moving average algorithm, ensuring that the frequency estimate is updated every 1ms, and an adaptive hysteresis interval of ±5Hz is set to effectively filter high-frequency noise interference and suppress signal jitter.

[0096] For pattern recognition, a multi-dimensional pulse timing analysis method is employed to accurately identify various complex pulse patterns, such as burst and gradual patterns. By deeply analyzing the coordinated firing characteristics of neuronal populations, the system achieves precise capture of the network's dynamic state. Specifically, the system can track minute changes in the network state in real time, including the precise location of abrupt changes and gradual trends. For decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in both time and space dimensions within a 100ms sliding window. The time dimension captures dynamic evolution characteristics through the sliding window, while the spatial dimension performs correlation analysis on the outputs of relevant neurons. Finally, a reliability-based weighted decision mechanism comprehensively considers the confidence levels of each neuron to generate more stable and accurate linkage control commands.

[0097] The control mapping unit is used for control mapping. It employs a multi-channel output structure to improve output flexibility, specifically including 8 relay outputs (DC24V / 5A) for critical control, 40-channel partition controller, and 200-channel intelligent detector CAN bus communication. In practical implementation, a dual CAN bus redundancy design can be used to ensure communication reliability. The decoding results are mapped to specific linkage control commands through a four-level priority scheduling strategy. The four priority levels are emergency, high, medium, and low. This invention, through its four-level priority dynamic scheduling management mechanism, has the following two key features: first, high-priority tasks can interrupt low-priority operations in real time, ensuring immediate response to critical tasks; second, through a priority inversion protection mechanism, it can effectively prevent deadlock risks that may occur in traditional real-time systems, significantly improving the system's real-time performance and reliability. Meanwhile, four mapping modes are supported during mapping. The mapping strategy of this invention adopts multimodal intelligent mapping technology: (1) direct mapping of a single input to a single output to ensure one-to-one correspondence of signal transmission; (2) complex logic mapping of multiple input condition combinations to handle complex condition associations; (3) sequential mapping based on predetermined timing to achieve precise control of time-dependent tasks; and multi-output collaborative linkage mapping to support system-level linkage control.

[0098] The execution control unit is used to perform control operations: when executing linkage control operations according to linkage control commands, a bistable magnetic latching relay and multiple protection mechanisms are adopted to achieve timing control with an accuracy of 0.1ms. In specific implementations of this invention, the drive control adopts low-power bistable magnetic latching relay technology, which supports the CAN communication protocol and achieves ultra-high precision timing control of 0.1ms at the sub-millisecond level. This not only significantly reduces system energy consumption but also has unique power-off state retention and automatic power-on recovery functions, ensuring stable operation of the system under complex working conditions. The multiple protection mechanisms include a safety mechanism, a monitoring feedback mechanism, and data management. Among them, the safety mechanism adopts a multi-layered, three-dimensional protection architecture. By constructing a system-level linkage protection system, it achieves proactive risk identification and intelligent isolation, greatly improving the reliability and security of the system. The monitoring feedback mechanism adopts an intelligent design, establishing a full-link state perception and predictive early warning model to achieve precise optimization of the control process. In terms of data management, it provides real-time recording of the control process, detailed logs of key events, statistical analysis of control effects, and provides strong support for system maintenance and continuous optimization through a secure remote data access interface.

[0099] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A smart linkage method for energy storage fire protection based on hybrid control, characterized in that, The method includes: Step S1: Acquire and standardize the input multi-source data signal using a preset sampling frequency, and convert the processed data signal into a pulse sequence; specifically including: Multi-source data input: It can access 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 zone controllers, and analog signals from 200 smart detectors. Each signal is equipped with an independent opto-isolation circuit and EMC protection circuit. It also includes a preset parameter matrix with a hierarchical classification storage architecture, including a system global configuration layer, a device type configuration layer, an alarm threshold layer, and a linkage rule layer. The system global configuration layer is used to store system configuration parameters, the device type configuration layer is used to store device type parameters, the alarm threshold layer is used to store alarm threshold parameters, and the linkage rule layer is used to store linkage rule parameters. The parameter matrix supports online modification and power failure saving, and the parameter matrix also has a built-in security verification mechanism. Step S2: Construct a three-layer network using an improved LIF neuron model. Utilize this three-layer network to extract features and generate decisions from the input pulse sequence. Specifically, this includes: Input layer: Configured with at least 512 neurons to receive input signals, with each input signal mapped to 8 neurons by default to form redundant encoding; the neurons employ an improved LIF neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. The input current is represented; a dynamic threshold adjustment mechanism is adopted for neurons, including: using a sliding window-based real-time threshold update strategy, adaptively adjusting the activation threshold of neurons according to the intensity of the input signal, and calculating the threshold offset through an exponential weighted average algorithm; Hidden layer: The network adopts a hierarchical functional sub-network, containing 64 feature extraction neurons, 128 pattern recognition neurons and 32 decision neurons. The network connection adopts a variety of mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections with a range of ±3 neurons, and high-level to low-level feedback connections with an intensity coefficient of 0.

2. Output layer: Configured with multiple types of output neurons, including main control neuron group, auxiliary control neuron group and state feedback neuron group. Multiple guarantee mechanisms are adopted in the decision-making process, including 2 / 3 voting mechanism, adaptive threshold control method and pulse sequence analysis with 100ms time window; Step S3: Decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control commands, and execute linkage control operations according to the linkage control commands, specifically including: Pulse decoding: For pulse frequency decoding, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window, and the instantaneous frequency is calculated by an exponential moving average algorithm, with an adaptive lag interval of ±5Hz set; for pattern recognition, a multi-dimensional pulse time series analysis method is used for analysis and recognition; for decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in time and space dimensions within a 100ms sliding window. Control mapping: It adopts a multi-channel output structure, including 8 relay outputs, 40 partition controllers and 200 intelligent detectors for CAN bus communication; the decoding results are mapped into specific linkage control commands through a scheduling strategy based on four-level priority. At the same time, it supports four mapping modes, including direct mapping of a single input to a single output, complex logic mapping of multiple input conditions, sequential mapping based on a predetermined timing, and linkage mapping of multiple outputs in coordination.

2. The intelligent linkage method for energy storage fire protection based on hybrid control as described in claim 1, characterized in that, Step S1 specifically also includes: Data standardization processing: For digital signals, a Schmitt trigger is used for signal shaping, and the shaped standard signal is stored in a 32-bit status word; for analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning, with a sampling frequency of 10kHz, and all analog signals are uniformly normalized to the [0, 1] interval; in terms of timing processing, a uniform 100MHz main frequency clock source is used, and all input signals are synchronously sampled at a period of 50ms. Pulse sequence coding: Frequency coding adopts an adaptive probability distribution method to map the signal strength to a pulse frequency range of 1Hz~200Hz; Timing coding adopts a sliding mechanism with a 100ms time window to reflect the order of events through the phase difference of the pulses; Data buffering adopts a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, and is divided into real-time data area and historical data area.

3. The intelligent linkage method for energy storage fire protection based on hybrid control as described in claim 2, characterized in that, The data standardization process also includes a three-level anomaly handling mechanism: a hardware-level protection mechanism, a signal-level detection mechanism, and a data-level processing mechanism. Specifically, the hardware-level protection mechanism uses opto-isolators and EMC protection circuits to achieve overvoltage and overcurrent protection. The signal-level detection mechanism uses real-time monitoring of current and voltage thresholds to detect open circuits and short circuits. The data-level processing mechanism uses wavelet transform and digital filtering to eliminate signal interference.

4. The intelligent linkage method for energy storage fire protection based on hybrid control as described in claim 1, characterized in that, Feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, pattern recognition neurons are responsible for identifying typical alarm patterns, and decision neurons are responsible for information integration and preliminary decision-making; at the same time, synaptic plasticity is realized based on the STDP mechanism. The main control neuron group is used to generate key control commands, the auxiliary control neuron group is used to generate secondary control commands, and the state feedback neuron group is used to monitor the execution results and provide feedback adjustment.

5. The intelligent linkage method for energy storage fire protection based on hybrid control as described in claim 4, characterized in that, The input layer further includes: before the input signal enters the neuron, the input signal is processed by a presynaptic processing circuit, including weight modulation and delay control, with a weight range of [0,1] and a delay range of 0.5ms.

6. The intelligent linkage method for energy storage fire protection based on hybrid control as described in claim 1, characterized in that, Step S3 specifically also includes: Execution control: When executing linkage control operations according to linkage control commands, a bistable magnetic latching relay and multiple protection mechanisms are adopted to achieve timing control with an accuracy of 0.1ms.

7. A smart linkage system for energy storage fire protection based on hybrid control, characterized in that, The system includes an input data processing and encoding module, a spiking neural network module, and an output decoding and control module. The input data processing and encoding module is used to acquire and standardize the input multi-source data signals using a preset sampling frequency, and convert the processed data signals into pulse sequences; the input data processing and encoding module specifically includes a multi-source input interface: For multi-source data input: it can access 8 digital signals from the host computer in the warehouse, 3 signals from up to 40 zone controllers, and analog signals from 200 smart detectors. Each signal is equipped with an independent opto-isolation circuit and EMC protection circuit. It also includes a preset parameter matrix, which adopts a hierarchical classification storage architecture, including a system global configuration layer, a device type configuration layer, an alarm threshold layer, and a linkage rule layer. The system global configuration layer is used to store system configuration parameters, the device type configuration layer is used to store device type parameters, the alarm threshold layer is used to store alarm threshold parameters, and the linkage rule layer is used to store linkage rule parameters. The parameter matrix supports online modification and power failure saving, and the parameter matrix also has a built-in security verification mechanism. The spiking neural network module is used to construct a three-layer network using an improved LIF neuron model. This three-layer network is then used to extract features and generate decisions from the input spiking sequence. Specifically, it includes: Input layer: Configured with at least 512 neurons to receive input signals, with each input signal mapped to 8 neurons by default to form redundant encoding; the neurons employ an improved LIF neuron model, with the following dynamic equation: membrane time constant Set to 20ms, resting potential 65mV, film resistance It is 100MΩ. Represents membrane potential. The input current is represented; a dynamic threshold adjustment mechanism is adopted for neurons, including: using a sliding window-based real-time threshold update strategy, adaptively adjusting the activation threshold of neurons according to the intensity of the input signal, and calculating the threshold offset through an exponential weighted average algorithm; Hidden layer: The network adopts a hierarchical functional sub-network, containing 64 feature extraction neurons, 128 pattern recognition neurons and 32 decision neurons. The network connection adopts a variety of mechanisms, including sparse feedforward connections with a probability of 0.3, lateral inhibition connections with a range of ±3 neurons, and high-level to low-level feedback connections with an intensity coefficient of 0.

2. Output layer: Configured with multiple types of output neurons, including main control neuron group, auxiliary control neuron group and state feedback neuron group. Multiple guarantee mechanisms are adopted in the decision-making process, including 2 / 3 voting mechanism, adaptive threshold control method and pulse sequence analysis with 100ms time window; The output decoding and control module is used to decode the pulse sequence output by the three-layer architecture network, map the decoding result into specific linkage control commands, and execute linkage control operations according to the linkage control commands; specifically, it includes a pulse decoder and a control mapping unit: The pulse decoder is used for pulse decoding: For pulse frequency decoding, a 32-bit hardware counter is used to count the number of pulses within a 20ms time window, the instantaneous frequency is calculated by an exponential moving average algorithm, and an adaptive hysteresis interval of ±5Hz is set; For pattern recognition, a multi-dimensional pulse time series analysis method is used for analysis and recognition; For decision integration, a spatiotemporal dual integration mechanism is used to simultaneously fuse information in time and space dimensions within a 100ms sliding window. The control mapping unit is used for control mapping: it adopts a multi-channel output structure, including 8 relay outputs, 40 partition controllers and 200 intelligent detectors for CAN bus communication; it maps the decoding results into specific linkage control commands through a scheduling strategy based on four-level priority, and supports four mapping modes during mapping, including direct mapping of a single input to a single output, complex logic mapping of multiple input conditions, sequential mapping based on a predetermined timing, and linkage mapping of multiple outputs in coordination.

8. The intelligent linkage system for energy storage fire protection based on hybrid control as described in claim 7, characterized in that, The input data processing and encoding module further includes a standardization processing unit and a pulse sequence encoder; The standardization processing unit is used for data standardization processing: for digital signals, a Schmitt trigger is used for signal shaping, and the shaped standard signal is stored in a 32-bit status word; for analog signals, a 24-bit Σ-Δ ADC is used for signal sampling and conditioning, with a sampling frequency of 10kHz, and all analog signals are uniformly normalized to the [0, 1] interval; in terms of timing processing, a uniform 100MHz main frequency clock source is used, and all input signals are synchronously sampled at a period of 50ms. The pulse sequence encoder is used for pulse sequence encoding: frequency encoding adopts an adaptive probability distribution method to map the signal strength to a pulse frequency range of 1Hz~200Hz; timing encoding adopts a sliding mechanism with a 100ms time window to reflect the order of events through the phase difference of the pulses; data buffering adopts a double-ended queue to implement a circular buffer with a buffer depth of 100 sampling points, and is divided into a real-time data area and a historical data area.

9. The intelligent linkage system for energy storage fire protection based on hybrid control as described in claim 7, characterized in that, Feature extraction neurons are responsible for extracting the spatiotemporal features of the input signal, pattern recognition neurons are responsible for identifying typical alarm patterns, and decision neurons are responsible for information integration and preliminary decision-making; at the same time, synaptic plasticity is realized based on the STDP mechanism. The main control neuron group is used to generate key control commands, the auxiliary control neuron group is used to generate secondary control commands, and the state feedback neuron group is used to monitor the execution results and provide feedback adjustment.

10. The intelligent linkage system for energy storage fire protection based on hybrid control as described in claim 7, characterized in that, The output decoding and control module further includes an execution control unit; The execution control unit is used to perform control: when performing linkage control operations according to linkage control instructions, it adopts a bistable magnetic latching relay and multiple protection mechanisms to achieve timing control with an accuracy of 0.1ms.

Images